From 48fe111c12f6599fafb694946d86c9d1e1dd77a7 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Mon, 13 Jul 2026 17:47:48 +0000 Subject: [PATCH 001/141] fix(responses): stop managed Responses WS from leaking litellm_params into provider body Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com> --- litellm/responses/streaming_iterator.py | 3 +- .../test_responses_websocket_all_providers.py | 53 +++++++++++++++++++ 2 files changed, 54 insertions(+), 2 deletions(-) diff --git a/litellm/responses/streaming_iterator.py b/litellm/responses/streaming_iterator.py index eb78e6f9c8d..616a6659a55 100644 --- a/litellm/responses/streaming_iterator.py +++ b/litellm/responses/streaming_iterator.py @@ -2097,8 +2097,7 @@ class ManagedResponsesWebSocketHandler: if "litellm_metadata" not in call_kwargs: call_kwargs["litellm_metadata"] = {} call_kwargs["litellm_metadata"]["proxy_server_request"] = proxy_server_request - call_kwargs.setdefault("litellm_params", {}) - call_kwargs["litellm_params"]["proxy_server_request"] = proxy_server_request + call_kwargs["proxy_server_request"] = proxy_server_request async def _stream_and_forward(self, model: str, call_kwargs: Dict[str, Any]) -> Optional[Dict[str, Any]]: """ diff --git a/tests/test_litellm/responses/test_responses_websocket_all_providers.py b/tests/test_litellm/responses/test_responses_websocket_all_providers.py index 4509abc7749..7557757ded1 100644 --- a/tests/test_litellm/responses/test_responses_websocket_all_providers.py +++ b/tests/test_litellm/responses/test_responses_websocket_all_providers.py @@ -660,6 +660,59 @@ class TestChunkTransformation: assert ManagedResponsesWebSocketHandler._input_to_messages({}) == [] +class TestUpdateProxyRequest: + """Regression tests for ManagedResponsesWebSocketHandler._update_proxy_request. + + The managed WebSocket path calls ``litellm.aresponses(model=..., **call_kwargs)``. + ``litellm_params`` is not a Responses API request field, so passing it as a + top-level kwarg leaks it into the provider request body and providers that + forbid extra inputs (e.g. Anthropic) reject the call with + ``litellm_params: Extra inputs are not permitted``. The request-tracking data + must ride along as ``proxy_server_request`` instead, which litellm consumes + internally and never forwards to the provider. + """ + + def test_does_not_inject_litellm_params_kwarg(self): + from litellm.responses.streaming_iterator import ( + ManagedResponsesWebSocketHandler, + ) + + call_kwargs = { + "input": "hello", + "store": True, + "litellm_metadata": { + "proxy_server_request": {"headers": {}, "body": {}}, + }, + } + + ManagedResponsesWebSocketHandler._update_proxy_request( + call_kwargs, "anthropic/claude-sonnet-4-5" + ) + + assert "litellm_params" not in call_kwargs + assert call_kwargs["proxy_server_request"]["body"]["model"] == ( + "anthropic/claude-sonnet-4-5" + ) + assert call_kwargs["proxy_server_request"]["body"]["input"] == "hello" + + def test_proxy_server_request_matches_metadata(self): + from litellm.responses.streaming_iterator import ( + ManagedResponsesWebSocketHandler, + ) + + call_kwargs = { + "input": "hi", + "litellm_metadata": {"proxy_server_request": {"body": {}}}, + } + + ManagedResponsesWebSocketHandler._update_proxy_request(call_kwargs, "gpt-4o") + + assert ( + call_kwargs["proxy_server_request"] + == call_kwargs["litellm_metadata"]["proxy_server_request"] + ) + + class TestWebSocketEventTypes: """Test that all WebSocket event types are properly handled with dict-based chunks""" From 52f3ff13f0a7da9f5a2cbe7333e9b7dcd8643780 Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Thu, 6 Aug 2026 13:26:11 -0700 Subject: [PATCH 002/141] feat(proxy): limit repeated failed Admin UI sign-in attempts The Admin UI sign-in endpoints accept an unbounded number of password attempts. All three call authenticate_user, and none of them keeps any record of how many times a given caller has already been refused, so a misbehaving or misconfigured client can retry indefinitely at full speed. A LoginThrottle is now a required argument to authenticate_user, so the accounting lives at the one function all three endpoints share and a fourth endpoint cannot be added without deciding what to pass. Failures are counted per username and source address over a fixed window and further attempts are refused with 429 and a Retry-After header. The check runs before the database lookup and before the password comparison, so a refused caller does no further work. Only genuine credential rejections count. Configuration errors do not, a refused attempt does not extend the window, and a successful sign-in clears the bucket. The username is case folded because the user lookup is case insensitive, so casing cannot multiply the allowance. Both credential rejections now return one identical message. SSO is unaffected; it never calls this function. max_failed_login_attempts (10) and failed_login_window_seconds (900) are read from config.yaml, with LITELLM_DISABLE_LOGIN_RATE_LIMIT to turn the accounting off. They are deliberately not database backed, so editing YAML always wins and an operator refused by a bad value can recover. --- litellm/proxy/_types.py | 10 + litellm/proxy/auth/login_throttle.py | 177 +++++++ litellm/proxy/auth/login_utils.py | 18 +- litellm/proxy/proxy_server.py | 7 +- .../proxy/auth/test_login_utils.py | 477 ++++++++++++++++++ .../proxy/proxy_server/conftest.py | 25 + .../proxy_server/test_routes_login_sso.py | 92 +++- tests/test_litellm/proxy/test_proxy_server.py | 39 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 10 + 9 files changed, 843 insertions(+), 12 deletions(-) create mode 100644 litellm/proxy/auth/login_throttle.py diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index ed35691c6ec..1fd2781bd95 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2525,6 +2525,16 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): description="sends alerts if requests hang for 5min+", ) ui_access_mode: Literal["admin_only", "all"] | None = Field("all", description="Control access to the Proxy UI") + max_failed_login_attempts: int | None = Field( + None, + ge=1, + description="Number of failed Admin UI sign-in attempts allowed for one username from one source address within `failed_login_window_seconds`, before further attempts are refused with 429. Configurable from config.yaml only. Defaults to 10", + ) + failed_login_window_seconds: int | None = Field( + None, + ge=1, + description="Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Configurable from config.yaml only. Defaults to 900", + ) allowed_routes: list | None = Field(None, description="Proxy API Endpoints you want users to be able to access") reject_clientside_metadata_tags: bool | None = Field( None, diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py new file mode 100644 index 00000000000..e2befd15a35 --- /dev/null +++ b/litellm/proxy/auth/login_throttle.py @@ -0,0 +1,177 @@ +"""Failed-login accounting for the Admin UI sign-in path. + +Counts failed credential checks per (username, source address) over a fixed window and +denies further attempts with 429 once the count reaches the limit. Built per request by +``LoginThrottle.from_request`` because it carries that request's resolved source address, +and because the coordination cache is assigned at startup and can be reassigned later. +""" + +import hashlib +from collections.abc import Awaitable +from dataclasses import dataclass +from types import MappingProxyType +from typing import Final + +from fastapi import Request + +from litellm._logging import verbose_proxy_logger +from litellm.caching.dual_cache import DualCache +from litellm.caching.in_memory_cache import InMemoryCache +from litellm.caching.redis_cache import RedisCache +from litellm.proxy._types import ProxyErrorTypes, ProxyException +from litellm.proxy.auth.network import TrustedProxyConfig, normalize_cidr_ranges, resolve_client_ip +from litellm.proxy.auth.trusted_proxy_utils import TRUSTED_PROXY_RANGES_KEY +from litellm.secret_managers.main import get_secret_bool + +DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS: Final = 10 +DEFAULT_FAILED_LOGIN_WINDOW_SECONDS: Final = 900 + +_CACHE_KEY_PREFIX: Final = "login_fail" +_UNKNOWN_SOURCE: Final = "unknown" +_MAX_LOGGED_USERNAME_CHARS: Final = 128 + +_MAX_TRACKED_LOGIN_SOURCES: Final = 10_000 + +_FAILED_LOGIN_CACHE: Final = DualCache( + in_memory_cache=InMemoryCache(max_size_in_memory=_MAX_TRACKED_LOGIN_SOURCES), + default_in_memory_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS, +) +_NO_SETTINGS: Final = MappingProxyType({}) + + +def _int_setting(name: str, value: object, default: int, minimum: int) -> int: + if value is None: + return default + if isinstance(value, bool) or not isinstance(value, int) or value < minimum: + verbose_proxy_logger.warning( + "general_settings.%s=%r is not an integer >= %s; using the default of %s", name, value, minimum, default + ) + return default + return value + + +def _as_count(cached: object) -> int: + return int(cached) if isinstance(cached, int | float) and not isinstance(cached, bool) else 0 + + +@dataclass(frozen=True, slots=True) +class LoginThrottle: + """Fixed-window failed-login accounting for one request's source address.""" + + client_ip: str + max_attempts: int + window_seconds: int + cache: DualCache + redis_cache: RedisCache | None = None + enabled: bool = True + + @classmethod + def from_request(cls, request: Request) -> "LoginThrottle": + """Build the throttle for this request from the live proxy settings and caches.""" + from litellm.proxy.proxy_server import general_settings, redis_usage_cache + + settings: Final = general_settings or _NO_SETTINGS + cidrs: Final = normalize_cidr_ranges( + settings.get(TRUSTED_PROXY_RANGES_KEY), setting_name=TRUSTED_PROXY_RANGES_KEY + ) + resolved, _ = resolve_client_ip( + request, TrustedProxyConfig(use_forwarded_for=bool(cidrs), trusted_proxy_cidrs=cidrs) + ) + return cls( + client_ip=resolved or _UNKNOWN_SOURCE, + max_attempts=_int_setting( + "max_failed_login_attempts", + settings.get("max_failed_login_attempts"), + DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS, + 1, + ), + window_seconds=_int_setting( + "failed_login_window_seconds", + settings.get("failed_login_window_seconds"), + DEFAULT_FAILED_LOGIN_WINDOW_SECONDS, + 1, + ), + cache=_FAILED_LOGIN_CACHE, + redis_cache=redis_usage_cache, + enabled=not get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT"), + ) + + @staticmethod + def _loggable(username: str) -> str: + """The username with anything that could forge a log line removed.""" + return "".join(c for c in username if c.isprintable())[:_MAX_LOGGED_USERNAME_CHARS] + + def _key(self, username: str) -> str: + identity: Final = hashlib.sha256(username.casefold().encode("utf-8")).hexdigest() + return f"{_CACHE_KEY_PREFIX}:{identity}:{self.client_ip}" + + async def _outcome(self, work: Awaitable[object]) -> object: + try: + return await work + except Exception as exc: # noqa: BLE001 # an unreachable cache must never deny a valid credential + verbose_proxy_logger.warning("login attempt accounting unavailable: %s", exc) + return None + + async def _failures(self, key: str) -> int: + store: Final = self.cache if self.redis_cache is None else self.redis_cache + return _as_count(await self._outcome(store.async_get_cache(key=key))) + + async def _ensure_expiry(self, key: str) -> None: + """Give the counter an expiry if it somehow has none. + + Redis commits the increment before setting the TTL, so a failure in between can + leave a counter that never expires. Nothing increments the key again once the + limit is reached, so without this the pair would stay refused indefinitely. + """ + redis_cache: Final = self.redis_cache + if redis_cache is None: + return + if isinstance(await self._outcome(redis_cache.async_get_ttl(key)), int): + return + await self._outcome(redis_cache.async_increment(key, 0, ttl=self.window_seconds)) + + async def raise_if_blocked(self, username: str) -> None: + """Deny before the database lookup and before the password comparison.""" + if not self.enabled: + return + key: Final = self._key(username) + if await self._failures(key) < self.max_attempts: + return + await self._ensure_expiry(key) + verbose_proxy_logger.warning( + "Admin UI sign-in attempts exhausted for username=%s source=%s; %s attempts in %ss, retry after %ss", + self._loggable(username), + self.client_ip, + self.max_attempts, + self.window_seconds, + self.window_seconds, + ) + raise ProxyException( + message="Too many failed sign-in attempts. Try again later.", + type=ProxyErrorTypes.auth_error, + param="max_failed_login_attempts", + code=429, + headers={"Retry-After": str(self.window_seconds)}, # mutable-ok: ProxyException coerces header values + ) + + async def record_failure(self, username: str) -> None: + """Count one rejected credential guess against this username and source.""" + if not self.enabled: + return + key: Final = self._key(username) + redis_cache: Final = self.redis_cache + if redis_cache is not None: + await self._outcome(redis_cache.async_increment(key, 1, ttl=self.window_seconds)) + await self._ensure_expiry(key) + return + await self._outcome(self.cache.async_increment_cache(key=key, value=1, ttl=self.window_seconds)) + + async def clear(self, username: str) -> None: + """Drop the bucket after a successful sign-in.""" + if not self.enabled: + return + key: Final = self._key(username) + redis_cache: Final = self.redis_cache + if redis_cache is not None: + await self._outcome(redis_cache.async_delete_cache(key)) + await self._outcome(self.cache.async_delete_cache(key=key)) diff --git a/litellm/proxy/auth/login_utils.py b/litellm/proxy/auth/login_utils.py index fba95972944..98780a4692a 100644 --- a/litellm/proxy/auth/login_utils.py +++ b/litellm/proxy/auth/login_utils.py @@ -24,6 +24,7 @@ from litellm.proxy._types import ( UpdateUserRequest, UserAPIKeyAuth, ) +from litellm.proxy.auth.login_throttle import LoginThrottle from litellm.proxy.management_endpoints.internal_user_endpoints import user_update from litellm.proxy.management_endpoints.key_management_endpoints import ( generate_key_helper_fn, @@ -41,6 +42,10 @@ from litellm.repositories.user_repository import UserRepository from litellm.secret_managers.main import get_secret_bool from litellm.types.proxy.ui_sso import ReturnedUITokenObject +INVALID_UI_CREDENTIALS_MESSAGE: Final = ( + "Invalid credentials used to access UI. Check 'UI_USERNAME' and 'UI_PASSWORD', or the password set for your user" +) + async def _rehash_password_if_needed(user_id: str, password: str, stored: str) -> None: """Rehash legacy password (SHA256) to scrypt on successful login.""" @@ -111,6 +116,7 @@ async def authenticate_user( password: str, master_key: str | None, prisma_client: PrismaClient | None, + throttle: LoginThrottle, ) -> LoginResult: """ Authenticate a user and generate an API key for UI access. @@ -139,6 +145,8 @@ async def authenticate_user( code=500, ) + await throttle.raise_if_blocked(username) + ui_username, ui_password = get_ui_credentials(master_key) # Check if we can find the `username` in the db. On the UI, users can enter username=their email @@ -240,6 +248,8 @@ async def authenticate_user( key = ExperimentalUIJWTToken.get_experimental_ui_login_jwt_auth_token(user_info) + await throttle.clear(username) + return LoginResult( user_id=user_id, key=key, @@ -294,6 +304,8 @@ async def authenticate_user( key = response["token"] + await throttle.clear(username) + return LoginResult( user_id=user_id, key=key, @@ -302,15 +314,17 @@ async def authenticate_user( login_method="username_password", ) else: + await throttle.record_failure(username) raise ProxyException( - message=f"Invalid credentials used to access UI.\nNot valid credentials for {username}", + message=INVALID_UI_CREDENTIALS_MESSAGE, type=ProxyErrorTypes.auth_error, param="invalid_credentials", code=401, ) else: + await throttle.record_failure(username) raise ProxyException( - message="Invalid credentials used to access UI.\nCheck 'UI_USERNAME', 'UI_PASSWORD' in .env file", + message=INVALID_UI_CREDENTIALS_MESSAGE, type=ProxyErrorTypes.auth_error, param="invalid_credentials", code=401, diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 9b5d1b9bfea..39a419df73f 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -293,6 +293,7 @@ from litellm.proxy.auth.auth_utils import ( ) from litellm.proxy.auth.handle_jwt import JWTHandler from litellm.proxy.auth.litellm_license import LicenseCheck +from litellm.proxy.auth.login_throttle import LoginThrottle from litellm.proxy.auth.model_checks import ( expand_wildcard_deployments_for_model_info, get_all_fallbacks, @@ -723,6 +724,7 @@ from fastapi.openapi.docs import get_swagger_ui_html from fastapi.openapi.utils import get_openapi from fastapi.responses import ( FileResponse, + HTMLResponse, JSONResponse, ORJSONResponse, RedirectResponse, @@ -14730,8 +14732,6 @@ async def fallback_login(request: Request): else: redirect_url += "/sso/callback" - from fastapi.responses import HTMLResponse - hide_default_credentials_hint: Final = ( os.getenv("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", "false").lower() == "true" or general_settings.get("hide_default_credentials_hint", False) is True @@ -14761,6 +14761,7 @@ async def login(request: Request): password=password, master_key=master_key, prisma_client=prisma_client, + throttle=LoginThrottle.from_request(request), ) # Create UI token object @@ -14835,6 +14836,7 @@ async def login_v2(request: Request): password=password, master_key=master_key, prisma_client=prisma_client, + throttle=LoginThrottle.from_request(request), ) returned_ui_token_object: Final = create_ui_token_object( @@ -14905,6 +14907,7 @@ async def login_v3(request: Request): password=password, master_key=master_key, prisma_client=prisma_client, + throttle=LoginThrottle.from_request(request), ) returned_ui_token_object: Final = create_ui_token_object( diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 1c66acf8678..1e3fcaf5d78 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -10,6 +10,16 @@ from unittest.mock import AsyncMock, MagicMock, patch import pytest + +def _unlimited_throttle(): + """A throttle wired to a real in-memory store with a limit no test can reach.""" + from litellm.caching.dual_cache import DualCache + from litellm.proxy.auth.login_throttle import LoginThrottle + + return LoginThrottle(client_ip="1.2.3.4", max_attempts=10_000, window_seconds=900, cache=DualCache()) + + + from litellm.constants import LITELLM_PROXY_ADMIN_NAME from litellm.proxy._types import ( LiteLLM_UserTable, @@ -98,6 +108,7 @@ async def test_authenticate_user_admin_login_with_ui_credentials(): password=ui_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert isinstance(result, LoginResult) @@ -155,6 +166,7 @@ async def test_authenticate_user_admin_login_with_master_key_as_password(monkeyp password=master_key, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert isinstance(result, LoginResult) @@ -179,6 +191,7 @@ async def test_authenticate_user_invalid_credentials(): password=wrong_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert exc_info.value.type == ProxyErrorTypes.auth_error @@ -197,6 +210,7 @@ async def test_authenticate_user_missing_master_key(): password="password", master_key=None, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert exc_info.value.type == ProxyErrorTypes.auth_error @@ -237,6 +251,7 @@ async def test_authenticate_user_wrong_password(): password=wrong_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert exc_info.value.type == ProxyErrorTypes.auth_error @@ -295,12 +310,14 @@ async def test_authenticate_user_email_case_insensitive_login(): password=correct_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) result_lower = await authenticate_user( username=stored_email, password=correct_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert result_mixed.user_id == result_lower.user_id == "test-user-123" @@ -342,6 +359,7 @@ async def test_authenticate_user_database_required_for_admin(monkeypatch): password=ui_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert exc_info.value.type == ProxyErrorTypes.auth_error @@ -393,6 +411,7 @@ async def test_authenticate_user_admin_login_with_non_ascii_characters(): password=ui_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert isinstance(result, LoginResult) @@ -469,18 +488,21 @@ async def test_authenticate_user_multiple_logins_generate_unique_tokens(): password=ui_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) result2 = await authenticate_user( username=ui_username, password=ui_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) result3 = await authenticate_user( username=ui_username, password=ui_password, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) # Each login should return a unique token @@ -538,6 +560,7 @@ async def test_authenticate_user_database_login_with_non_ascii_password(): password=password_with_special_char, master_key=master_key, prisma_client=mock_prisma_client, + throttle=_unlimited_throttle(), ) assert isinstance(result, LoginResult) @@ -598,3 +621,457 @@ class TestEncodeUiSessionJwt: request.cookies = {"token": token} with patch("litellm.proxy.proxy_server.master_key", "sk-master-for-tests"): assert _user_id_from_session_cookie(request) == "cornell-user" + + +# --------------------------------------------------------------------------- +# Failed-login accounting (LIT-5285) +# --------------------------------------------------------------------------- + + +def _throttle(max_attempts: int = 3, window_seconds: int = 900, client_ip: str = "1.2.3.4", cache=None, redis_cache=None): + """A throttle over a real in-memory store, so the tests exercise the true counters.""" + from litellm.caching.dual_cache import DualCache + from litellm.proxy.auth.login_throttle import LoginThrottle + + return LoginThrottle( + client_ip=client_ip, + max_attempts=max_attempts, + window_seconds=window_seconds, + cache=cache if cache is not None else DualCache(), + redis_cache=redis_cache, + ) + + +async def _guess(throttle, username: str = "admin", password: str = "wrong"): + from litellm.proxy.auth.login_utils import authenticate_user + + return await authenticate_user( + username=username, + password=password, + master_key="sk-master", + prisma_client=None, + throttle=throttle, + ) + + +@pytest.mark.asyncio +async def test_attempts_are_refused_once_the_limit_is_reached(monkeypatch): + """The limit denies further attempts for the window, and the denial carries Retry-After.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=3, window_seconds=77) + + for _ in range(3): + with pytest.raises(ProxyException) as first: + await _guess(throttle) + assert first.value.code == "401" + + with pytest.raises(ProxyException) as blocked: + await _guess(throttle) + assert blocked.value.code == "429" + assert blocked.value.headers.get("Retry-After") == "77" + + +@pytest.mark.asyncio +async def test_a_correct_password_is_refused_while_blocked(monkeypatch): + """The check precedes the credential comparison, so being over the limit wins.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=2) + + for _ in range(2): + with pytest.raises(ProxyException): + await _guess(throttle) + + with pytest.raises(ProxyException) as blocked: + await _guess(throttle, password="right") + assert blocked.value.code == "429" + + +@pytest.mark.asyncio +async def test_a_blocked_attempt_does_not_extend_the_window(monkeypatch): + """Hammering while blocked must not push the counter or refresh its TTL.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=2) + key = throttle._key("admin") + + for _ in range(2): + with pytest.raises(ProxyException): + await _guess(throttle) + counted_at_limit = await throttle._failures(key) + + for _ in range(5): + with pytest.raises(ProxyException): + await _guess(throttle) + + assert await throttle._failures(key) == counted_at_limit == 2 + + +@pytest.mark.asyncio +async def test_a_successful_sign_in_clears_the_bucket(monkeypatch): + """Success resets the budget rather than leaving the operator near the limit.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + throttle = _throttle(max_attempts=3) + + for _ in range(2): + with pytest.raises(ProxyException): + await _guess(throttle) + + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ): + await _guess(throttle, password="right") + + assert await throttle._failures(throttle._key("admin")) == 0 + + +@pytest.mark.asyncio +async def test_a_configuration_error_never_counts(monkeypatch): + """A 500 from an unset master key is not a guess and must not consume the budget.""" + from litellm.proxy._types import ProxyException + + throttle = _throttle(max_attempts=2) + for _ in range(5): + with pytest.raises(ProxyException) as exc: + await authenticate_user( + username="admin", password="x", master_key=None, prisma_client=None, throttle=throttle + ) + assert exc.value.code == "500" + + assert await throttle._failures(throttle._key("admin")) == 0 + + +@pytest.mark.asyncio +async def test_the_username_is_case_folded_into_one_bucket(monkeypatch): + """The DB lookup is case-insensitive, so casing must not multiply the budget.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=4) + + for name in ("admin@corp.com", "ADMIN@corp.com", "Admin@corp.com", "aDmIn@corp.com"): + with pytest.raises(ProxyException) as exc: + await _guess(throttle, username=name) + assert exc.value.code == "401" + + with pytest.raises(ProxyException) as blocked: + await _guess(throttle, username="admin@CORP.com") + assert blocked.value.code == "429" + + +@pytest.mark.asyncio +async def test_a_different_username_from_the_same_source_is_unaffected(monkeypatch): + """The bucket is the pair, so one username's failures do not block another.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=2) + + for _ in range(3): + with pytest.raises(ProxyException): + await _guess(throttle, username="admin") + + with pytest.raises(ProxyException) as other: + await _guess(throttle, username="someone-else@example.com") + assert other.value.code == "401", "a second username must still reach the credential check" + + +@pytest.mark.asyncio +async def test_both_credential_rejections_are_indistinguishable(monkeypatch): + """One message for the known and the unknown username, so responses do not enumerate.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + + with pytest.raises(ProxyException) as unknown: + await _guess(_throttle(max_attempts=99), username="nobody@example.com") + + fake_user = MagicMock() + fake_user.user_id = "u-1" + fake_user.user_email = "known@example.com" + fake_user.user_role = "internal_user" + fake_user.password = "scrypt:fake" + repo = MagicMock() + repo.return_value.table.find_first = AsyncMock(return_value=fake_user) + with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( + "litellm.proxy.auth.login_utils.verify_password", return_value=False + ): + with pytest.raises(ProxyException) as known: + await authenticate_user( + username="known@example.com", + password="wrong", + master_key="sk-master", + prisma_client=MagicMock(), + throttle=_throttle(max_attempts=99), + ) + + assert unknown.value.message == known.value.message + assert "known@example.com" not in unknown.value.message + known.value.message + + +@pytest.mark.asyncio +async def test_a_user_with_no_password_set_does_not_consume_the_budget(monkeypatch): + """That 401 is deterministic and guards no secret, so counting it would only let + someone burn a passwordless account's bucket.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=2) + + passwordless = MagicMock() + passwordless.user_id = "u-2" + passwordless.user_email = "nopass@example.com" + passwordless.user_role = "internal_user" + passwordless.password = None + repo = MagicMock() + repo.return_value.table.find_first = AsyncMock(return_value=passwordless) + + with patch("litellm.proxy.auth.login_utils.UserRepository", repo): + for _ in range(5): + with pytest.raises(ProxyException) as exc: + await authenticate_user( + username="nopass@example.com", + password="x", + master_key="sk-master", + prisma_client=MagicMock(), + throttle=throttle, + ) + assert exc.value.code == "401" + + assert await throttle._failures(throttle._key("nopass@example.com")) == 0 + + +@pytest.mark.asyncio +async def test_a_wrong_password_for_a_known_user_also_counts(monkeypatch): + """The database-user branch must charge the bucket too, not just the unknown-user branch.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=3) + + known = MagicMock() + known.user_id = "u-1" + known.user_email = "known@example.com" + known.user_role = "internal_user" + known.password = "scrypt:stored" + repo = MagicMock() + repo.return_value.table.find_first = AsyncMock(return_value=known) + + async def _attempt(): + return await authenticate_user( + username="known@example.com", + password="wrong", + master_key="sk-master", + prisma_client=MagicMock(), + throttle=throttle, + ) + + with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( + "litellm.proxy.auth.login_utils.verify_password", return_value=False + ): + for _ in range(3): + with pytest.raises(ProxyException) as rejected: + await _attempt() + assert rejected.value.code == "401" + + with pytest.raises(ProxyException) as blocked: + await _attempt() + assert blocked.value.code == "429" + + +@pytest.mark.asyncio +async def test_two_source_addresses_do_not_share_a_bucket(monkeypatch): + """The key is the pair, so one address exhausting its budget must not block another. + + Dropping the address from the key would turn this into the username-only counter the + design rejects, where anyone can lock a named admin out from anywhere. + """ + from litellm.caching.dual_cache import DualCache + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + shared_store = DualCache() + attacker = _throttle(max_attempts=2, client_ip="203.0.113.9", cache=shared_store) + operator = _throttle(max_attempts=2, client_ip="198.51.100.7", cache=shared_store) + + for _ in range(3): + with pytest.raises(ProxyException): + await _guess(attacker, username="admin") + + with pytest.raises(ProxyException) as blocked: + await _guess(attacker, username="admin") + assert blocked.value.code == "429" + + with pytest.raises(ProxyException) as unaffected: + await _guess(operator, username="admin") + assert unaffected.value.code == "401", "the real operator must still reach the credential check" + + +class _NoExpiryRedis: + """Redis that stores the counter but never records an expiry for it. + + Models the window between INCRBYFLOAT committing and the TTL call failing. + """ + + def __init__(self): + self.values: dict = {} + self.expiry_repairs = 0 + + async def async_get_cache(self, key, **kwargs): + return self.values.get(key) + + async def async_increment(self, key, value, ttl=None, **kwargs): + if int(value) == 0: + self.expiry_repairs += 1 + self.values[key] = self.values.get(key, 0) + int(value) + return self.values[key] + + async def async_get_ttl(self, key): + return None + + async def async_delete_cache(self, key): + self.values.pop(key, None) + + +@pytest.mark.asyncio +async def test_a_counter_left_without_an_expiry_is_repaired(monkeypatch): + """Regression: a counter with no TTL would refuse the pair forever. + + Redis commits the increment before setting the expiry, and nothing increments the key + again once the limit is reached, so a TTL that never landed is never repaired on its + own and the username and source pair stays refused with no way back. + """ + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + redis = _NoExpiryRedis() + throttle = _throttle(max_attempts=2, redis_cache=redis) + + for _ in range(2): + with pytest.raises(ProxyException): + await _guess(throttle) + + assert redis.expiry_repairs >= 1, "each recorded failure must leave the counter with an expiry" + + repairs_before_block = redis.expiry_repairs + with pytest.raises(ProxyException) as blocked: + await _guess(throttle) + assert blocked.value.code == "429" + assert redis.expiry_repairs > repairs_before_block, "the refusal path must repair a missing expiry too" + + +@pytest.mark.asyncio +async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): + """Regression: throttle entries must not evict cached credentials. + + user_api_key_cache holds at most 200 in-memory entries and evicts the soonest to + expire first, so parking 900s sign-in counters there let a stream of made-up usernames + push out the much shorter lived credential entries, sending every ordinary API request + back to the database. + """ + from litellm.proxy import proxy_server as ps + from litellm.proxy.auth.login_throttle import _CACHE_KEY_PREFIX, LoginThrottle + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setattr(ps, "redis_usage_cache", None) + + auth_cache_keys_before = set(ps.user_api_key_cache.in_memory_cache.cache_dict) + + request = MagicMock() + request.headers = {} + request.client = MagicMock() + request.client.host = "1.2.3.4" + throttle = LoginThrottle.from_request(request) + + for i in range(25): + with pytest.raises(Exception): + await _guess(throttle, username=f"made-up-{i}@example.com") + + added = set(ps.user_api_key_cache.in_memory_cache.cache_dict) - auth_cache_keys_before + assert not [k for k in added if str(k).startswith(_CACHE_KEY_PREFIX)], ( + "sign-in counters must live in their own cache, not the key-authentication cache" + ) + + +@pytest.mark.asyncio +async def test_a_refused_username_cannot_forge_log_lines(monkeypatch): + """The username reaches a warning log, so it must not carry newlines or control bytes.""" + import logging + + from litellm._logging import verbose_proxy_logger + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=1) + forged = "victim@example.com\nWARNING: sign-in succeeded for attacker\x00" + + with pytest.raises(ProxyException): + await _guess(throttle, username=forged) + + records: list[logging.LogRecord] = [] + handler = logging.Handler() + handler.emit = records.append + verbose_proxy_logger.addHandler(handler) + try: + with pytest.raises(ProxyException) as blocked: + await _guess(throttle, username=forged) + finally: + verbose_proxy_logger.removeHandler(handler) + + assert blocked.value.code == "429" + emitted = [r.getMessage() for r in records if "sign-in attempts exhausted" in r.getMessage()] + assert emitted, "the refusal must be logged" + assert "\n" not in emitted[0] and "\x00" not in emitted[0] + assert "victim@example.com" in emitted[0] + + +@pytest.mark.asyncio +async def test_a_username_spray_cannot_evict_an_existing_counter(monkeypatch): + """Regression: the in-memory tier must hold more counters than a spray can create. + + The default in-memory cache keeps 200 entries and evicts the soonest to expire, and + every counter shares one window, so eviction was effectively oldest-first. A few + hundred made-up usernames therefore pushed out the attacker's own counter and handed + back a fresh allowance against the real account. + """ + from litellm.proxy._types import ProxyException + from litellm.proxy.auth.login_throttle import _FAILED_LOGIN_CACHE, _MAX_TRACKED_LOGIN_SOURCES + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + assert _MAX_TRACKED_LOGIN_SOURCES >= 10_000 + assert _FAILED_LOGIN_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_SOURCES + + throttle = _throttle(max_attempts=3, cache=_FAILED_LOGIN_CACHE, client_ip="10.9.9.9") + victim = "spray-victim@corp.com" + for _ in range(3): + with pytest.raises(ProxyException): + await _guess(throttle, username=victim) + + for i in range(500): + await throttle.record_failure(f"spray-filler-{i}@corp.com") + + assert await throttle._failures(throttle._key(victim)) == 3, "the counter must survive a spray" + with pytest.raises(ProxyException) as blocked: + await _guess(throttle, username=victim) + assert blocked.value.code == "429" diff --git a/tests/test_litellm/proxy/proxy_server/conftest.py b/tests/test_litellm/proxy/proxy_server/conftest.py index c545965f9a9..cdfd3549544 100644 --- a/tests/test_litellm/proxy/proxy_server/conftest.py +++ b/tests/test_litellm/proxy/proxy_server/conftest.py @@ -511,3 +511,28 @@ def make_key( max_budget=max_budget, **kwargs, ) + + +@pytest.fixture(autouse=True) +def reset_login_throttle(monkeypatch): + """Clear the Admin UI failed-login counters between tests. + + `client` is session scoped and the counters live in the shared `user_api_key_cache` + with a 900s window, so without this any test that fails a sign-in enough times would + start returning 429 from unrelated tests later in the same process. Only the throttle's + own keys are removed, so nothing else in that cache is disturbed. + """ + from litellm.proxy import proxy_server as ps + from litellm.proxy.auth.login_throttle import _FAILED_LOGIN_CACHE, _CACHE_KEY_PREFIX + + def _drop_throttle_keys() -> None: + in_memory = getattr(_FAILED_LOGIN_CACHE, "in_memory_cache", None) + cache_dict = getattr(in_memory, "cache_dict", None) + if isinstance(cache_dict, dict): + for key in [k for k in cache_dict if str(k).startswith(_CACHE_KEY_PREFIX)]: + cache_dict.pop(key, None) + + monkeypatch.setattr(ps, "redis_usage_cache", None) + _drop_throttle_keys() + yield _drop_throttle_keys + _drop_throttle_keys() diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py index af37dbe85fe..faeb7750e03 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py @@ -10,6 +10,7 @@ Routes covered: from __future__ import annotations +from concurrent.futures import ThreadPoolExecutor from unittest.mock import AsyncMock, MagicMock import pytest @@ -29,7 +30,7 @@ def _install_login_mocks(monkeypatch, raise_on_auth: bool = False) -> None: """ from litellm.proxy import proxy_server as ps - async def _fake_auth(username, password, master_key, prisma_client): + async def _fake_auth(username, password, master_key, prisma_client, throttle=None): if raise_on_auth: raise Exception("boom-auth-failure") fake = MagicMock() @@ -460,3 +461,92 @@ def test_login_form_ignores_open_redirect_return_to(client, monkeypatch): location = response.headers.get("location", "") assert "evil.example.com" not in location assert "/ui" in location # dashboard fallback + + +# --------------------------------------------------------------------------- +# Failed-login accounting across the login routes (LIT-5285) +# --------------------------------------------------------------------------- + + +def _install_real_auth(monkeypatch, **settings): + """Run the real authenticate_user so the throttle inside it is exercised. + + prisma_client stays None, so every guess falls through to the credential rejection. + """ + from litellm.proxy import proxy_server as ps + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right-password") + monkeypatch.setattr(ps, "master_key", "sk-test-master") + monkeypatch.setattr(ps, "prisma_client", None) + monkeypatch.setattr(ps, "premium_user", False) + monkeypatch.setattr(ps, "general_settings", dict(settings)) + + +def _form_login(client, username="admin", password="wrong"): + return client.post( + "/login", data={"username": username, "password": password}, follow_redirects=False + ).status_code + + +def _json_login(client, path, username="admin", password="wrong"): + return client.post(path, json={"username": username, "password": password}).status_code + + +def test_budget_is_shared_across_every_login_endpoint(client, monkeypatch, reset_login_throttle): + """The endpoint is not part of the key, so spending the budget on one route blocks the rest. + + Partitioning the counter per endpoint would silently triple the real allowance. + """ + _install_real_auth( + monkeypatch, + max_failed_login_attempts=10, + control_plane_url="https://cp.example.com", + ) + + assert [_form_login(client) for _ in range(5)] == [401] * 5 + assert [_json_login(client, "/v2/login") for _ in range(5)] == [401] * 5 + + assert _json_login(client, "/v3/login") == 429, "the eleventh attempt must be refused on a third route" + + +def test_budget_is_shared_across_username_casing(client, monkeypatch, reset_login_throttle): + """The database lookup is case-insensitive, so casing must not partition the counter.""" + _install_real_auth(monkeypatch, max_failed_login_attempts=10) + + assert [_json_login(client, "/v2/login", username="admin@corp.com") for _ in range(5)] == [401] * 5 + assert [_json_login(client, "/v2/login", username="ADMIN@corp.com") for _ in range(5)] == [401] * 5 + + assert _json_login(client, "/v2/login", username="Admin@corp.com") == 429 + + +def test_a_refused_attempt_carries_retry_after(client, monkeypatch, reset_login_throttle): + """The 429 tells the caller how long the window has left.""" + _install_real_auth(monkeypatch, max_failed_login_attempts=2, failed_login_window_seconds=77) + + assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] + + refused = client.post("/v2/login", json={"username": "admin", "password": "wrong"}) + assert refused.status_code == 429 + assert refused.headers.get("retry-after") == "77" + + +def test_a_second_username_from_the_same_source_still_gets_through(client, monkeypatch, reset_login_throttle): + """The bucket is the username and source pair, so one account cannot block another.""" + _install_real_auth(monkeypatch, max_failed_login_attempts=2) + + for _ in range(3): + _json_login(client, "/v2/login", username="admin") + + assert _json_login(client, "/v2/login", username="someone-else@example.com") == 401 + + +def test_sign_in_succeeds_again_once_the_budget_is_restored(client, monkeypatch, reset_login_throttle): + """A cleared bucket lets the same username straight back in.""" + _install_real_auth(monkeypatch, max_failed_login_attempts=2) + + assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] + assert _json_login(client, "/v2/login") == 429 + + reset_login_throttle() + assert _json_login(client, "/v2/login") == 401 diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index afc42e8db45..ac2e84b8474 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -19,7 +19,6 @@ from fastapi import FastAPI from fastapi.staticfiles import StaticFiles from fastapi.testclient import TestClient - import litellm import litellm.proxy.proxy_server as proxy_server_module from litellm.caching.caching import RedisCache @@ -27,6 +26,7 @@ from litellm.caching.redis_cluster_cache import RedisClusterCache from litellm.litellm_core_utils.get_model_cost_map import ModelCostMapReloaded from litellm.caching.dual_cache import DualCache from litellm.proxy._types import LitellmUserRoles, UserAPIKeyAuth +from litellm.proxy.auth.login_throttle import LoginThrottle from litellm.proxy.auth.user_api_key_auth import user_api_key_auth from litellm.proxy.proxy_server import app, initialize from litellm.utils import _invalidate_model_cost_lowercase_map @@ -124,12 +124,13 @@ def test_login_v2_returns_redirect_url_and_sets_cookie(monkeypatch): } assert response.cookies.get("token") == "signed-token" - mock_authenticate_user.assert_awaited_once_with( - username="alice", - password="secret", - master_key="test-master-key", - prisma_client=mock_prisma_client, - ) + mock_authenticate_user.assert_awaited_once() + auth_kwargs = mock_authenticate_user.call_args.kwargs + assert auth_kwargs["username"] == "alice" + assert auth_kwargs["password"] == "secret" + assert auth_kwargs["master_key"] == "test-master-key" + assert auth_kwargs["prisma_client"] is mock_prisma_client + assert isinstance(auth_kwargs["throttle"], LoginThrottle), "the endpoint must thread a throttle through" mock_create_ui_token_object.assert_called_once_with( login_result=mock_login_result, general_settings={}, @@ -11424,3 +11425,27 @@ async def test_authoritative_floor_spend_keeps_a_reset_marker_written_during_the assert real_spend_counter_cache.in_memory_cache.get_cache(key=marker_key) == 0.0, ( "the in-flight DB read clobbered the post-reset floor marker with the stale pre-reset value" ) + + +@pytest.mark.asyncio +async def test_login_throttle_settings_are_not_overridable_from_the_database(): + """LIT-5285: the sign-in limits stay config.yaml only. + + _update_general_settings copies an allowlist of keys out of the DB row. Adding these + to it would let a stored value outrank config.yaml, so an operator refused by a bad + value could not fix it by editing YAML and restarting. + """ + import litellm.proxy.proxy_server as ps + from litellm.proxy.proxy_server import ProxyConfig + + original = dict(ps.general_settings) + try: + ps.general_settings.clear() + await ProxyConfig()._update_general_settings( + db_general_settings={"max_failed_login_attempts": 999, "failed_login_window_seconds": 1} + ) + assert "max_failed_login_attempts" not in ps.general_settings + assert "failed_login_window_seconds" not in ps.general_settings + finally: + ps.general_settings.clear() + ps.general_settings.update(original) diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index e0b8cf19159..bb2841b6388 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -24150,6 +24150,11 @@ export interface components { * @default false */ enable_public_model_hub: boolean; + /** + * Failed Login Window Seconds + * @description Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Configurable from config.yaml only. Defaults to 900 + */ + failed_login_window_seconds?: number | null; /** * Forward Client Headers To Llm Api * @description If True, forwards client headers (e.g. Authorization) to the LLM API. Required for Claude Code with Max subscription. @@ -24194,6 +24199,11 @@ export interface components { * @description max batch input file size in MB for /v1/files uploads with purpose=batch, if a file is larger than this size it will be rejected before being forwarded to the provider */ max_batch_file_size_mb?: number | null; + /** + * Max Failed Login Attempts + * @description Number of failed Admin UI sign-in attempts allowed for one username from one source address within `failed_login_window_seconds`, before further attempts are refused with 429. Configurable from config.yaml only. Defaults to 10 + */ + max_failed_login_attempts?: number | null; /** * Max Parallel Requests * @description maximum parallel requests for each api key From a39efb1a1d59fbf122e8ddfc5929f256bda9f9c0 Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Wed, 26 Aug 2026 10:24:32 -0700 Subject: [PATCH 003/141] test(proxy): clear failed-login TTLs when resetting the throttle between tests --- tests/test_litellm/proxy/proxy_server/conftest.py | 15 +++++++++++---- 1 file changed, 11 insertions(+), 4 deletions(-) diff --git a/tests/test_litellm/proxy/proxy_server/conftest.py b/tests/test_litellm/proxy/proxy_server/conftest.py index cdfd3549544..22f2e5b3feb 100644 --- a/tests/test_litellm/proxy/proxy_server/conftest.py +++ b/tests/test_litellm/proxy/proxy_server/conftest.py @@ -527,10 +527,17 @@ def reset_login_throttle(monkeypatch): def _drop_throttle_keys() -> None: in_memory = getattr(_FAILED_LOGIN_CACHE, "in_memory_cache", None) - cache_dict = getattr(in_memory, "cache_dict", None) - if isinstance(cache_dict, dict): - for key in [k for k in cache_dict if str(k).startswith(_CACHE_KEY_PREFIX)]: - cache_dict.pop(key, None) + if in_memory is None: + return + tracked = tuple( + key + for store in (getattr(in_memory, "cache_dict", None), getattr(in_memory, "ttl_dict", None)) + if isinstance(store, dict) + for key in tuple(store) + if str(key).startswith(_CACHE_KEY_PREFIX) + ) + for key in tracked: + in_memory.delete_cache(key) monkeypatch.setattr(ps, "redis_usage_cache", None) _drop_throttle_keys() From 2d33c949cb15a6007793bd2aeed42ec52d3d9fce Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Tue, 1 Sep 2026 13:03:43 -0700 Subject: [PATCH 004/141] feat(proxy): harden Admin UI login throttling --- litellm/proxy/_types.py | 7 +- litellm/proxy/auth/login_throttle.py | 236 ++++++++++---- litellm/proxy/auth/login_utils.py | 17 +- litellm/proxy/proxy_cli.py | 4 + litellm/proxy/proxy_server.py | 43 ++- .../proxy/auth/test_login_utils.py | 294 ++++++++++++++++-- .../proxy/proxy_server/conftest.py | 45 ++- .../proxy_server/test_routes_login_sso.py | 41 ++- tests/test_litellm/proxy/test_proxy_server.py | 7 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 7 +- 10 files changed, 581 insertions(+), 120 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 1fd2781bd95..0071ad4603d 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2528,7 +2528,12 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): max_failed_login_attempts: int | None = Field( None, ge=1, - description="Number of failed Admin UI sign-in attempts allowed for one username from one source address within `failed_login_window_seconds`, before further attempts are refused with 429. Configurable from config.yaml only. Defaults to 10", + description="Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Configurable from config.yaml only. Defaults to 50", + ) + max_failed_login_attempts_per_source: int | None = Field( + None, + ge=1, + description="Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Configurable from config.yaml only. Defaults to 250", ) failed_login_window_seconds: int | None = Field( None, diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index e2befd15a35..ed53d9e7bc7 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -1,16 +1,20 @@ """Failed-login accounting for the Admin UI sign-in path. -Counts failed credential checks per (username, source address) over a fixed window and -denies further attempts with 429 once the count reaches the limit. Built per request by -``LoginThrottle.from_request`` because it carries that request's resolved source address, -and because the coordination cache is assigned at startup and can be reassigned later. +Counts failed credential checks over a fixed window against two independent keys, the +username on its own and the source address on its own, so that one username attacked from +many sources and one source spraying many usernames are both counted. Repeated failures +are answered slowly, doubling from one second, and refused with 429 once either counter +reaches its limit. Built per request by ``LoginThrottle.from_request`` because it carries +that request's resolved source address, and because the coordination cache is assigned at +startup and can be reassigned later. """ +import asyncio import hashlib from collections.abc import Awaitable from dataclasses import dataclass from types import MappingProxyType -from typing import Final +from typing import Final, NamedTuple, NoReturn from fastapi import Request @@ -23,21 +27,53 @@ from litellm.proxy.auth.network import TrustedProxyConfig, normalize_cidr_ranges from litellm.proxy.auth.trusted_proxy_utils import TRUSTED_PROXY_RANGES_KEY from litellm.secret_managers.main import get_secret_bool -DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS: Final = 10 +DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS: Final = 50 +DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE: Final = 250 DEFAULT_FAILED_LOGIN_WINDOW_SECONDS: Final = 900 +USERNAME_DELAY_ONSET: Final = 3 +SOURCE_DELAY_ONSET: Final = 25 +FIRST_DELAY_SECONDS: Final = 1.0 +MAX_DELAY_SECONDS: Final = 30.0 +MAX_CONCURRENT_DELAYS_PER_SOURCE: Final = 5 + +_MAX_DELAY_DOUBLINGS: Final = 16 + _CACHE_KEY_PREFIX: Final = "login_fail" _UNKNOWN_SOURCE: Final = "unknown" _MAX_LOGGED_USERNAME_CHARS: Final = 128 +_MAX_TRACKED_LOGIN_USERNAMES: Final = 10_000 _MAX_TRACKED_LOGIN_SOURCES: Final = 10_000 -_FAILED_LOGIN_CACHE: Final = DualCache( - in_memory_cache=InMemoryCache(max_size_in_memory=_MAX_TRACKED_LOGIN_SOURCES), - default_in_memory_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS, -) + +def _bounded_store(max_entries: int) -> DualCache: + return DualCache( + in_memory_cache=InMemoryCache(max_size_in_memory=max_entries), + default_in_memory_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS, + ) + + +# Separate stores: eviction is earliest-expiring-first, so in one shared store a spray of +# fresh usernames would evict the source counter that is meant to stop that same spray. +_FAILED_LOGIN_USERNAME_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_USERNAMES) +_FAILED_LOGIN_SOURCE_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_SOURCES) _NO_SETTINGS: Final = MappingProxyType({}) +_DELAYS_IN_FLIGHT: Final[dict[str, int]] = {} + + +async def _sleep(seconds: float) -> None: + """The wait a rejected sign-in is held for. Replaced in tests so the suite pays no wall clock.""" + await asyncio.sleep(seconds) + + +class FailureCounts(NamedTuple): + """Failures recorded so far in this window against each of the two keys.""" + + username: int + source: int + def _int_setting(name: str, value: object, default: int, minimum: int) -> int: if value is None: @@ -56,12 +92,14 @@ def _as_count(cached: object) -> int: @dataclass(frozen=True, slots=True) class LoginThrottle: - """Fixed-window failed-login accounting for one request's source address.""" + """Fixed-window failed-login accounting for one request's username and source address.""" client_ip: str max_attempts: int + max_attempts_per_source: int window_seconds: int - cache: DualCache + username_cache: DualCache + source_cache: DualCache redis_cache: RedisCache | None = None enabled: bool = True @@ -85,13 +123,20 @@ class LoginThrottle: DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS, 1, ), + max_attempts_per_source=_int_setting( + "max_failed_login_attempts_per_source", + settings.get("max_failed_login_attempts_per_source"), + DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE, + 1, + ), window_seconds=_int_setting( "failed_login_window_seconds", settings.get("failed_login_window_seconds"), DEFAULT_FAILED_LOGIN_WINDOW_SECONDS, 1, ), - cache=_FAILED_LOGIN_CACHE, + username_cache=_FAILED_LOGIN_USERNAME_CACHE, + source_cache=_FAILED_LOGIN_SOURCE_CACHE, redis_cache=redis_usage_cache, enabled=not get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT"), ) @@ -101,9 +146,13 @@ class LoginThrottle: """The username with anything that could forge a log line removed.""" return "".join(c for c in username if c.isprintable())[:_MAX_LOGGED_USERNAME_CHARS] - def _key(self, username: str) -> str: + @staticmethod + def _username_key(username: str) -> str: identity: Final = hashlib.sha256(username.casefold().encode("utf-8")).hexdigest() - return f"{_CACHE_KEY_PREFIX}:{identity}:{self.client_ip}" + return f"{_CACHE_KEY_PREFIX}:user:{identity}" + + def _source_key(self) -> str: + return f"{_CACHE_KEY_PREFIX}:source:{self.client_ip}" async def _outcome(self, work: Awaitable[object]) -> object: try: @@ -112,66 +161,147 @@ class LoginThrottle: verbose_proxy_logger.warning("login attempt accounting unavailable: %s", exc) return None - async def _failures(self, key: str) -> int: - store: Final = self.cache if self.redis_cache is None else self.redis_cache - return _as_count(await self._outcome(store.async_get_cache(key=key))) + async def _failures(self, store: DualCache, key: str) -> int: + """The larger of the shared and the process-local count, so a Redis outage degrades + to per-worker accounting instead of switching the control off.""" + local: Final = _as_count(await self._outcome(store.async_get_cache(key=key))) + redis_cache: Final = self.redis_cache + if redis_cache is None: + return local + return max(local, _as_count(await self._outcome(redis_cache.async_get_cache(key)))) - async def _ensure_expiry(self, key: str) -> None: - """Give the counter an expiry if it somehow has none. + async def _remaining_window(self, key: str) -> int: + """Seconds until this counter expires, repairing a counter left without an expiry. Redis commits the increment before setting the TTL, so a failure in between can leave a counter that never expires. Nothing increments the key again once the - limit is reached, so without this the pair would stay refused indefinitely. + limit is reached, so without the repair the key would stay refused indefinitely. """ redis_cache: Final = self.redis_cache if redis_cache is None: - return - if isinstance(await self._outcome(redis_cache.async_get_ttl(key)), int): - return + return self.window_seconds + ttl: Final = await self._outcome(redis_cache.async_get_ttl(key)) + if isinstance(ttl, int) and ttl > 0: + return min(ttl, self.window_seconds) await self._outcome(redis_cache.async_increment(key, 0, ttl=self.window_seconds)) + return self.window_seconds - async def raise_if_blocked(self, username: str) -> None: - """Deny before the database lookup and before the password comparison.""" - if not self.enabled: - return - key: Final = self._key(username) - if await self._failures(key) < self.max_attempts: - return - await self._ensure_expiry(key) - verbose_proxy_logger.warning( - "Admin UI sign-in attempts exhausted for username=%s source=%s; %s attempts in %ss, retry after %ss", - self._loggable(username), - self.client_ip, - self.max_attempts, - self.window_seconds, - self.window_seconds, - ) - raise ProxyException( + def _refused(self, retry_after: int, param: str) -> ProxyException: + return ProxyException( message="Too many failed sign-in attempts. Try again later.", type=ProxyErrorTypes.auth_error, - param="max_failed_login_attempts", + param=param, code=429, - headers={"Retry-After": str(self.window_seconds)}, # mutable-ok: ProxyException coerces header values + headers={"Retry-After": str(retry_after)}, # mutable-ok: ProxyException coerces header values ) - async def record_failure(self, username: str) -> None: - """Count one rejected credential guess against this username and source.""" + async def _refuse(self, key: str, scope: str, param: str, username: str, failures: int, limit: int) -> NoReturn: + retry_after: Final = await self._remaining_window(key) + verbose_proxy_logger.warning( + "Admin UI sign-in attempts exhausted for %s; username=%s source=%s failures=%s limit=%s window=%ss", + scope, + self._loggable(username), + self.client_ip, + failures, + limit, + self.window_seconds, + ) + raise self._refused(retry_after, param) + + async def raise_if_blocked(self, username: str) -> None: + """Refuse before the database lookup and before the invite-link password hash.""" if not self.enabled: return - key: Final = self._key(username) + username_key: Final = self._username_key(username) + source_key: Final = self._source_key() + username_failures: Final = await self._failures(self.username_cache, username_key) + if username_failures >= self.max_attempts: + await self._refuse( + username_key, "username", "max_failed_login_attempts", username, username_failures, self.max_attempts + ) + source_failures: Final = await self._failures(self.source_cache, source_key) + if source_failures >= self.max_attempts_per_source: + await self._refuse( + source_key, + "source address", + "max_failed_login_attempts_per_source", + username, + source_failures, + self.max_attempts_per_source, + ) + + async def _bump(self, store: DualCache, key: str) -> int: + local: Final = _as_count( + await self._outcome(store.async_increment_cache(key=key, value=1, ttl=self.window_seconds)) + ) redis_cache: Final = self.redis_cache - if redis_cache is not None: - await self._outcome(redis_cache.async_increment(key, 1, ttl=self.window_seconds)) - await self._ensure_expiry(key) + if redis_cache is None: + return local + shared: Final = _as_count(await self._outcome(redis_cache.async_increment(key, 1, ttl=self.window_seconds))) + await self._remaining_window(key) + return max(local, shared) + + async def record_failure(self, username: str) -> FailureCounts: + """Count one rejected credential guess against this username and against this source.""" + if not self.enabled: + return FailureCounts(username=0, source=0) + return FailureCounts( + username=await self._bump(self.username_cache, self._username_key(username)), + source=await self._bump(self.source_cache, self._source_key()), + ) + + @staticmethod + def delay_seconds(counts: FailureCounts) -> float: + """Seconds to hold a rejected attempt for, doubling per failure past whichever onset is further along.""" + steps: Final = min( + max(counts.username - USERNAME_DELAY_ONSET, counts.source - SOURCE_DELAY_ONSET), + _MAX_DELAY_DOUBLINGS, + ) + if steps < 0: + return 0.0 + return min(FIRST_DELAY_SECONDS * float(2**steps), MAX_DELAY_SECONDS) + + async def delay_for(self, username: str, counts: FailureCounts) -> None: + """Hold this rejected attempt open before answering it, so guessing costs wall-clock time. + + Only ever reached once the credentials are known to be wrong, so a valid password is + never delayed. Sources are capped at ``MAX_CONCURRENT_DELAYS_PER_SOURCE`` held + connections; over that, the attempt is refused immediately instead of parking a socket. + """ + if not self.enabled: return - await self._outcome(self.cache.async_increment_cache(key=key, value=1, ttl=self.window_seconds)) + delay: Final = self.delay_seconds(counts) + if delay <= 0: + return + in_flight: Final = _DELAYS_IN_FLIGHT.get(self.client_ip, 0) + if in_flight >= MAX_CONCURRENT_DELAYS_PER_SOURCE: + verbose_proxy_logger.warning( + "Admin UI sign-in attempts held concurrently exhausted; username=%s source=%s in_flight=%s", + self._loggable(username), + self.client_ip, + in_flight, + ) + raise self._refused(int(MAX_DELAY_SECONDS), "concurrent_failed_logins") + _DELAYS_IN_FLIGHT[self.client_ip] = in_flight + 1 + try: + await _sleep(delay) + finally: + remaining: Final = _DELAYS_IN_FLIGHT.get(self.client_ip, 1) - 1 + if remaining > 0: + _DELAYS_IN_FLIGHT[self.client_ip] = remaining + else: + _DELAYS_IN_FLIGHT.pop(self.client_ip, None) async def clear(self, username: str) -> None: - """Drop the bucket after a successful sign-in.""" + """Drop the username counter after a successful sign-in. + + The source counter is left alone. It is shared by every account behind that address, + so one success there says nothing about the other attempts it is counting. + """ if not self.enabled: return - key: Final = self._key(username) + key: Final = self._username_key(username) redis_cache: Final = self.redis_cache if redis_cache is not None: await self._outcome(redis_cache.async_delete_cache(key)) - await self._outcome(self.cache.async_delete_cache(key=key)) + await self._outcome(self.username_cache.async_delete_cache(key=key)) diff --git a/litellm/proxy/auth/login_utils.py b/litellm/proxy/auth/login_utils.py index 98780a4692a..8835227cd1c 100644 --- a/litellm/proxy/auth/login_utils.py +++ b/litellm/proxy/auth/login_utils.py @@ -145,10 +145,15 @@ async def authenticate_user( code=500, ) - await throttle.raise_if_blocked(username) - ui_username, ui_password = get_ui_credentials(master_key) + admin_credentials_match: Final = secrets.compare_digest( + username.encode("utf-8"), ui_username.encode("utf-8") + ) and secrets.compare_digest(password.encode("utf-8"), ui_password.encode("utf-8")) + + if not admin_credentials_match: + await throttle.raise_if_blocked(username) + # Check if we can find the `username` in the db. On the UI, users can enter username=their email _user_row: LiteLLM_UserTable | None = None user_role: ( @@ -174,9 +179,7 @@ async def authenticate_user( - Login with UI_USERNAME and UI_PASSWORD - Login with Invite Link `user_email` and `password` combination """ - if secrets.compare_digest(username.encode("utf-8"), ui_username.encode("utf-8")) and secrets.compare_digest( - password.encode("utf-8"), ui_password.encode("utf-8") - ): + if admin_credentials_match: # Non SSO -> If user is using UI_USERNAME and UI_PASSWORD they are Proxy admin user_role = LitellmUserRoles.PROXY_ADMIN user_id = LITELLM_PROXY_ADMIN_NAME @@ -314,7 +317,7 @@ async def authenticate_user( login_method="username_password", ) else: - await throttle.record_failure(username) + await throttle.delay_for(username, await throttle.record_failure(username)) raise ProxyException( message=INVALID_UI_CREDENTIALS_MESSAGE, type=ProxyErrorTypes.auth_error, @@ -322,7 +325,7 @@ async def authenticate_user( code=401, ) else: - await throttle.record_failure(username) + await throttle.delay_for(username, await throttle.record_failure(username)) raise ProxyException( message=INVALID_UI_CREDENTIALS_MESSAGE, type=ProxyErrorTypes.auth_error, diff --git a/litellm/proxy/proxy_cli.py b/litellm/proxy/proxy_cli.py index 0449802abae..a2e3a649586 100644 --- a/litellm/proxy/proxy_cli.py +++ b/litellm/proxy/proxy_cli.py @@ -1358,6 +1358,10 @@ def run_server( # DO NOT DELETE - enables global variables to work across files from litellm.proxy.proxy_server import app + # Write the resolved --num_workers back to its env var so worker processes can read + # the fleet size at startup (the failed-login accounting warning keys off it) + os.environ["NUM_WORKERS"] = str(num_workers) + # Auto-create PROMETHEUS_MULTIPROC_DIR for multi-worker setups ProxyInitializationHelpers._maybe_setup_prometheus_multiproc_dir( num_workers=num_workers, diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 39a419df73f..c6a1dc4cdb7 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -2264,6 +2264,7 @@ user_custom_key_generate = None # Tests that need to reset it can patch 'litellm.proxy.proxy_server._pkce_no_redis_warning_emitted'. _pkce_no_redis_warning_emitted: bool = False _cp_no_redis_warning_emitted: bool = False +_login_throttle_no_redis_warning_emitted: bool = False user_custom_key_update = None user_custom_sso = None user_custom_ui_sso_sign_in_handler = None @@ -5317,6 +5318,21 @@ class ProxyConfig: "or ensure sticky sessions for single-instance deployments." ) + ### FAILED-LOGIN ACCOUNTING MULTI-INSTANCE PREREQUISITE CHECK ### + # Failed Admin UI sign-in counters live in redis_usage_cache when available so a + # brute-force run is counted once across workers instead of once per worker. + if os.getenv("NUM_WORKERS", "1") != "1" and redis_usage_cache is None: + global _login_throttle_no_redis_warning_emitted + if not _login_throttle_no_redis_warning_emitted: + _login_throttle_no_redis_warning_emitted = True + verbose_proxy_logger.warning( + "Running %s workers but Redis is not configured for LiteLLM caching. " + "Failed Admin UI sign-in attempts are counted per worker, so an attacker " + "gets max_failed_login_attempts guesses per worker instead of overall. " + "Configure Redis via the 'cache' section in your proxy config.", + os.getenv("NUM_WORKERS", "1"), + ) + ### STORE MODEL IN DB ### feature flag for `/model/new` store_model_in_db = general_settings.get("store_model_in_db", False) if store_model_in_db is None: @@ -14756,13 +14772,26 @@ async def login(request: Request): password: Final = str(form.get("password")) # Authenticate user and get login result - login_result: Final = await authenticate_user( - username=username, - password=password, - master_key=master_key, - prisma_client=prisma_client, - throttle=LoginThrottle.from_request(request), - ) + try: + login_result: Final = await authenticate_user( + username=username, + password=password, + master_key=master_key, + prisma_client=prisma_client, + throttle=LoginThrottle.from_request(request), + ) + except ProxyException as exc: + if int(exc.code) != status.HTTP_429_TOO_MANY_REQUESTS: + raise + retry_after: Final = exc.headers.get("Retry-After", "30") + return HTMLResponse( + content=( + "

Too many sign-in attempts

" + f"

Try again in about {retry_after} seconds

" + ), + status_code=status.HTTP_429_TOO_MANY_REQUESTS, + headers=exc.headers, + ) # Create UI token object returned_ui_token_object: Final = create_ui_token_object( diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 1e3fcaf5d78..1b85c9f4046 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -6,17 +6,48 @@ to login_utils.py for better reusability. """ import os +from typing import Final from unittest.mock import AsyncMock, MagicMock, patch import pytest +class _RecordedSleeps: + """A sleep that records what it was asked to wait for instead of waiting.""" + + def __init__(self): + self.seconds: list[float] = [] + + async def __call__(self, seconds: float) -> None: + self.seconds.append(seconds) + + +@pytest.fixture(autouse=True) +def login_delays(monkeypatch): + """Replace the failed-login wait, so the suite pays no wall clock and can read it back.""" + from litellm.proxy.auth import login_throttle + + recorded = _RecordedSleeps() + monkeypatch.setattr(login_throttle, "_sleep", recorded) + login_throttle._DELAYS_IN_FLIGHT.clear() + yield recorded + login_throttle._DELAYS_IN_FLIGHT.clear() + + def _unlimited_throttle(): """A throttle wired to a real in-memory store with a limit no test can reach.""" from litellm.caching.dual_cache import DualCache from litellm.proxy.auth.login_throttle import LoginThrottle - return LoginThrottle(client_ip="1.2.3.4", max_attempts=10_000, window_seconds=900, cache=DualCache()) + store: Final = DualCache() + return LoginThrottle( + client_ip="1.2.3.4", + max_attempts=10_000, + max_attempts_per_source=10_000, + window_seconds=900, + username_cache=store, + source_cache=store, + ) @@ -628,16 +659,26 @@ class TestEncodeUiSessionJwt: # --------------------------------------------------------------------------- -def _throttle(max_attempts: int = 3, window_seconds: int = 900, client_ip: str = "1.2.3.4", cache=None, redis_cache=None): +def _throttle( + max_attempts: int = 3, + window_seconds: int = 900, + client_ip: str = "1.2.3.4", + cache=None, + redis_cache=None, + max_attempts_per_source: int = 10_000, +): """A throttle over a real in-memory store, so the tests exercise the true counters.""" from litellm.caching.dual_cache import DualCache from litellm.proxy.auth.login_throttle import LoginThrottle + store: Final = cache if cache is not None else DualCache() return LoginThrottle( client_ip=client_ip, max_attempts=max_attempts, + max_attempts_per_source=max_attempts_per_source, window_seconds=window_seconds, - cache=cache if cache is not None else DualCache(), + username_cache=store, + source_cache=store, redis_cache=redis_cache, ) @@ -675,21 +716,32 @@ async def test_attempts_are_refused_once_the_limit_is_reached(monkeypatch): @pytest.mark.asyncio -async def test_a_correct_password_is_refused_while_blocked(monkeypatch): - """The check precedes the credential comparison, so being over the limit wins.""" +async def test_a_correct_admin_password_is_accepted_while_blocked(monkeypatch): + """The configured admin credentials are compared before the gate, so the operator gets in. + + A throttle that refuses a valid password hands anyone who can reach the login form a + denial of service against the one account that can fix it. + """ from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") throttle = _throttle(max_attempts=2) for _ in range(2): with pytest.raises(ProxyException): await _guess(throttle) - with pytest.raises(ProxyException) as blocked: - await _guess(throttle, password="right") - assert blocked.value.code == "429" + with pytest.raises(ProxyException) as still_blocked: + await _guess(throttle) + assert still_blocked.value.code == "429", "a wrong password is still refused" + + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ): + result = await _guess(throttle, password="right") + assert result.key == "sk-ui" @pytest.mark.asyncio @@ -700,18 +752,18 @@ async def test_a_blocked_attempt_does_not_extend_the_window(monkeypatch): monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") throttle = _throttle(max_attempts=2) - key = throttle._key("admin") + key = throttle._username_key("admin") for _ in range(2): with pytest.raises(ProxyException): await _guess(throttle) - counted_at_limit = await throttle._failures(key) + counted_at_limit = await throttle._failures(throttle.username_cache, key) for _ in range(5): with pytest.raises(ProxyException): await _guess(throttle) - assert await throttle._failures(key) == counted_at_limit == 2 + assert await throttle._failures(throttle.username_cache, key) == counted_at_limit == 2 @pytest.mark.asyncio @@ -733,7 +785,7 @@ async def test_a_successful_sign_in_clears_the_bucket(monkeypatch): ): await _guess(throttle, password="right") - assert await throttle._failures(throttle._key("admin")) == 0 + assert await throttle._failures(throttle.username_cache, throttle._username_key("admin")) == 0 @pytest.mark.asyncio @@ -749,7 +801,7 @@ async def test_a_configuration_error_never_counts(monkeypatch): ) assert exc.value.code == "500" - assert await throttle._failures(throttle._key("admin")) == 0 + assert await throttle._failures(throttle.username_cache, throttle._username_key("admin")) == 0 @pytest.mark.asyncio @@ -773,7 +825,7 @@ async def test_the_username_is_case_folded_into_one_bucket(monkeypatch): @pytest.mark.asyncio async def test_a_different_username_from_the_same_source_is_unaffected(monkeypatch): - """The bucket is the pair, so one username's failures do not block another.""" + """The counters are independent, so one username's failures do not exhaust another's.""" from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") @@ -853,7 +905,7 @@ async def test_a_user_with_no_password_set_does_not_consume_the_budget(monkeypat ) assert exc.value.code == "401" - assert await throttle._failures(throttle._key("nopass@example.com")) == 0 + assert await throttle._failures(throttle.username_cache, throttle._username_key("nopass@example.com")) == 0 @pytest.mark.asyncio @@ -896,11 +948,40 @@ async def test_a_wrong_password_for_a_known_user_also_counts(monkeypatch): @pytest.mark.asyncio -async def test_two_source_addresses_do_not_share_a_bucket(monkeypatch): - """The key is the pair, so one address exhausting its budget must not block another. +async def test_one_source_exhausting_its_own_budget_does_not_refuse_another_source(monkeypatch): + """The source counter is per address, so a noisy office does not take its neighbour down.""" + from litellm.caching.dual_cache import DualCache + from litellm.proxy._types import ProxyException - Dropping the address from the key would turn this into the username-only counter the - design rejects, where anyone can lock a named admin out from anywhere. + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + shared_store = DualCache() + attacker = _throttle( + max_attempts=10_000, max_attempts_per_source=2, client_ip="203.0.113.9", cache=shared_store + ) + operator = _throttle( + max_attempts=10_000, max_attempts_per_source=2, client_ip="198.51.100.7", cache=shared_store + ) + + for i in range(2): + with pytest.raises(ProxyException): + await _guess(attacker, username=f"target-{i}@corp.com") + + with pytest.raises(ProxyException) as blocked: + await _guess(attacker, username="target-2@corp.com") + assert blocked.value.code == "429" + + with pytest.raises(ProxyException) as unaffected: + await _guess(operator, username="target-3@corp.com") + assert unaffected.value.code == "401", "the other address must still reach the credential check" + + +@pytest.mark.asyncio +async def test_a_username_exhausted_from_one_source_is_refused_from_another(monkeypatch): + """The username counter carries no address, so spreading the guesses buys nothing. + + The pair key this replaced reset the budget for every new address, which is exactly the + shape of a credential-stuffing run from a proxy pool. """ from litellm.caching.dual_cache import DualCache from litellm.proxy._types import ProxyException @@ -908,20 +989,154 @@ async def test_two_source_addresses_do_not_share_a_bucket(monkeypatch): monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") shared_store = DualCache() - attacker = _throttle(max_attempts=2, client_ip="203.0.113.9", cache=shared_store) - operator = _throttle(max_attempts=2, client_ip="198.51.100.7", cache=shared_store) + first_hop = _throttle(max_attempts=2, client_ip="203.0.113.9", cache=shared_store) + second_hop = _throttle(max_attempts=2, client_ip="198.51.100.7", cache=shared_store) - for _ in range(3): + for _ in range(2): with pytest.raises(ProxyException): - await _guess(attacker, username="admin") + await _guess(first_hop, username="victim@corp.com") + + with pytest.raises(ProxyException) as rotated: + await _guess(second_hop, username="victim@corp.com") + assert rotated.value.code == "429" + + +@pytest.mark.asyncio +async def test_a_source_wide_spray_is_counted_even_though_each_username_is_fresh(monkeypatch): + """One guess against each of many usernames never trips a username counter, only the source one.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=10_000, max_attempts_per_source=6, client_ip="203.0.113.11") + + for i in range(6): + with pytest.raises(ProxyException) as rejected: + await _guess(throttle, username=f"sprayed-{i}@corp.com") + assert rejected.value.code == "401" with pytest.raises(ProxyException) as blocked: - await _guess(attacker, username="admin") + await _guess(throttle, username="sprayed-7@corp.com") assert blocked.value.code == "429" + assert await throttle._failures(throttle.username_cache, throttle._username_key("sprayed-7@corp.com")) == 0 - with pytest.raises(ProxyException) as unaffected: - await _guess(operator, username="admin") - assert unaffected.value.code == "401", "the real operator must still reach the credential check" + +@pytest.mark.asyncio +async def test_a_successful_sign_in_leaves_the_source_counter_alone(monkeypatch): + """One account's success says nothing about the other attempts the address is making.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + throttle = _throttle(max_attempts=10) + + for _ in range(2): + with pytest.raises(ProxyException): + await _guess(throttle) + + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ): + await _guess(throttle, password="right") + + assert await throttle._failures(throttle.username_cache, throttle._username_key("admin")) == 0 + assert await throttle._failures(throttle.source_cache, throttle._source_key()) == 2 + + +@pytest.mark.asyncio +async def test_the_delay_doubles_from_one_second_and_is_capped(monkeypatch, login_delays): + """Guessing has to cost wall clock, and the cost has to stop short of an unbounded hang.""" + from litellm.proxy._types import ProxyException + from litellm.proxy.auth.login_throttle import MAX_DELAY_SECONDS + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=10_000) + + for _ in range(9): + with pytest.raises(ProxyException): + await _guess(throttle) + + assert login_delays.seconds == [1.0, 2.0, 4.0, 8.0, 16.0, 30.0, 30.0], ( + "the first two failures answer immediately, then the wait doubles up to the cap" + ) + assert max(login_delays.seconds) == MAX_DELAY_SECONDS + + +@pytest.mark.asyncio +async def test_the_delay_tracks_whichever_counter_is_further_past_its_onset(monkeypatch): + """A source deep into a spray must not be answered instantly just because the username is fresh.""" + from litellm.proxy.auth.login_throttle import FailureCounts, LoginThrottle + + assert LoginThrottle.delay_seconds(FailureCounts(username=1, source=1)) == 0.0 + assert LoginThrottle.delay_seconds(FailureCounts(username=2, source=24)) == 0.0 + assert LoginThrottle.delay_seconds(FailureCounts(username=3, source=1)) == 1.0 + assert LoginThrottle.delay_seconds(FailureCounts(username=1, source=25)) == 1.0 + assert LoginThrottle.delay_seconds(FailureCounts(username=4, source=28)) == 8.0 + + +@pytest.mark.asyncio +async def test_held_attempts_from_one_source_are_capped(monkeypatch): + """Holding a rejected attempt open must not let one address park unlimited sockets.""" + import asyncio + + from litellm.proxy._types import ProxyException + from litellm.proxy.auth import login_throttle as lt + from litellm.proxy.auth.login_throttle import MAX_CONCURRENT_DELAYS_PER_SOURCE + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + release = asyncio.Event() + + async def _park(_seconds: float) -> None: + await release.wait() + + monkeypatch.setattr(lt, "_sleep", _park) + throttle = _throttle(max_attempts=10_000, client_ip="203.0.113.44") + await throttle.record_failure("admin") + await throttle.record_failure("admin") + + held = [asyncio.create_task(_guess(throttle)) for _ in range(MAX_CONCURRENT_DELAYS_PER_SOURCE)] + for _ in range(1000): + if lt._DELAYS_IN_FLIGHT.get("203.0.113.44") == MAX_CONCURRENT_DELAYS_PER_SOURCE: + break + await asyncio.sleep(0) + assert lt._DELAYS_IN_FLIGHT.get("203.0.113.44") == MAX_CONCURRENT_DELAYS_PER_SOURCE + + try: + with pytest.raises(ProxyException) as over_cap: + await _guess(throttle) + assert over_cap.value.code == "429" + assert over_cap.value.headers.get("Retry-After") == "30" + finally: + release.set() + for task in held: + with pytest.raises(ProxyException): + await task + + with pytest.raises(ProxyException) as after_drain: + await _guess(throttle) + assert after_drain.value.code == "401", "the cap must release once the held attempts answer" + + +@pytest.mark.asyncio +async def test_disabling_the_control_removes_the_delay_as_well(monkeypatch, login_delays): + """The escape hatch has to turn off the whole control, not only the refusal.""" + import dataclasses + + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = dataclasses.replace(_throttle(max_attempts=2), enabled=False) + + for _ in range(6): + with pytest.raises(ProxyException) as rejected: + await _guess(throttle) + assert rejected.value.code == "401" + + assert login_delays.seconds == [] class _NoExpiryRedis: @@ -1003,7 +1218,7 @@ async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): throttle = LoginThrottle.from_request(request) for i in range(25): - with pytest.raises(Exception): + with pytest.raises(ProxyException, match="Invalid credentials"): await _guess(throttle, username=f"made-up-{i}@example.com") added = set(ps.user_api_key_cache.in_memory_cache.cache_dict) - auth_cache_keys_before @@ -1055,14 +1270,29 @@ async def test_a_username_spray_cannot_evict_an_existing_counter(monkeypatch): back a fresh allowance against the real account. """ from litellm.proxy._types import ProxyException - from litellm.proxy.auth.login_throttle import _FAILED_LOGIN_CACHE, _MAX_TRACKED_LOGIN_SOURCES + from litellm.proxy.auth.login_throttle import ( + LoginThrottle, + _FAILED_LOGIN_SOURCE_CACHE, + _FAILED_LOGIN_USERNAME_CACHE, + _MAX_TRACKED_LOGIN_SOURCES, + _MAX_TRACKED_LOGIN_USERNAMES, + ) monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") assert _MAX_TRACKED_LOGIN_SOURCES >= 10_000 - assert _FAILED_LOGIN_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_SOURCES + assert _MAX_TRACKED_LOGIN_USERNAMES >= 10_000 + assert _FAILED_LOGIN_SOURCE_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_SOURCES + assert _FAILED_LOGIN_USERNAME_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_USERNAMES - throttle = _throttle(max_attempts=3, cache=_FAILED_LOGIN_CACHE, client_ip="10.9.9.9") + throttle = LoginThrottle( + client_ip="10.9.9.9", + max_attempts=3, + max_attempts_per_source=10_000, + window_seconds=900, + username_cache=_FAILED_LOGIN_USERNAME_CACHE, + source_cache=_FAILED_LOGIN_SOURCE_CACHE, + ) victim = "spray-victim@corp.com" for _ in range(3): with pytest.raises(ProxyException): @@ -1071,7 +1301,7 @@ async def test_a_username_spray_cannot_evict_an_existing_counter(monkeypatch): for i in range(500): await throttle.record_failure(f"spray-filler-{i}@corp.com") - assert await throttle._failures(throttle._key(victim)) == 3, "the counter must survive a spray" + assert await throttle._failures(throttle.username_cache, throttle._username_key(victim)) == 3, "the counter must survive a spray" with pytest.raises(ProxyException) as blocked: await _guess(throttle, username=victim) assert blocked.value.code == "429" diff --git a/tests/test_litellm/proxy/proxy_server/conftest.py b/tests/test_litellm/proxy/proxy_server/conftest.py index 22f2e5b3feb..56aa87f1e50 100644 --- a/tests/test_litellm/proxy/proxy_server/conftest.py +++ b/tests/test_litellm/proxy/proxy_server/conftest.py @@ -517,27 +517,38 @@ def make_key( def reset_login_throttle(monkeypatch): """Clear the Admin UI failed-login counters between tests. - `client` is session scoped and the counters live in the shared `user_api_key_cache` - with a 900s window, so without this any test that fails a sign-in enough times would - start returning 429 from unrelated tests later in the same process. Only the throttle's - own keys are removed, so nothing else in that cache is disturbed. + `client` is session scoped and the counters live in shared module stores with a 900s + window, so without this a failed sign-in test could return 429 in unrelated tests later. + Only the throttle's own keys are removed, so other cache entries remain untouched. """ from litellm.proxy import proxy_server as ps - from litellm.proxy.auth.login_throttle import _FAILED_LOGIN_CACHE, _CACHE_KEY_PREFIX + from litellm.proxy.auth import login_throttle + from litellm.proxy.auth.login_throttle import ( + _CACHE_KEY_PREFIX, + _FAILED_LOGIN_SOURCE_CACHE, + _FAILED_LOGIN_USERNAME_CACHE, + ) + + async def _no_delay(_seconds: float) -> None: + """The escalating wait on a rejected sign-in, replaced so the route tests stay fast.""" + + monkeypatch.setattr(login_throttle, "_sleep", _no_delay) def _drop_throttle_keys() -> None: - in_memory = getattr(_FAILED_LOGIN_CACHE, "in_memory_cache", None) - if in_memory is None: - return - tracked = tuple( - key - for store in (getattr(in_memory, "cache_dict", None), getattr(in_memory, "ttl_dict", None)) - if isinstance(store, dict) - for key in tuple(store) - if str(key).startswith(_CACHE_KEY_PREFIX) - ) - for key in tracked: - in_memory.delete_cache(key) + login_throttle._DELAYS_IN_FLIGHT.clear() + for cache in (_FAILED_LOGIN_USERNAME_CACHE, _FAILED_LOGIN_SOURCE_CACHE): + in_memory = getattr(cache, "in_memory_cache", None) + if in_memory is None: + continue + tracked = tuple( + key + for store in (getattr(in_memory, "cache_dict", None), getattr(in_memory, "ttl_dict", None)) + if isinstance(store, dict) + for key in tuple(store) + if str(key).startswith(_CACHE_KEY_PREFIX) + ) + for key in tracked: + in_memory.delete_cache(key) monkeypatch.setattr(ps, "redis_usage_cache", None) _drop_throttle_keys() diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py index faeb7750e03..c8185cdb886 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py @@ -531,8 +531,21 @@ def test_a_refused_attempt_carries_retry_after(client, monkeypatch, reset_login_ assert refused.headers.get("retry-after") == "77" +def test_the_form_returns_a_human_readable_lockout_page(client, monkeypatch, reset_login_throttle): + """The no-JavaScript form must render a wait page when its POST is throttled.""" + _install_real_auth(monkeypatch, max_failed_login_attempts=2, failed_login_window_seconds=77) + + assert [_form_login(client) for _ in range(2)] == [401, 401] + + refused = client.post("/login", data={"username": "admin", "password": "wrong"}) + assert refused.status_code == 429 + assert refused.headers.get("content-type", "").startswith("text/html") + assert "Try again in about 77 seconds" in refused.text + assert refused.headers.get("retry-after") == "77" + + def test_a_second_username_from_the_same_source_still_gets_through(client, monkeypatch, reset_login_throttle): - """The bucket is the username and source pair, so one account cannot block another.""" + """The username counter carries no address, so one account exhausting it cannot block another.""" _install_real_auth(monkeypatch, max_failed_login_attempts=2) for _ in range(3): @@ -541,6 +554,32 @@ def test_a_second_username_from_the_same_source_still_gets_through(client, monke assert _json_login(client, "/v2/login", username="someone-else@example.com") == 401 +def test_a_spray_across_usernames_is_refused_on_the_source_counter(client, monkeypatch, reset_login_throttle): + """A fresh username per guess keeps every username counter at one, so the address is what stops it.""" + _install_real_auth(monkeypatch, max_failed_login_attempts=100, max_failed_login_attempts_per_source=4) + + sprayed = [_json_login(client, "/v2/login", username=f"sprayed-{i}@corp.com") for i in range(4)] + assert sprayed == [401] * 4 + + assert _json_login(client, "/v2/login", username="sprayed-5@corp.com") == 429 + + +def test_the_configured_admin_password_still_signs_in_while_refused(client, monkeypatch, reset_login_throttle): + """The operator must never be locked out of the console by traffic aimed at it.""" + from unittest.mock import AsyncMock, patch + + _install_real_auth(monkeypatch, max_failed_login_attempts=2) + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + + assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] + assert _json_login(client, "/v2/login") == 429 + + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ): + assert _json_login(client, "/v2/login", password="right-password") == 200 + + def test_sign_in_succeeds_again_once_the_budget_is_restored(client, monkeypatch, reset_login_throttle): """A cleared bucket lets the same username straight back in.""" _install_real_auth(monkeypatch, max_failed_login_attempts=2) diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index ac2e84b8474..17563f60a4f 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -11442,9 +11442,14 @@ async def test_login_throttle_settings_are_not_overridable_from_the_database(): try: ps.general_settings.clear() await ProxyConfig()._update_general_settings( - db_general_settings={"max_failed_login_attempts": 999, "failed_login_window_seconds": 1} + db_general_settings={ + "max_failed_login_attempts": 999, + "max_failed_login_attempts_per_source": 999, + "failed_login_window_seconds": 1, + } ) assert "max_failed_login_attempts" not in ps.general_settings + assert "max_failed_login_attempts_per_source" not in ps.general_settings assert "failed_login_window_seconds" not in ps.general_settings finally: ps.general_settings.clear() diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index bb2841b6388..9d66320041f 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -24201,9 +24201,14 @@ export interface components { max_batch_file_size_mb?: number | null; /** * Max Failed Login Attempts - * @description Number of failed Admin UI sign-in attempts allowed for one username from one source address within `failed_login_window_seconds`, before further attempts are refused with 429. Configurable from config.yaml only. Defaults to 10 + * @description Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Configurable from config.yaml only. Defaults to 50 */ max_failed_login_attempts?: number | null; + /** + * Max Failed Login Attempts Per Source + * @description Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Configurable from config.yaml only. Defaults to 250 + */ + max_failed_login_attempts_per_source?: number | null; /** * Max Parallel Requests * @description maximum parallel requests for each api key From fb7ca1eda16ef666035b22b6c003bfb11f9ed6d4 Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Tue, 1 Sep 2026 13:21:44 -0700 Subject: [PATCH 005/141] docs(proxy): clarify login throttle configuration --- litellm/proxy/_types.py | 6 +++--- tests/test_litellm/proxy/test_proxy_server.py | 10 +++++----- ui/litellm-dashboard/src/lib/http/schema.d.ts | 6 +++--- 3 files changed, 11 insertions(+), 11 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 0071ad4603d..756a878f7f8 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2528,17 +2528,17 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): max_failed_login_attempts: int | None = Field( None, ge=1, - description="Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Configurable from config.yaml only. Defaults to 50", + description="Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Set under `general_settings` in config.yaml. Defaults to 50", ) max_failed_login_attempts_per_source: int | None = Field( None, ge=1, - description="Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Configurable from config.yaml only. Defaults to 250", + description="Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Set under `general_settings` in config.yaml. Defaults to 250", ) failed_login_window_seconds: int | None = Field( None, ge=1, - description="Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Configurable from config.yaml only. Defaults to 900", + description="Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Set under `general_settings` in config.yaml. Defaults to 900", ) allowed_routes: list | None = Field(None, description="Proxy API Endpoints you want users to be able to access") reject_clientside_metadata_tags: bool | None = Field( diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 17563f60a4f..a39fec7f523 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -11428,12 +11428,12 @@ async def test_authoritative_floor_spend_keeps_a_reset_marker_written_during_the @pytest.mark.asyncio -async def test_login_throttle_settings_are_not_overridable_from_the_database(): - """LIT-5285: the sign-in limits stay config.yaml only. +async def test_login_throttle_settings_are_not_hot_applied_from_the_database(): + """LIT-5285: a stored sign-in limit does not take effect on a live worker. - _update_general_settings copies an allowlist of keys out of the DB row. Adding these - to it would let a stored value outrank config.yaml, so an operator refused by a bad - value could not fix it by editing YAML and restarting. + _update_general_settings copies an allowlist of keys out of the DB row on every config + poll. Adding these to it would let a stored value outrank config.yaml without a restart, + so an operator locked out by a bad value could not fix it by editing YAML and restarting. """ import litellm.proxy.proxy_server as ps from litellm.proxy.proxy_server import ProxyConfig diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 9d66320041f..62ba7ce8b95 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -24152,7 +24152,7 @@ export interface components { enable_public_model_hub: boolean; /** * Failed Login Window Seconds - * @description Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Configurable from config.yaml only. Defaults to 900 + * @description Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Set under `general_settings` in config.yaml. Defaults to 900 */ failed_login_window_seconds?: number | null; /** @@ -24201,12 +24201,12 @@ export interface components { max_batch_file_size_mb?: number | null; /** * Max Failed Login Attempts - * @description Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Configurable from config.yaml only. Defaults to 50 + * @description Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Set under `general_settings` in config.yaml. Defaults to 50 */ max_failed_login_attempts?: number | null; /** * Max Failed Login Attempts Per Source - * @description Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Configurable from config.yaml only. Defaults to 250 + * @description Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Set under `general_settings` in config.yaml. Defaults to 250 */ max_failed_login_attempts_per_source?: number | null; /** From 65141a5fd89769dc1ec64873637b5992563f9c41 Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Tue, 1 Sep 2026 13:39:19 -0700 Subject: [PATCH 006/141] fix(proxy): keep the login throttle inside the type budget and fail safe on secret errors --- litellm/proxy/auth/login_throttle.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index ed53d9e7bc7..91e165006c3 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -60,7 +60,7 @@ _FAILED_LOGIN_USERNAME_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_USERNAME _FAILED_LOGIN_SOURCE_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_SOURCES) _NO_SETTINGS: Final = MappingProxyType({}) -_DELAYS_IN_FLIGHT: Final[dict[str, int]] = {} +_DELAYS_IN_FLIGHT: Final[dict[str, int]] = {} # mutable-ok: per-source slots taken and released around each held delay async def _sleep(seconds: float) -> None: @@ -138,7 +138,7 @@ class LoginThrottle: username_cache=_FAILED_LOGIN_USERNAME_CACHE, source_cache=_FAILED_LOGIN_SOURCE_CACHE, redis_cache=redis_usage_cache, - enabled=not get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT"), + enabled=not get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT", False), ) @staticmethod From e9166019a523204e6964c976eb76791ffcb6e47a Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Tue, 1 Sep 2026 14:02:29 -0700 Subject: [PATCH 007/141] fix(proxy): warn about per-worker sign-in counters without a module global --- litellm/proxy/auth/login_throttle.py | 13 +++++++++++++ litellm/proxy/proxy_server.py | 14 ++------------ 2 files changed, 15 insertions(+), 12 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 91e165006c3..b84f944907b 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -13,6 +13,7 @@ import asyncio import hashlib from collections.abc import Awaitable from dataclasses import dataclass +from functools import cache from types import MappingProxyType from typing import Final, NamedTuple, NoReturn @@ -63,6 +64,18 @@ _NO_SETTINGS: Final = MappingProxyType({}) _DELAYS_IN_FLIGHT: Final[dict[str, int]] = {} # mutable-ok: per-source slots taken and released around each held delay +@cache +def warn_login_counters_are_per_worker(num_workers: str) -> None: + """Warn once per process that failed sign-in counters are not shared across workers.""" + verbose_proxy_logger.warning( + "Running %s workers but Redis is not configured for LiteLLM caching. " + "Failed Admin UI sign-in attempts are counted per worker, so an attacker " + "gets max_failed_login_attempts guesses per worker instead of overall. " + "Configure Redis via the 'cache' section in your proxy config.", + num_workers, + ) + + async def _sleep(seconds: float) -> None: """The wait a rejected sign-in is held for. Replaced in tests so the suite pays no wall clock.""" await asyncio.sleep(seconds) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index d9d647c5f18..56a9ba390c4 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -299,7 +299,7 @@ from litellm.proxy.auth.auth_utils import ( from litellm.proxy.auth.fallback_model_access import router_fallback_access_check from litellm.proxy.auth.handle_jwt import JWTHandler from litellm.proxy.auth.litellm_license import LicenseCheck -from litellm.proxy.auth.login_throttle import LoginThrottle +from litellm.proxy.auth.login_throttle import LoginThrottle, warn_login_counters_are_per_worker from litellm.proxy.auth.model_checks import ( expand_wildcard_deployments_for_model_info, get_all_fallbacks, @@ -2288,7 +2288,6 @@ user_custom_key_generate = None # Tests that need to reset it can patch 'litellm.proxy.proxy_server._pkce_no_redis_warning_emitted'. _pkce_no_redis_warning_emitted: bool = False _cp_no_redis_warning_emitted: bool = False -_login_throttle_no_redis_warning_emitted: bool = False user_custom_key_update = None user_custom_sso = None user_custom_ui_sso_sign_in_handler = None @@ -5512,16 +5511,7 @@ class ProxyConfig: # Failed Admin UI sign-in counters live in redis_usage_cache when available so a # brute-force run is counted once across workers instead of once per worker. if os.getenv("NUM_WORKERS", "1") != "1" and redis_usage_cache is None: - global _login_throttle_no_redis_warning_emitted - if not _login_throttle_no_redis_warning_emitted: - _login_throttle_no_redis_warning_emitted = True - verbose_proxy_logger.warning( - "Running %s workers but Redis is not configured for LiteLLM caching. " - "Failed Admin UI sign-in attempts are counted per worker, so an attacker " - "gets max_failed_login_attempts guesses per worker instead of overall. " - "Configure Redis via the 'cache' section in your proxy config.", - os.getenv("NUM_WORKERS", "1"), - ) + warn_login_counters_are_per_worker(os.getenv("NUM_WORKERS", "1")) ### STORE MODEL IN DB ### feature flag for `/model/new` store_model_in_db = general_settings.get("store_model_in_db", False) From 429a4f213547d8a442913ebd7e83f573011224e7 Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Tue, 1 Sep 2026 14:35:05 -0700 Subject: [PATCH 008/141] fix: honor environment login throttle settings --- litellm/proxy/auth/login_throttle.py | 11 +++- .../proxy/auth/test_login_utils.py | 63 +++++++++++++++++-- .../proxy_server/test_routes_login_sso.py | 2 +- 3 files changed, 67 insertions(+), 9 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index b84f944907b..4290cdbd62c 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -91,12 +91,19 @@ class FailureCounts(NamedTuple): def _int_setting(name: str, value: object, default: int, minimum: int) -> int: if value is None: return default - if isinstance(value, bool) or not isinstance(value, int) or value < minimum: + if isinstance(value, str): + try: + parsed: Final = int(value.strip()) + except ValueError: + parsed = value + else: + parsed = value + if isinstance(parsed, bool) or not isinstance(parsed, int) or parsed < minimum: verbose_proxy_logger.warning( "general_settings.%s=%r is not an integer >= %s; using the default of %s", name, value, minimum, default ) return default - return value + return parsed def _as_count(cached: object) -> int: diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 1b85c9f4046..969103add6f 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -737,7 +737,7 @@ async def test_a_correct_admin_password_is_accepted_while_blocked(monkeypatch): await _guess(throttle) assert still_blocked.value.code == "429", "a wrong password is still refused" - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ): result = await _guess(throttle, password="right") @@ -780,7 +780,7 @@ async def test_a_successful_sign_in_clears_the_bucket(monkeypatch): with pytest.raises(ProxyException): await _guess(throttle) - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ): await _guess(throttle, password="right") @@ -859,7 +859,7 @@ async def test_both_credential_rejections_are_indistinguishable(monkeypatch): fake_user.password = "scrypt:fake" repo = MagicMock() repo.return_value.table.find_first = AsyncMock(return_value=fake_user) - with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( + with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( # test-quality-ok: reaches the known-DB-user branch without a database "litellm.proxy.auth.login_utils.verify_password", return_value=False ): with pytest.raises(ProxyException) as known: @@ -893,7 +893,7 @@ async def test_a_user_with_no_password_set_does_not_consume_the_budget(monkeypat repo = MagicMock() repo.return_value.table.find_first = AsyncMock(return_value=passwordless) - with patch("litellm.proxy.auth.login_utils.UserRepository", repo): + with patch("litellm.proxy.auth.login_utils.UserRepository", repo): # test-quality-ok: reaches the passwordless-DB-user branch without a database for _ in range(5): with pytest.raises(ProxyException) as exc: await authenticate_user( @@ -934,7 +934,7 @@ async def test_a_wrong_password_for_a_known_user_also_counts(monkeypatch): throttle=throttle, ) - with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( + with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( # test-quality-ok: reaches the known-DB-user branch without a database "litellm.proxy.auth.login_utils.verify_password", return_value=False ): for _ in range(3): @@ -1035,7 +1035,7 @@ async def test_a_successful_sign_in_leaves_the_source_counter_alone(monkeypatch) with pytest.raises(ProxyException): await _guess(throttle) - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ): await _guess(throttle, password="right") @@ -1227,6 +1227,57 @@ async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): ) +def test_settings_that_arrive_as_environment_strings_are_honored(monkeypatch): + """An `os.environ/VAR` reference in general_settings resolves to a string, not an int. + + Regression: a digit string fell back to the default with only a log line, so an operator + tightening the limits through environment substitution silently kept the stock ceilings. + """ + from litellm.proxy import proxy_server as ps + from litellm.proxy.auth.login_throttle import LoginThrottle + + monkeypatch.setattr( + ps, + "general_settings", + { + "max_failed_login_attempts": "7", + "max_failed_login_attempts_per_source": " 70 ", + "failed_login_window_seconds": "not-a-number", + }, + ) + request = MagicMock() + request.headers = {} + request.client = MagicMock() + request.client.host = "1.2.3.4" + + throttle = LoginThrottle.from_request(request) + + assert throttle.max_attempts == 7 + assert throttle.max_attempts_per_source == 70 + assert throttle.window_seconds == 900, "garbage still falls back to the default" + + +def test_a_negative_or_boolean_setting_falls_back_to_the_default(monkeypatch): + """A limit below one would refuse everyone; a bool is a typo, not a count.""" + from litellm.proxy import proxy_server as ps + from litellm.proxy.auth.login_throttle import LoginThrottle + + monkeypatch.setattr( + ps, + "general_settings", + {"max_failed_login_attempts": "-7", "max_failed_login_attempts_per_source": True}, + ) + request = MagicMock() + request.headers = {} + request.client = MagicMock() + request.client.host = "1.2.3.4" + + throttle = LoginThrottle.from_request(request) + + assert throttle.max_attempts == 50 + assert throttle.max_attempts_per_source == 250 + + @pytest.mark.asyncio async def test_a_refused_username_cannot_forge_log_lines(monkeypatch): """The username reaches a warning log, so it must not carry newlines or control bytes.""" diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py index c8185cdb886..22d96fdc254 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py @@ -574,7 +574,7 @@ def test_the_configured_admin_password_still_signs_in_while_refused(client, monk assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] assert _json_login(client, "/v2/login") == 429 - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ): assert _json_login(client, "/v2/login", password="right-password") == 200 From 53ed7391ebcc0bb330ca12d06913f4fda6635fc8 Mon Sep 17 00:00:00 2001 From: Yucheng Zhu Date: Tue, 1 Sep 2026 14:44:28 -0700 Subject: [PATCH 009/141] fix: satisfy login setting type checks --- litellm/proxy/auth/login_throttle.py | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 4290cdbd62c..1a978d9c212 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -88,16 +88,19 @@ class FailureCounts(NamedTuple): source: int +def _parse_int_setting(value: object) -> object: + if not isinstance(value, str): + return value + try: + return int(value.strip()) + except ValueError: + return value + + def _int_setting(name: str, value: object, default: int, minimum: int) -> int: if value is None: return default - if isinstance(value, str): - try: - parsed: Final = int(value.strip()) - except ValueError: - parsed = value - else: - parsed = value + parsed: Final = _parse_int_setting(value) if isinstance(parsed, bool) or not isinstance(parsed, int) or parsed < minimum: verbose_proxy_logger.warning( "general_settings.%s=%r is not an integer >= %s; using the default of %s", name, value, minimum, default From 36b346d31ae2da6c1ffcfde835a27b833b583623 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 11 Sep 2026 01:02:44 +0000 Subject: [PATCH 010/141] refactor(proxy): keep authenticate_user within the C901 budget after merge Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_utils.py | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-) diff --git a/litellm/proxy/auth/login_utils.py b/litellm/proxy/auth/login_utils.py index 5b0d5e6edd7..d0bb9e3087d 100644 --- a/litellm/proxy/auth/login_utils.py +++ b/litellm/proxy/auth/login_utils.py @@ -98,6 +98,14 @@ def _matches_env_credentials(username: str, password: str, master_key: str | Non ) +def _admin_credentials_match( + username: str, password: str, master_key: str, general_settings: Mapping[str, object] +) -> bool: + return general_settings.get("disable_env_credential_login") is not True and _matches_env_credentials( + username, password, master_key + ) + + def _invalid_credentials_message(general_settings: Mapping[str, object]) -> str: """One rejection message for unknown usernames and wrong passwords alike, so neither can be enumerated.""" if is_env_credential_login_enabled(general_settings): @@ -209,9 +217,7 @@ async def authenticate_user( code=500, ) - admin_credentials_match: Final = general_settings.get("disable_env_credential_login") is not True and ( - _matches_env_credentials(username, password, master_key) - ) + admin_credentials_match: Final = _admin_credentials_match(username, password, master_key, general_settings) if not admin_credentials_match: await throttle.raise_if_blocked(username) @@ -247,12 +253,7 @@ async def authenticate_user( user_id = LITELLM_PROXY_ADMIN_NAME # we want the key created to have PROXY_ADMIN_PERMISSIONS - key_user_id = LITELLM_PROXY_ADMIN_NAME - if ( - os.getenv("PROXY_ADMIN_ID", None) is not None and os.environ["PROXY_ADMIN_ID"] == user_id - ) or user_id == LITELLM_PROXY_ADMIN_NAME: - # checks if user is admin - key_user_id = os.getenv("PROXY_ADMIN_ID", LITELLM_PROXY_ADMIN_NAME) + key_user_id: Final = os.getenv("PROXY_ADMIN_ID", LITELLM_PROXY_ADMIN_NAME) # Admin is Authe'd in - generate key for the UI to access Proxy From 82902e83c2d7a8f909ee8ee49a5123b0618de4de Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 11 Sep 2026 07:54:19 +0000 Subject: [PATCH 011/141] fix(proxy): write failed-login counters and their expiry in one Redis call Use RedisCache.async_increment_with_floor (a single Lua INCRBY + EXPIRE) for the shared login counters instead of the two-step INCRBYFLOAT then EXPIRE, so a counter can never be committed to Redis without its expiry. The repair in _remaining_window now only covers expiries stripped out of band (PERSIST, a restore) and uses the same atomic call Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 15 ++++--- .../proxy/auth/test_login_utils.py | 43 +++++++++++-------- 2 files changed, 33 insertions(+), 25 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 1a978d9c212..926b7a0b9ce 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -194,11 +194,11 @@ class LoginThrottle: return max(local, _as_count(await self._outcome(redis_cache.async_get_cache(key)))) async def _remaining_window(self, key: str) -> int: - """Seconds until this counter expires, repairing a counter left without an expiry. + """Seconds until this counter expires. - Redis commits the increment before setting the TTL, so a failure in between can - leave a counter that never expires. Nothing increments the key again once the - limit is reached, so without the repair the key would stay refused indefinitely. + Counters are only ever written together with their expiry, so a counter without one + was stripped out of band (PERSIST, a restore). It is given the full window again, + since nothing increments a key once the limit is reached. """ redis_cache: Final = self.redis_cache if redis_cache is None: @@ -206,7 +206,7 @@ class LoginThrottle: ttl: Final = await self._outcome(redis_cache.async_get_ttl(key)) if isinstance(ttl, int) and ttl > 0: return min(ttl, self.window_seconds) - await self._outcome(redis_cache.async_increment(key, 0, ttl=self.window_seconds)) + await self._outcome(redis_cache.async_increment_with_floor(key, 0, self.window_seconds)) return self.window_seconds def _refused(self, retry_after: int, param: str) -> ProxyException: @@ -260,8 +260,9 @@ class LoginThrottle: redis_cache: Final = self.redis_cache if redis_cache is None: return local - shared: Final = _as_count(await self._outcome(redis_cache.async_increment(key, 1, ttl=self.window_seconds))) - await self._remaining_window(key) + shared: Final = _as_count( + await self._outcome(redis_cache.async_increment_with_floor(key, 1, self.window_seconds)) + ) return max(local, shared) async def record_failure(self, username: str) -> FailureCounts: diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 73154c5ecb0..f5c8b02b834 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1142,58 +1142,65 @@ async def test_disabling_the_control_removes_the_delay_as_well(monkeypatch, logi assert login_delays.seconds == [] -class _NoExpiryRedis: - """Redis that stores the counter but never records an expiry for it. +class _FakeRedis: + """Redis whose only counter write is the atomic INCRBY-plus-EXPIRE Lua call. - Models the window between INCRBYFLOAT committing and the TTL call failing. + `async_increment` is deliberately absent: a two-step increment would fail the test + with AttributeError, because Redis could then commit a count without its expiry. """ def __init__(self): self.values: dict = {} - self.expiry_repairs = 0 + self.ttls: dict = {} async def async_get_cache(self, key, **kwargs): return self.values.get(key) - async def async_increment(self, key, value, ttl=None, **kwargs): - if int(value) == 0: - self.expiry_repairs += 1 - self.values[key] = self.values.get(key, 0) + int(value) + async def async_increment_with_floor(self, key, value, ttl): + self.values[key] = self.values.get(key, 0) + value + self.ttls.setdefault(key, ttl) return self.values[key] async def async_get_ttl(self, key): - return None + return self.ttls.get(key) async def async_delete_cache(self, key): self.values.pop(key, None) + self.ttls.pop(key, None) + + def persist(self): + self.ttls.clear() @pytest.mark.asyncio -async def test_a_counter_left_without_an_expiry_is_repaired(monkeypatch): +async def test_counters_are_written_with_their_expiry_and_re_armed_if_stripped(monkeypatch): """Regression: a counter with no TTL would refuse the pair forever. - Redis commits the increment before setting the expiry, and nothing increments the key - again once the limit is reached, so a TTL that never landed is never repaired on its - own and the username and source pair stays refused with no way back. + Nothing increments a key once the limit is reached, so a counter that ever exists + without an expiry stays refused with no way back. Every write must therefore carry the + expiry, and a refusal that finds it stripped (PERSIST) must put the window back. """ from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - redis = _NoExpiryRedis() - throttle = _throttle(max_attempts=2, redis_cache=redis) + redis = _FakeRedis() + throttle = _throttle(max_attempts=2, window_seconds=77, redis_cache=redis) for _ in range(2): with pytest.raises(ProxyException): await _guess(throttle) - assert redis.expiry_repairs >= 1, "each recorded failure must leave the counter with an expiry" + assert redis.values, "failures must land in the shared counter" + assert set(redis.ttls) == set(redis.values), "no counter may exist without its expiry" + assert set(redis.ttls.values()) == {77} - repairs_before_block = redis.expiry_repairs + redis.persist() with pytest.raises(ProxyException) as blocked: await _guess(throttle) assert blocked.value.code == "429" - assert redis.expiry_repairs > repairs_before_block, "the refusal path must repair a missing expiry too" + assert blocked.value.headers.get("Retry-After") == "77" + assert set(redis.ttls) >= {k for k in redis.values if ":user:" in k}, "the refusal must re-arm a stripped expiry" @pytest.mark.asyncio From ec9a926e84df7e477e16664d6f517dd34612a8a7 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 11 Sep 2026 08:34:11 +0000 Subject: [PATCH 012/141] fix(proxy): resolve the login rate limit kill switch once per process Reading LITELLM_DISABLE_LOGIN_RATE_LIMIT through get_secret_bool on every unauthenticated sign-in attempt meant a hosted secret manager in read mode was queried once per password guess, before any counter was checked Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 8 +++++- .../proxy/auth/test_login_utils.py | 25 +++++++++++++++++++ 2 files changed, 32 insertions(+), 1 deletion(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 926b7a0b9ce..11290ef6d40 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -76,6 +76,12 @@ def warn_login_counters_are_per_worker(num_workers: str) -> None: ) +@cache +def _rate_limit_disabled() -> bool: + """Resolved once per process so an unauthenticated flood never reaches the secret manager.""" + return bool(get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT", False)) + + async def _sleep(seconds: float) -> None: """The wait a rejected sign-in is held for. Replaced in tests so the suite pays no wall clock.""" await asyncio.sleep(seconds) @@ -161,7 +167,7 @@ class LoginThrottle: username_cache=_FAILED_LOGIN_USERNAME_CACHE, source_cache=_FAILED_LOGIN_SOURCE_CACHE, redis_cache=redis_usage_cache, - enabled=not get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT", False), + enabled=not _rate_limit_disabled(), ) @staticmethod diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index f5c8b02b834..8afac8a622d 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1267,6 +1267,31 @@ def test_settings_that_arrive_as_environment_strings_are_honored(monkeypatch): assert throttle.window_seconds == 900, "garbage still falls back to the default" +def test_the_disable_flag_is_read_once_not_per_login_attempt(monkeypatch): + """Regression: the kill switch was read through the secret manager on every unauthenticated request. + + With a hosted secret manager in read mode that is a synchronous network call per guess, so a + flood of wrong passwords could exhaust the secret manager even after the source was refused. + """ + from litellm.proxy import proxy_server as ps + from litellm.proxy.auth import login_throttle + + reads: Final[list[str]] = [] # mutable-ok: test-only call recorder + monkeypatch.setattr(login_throttle, "get_secret_bool", lambda name, default: reads.append(name) or default) + login_throttle._rate_limit_disabled.cache_clear() + monkeypatch.setattr(ps, "general_settings", {}) + request = MagicMock() + request.headers = {} + request.client = MagicMock() + request.client.host = "1.2.3.4" + + for _ in range(50): + assert login_throttle.LoginThrottle.from_request(request).enabled is True + + login_throttle._rate_limit_disabled.cache_clear() + assert reads == ["LITELLM_DISABLE_LOGIN_RATE_LIMIT"] + + def test_a_negative_or_boolean_setting_falls_back_to_the_default(monkeypatch): """A limit below one would refuse everyone; a bool is a typo, not a count.""" from litellm.proxy import proxy_server as ps From 9c541d9ce3ecf31ca8673ebc4472250f477b55ae Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 11 Sep 2026 15:27:33 +0000 Subject: [PATCH 013/141] fix(ui): show internal user email in logs table and log detail drawer Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../(dashboard)/hooks/users/useUsers.test.ts | 61 ++++++++++++++++++- .../app/(dashboard)/hooks/users/useUsers.ts | 18 ++++++ .../LogDetailContent.integration.test.tsx | 23 +++++++ .../LogDetailsDrawer/LogDetailContent.tsx | 23 ++++++- .../LogDetailsDrawer.test.tsx | 39 +++++++++++- .../LogDetailsDrawer/LogDetailsDrawer.tsx | 3 + .../components/view_logs/RequestLogsTable.tsx | 10 ++- .../RequestLogsTableColumns.test.tsx | 21 +++++++ .../view_logs/RequestLogsTableColumns.tsx | 23 ++++++- 9 files changed, 212 insertions(+), 9 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts index dd7209140a9..2e8471ba84f 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts @@ -2,7 +2,7 @@ import { describe, it, expect, vi, beforeEach } from "vitest"; import { renderHook, waitFor } from "@testing-library/react"; import { QueryClient, QueryClientProvider } from "@tanstack/react-query"; import React, { ReactNode } from "react"; -import { useInfiniteUsers, useUserLookup } from "./useUsers"; +import { useInfiniteUsers, useUserEmailLookup, useUserLookup } from "./useUsers"; import { userListCall } from "@/components/networking"; import type { UserListResponse } from "@/components/networking"; @@ -335,3 +335,62 @@ describe("useUserLookup", () => { expect(userListCall).not.toHaveBeenCalled(); }); }); + +describe("useUserEmailLookup", () => { + let queryClient: QueryClient; + + beforeEach(() => { + queryClient = new QueryClient({ defaultOptions: { queries: { retry: false } } }); + vi.clearAllMocks(); + mockUseAuthorized.mockReturnValue(DEFAULT_AUTH); + }); + + const wrapper = ({ children }: { children: ReactNode }) => + React.createElement(QueryClientProvider, { client: queryClient }, children); + + it("fetches the distinct ids in one call and maps each id to its email", async () => { + const response = buildUserListResponse(1, 1, 2); + vi.mocked(userListCall).mockResolvedValue(response); + + const { result } = renderHook(() => useUserEmailLookup(["user-1-1", "user-1-0", "user-1-1", ""]), { wrapper }); + + await waitFor(() => expect(result.current.isSuccess).toBe(true)); + expect(userListCall).toHaveBeenCalledTimes(1); + expect(userListCall).toHaveBeenCalledWith("test-access-token", ["user-1-0", "user-1-1"], 1, 2); + expect(result.current.data).toEqual({ + "user-1-0": "user-1-0@example.com", + "user-1-1": "user-1-1@example.com", + }); + }); + + it("omits users that have no email so callers fall back to the id", async () => { + const response = buildUserListResponse(1, 1, 2); + vi.mocked(userListCall).mockResolvedValue({ + ...response, + users: [{ ...response.users[0], user_email: "" }, response.users[1]], + }); + + const { result } = renderHook(() => useUserEmailLookup(["user-1-0", "user-1-1"]), { wrapper }); + + await waitFor(() => expect(result.current.isSuccess).toBe(true)); + expect(result.current.data).toEqual({ "user-1-1": "user-1-1@example.com" }); + }); + + it("does not query with no ids", async () => { + const { result } = renderHook(() => useUserEmailLookup([]), { wrapper }); + + await new Promise((resolve) => setTimeout(resolve, 20)); + expect(result.current.fetchStatus).toBe("idle"); + expect(userListCall).not.toHaveBeenCalled(); + }); + + it("does not query for a non-admin role", async () => { + mockUseAuthorized.mockReturnValue({ ...DEFAULT_AUTH, userRole: "Internal User" }); + + const { result } = renderHook(() => useUserEmailLookup(["user-1-0"]), { wrapper }); + + await new Promise((resolve) => setTimeout(resolve, 20)); + expect(result.current.fetchStatus).toBe("idle"); + expect(userListCall).not.toHaveBeenCalled(); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts index 011e43777b5..4a28e7ff3f2 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts @@ -38,6 +38,24 @@ export const useInfiniteUsers = (pageSize: number = DEFAULT_PAGE_SIZE, searchEma }); }; +const USER_LIST_MAX_PAGE_SIZE = 100; + +export const useUserEmailLookup = (userIds: readonly string[]) => { + const { accessToken, userRole } = useAuthorized(); + const distinctIds = Array.from(new Set(userIds.filter((id) => id !== ""))).sort(); + return useQuery>({ + queryKey: userLookupKeys.list({ filters: { ids: distinctIds.join(",") } }), + queryFn: async () => { + const ids = distinctIds.slice(0, USER_LIST_MAX_PAGE_SIZE); + const response = await userListCall(accessToken!, ids, 1, ids.length); + return Object.fromEntries( + response.users.filter((user) => Boolean(user.user_email)).map((user) => [user.user_id, user.user_email]), + ); + }, + enabled: Boolean(accessToken) && distinctIds.length > 0 && all_admin_roles.includes(userRole!), + }); +}; + export const useUserLookup = (userId: string | null) => { const { accessToken, userRole } = useAuthorized(); return useQuery({ diff --git a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailContent.integration.test.tsx b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailContent.integration.test.tsx index 721525268b3..793bfa7c246 100644 --- a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailContent.integration.test.tsx +++ b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailContent.integration.test.tsx @@ -56,6 +56,29 @@ describe("LogDetailContent", () => { expect(screen.getByText("completion")).toBeInTheDocument(); }); + it("shows the requesting user's email and id in Request Details when the email is resolved", () => { + render( + , + ); + + expect(screen.getByText("User")).toBeInTheDocument(); + expect(screen.getByText("alice@example.com")).toBeInTheDocument(); + expect(screen.getByText("106514937785257944828")).toBeInTheDocument(); + }); + + it("falls back to the user id in Request Details when no email is resolved", () => { + render(); + + expect(screen.getByText("User")).toBeInTheDocument(); + expect(screen.getByText("106514937785257944828")).toBeInTheDocument(); + }); + + it("omits the User row when the log has no internal user", () => { + render(); + + expect(screen.queryByText("User")).not.toBeInTheDocument(); + }); + it("should display error alert when request has failed", () => { render( {logEntry.model} {logEntry.custom_llm_provider || "-"} {logEntry.call_type} + {logEntry.user && ( + + + + )} @@ -333,6 +344,16 @@ function TagsSection({ tags }: { tags: Record }) { ); } +function UserIdentity({ userId, email }: { userId: string; email?: string }) { + if (!email || email === userId) return ; + return ( + + {email} + + + ); +} + function GuardrailLabel({ label, maskedCount }: { label: string; maskedCount: number }) { const handleClick = () => { const el = document.getElementById("guardrail-section"); diff --git a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.test.tsx b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.test.tsx index b96e5279722..f5bf7cda951 100644 --- a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.test.tsx +++ b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.test.tsx @@ -14,8 +14,13 @@ vi.mock("@/app/(dashboard)/hooks/logDetails/useLogDetails", () => ({ useLogDetails: () => ({ data: null, isLoading: false }), })); +const mockUseUserLookup = vi.fn(() => ({ data: undefined })); +vi.mock("@/app/(dashboard)/hooks/users/useUsers", () => ({ + useUserLookup: (userId: string | null) => mockUseUserLookup(userId), +})); + vi.mock("./LogDetailContent", () => ({ - LogDetailContent: () => null, + LogDetailContent: ({ userEmail }: { userEmail?: string }) => user-email:{userEmail ?? "none"}, GuardrailJumpLink: () => null, })); @@ -124,6 +129,38 @@ describe("LogDetailsDrawer session sidebar sorting", () => { }); }); +describe("LogDetailsDrawer internal user email", () => { + const renderSingleLog = (user: string | undefined) => { + const queryClient = new QueryClient({ defaultOptions: { queries: { retry: false } } }); + return render( + + {}} + logEntry={makeLog({ request_id: "single", user })} + accessToken="token" + /> + , + ); + }; + + it("looks up the log's internal user and hands the resolved email to the detail content", () => { + mockUseUserLookup.mockReturnValue({ data: { user_id: "u-1", user_email: "alice@example.com" } }); + renderSingleLog("u-1"); + + expect(mockUseUserLookup).toHaveBeenCalledWith("u-1"); + expect(screen.getByText("user-email:alice@example.com")).toBeInTheDocument(); + }); + + it("skips the lookup and passes no email when the log has no internal user", () => { + mockUseUserLookup.mockReturnValue({ data: undefined }); + renderSingleLog(undefined); + + expect(mockUseUserLookup).toHaveBeenCalledWith(null); + expect(screen.getByText("user-email:none")).toBeInTheDocument(); + }); +}); + describe("LogDetailsDrawer session sidebar auto-router icon", () => { const routedSessionLogs = [ makeLog({ request_id: "routed", model: "claude-opus-4-8", model_group: "smart-router" }), diff --git a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.tsx b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.tsx index ddd0a650c04..dc9207d59eb 100644 --- a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.tsx +++ b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.tsx @@ -17,6 +17,7 @@ import { getSpendString } from "@/utils/dataUtils"; import { normalizeGuardrailEntries, sortSessionLogs, SessionLogSortMode } from "./utils"; import { DRAWER_WIDTH } from "./constants"; import { useLogDetails } from "@/app/(dashboard)/hooks/logDetails/useLogDetails"; +import { useUserLookup } from "@/app/(dashboard)/hooks/users/useUsers"; export interface LogDetailsDrawerProps { open: boolean; @@ -245,6 +246,7 @@ export function LogDetailsDrawer({ const logDetails = useLogDetails(currentLog?.request_id, startTime, open && !!currentLog?.request_id); const detailsData = logDetails.data as any; const isLoadingDetails = logDetails.isLoading; + const { data: logUser } = useUserLookup(open && currentLog?.user ? currentLog.user : null); // Build an enriched log entry that merges lazy-loaded details. // The list endpoint may already include messages/response when store_prompts_in_spend_logs is enabled, @@ -465,6 +467,7 @@ export function LogDetailsDrawer({ logEntry={enrichedLog} isLoadingDetails={isLoadingDetails} accessToken={accessToken ?? null} + userEmail={logUser?.user_email || undefined} /> diff --git a/ui/litellm-dashboard/src/components/view_logs/RequestLogsTable.tsx b/ui/litellm-dashboard/src/components/view_logs/RequestLogsTable.tsx index c3d204e1ae3..5a9bac428c8 100644 --- a/ui/litellm-dashboard/src/components/view_logs/RequestLogsTable.tsx +++ b/ui/litellm-dashboard/src/components/view_logs/RequestLogsTable.tsx @@ -4,6 +4,7 @@ import type { ColumnFiltersState, OnChangeFn, PaginationState, SortingState } fr import { ScrollText } from "lucide-react"; import { useMemo, useState, type ReactNode } from "react"; +import { useUserEmailLookup } from "@/app/(dashboard)/hooks/users/useUsers"; import { DataTable, DataTableFilterDrawer, DataTableToolbar } from "@/components/shared/DataTable"; import type { Team } from "../key_team_helpers/key_list"; @@ -73,10 +74,13 @@ export function RequestLogsTable({ }: RequestLogsTableProps) { const [filtersOpen, setFiltersOpen] = useState(false); + const userIds = useMemo(() => data.flatMap((log) => (log.user ? [log.user] : [])), [data]); + const { data: emailByUserId } = useUserEmailLookup(userIds); + const columns = useMemo(() => { - const deps = { onKeyHashClick, onSessionClick }; - return getRequestLogsTableColumns(deps); - }, [onKeyHashClick, onSessionClick]); + const resolveUserEmail = (userId: string) => emailByUserId?.[userId]; + return getRequestLogsTableColumns({ onKeyHashClick, onSessionClick, resolveUserEmail }); + }, [onKeyHashClick, onSessionClick, emailByUserId]); const isFiltered = columnFilters.length > 0 || searchValue !== ""; diff --git a/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.test.tsx b/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.test.tsx index 9f0e659cb1f..08853814269 100644 --- a/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.test.tsx +++ b/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.test.tsx @@ -75,6 +75,27 @@ describe("Cost column", () => { }); }); +describe("Internal User column", () => { + const emailById: Record = { "106514937785257944828": "alice@example.com" }; + const deps = { ...noopDeps, resolveUserEmail: (userId: string) => emailById[userId] }; + + it("shows the user's email instead of the raw id, with both in the tooltip", async () => { + const user = userEvent.setup(); + renderRows([logEntry({ request_id: "req-known-user", user: "106514937785257944828" })], deps); + + const emailCell = screen.getByText("alice@example.com"); + expect(screen.queryByText("106514937785257944828")).not.toBeInTheDocument(); + await user.hover(emailCell); + expect(await screen.findByText("alice@example.com (106514937785257944828)")).toBeInTheDocument(); + }); + + it("falls back to the raw id when no email is known for the user", () => { + renderRows([logEntry({ request_id: "req-unknown-user", user: "unknown-user-id" })], deps); + + expect(screen.getByText("unknown-user-id")).toBeInTheDocument(); + }); +}); + describe("Tokens column", () => { const sessionRow: Partial = { request_id: "req-session-tokens", diff --git a/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.tsx b/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.tsx index 1ec1087a1a4..dd83ad6eb05 100644 --- a/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.tsx +++ b/ui/litellm-dashboard/src/components/view_logs/RequestLogsTableColumns.tsx @@ -15,6 +15,7 @@ import { AgentBadge, AgentIcon, BatchBadge, LlmBadge, McpBadge, SparkleIcon, Wre export interface RequestLogsTableColumnsDeps { onKeyHashClick: (keyHash: string) => void; onSessionClick: (log: LogEntry) => void; + resolveUserEmail?: (userId: string) => string | undefined; } const readMetaString = (metadata: Record | undefined, key: string): string | undefined => { @@ -32,14 +33,25 @@ const readMcpLogoUrl = (metadata: Record | undefined): string | const getLogoUrl = (row: LogEntry, provider: string): string => readMcpLogoUrl(row.metadata) ?? (provider ? getProviderLogoAndName(provider).logo : ""); -function TruncatedText({ value }: { value: string | undefined }) { +function TruncatedText({ value, tooltip }: { value: string | undefined; tooltip?: string }) { const display = value ?? "-"; - return {display}} />; + return ( + {display}} + /> + ); +} + +function UserCell({ userId, email }: { userId: string | undefined; email: string | undefined }) { + if (!userId || !email || email === userId) return ; + return ; } export const getRequestLogsTableColumns = ({ onKeyHashClick, onSessionClick, + resolveUserEmail = () => undefined, }: RequestLogsTableColumnsDeps): ColumnDef[] => [ { id: "startTime", @@ -313,7 +325,12 @@ export const getRequestLogsTableColumns = ({ header: "Internal User", size: 150, enableSorting: false, - cell: ({ row }) => , + cell: ({ row }) => ( + + ), }, { id: "end_user", From bde96e3197bf1d1d9af07fa238f63649b543adf0 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 11 Sep 2026 15:36:29 +0000 Subject: [PATCH 014/141] fix(ui): keep user id boundaries in email lookup query key Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../src/app/(dashboard)/hooks/users/useUsers.ts | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts index 4a28e7ff3f2..84aeba90ae2 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts @@ -44,7 +44,7 @@ export const useUserEmailLookup = (userIds: readonly string[]) => { const { accessToken, userRole } = useAuthorized(); const distinctIds = Array.from(new Set(userIds.filter((id) => id !== ""))).sort(); return useQuery>({ - queryKey: userLookupKeys.list({ filters: { ids: distinctIds.join(",") } }), + queryKey: userLookupKeys.list({ filters: { ids: JSON.stringify(distinctIds) } }), queryFn: async () => { const ids = distinctIds.slice(0, USER_LIST_MAX_PAGE_SIZE); const response = await userListCall(accessToken!, ids, 1, ids.length); From 3630642110e0055040f5c5666cb2ff1c195d519e Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 11 Sep 2026 16:21:20 +0000 Subject: [PATCH 015/141] fix(ui): allow Org Admin session role to resolve user emails in logs Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../src/app/(dashboard)/hooks/users/useUsers.test.ts | 12 +++++++++++- .../src/app/(dashboard)/hooks/users/useUsers.ts | 8 ++++---- ui/litellm-dashboard/src/utils/roles.ts | 6 ++++++ 3 files changed, 21 insertions(+), 5 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts index 2e8471ba84f..f49c728446b 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.test.ts @@ -235,7 +235,7 @@ describe("useInfiniteUsers", () => { }); it("should execute query for each admin role", async () => { - const adminRoles = ["Admin", "Admin Viewer", "proxy_admin", "proxy_admin_viewer", "org_admin"]; + const adminRoles = ["Admin", "Admin Viewer", "proxy_admin", "proxy_admin_viewer", "org_admin", "Org Admin"]; for (const role of adminRoles) { vi.clearAllMocks(); @@ -384,6 +384,16 @@ describe("useUserEmailLookup", () => { expect(userListCall).not.toHaveBeenCalled(); }); + it("queries for the formatted Org Admin session role", async () => { + mockUseAuthorized.mockReturnValue({ ...DEFAULT_AUTH, userRole: "Org Admin" }); + vi.mocked(userListCall).mockResolvedValue(buildUserListResponse(1, 1, 1)); + + const { result } = renderHook(() => useUserEmailLookup(["user-1-0"]), { wrapper }); + + await waitFor(() => expect(result.current.isSuccess).toBe(true)); + expect(result.current.data).toEqual({ "user-1-0": "user-1-0@example.com" }); + }); + it("does not query for a non-admin role", async () => { mockUseAuthorized.mockReturnValue({ ...DEFAULT_AUTH, userRole: "Internal User" }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts index 84aeba90ae2..3b7f9fbeb02 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/users/useUsers.ts @@ -1,7 +1,7 @@ import { userListCall, UserInfo, UserListResponse } from "@/components/networking"; import { useInfiniteQuery, useQuery } from "@tanstack/react-query"; import { createQueryKeys } from "../common/queryKeysFactory"; -import { all_admin_roles } from "@/utils/roles"; +import { canListUsers } from "@/utils/roles"; import useAuthorized from "@/app/(dashboard)/hooks/useAuthorized"; const infiniteUsersKeys = createQueryKeys("infiniteUsers"); @@ -34,7 +34,7 @@ export const useInfiniteUsers = (pageSize: number = DEFAULT_PAGE_SIZE, searchEma } return undefined; }, - enabled: Boolean(accessToken) && all_admin_roles.includes(userRole!), + enabled: Boolean(accessToken) && canListUsers(userRole), }); }; @@ -52,7 +52,7 @@ export const useUserEmailLookup = (userIds: readonly string[]) => { response.users.filter((user) => Boolean(user.user_email)).map((user) => [user.user_id, user.user_email]), ); }, - enabled: Boolean(accessToken) && distinctIds.length > 0 && all_admin_roles.includes(userRole!), + enabled: Boolean(accessToken) && distinctIds.length > 0 && canListUsers(userRole), }); }; @@ -64,6 +64,6 @@ export const useUserLookup = (userId: string | null) => { const response = await userListCall(accessToken!, [userId!], 1, 1); return response.users.find((user) => user.user_id === userId) ?? null; }, - enabled: Boolean(accessToken) && Boolean(userId) && all_admin_roles.includes(userRole!), + enabled: Boolean(accessToken) && Boolean(userId) && canListUsers(userRole), }); }; diff --git a/ui/litellm-dashboard/src/utils/roles.ts b/ui/litellm-dashboard/src/utils/roles.ts index 62a5f02cc39..1cb7e75c19b 100644 --- a/ui/litellm-dashboard/src/utils/roles.ts +++ b/ui/litellm-dashboard/src/utils/roles.ts @@ -28,6 +28,12 @@ export const isAdminRole = (role: string): boolean => { return all_admin_roles.includes(role); }; +// /user/list admits proxy admins and org admins; the session role for the latter is the formatted +// "Org Admin", which all_admin_roles does not carry +const rolesAllowedToListUsers: string[] = [...all_admin_roles, "Org Admin"]; + +export const canListUsers = (role: string | null): boolean => rolesAllowedToListUsers.includes(role ?? ""); + export const isProxyAdminRole = (role: string): boolean => { return role === "proxy_admin" || role === "Admin"; }; From b7596d6fba97f712a2dca3ead9ed280af3654292 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 11 Sep 2026 17:16:30 +0000 Subject: [PATCH 016/141] refactor(ui): drop redundant comment on canListUsers role list Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ui/litellm-dashboard/src/utils/roles.ts | 2 -- 1 file changed, 2 deletions(-) diff --git a/ui/litellm-dashboard/src/utils/roles.ts b/ui/litellm-dashboard/src/utils/roles.ts index 1cb7e75c19b..066a83992b4 100644 --- a/ui/litellm-dashboard/src/utils/roles.ts +++ b/ui/litellm-dashboard/src/utils/roles.ts @@ -28,8 +28,6 @@ export const isAdminRole = (role: string): boolean => { return all_admin_roles.includes(role); }; -// /user/list admits proxy admins and org admins; the session role for the latter is the formatted -// "Org Admin", which all_admin_roles does not carry const rolesAllowedToListUsers: string[] = [...all_admin_roles, "Org Admin"]; export const canListUsers = (role: string | null): boolean => rolesAllowedToListUsers.includes(role ?? ""); From fcca3c239e1683a6f1a960e87fff2e7ce34eaffc Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 11 Sep 2026 22:50:52 +0000 Subject: [PATCH 017/141] fix(proxy): count failed sign-ins in Redis alone while it answers Every worker spends one shared budget and a successful sign-in clears it for all of them. This worker's own counter is only consulted while Redis raises, so an outage degrades to per-worker accounting instead of switching the control off Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 48 ++++++------ .../proxy/auth/test_login_utils.py | 77 +++++++++++++++++++ 2 files changed, 101 insertions(+), 24 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 11290ef6d40..0f5b7e64a57 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -11,7 +11,7 @@ startup and can be reassigned later. import asyncio import hashlib -from collections.abc import Awaitable +from collections.abc import Awaitable, Callable from dataclasses import dataclass from functools import cache from types import MappingProxyType @@ -60,6 +60,7 @@ def _bounded_store(max_entries: int) -> DualCache: _FAILED_LOGIN_USERNAME_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_USERNAMES) _FAILED_LOGIN_SOURCE_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_SOURCES) _NO_SETTINGS: Final = MappingProxyType({}) +_UNAVAILABLE: Final = object() _DELAYS_IN_FLIGHT: Final[dict[str, int]] = {} # mutable-ok: per-source slots taken and released around each held delay @@ -188,16 +189,22 @@ class LoginThrottle: return await work except Exception as exc: # noqa: BLE001 # an unreachable cache must never deny a valid credential verbose_proxy_logger.warning("login attempt accounting unavailable: %s", exc) - return None + return _UNAVAILABLE - async def _failures(self, store: DualCache, key: str) -> int: - """The larger of the shared and the process-local count, so a Redis outage degrades - to per-worker accounting instead of switching the control off.""" - local: Final = _as_count(await self._outcome(store.async_get_cache(key=key))) + async def _shared(self, work: Callable[[RedisCache], Awaitable[object]]) -> object: + """The Redis result, or ``_UNAVAILABLE`` when Redis is not configured or the call raised.""" redis_cache: Final = self.redis_cache if redis_cache is None: - return local - return max(local, _as_count(await self._outcome(redis_cache.async_get_cache(key)))) + return _UNAVAILABLE + return await self._outcome(work(redis_cache)) + + async def _failures(self, store: DualCache, key: str) -> int: + """The shared count while Redis answers, so every worker sees one budget; this worker's + own count only while it does not, so an outage degrades to per-worker accounting.""" + shared: Final = await self._shared(lambda redis_cache: redis_cache.async_get_cache(key)) + if shared is not _UNAVAILABLE: + return _as_count(shared) + return _as_count(await self._outcome(store.async_get_cache(key=key))) async def _remaining_window(self, key: str) -> int: """Seconds until this counter expires. @@ -206,13 +213,12 @@ class LoginThrottle: was stripped out of band (PERSIST, a restore). It is given the full window again, since nothing increments a key once the limit is reached. """ - redis_cache: Final = self.redis_cache - if redis_cache is None: + if self.redis_cache is None: return self.window_seconds - ttl: Final = await self._outcome(redis_cache.async_get_ttl(key)) + ttl: Final = await self._shared(lambda redis_cache: redis_cache.async_get_ttl(key)) if isinstance(ttl, int) and ttl > 0: return min(ttl, self.window_seconds) - await self._outcome(redis_cache.async_increment_with_floor(key, 0, self.window_seconds)) + await self._shared(lambda redis_cache: redis_cache.async_increment_with_floor(key, 0, self.window_seconds)) return self.window_seconds def _refused(self, retry_after: int, param: str) -> ProxyException: @@ -260,16 +266,12 @@ class LoginThrottle: ) async def _bump(self, store: DualCache, key: str) -> int: - local: Final = _as_count( - await self._outcome(store.async_increment_cache(key=key, value=1, ttl=self.window_seconds)) + shared: Final = await self._shared( + lambda redis_cache: redis_cache.async_increment_with_floor(key, 1, self.window_seconds) ) - redis_cache: Final = self.redis_cache - if redis_cache is None: - return local - shared: Final = _as_count( - await self._outcome(redis_cache.async_increment_with_floor(key, 1, self.window_seconds)) - ) - return max(local, shared) + if shared is not _UNAVAILABLE: + return _as_count(shared) + return _as_count(await self._outcome(store.async_increment_cache(key=key, value=1, ttl=self.window_seconds))) async def record_failure(self, username: str) -> FailureCounts: """Count one rejected credential guess against this username and against this source.""" @@ -331,7 +333,5 @@ class LoginThrottle: if not self.enabled: return key: Final = self._username_key(username) - redis_cache: Final = self.redis_cache - if redis_cache is not None: - await self._outcome(redis_cache.async_delete_cache(key)) + await self._shared(lambda redis_cache: redis_cache.async_delete_cache(key)) await self._outcome(self.username_cache.async_delete_cache(key=key)) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 8afac8a622d..70e7125485c 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1203,6 +1203,83 @@ async def test_counters_are_written_with_their_expiry_and_re_armed_if_stripped(m assert set(redis.ttls) >= {k for k in redis.values if ":user:" in k}, "the refusal must re-arm a stripped expiry" +class _DownRedis(_FakeRedis): + """Redis whose every call raises, as during an outage or an open circuit breaker.""" + + async def async_get_cache(self, key, **kwargs): + raise ConnectionError("redis is down") + + async def async_increment_with_floor(self, key, value, ttl): + raise ConnectionError("redis is down") + + async def async_get_ttl(self, key): + raise ConnectionError("redis is down") + + async def async_delete_cache(self, key): + raise ConnectionError("redis is down") + + +@pytest.mark.asyncio +async def test_redis_is_the_only_counter_while_it_answers(monkeypatch): + """Regression: every worker must spend the same budget, and a success must clear it for all. + + Counting in this worker's memory as well as in Redis let the two drift apart: a worker + whose Redis write failed kept its own count while the others gave the attacker fresh + guesses, and a stale local count outlived the shared clear after a correct password. + """ + from litellm.caching.dual_cache import DualCache + from litellm.proxy._types import ProxyException + from litellm.proxy.auth.login_throttle import _CACHE_KEY_PREFIX + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + redis = _FakeRedis() + first_worker_store = DualCache() + second_worker_store = DualCache() + first_worker = _throttle(max_attempts=2, cache=first_worker_store, redis_cache=redis) + second_worker = _throttle(max_attempts=2, cache=second_worker_store, redis_cache=redis) + + for _ in range(2): + with pytest.raises(ProxyException, match="Invalid credentials"): + await _guess(first_worker) + + assert not [k for k in first_worker_store.in_memory_cache.cache_dict if str(k).startswith(_CACHE_KEY_PREFIX)], ( + "with Redis answering, no worker may keep a counter of its own" + ) + with pytest.raises(ProxyException) as blocked: + await _guess(second_worker) + assert blocked.value.code == "429", "the second worker must see the budget the first one spent" + + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ): + await _guess(second_worker, password="right") + + assert not [k for k in redis.values if ":user:" in k], "a success must clear the shared username counter" + with pytest.raises(ProxyException, match="Invalid credentials"): + await _guess(first_worker) + + +@pytest.mark.asyncio +async def test_a_redis_outage_falls_back_to_this_workers_own_counter(monkeypatch): + """With Redis raising, guesses are still counted and refused, per worker, instead of unbounded.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(max_attempts=2, redis_cache=_DownRedis()) + + for _ in range(2): + with pytest.raises(ProxyException, match="Invalid credentials"): + await _guess(throttle) + + with pytest.raises(ProxyException) as blocked: + await _guess(throttle) + assert blocked.value.code == "429" + assert blocked.value.headers.get("Retry-After") == "900" + + @pytest.mark.asyncio async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): """Regression: throttle entries must not evict cached credentials. From 383edfe9532dfbff6827e579fca5e97d0feb8e23 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 11 Sep 2026 23:25:33 +0000 Subject: [PATCH 018/141] fix(proxy): keep counting the failed sign-ins Redis missed once it answers again A guess is recorded in exactly one place, Redis or this worker's own store when Redis refused it, so the count is the sum of the two. Redis is read through async_batch_get_counts, which raises on failure, instead of async_get_cache, which swallows it into None and read as an empty counter Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 26 ++++++--- .../proxy/auth/test_login_utils.py | 56 ++++++++++++++++++- 2 files changed, 72 insertions(+), 10 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 0f5b7e64a57..ca73bb4f26b 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -199,12 +199,18 @@ class LoginThrottle: return await self._outcome(work(redis_cache)) async def _failures(self, store: DualCache, key: str) -> int: - """The shared count while Redis answers, so every worker sees one budget; this worker's - own count only while it does not, so an outage degrades to per-worker accounting.""" - shared: Final = await self._shared(lambda redis_cache: redis_cache.async_get_cache(key)) - if shared is not _UNAVAILABLE: - return _as_count(shared) - return _as_count(await self._outcome(store.async_get_cache(key=key))) + """The shared count plus this worker's own. + + A failure is written to exactly one of the two: Redis, or this worker's store when Redis + refused it. So the local store is empty while Redis is healthy, and once Redis answers + again the guesses it missed still count. Read through ``async_batch_get_counts`` because + ``async_get_cache`` turns a failed GET into ``None``, which would pass as an empty counter. + """ + local: Final = _as_count(await self._outcome(store.async_get_cache(key=key))) + shared: Final = await self._shared(lambda redis_cache: redis_cache.async_batch_get_counts([key])) + if not isinstance(shared, tuple): + return local + return _as_count(shared[0]) + local async def _remaining_window(self, key: str) -> int: """Seconds until this counter expires. @@ -269,9 +275,11 @@ class LoginThrottle: shared: Final = await self._shared( lambda redis_cache: redis_cache.async_increment_with_floor(key, 1, self.window_seconds) ) - if shared is not _UNAVAILABLE: - return _as_count(shared) - return _as_count(await self._outcome(store.async_increment_cache(key=key, value=1, ttl=self.window_seconds))) + if shared is _UNAVAILABLE: + return _as_count( + await self._outcome(store.async_increment_cache(key=key, value=1, ttl=self.window_seconds)) + ) + return _as_count(shared) + _as_count(await self._outcome(store.async_get_cache(key=key))) async def record_failure(self, username: str) -> FailureCounts: """Count one rejected credential guess against this username and against this source.""" diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 70e7125485c..14f8dd19683 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1156,6 +1156,9 @@ class _FakeRedis: async def async_get_cache(self, key, **kwargs): return self.values.get(key) + async def async_batch_get_counts(self, key_list): + return tuple(self.values.get(key) for key in key_list) + async def async_increment_with_floor(self, key, value, ttl): self.values[key] = self.values.get(key, 0) + value self.ttls.setdefault(key, ttl) @@ -1204,9 +1207,16 @@ async def test_counters_are_written_with_their_expiry_and_re_armed_if_stripped(m class _DownRedis(_FakeRedis): - """Redis whose every call raises, as during an outage or an open circuit breaker.""" + """Redis whose every call fails, as during an outage or an open circuit breaker. + + `async_get_cache` returns None rather than raising, as the real one does: it swallows the + error, so a failed GET is indistinguishable from an empty key to anyone reading through it. + """ async def async_get_cache(self, key, **kwargs): + return None + + async def async_batch_get_counts(self, key_list): raise ConnectionError("redis is down") async def async_increment_with_floor(self, key, value, ttl): @@ -1280,6 +1290,50 @@ async def test_a_redis_outage_falls_back_to_this_workers_own_counter(monkeypatch assert blocked.value.headers.get("Retry-After") == "900" +class _WriteRefusingRedis(_FakeRedis): + """Redis that answers reads but raises on writes until `recover()` is called.""" + + def __init__(self): + super().__init__() + self.writable = False + + def recover(self): + self.writable = True + + async def async_increment_with_floor(self, key, value, ttl): + if not self.writable: + raise ConnectionError("redis write failed") + return await super().async_increment_with_floor(key, value, ttl) + + +@pytest.mark.asyncio +async def test_failures_redis_refused_still_count_once_redis_recovers(monkeypatch): + """Regression: a guess Redis could not record must not be forgotten when Redis comes back. + + Such a guess lands in this worker's own store. Reading only Redis afterwards handed the + attacker that guess again, so the budget was the limit plus however many writes failed. + """ + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + redis = _WriteRefusingRedis() + throttle = _throttle(max_attempts=2, redis_cache=redis) + + with pytest.raises(ProxyException, match="Invalid credentials"): + await _guess(throttle) + assert not redis.values, "the refused write must not have reached Redis" + + redis.recover() + with pytest.raises(ProxyException, match="Invalid credentials"): + await _guess(throttle) + assert [v for k, v in redis.values.items() if ":user:" in k] == [1], "only the recorded guess is in Redis" + + with pytest.raises(ProxyException) as blocked: + await _guess(throttle) + assert blocked.value.code == "429", "the guess Redis missed and the one it took must add up to the limit" + + @pytest.mark.asyncio async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): """Regression: throttle entries must not evict cached credentials. From 04b7b716561551814b641bd8ce0fa67b4cddad1f Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 11 Sep 2026 23:37:22 +0000 Subject: [PATCH 019/141] refactor(proxy): read the shared sign-in counter through a tuple of keys Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/caching/redis_cache.py | 4 ++-- litellm/proxy/auth/login_throttle.py | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/litellm/caching/redis_cache.py b/litellm/caching/redis_cache.py index 6b93529e456..f37ca8a23a1 100644 --- a/litellm/caching/redis_cache.py +++ b/litellm/caching/redis_cache.py @@ -791,7 +791,7 @@ class RedisCache(BaseCache): return _LUA_COUNT.validate_python(count) @_redis_circuit_breaker_guard_sync - def batch_get_counts(self, key_list: list[str]) -> tuple[int | None, ...]: + def batch_get_counts(self, key_list: Sequence[str]) -> tuple[int | None, ...]: """Read integer counters for ``key_list``, in order, raising when Redis cannot answer. ``batch_get_cache`` swallows every failure and returns an empty dict, which the caller @@ -802,7 +802,7 @@ class RedisCache(BaseCache): return _decoded_counts(self._run_redis_mget_operation(keys=namespaced_keys)) @_redis_circuit_breaker_guard - async def async_batch_get_counts(self, key_list: list[str]) -> tuple[int | None, ...]: + async def async_batch_get_counts(self, key_list: Sequence[str]) -> tuple[int | None, ...]: """Async twin of ``batch_get_counts``, raising on failure the same way.""" namespaced_keys: Final = [self.check_and_fix_namespace(key=key) for key in key_list] return _decoded_counts(await self._async_run_redis_mget_operation(keys=namespaced_keys)) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index ca73bb4f26b..5a56628bed2 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -207,7 +207,7 @@ class LoginThrottle: ``async_get_cache`` turns a failed GET into ``None``, which would pass as an empty counter. """ local: Final = _as_count(await self._outcome(store.async_get_cache(key=key))) - shared: Final = await self._shared(lambda redis_cache: redis_cache.async_batch_get_counts([key])) + shared: Final = await self._shared(lambda redis_cache: redis_cache.async_batch_get_counts((key,))) if not isinstance(shared, tuple): return local return _as_count(shared[0]) + local From 49d49b70508a3e15f360a5da4e754001e8eb36eb Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 11 Sep 2026 23:48:32 +0000 Subject: [PATCH 020/141] refactor(caching): spell out the key collections batch_get_counts accepts Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/caching/redis_cache.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/litellm/caching/redis_cache.py b/litellm/caching/redis_cache.py index f37ca8a23a1..f1b723de625 100644 --- a/litellm/caching/redis_cache.py +++ b/litellm/caching/redis_cache.py @@ -791,7 +791,7 @@ class RedisCache(BaseCache): return _LUA_COUNT.validate_python(count) @_redis_circuit_breaker_guard_sync - def batch_get_counts(self, key_list: Sequence[str]) -> tuple[int | None, ...]: + def batch_get_counts(self, key_list: list[str] | tuple[str, ...]) -> tuple[int | None, ...]: """Read integer counters for ``key_list``, in order, raising when Redis cannot answer. ``batch_get_cache`` swallows every failure and returns an empty dict, which the caller @@ -802,7 +802,7 @@ class RedisCache(BaseCache): return _decoded_counts(self._run_redis_mget_operation(keys=namespaced_keys)) @_redis_circuit_breaker_guard - async def async_batch_get_counts(self, key_list: Sequence[str]) -> tuple[int | None, ...]: + async def async_batch_get_counts(self, key_list: list[str] | tuple[str, ...]) -> tuple[int | None, ...]: """Async twin of ``batch_get_counts``, raising on failure the same way.""" namespaced_keys: Final = [self.check_and_fix_namespace(key=key) for key in key_list] return _decoded_counts(await self._async_run_redis_mget_operation(keys=namespaced_keys)) From c3a7b7c3eeb750e4fc1c7479fc502a2d143e2d2f Mon Sep 17 00:00:00 2001 From: yucheng Date: Sun, 13 Sep 2026 04:36:41 +0000 Subject: [PATCH 021/141] fix(proxy): warn about per-worker login counters even without general_settings Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/proxy_server.py | 13 +++++----- tests/test_litellm/proxy/test_proxy_server.py | 24 +++++++++++++++++++ 2 files changed, 31 insertions(+), 6 deletions(-) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 24532cc051c..138657fd5c3 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -5748,6 +5748,13 @@ class ProxyConfig: general_settings = config.get("general_settings", {}) if general_settings is None: general_settings = {} + + ### FAILED-LOGIN ACCOUNTING MULTI-INSTANCE PREREQUISITE CHECK ### + # Failed Admin UI sign-in counters live in redis_usage_cache when available so a + # brute-force run is counted once across workers instead of once per worker. + if os.getenv("NUM_WORKERS", "1") != "1" and redis_usage_cache is None: + warn_login_counters_are_per_worker(os.getenv("NUM_WORKERS", "1")) + _bg_hc_model_groups: Final = parse_background_health_check_model_groups(general_settings) _enable_hc_routing = False _hc_staleness = None @@ -5844,12 +5851,6 @@ class ProxyConfig: "or ensure sticky sessions for single-instance deployments." ) - ### FAILED-LOGIN ACCOUNTING MULTI-INSTANCE PREREQUISITE CHECK ### - # Failed Admin UI sign-in counters live in redis_usage_cache when available so a - # brute-force run is counted once across workers instead of once per worker. - if os.getenv("NUM_WORKERS", "1") != "1" and redis_usage_cache is None: - warn_login_counters_are_per_worker(os.getenv("NUM_WORKERS", "1")) - ### STORE MODEL IN DB ### feature flag for `/model/new` store_model_in_db = general_settings.get("store_model_in_db", False) if store_model_in_db is None: diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 2f8d5cbc43a..9a41b3a7006 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -3411,6 +3411,30 @@ async def test_load_config_user_url_validation_handles_null_and_string_false(tmp assert litellm.user_url_validation is False +@pytest.mark.asyncio +async def test_load_config_warns_per_worker_login_counters_without_general_settings(tmp_path, monkeypatch, caplog): + """Regression: the failed-login throttle is on by default, so a multi-worker proxy with no + Redis must hear that its counters are per worker even when the config has no general_settings.""" + import logging + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.auth.login_throttle import warn_login_counters_are_per_worker + from litellm.proxy.proxy_server import ProxyConfig + + for redis_var in ("REDIS_HOST", "REDIS_URL", "REDIS_CLUSTER_NODES", "REDIS_SENTINEL_NODES"): + monkeypatch.delenv(redis_var, raising=False) + monkeypatch.setenv("NUM_WORKERS", "4") + monkeypatch.setattr(proxy_server, "redis_usage_cache", None) + warn_login_counters_are_per_worker.cache_clear() + config_file = tmp_path / "config.yaml" + config_file.write_text("model_list: []\n") + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + await ProxyConfig().load_config(router=MagicMock(), config_file_path=str(config_file)) + + assert "Running 4 workers but Redis is not configured" in caplog.text + + @pytest.mark.asyncio async def test_load_environment_variables_direct_and_os_environ(): """ From 9ca6307b9d8fc60f557838ac6dc1b58d41f4c98d Mon Sep 17 00:00:00 2001 From: yucheng Date: Sun, 13 Sep 2026 08:28:06 +0000 Subject: [PATCH 022/141] chore(ui): regenerate schema.d.ts after merging main Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 -- 1 file changed, 2 deletions(-) diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index f1ef4351a8b..08f8f773a87 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -16781,7 +16781,6 @@ export interface paths { * - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking. * - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" } * - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x. - * - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests. * - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys) * - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}. * - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys) @@ -16887,7 +16886,6 @@ export interface paths { * - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking. * - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" } * - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x. - * - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests. * - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys) * - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}. * - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys) From ce82033f62c76a9d85332a171457a2b1fbbde757 Mon Sep 17 00:00:00 2001 From: yucheng Date: Sun, 13 Sep 2026 09:00:20 +0000 Subject: [PATCH 023/141] refactor(proxy): inject settings and Redis cache into LoginThrottle.from_request Removes the runtime import of proxy_server from login_throttle so the throttle module no longer participates in the import cycle CodeQL flagged (py/cyclic-import). Callers pass general_settings and redis_usage_cache explicitly; behavior is unchanged. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 12 +++--- litellm/proxy/proxy_server.py | 6 +-- .../proxy/auth/test_login_utils.py | 43 ++++++++----------- 3 files changed, 27 insertions(+), 34 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 5a56628bed2..b6ec37afc0e 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -11,7 +11,7 @@ startup and can be reassigned later. import asyncio import hashlib -from collections.abc import Awaitable, Callable +from collections.abc import Awaitable, Callable, Mapping from dataclasses import dataclass from functools import cache from types import MappingProxyType @@ -134,10 +134,10 @@ class LoginThrottle: enabled: bool = True @classmethod - def from_request(cls, request: Request) -> "LoginThrottle": - """Build the throttle for this request from the live proxy settings and caches.""" - from litellm.proxy.proxy_server import general_settings, redis_usage_cache - + def from_request( + cls, request: Request, general_settings: Mapping[str, object] | None, redis_cache: RedisCache | None + ) -> "LoginThrottle": + """Build the throttle for this request from the proxy's general_settings and shared Redis cache.""" settings: Final = general_settings or _NO_SETTINGS cidrs: Final = normalize_cidr_ranges( settings.get(TRUSTED_PROXY_RANGES_KEY), setting_name=TRUSTED_PROXY_RANGES_KEY @@ -167,7 +167,7 @@ class LoginThrottle: ), username_cache=_FAILED_LOGIN_USERNAME_CACHE, source_cache=_FAILED_LOGIN_SOURCE_CACHE, - redis_cache=redis_usage_cache, + redis_cache=redis_cache, enabled=not _rate_limit_disabled(), ) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 2e4417db696..37e7496610c 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -15853,7 +15853,7 @@ async def login(request: Request): password=password, master_key=master_key, prisma_client=prisma_client, - throttle=LoginThrottle.from_request(request), + throttle=LoginThrottle.from_request(request, general_settings, redis_usage_cache), general_settings=general_settings, ) except ProxyException as exc: @@ -15946,7 +15946,7 @@ async def login_v2(request: Request): password=password, master_key=master_key, prisma_client=prisma_client, - throttle=LoginThrottle.from_request(request), + throttle=LoginThrottle.from_request(request, general_settings, redis_usage_cache), general_settings=general_settings, ) @@ -16018,7 +16018,7 @@ async def login_v3(request: Request): password=password, master_key=master_key, prisma_client=prisma_client, - throttle=LoginThrottle.from_request(request), + throttle=LoginThrottle.from_request(request, general_settings, redis_usage_cache), general_settings=general_settings, ) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 14f8dd19683..7f5cd3e808f 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1348,7 +1348,6 @@ async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - monkeypatch.setattr(ps, "redis_usage_cache", None) auth_cache_keys_before = set(ps.user_api_key_cache.in_memory_cache.cache_dict) @@ -1356,7 +1355,7 @@ async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): request.headers = {} request.client = MagicMock() request.client.host = "1.2.3.4" - throttle = LoginThrottle.from_request(request) + throttle = LoginThrottle.from_request(request, general_settings={}, redis_cache=None) for i in range(25): with pytest.raises(ProxyException, match="Invalid credentials"): @@ -1368,30 +1367,28 @@ async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): ) -def test_settings_that_arrive_as_environment_strings_are_honored(monkeypatch): +def test_settings_that_arrive_as_environment_strings_are_honored(): """An `os.environ/VAR` reference in general_settings resolves to a string, not an int. Regression: a digit string fell back to the default with only a log line, so an operator tightening the limits through environment substitution silently kept the stock ceilings. """ - from litellm.proxy import proxy_server as ps from litellm.proxy.auth.login_throttle import LoginThrottle - monkeypatch.setattr( - ps, - "general_settings", - { - "max_failed_login_attempts": "7", - "max_failed_login_attempts_per_source": " 70 ", - "failed_login_window_seconds": "not-a-number", - }, - ) request = MagicMock() request.headers = {} request.client = MagicMock() request.client.host = "1.2.3.4" - throttle = LoginThrottle.from_request(request) + throttle = LoginThrottle.from_request( + request, + general_settings={ + "max_failed_login_attempts": "7", + "max_failed_login_attempts_per_source": " 70 ", + "failed_login_window_seconds": "not-a-number", + }, + redis_cache=None, + ) assert throttle.max_attempts == 7 assert throttle.max_attempts_per_source == 70 @@ -1404,41 +1401,37 @@ def test_the_disable_flag_is_read_once_not_per_login_attempt(monkeypatch): With a hosted secret manager in read mode that is a synchronous network call per guess, so a flood of wrong passwords could exhaust the secret manager even after the source was refused. """ - from litellm.proxy import proxy_server as ps from litellm.proxy.auth import login_throttle reads: Final[list[str]] = [] # mutable-ok: test-only call recorder monkeypatch.setattr(login_throttle, "get_secret_bool", lambda name, default: reads.append(name) or default) login_throttle._rate_limit_disabled.cache_clear() - monkeypatch.setattr(ps, "general_settings", {}) request = MagicMock() request.headers = {} request.client = MagicMock() request.client.host = "1.2.3.4" for _ in range(50): - assert login_throttle.LoginThrottle.from_request(request).enabled is True + assert login_throttle.LoginThrottle.from_request(request, general_settings={}, redis_cache=None).enabled is True login_throttle._rate_limit_disabled.cache_clear() assert reads == ["LITELLM_DISABLE_LOGIN_RATE_LIMIT"] -def test_a_negative_or_boolean_setting_falls_back_to_the_default(monkeypatch): +def test_a_negative_or_boolean_setting_falls_back_to_the_default(): """A limit below one would refuse everyone; a bool is a typo, not a count.""" - from litellm.proxy import proxy_server as ps from litellm.proxy.auth.login_throttle import LoginThrottle - monkeypatch.setattr( - ps, - "general_settings", - {"max_failed_login_attempts": "-7", "max_failed_login_attempts_per_source": True}, - ) request = MagicMock() request.headers = {} request.client = MagicMock() request.client.host = "1.2.3.4" - throttle = LoginThrottle.from_request(request) + throttle = LoginThrottle.from_request( + request, + general_settings={"max_failed_login_attempts": "-7", "max_failed_login_attempts_per_source": True}, + redis_cache=None, + ) assert throttle.max_attempts == 50 assert throttle.max_attempts_per_source == 250 From 05fe17e027325bfaa73086450240e2cc4a41700d Mon Sep 17 00:00:00 2001 From: yucheng Date: Sun, 13 Sep 2026 09:02:45 +0000 Subject: [PATCH 024/141] test(proxy): drop the section banner comment from the login tests Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/proxy/auth/test_login_utils.py | 5 ----- 1 file changed, 5 deletions(-) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 7f5cd3e808f..dadce1cd6c5 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -657,11 +657,6 @@ class TestEncodeUiSessionJwt: assert _user_id_from_session_cookie(request) == "cornell-user" -# --------------------------------------------------------------------------- -# Failed-login accounting (LIT-5285) -# --------------------------------------------------------------------------- - - def _throttle( max_attempts: int = 3, window_seconds: int = 900, From 685c6542985a6ae8cdb817f13a6a183a4111092b Mon Sep 17 00:00:00 2001 From: yucheng Date: Sun, 13 Sep 2026 09:18:39 +0000 Subject: [PATCH 025/141] refactor(proxy): assert separate login counter stores in the spray regression test instead of a comment Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 2 -- tests/test_litellm/proxy/auth/test_login_utils.py | 4 +++- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index b6ec37afc0e..50333ffad05 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -55,8 +55,6 @@ def _bounded_store(max_entries: int) -> DualCache: ) -# Separate stores: eviction is earliest-expiring-first, so in one shared store a spray of -# fresh usernames would evict the source counter that is meant to stop that same spray. _FAILED_LOGIN_USERNAME_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_USERNAMES) _FAILED_LOGIN_SOURCE_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_SOURCES) _NO_SETTINGS: Final = MappingProxyType({}) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index dadce1cd6c5..c53476f5148 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1472,7 +1472,8 @@ async def test_a_username_spray_cannot_evict_an_existing_counter(monkeypatch): The default in-memory cache keeps 200 entries and evicts the soonest to expire, and every counter shares one window, so eviction was effectively oldest-first. A few hundred made-up usernames therefore pushed out the attacker's own counter and handed - back a fresh allowance against the real account. + back a fresh allowance against the real account. Username and source counters must also + live in separate stores, or the same spray evicts the source counter meant to stop it. """ from litellm.proxy._types import ProxyException from litellm.proxy.auth.login_throttle import ( @@ -1489,6 +1490,7 @@ async def test_a_username_spray_cannot_evict_an_existing_counter(monkeypatch): assert _MAX_TRACKED_LOGIN_USERNAMES >= 10_000 assert _FAILED_LOGIN_SOURCE_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_SOURCES assert _FAILED_LOGIN_USERNAME_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_USERNAMES + assert _FAILED_LOGIN_SOURCE_CACHE.in_memory_cache is not _FAILED_LOGIN_USERNAME_CACHE.in_memory_cache throttle = LoginThrottle( client_ip="10.9.9.9", From c19a19999f86e3b1615e444714cdcf29be2f4349 Mon Sep 17 00:00:00 2001 From: Chloe Lu Date: Tue, 15 Sep 2026 15:34:24 +0800 Subject: [PATCH 026/141] fix(anthropic): register thinking-binding-controls-2026-08-01 in beta headers config Anthropic's preserved-thinking controls (`thinking.block_binding`, Claude Fable 5.1) are only accepted alongside the beta header `thinking-binding-controls-2026-08-01`. The proxy forwards the body field untouched but `filter_and_transform_beta_headers` drops the header because it has no entry in `anthropic_beta_headers_config.json`, so Bedrock and Vertex reject the request with "thinking.adaptive.block_binding: Extra inputs are not permitted". Map the header for anthropic, bedrock, bedrock_converse, vertex_ai and databricks (same beta name on all of them per Anthropic's docs). azure_ai is left null pending verification on Foundry. --- litellm/anthropic_beta_headers_config.json | 6 ++++++ .../test_anthropic_beta_headers_filtering.py | 16 ++++++++++++++++ 2 files changed, 22 insertions(+) diff --git a/litellm/anthropic_beta_headers_config.json b/litellm/anthropic_beta_headers_config.json index 3f6817f6e35..38fb9c9462f 100644 --- a/litellm/anthropic_beta_headers_config.json +++ b/litellm/anthropic_beta_headers_config.json @@ -27,6 +27,7 @@ "structured-outputs-2025-11-13": "structured-outputs-2025-11-13", "text_editor_20241022": null, "text_editor_20250124": null, + "thinking-binding-controls-2026-08-01": "thinking-binding-controls-2026-08-01", "token-efficient-tools-2025-02-19": "token-efficient-tools-2025-02-19", "web-fetch-2025-09-10": "web-fetch-2025-09-10", "web-search-2025-03-05": "web-search-2025-03-05" @@ -57,6 +58,7 @@ "structured-outputs-2025-11-13": "structured-outputs-2025-11-13", "text_editor_20241022": null, "text_editor_20250124": null, + "thinking-binding-controls-2026-08-01": null, "token-efficient-tools-2025-02-19": null, "web-fetch-2025-09-10": "web-fetch-2025-09-10", "web-search-2025-03-05": "web-search-2025-03-05" @@ -87,6 +89,7 @@ "structured-outputs-2025-11-13": "structured-outputs-2025-11-13", "text_editor_20241022": null, "text_editor_20250124": null, + "thinking-binding-controls-2026-08-01": "thinking-binding-controls-2026-08-01", "token-efficient-tools-2025-02-19": null, "tool-search-tool-2025-10-19": null, "web-fetch-2025-09-10": null, @@ -118,6 +121,7 @@ "structured-outputs-2025-11-13": null, "text_editor_20241022": null, "text_editor_20250124": null, + "thinking-binding-controls-2026-08-01": "thinking-binding-controls-2026-08-01", "token-efficient-tools-2025-02-19": null, "tool-search-tool-2025-10-19": "tool-search-tool-2025-10-19", "web-fetch-2025-09-10": null, @@ -149,6 +153,7 @@ "structured-outputs-2025-11-13": null, "text_editor_20241022": null, "text_editor_20250124": null, + "thinking-binding-controls-2026-08-01": "thinking-binding-controls-2026-08-01", "token-efficient-tools-2025-02-19": null, "tool-search-tool-2025-10-19": "tool-search-tool-2025-10-19", "web-fetch-2025-09-10": null, @@ -181,6 +186,7 @@ "structured-outputs-2025-11-13": "structured-outputs-2025-11-13", "text_editor_20241022": null, "text_editor_20250124": null, + "thinking-binding-controls-2026-08-01": "thinking-binding-controls-2026-08-01", "token-efficient-tools-2025-02-19": "token-efficient-tools-2025-02-19", "web-fetch-2025-09-10": "web-fetch-2025-09-10", "web-search-2025-03-05": "web-search-2025-03-05" diff --git a/tests/test_litellm/test_anthropic_beta_headers_filtering.py b/tests/test_litellm/test_anthropic_beta_headers_filtering.py index 59bab22de74..3c967283abf 100644 --- a/tests/test_litellm/test_anthropic_beta_headers_filtering.py +++ b/tests/test_litellm/test_anthropic_beta_headers_filtering.py @@ -426,6 +426,22 @@ class TestAnthropicBetaHeadersFiltering: assert filtered == ["fine-grained-tool-streaming-2025-05-14"] + @pytest.mark.parametrize( + "provider", ["anthropic", "bedrock", "bedrock_converse", "vertex_ai", "databricks"] + ) + def test_thinking_binding_controls_forwarded(self, provider): + """`thinking.block_binding` (preserved thinking, Claude Fable 5.1) is only + accepted alongside thinking-binding-controls-2026-08-01. The body field is + forwarded untouched, so stripping the header (previously unknown, hence + dropped) makes Bedrock and Vertex reject the request with + "thinking.adaptive.block_binding: Extra inputs are not permitted".""" + filtered = filter_and_transform_beta_headers( + beta_headers=["thinking-binding-controls-2026-08-01"], + provider=provider, + ) + + assert filtered == ["thinking-binding-controls-2026-08-01"] + def test_null_value_headers_filtered(self): """Test that headers with null values are always filtered out.""" for provider in [ From 163c0f3aee99ba61f12317453cd76741b9f1558b Mon Sep 17 00:00:00 2001 From: clonylu Date: Tue, 15 Sep 2026 15:49:06 +0800 Subject: [PATCH 027/141] fix(router): honor stream_timeout on the SDK-native passthrough route Anthropic /v1/messages and Bedrock /converse resolve their upstream timeout through resolve_llm_passthrough_timeout, which only reads timeout / request_timeout and then falls back to the 600s pass_through default. A stream_timeout set on the deployment or in router_settings was never consulted on that route, while /chat/completions honors it through Router._get_stream_timeout. For a streaming call the resolver now checks stream_timeout at each level before the non-stream key (kwargs -> litellm_params -> router), mirroring _get_stream_timeout; non-streaming resolution is unchanged. The router passes its stream_timeout alongside the explicit timeout. --- litellm/passthrough/timeout_utils.py | 23 ++++++-- litellm/router.py | 4 ++ .../test_pass_through_endpoints.py | 58 +++++++++++++++++++ tests/test_litellm/test_router.py | 58 +++++++++++++++++++ 4 files changed, 139 insertions(+), 4 deletions(-) diff --git a/litellm/passthrough/timeout_utils.py b/litellm/passthrough/timeout_utils.py index 39127d19183..fb649a9eeaf 100644 --- a/litellm/passthrough/timeout_utils.py +++ b/litellm/passthrough/timeout_utils.py @@ -34,22 +34,37 @@ def resolve_llm_passthrough_timeout( kwargs: dict | None = None, litellm_params: dict | None = None, router_timeout: float | None = None, + router_stream_timeout: float | None = None, ) -> float: """ - Resolve upstream httpx timeout for SDK native passthrough (e.g. Bedrock /converse). + Resolve upstream httpx timeout for SDK native passthrough (e.g. Bedrock /converse, + Anthropic /v1/messages). - Precedence: kwargs timeout/request_timeout -> litellm_params timeout/request_timeout - -> router_timeout -> general_settings.pass_through_request_timeout -> 600s default. + Non-streaming precedence: kwargs timeout/request_timeout -> litellm_params + timeout/request_timeout -> router_timeout -> general_settings.pass_through_request_timeout + -> 600s default. + + Streaming (``kwargs["stream"]`` truthy) additionally consults ``stream_timeout`` at each + level before the non-streaming key, matching ``Router._get_stream_timeout`` on the + completion route: kwargs stream_timeout -> kwargs timeout/request_timeout -> + litellm_params stream_timeout -> litellm_params timeout/request_timeout -> + router_stream_timeout -> router_timeout -> pass_through_request_timeout -> 600s. """ kwargs = kwargs or {} litellm_params = litellm_params or {} + is_stream: Final[bool] = bool(kwargs.get("stream", False)) + keys: Final[tuple[str, ...]] = ( + ("stream_timeout", "timeout", "request_timeout") if is_stream else ("timeout", "request_timeout") + ) for source in (kwargs, litellm_params): - for key in ("timeout", "request_timeout"): + for key in keys: val = source.get(key) if val is not None: return float(val) + if is_stream and router_stream_timeout is not None: + return float(router_stream_timeout) if router_timeout is not None: return float(router_timeout) diff --git a/litellm/router.py b/litellm/router.py index 1665583386f..7ce7ba30502 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -3879,10 +3879,14 @@ class Router: _router_timeout: Final = ( float(self._explicit_timeout) if isinstance(self._explicit_timeout, (int, float)) else None ) + _router_stream_timeout: Final = ( + float(self.stream_timeout) if isinstance(self.stream_timeout, (int, float)) else None + ) kwargs["timeout"] = resolve_llm_passthrough_timeout( kwargs=kwargs, litellm_params=deployment["litellm_params"], router_timeout=_router_timeout, + router_stream_timeout=_router_stream_timeout, ) else: kwargs["timeout"] = self._get_timeout(kwargs=kwargs, data=deployment["litellm_params"]) diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index d57bed430c1..d697a114613 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -1119,6 +1119,64 @@ def test_resolve_llm_passthrough_timeout_precedence(): assert resolve_llm_passthrough_timeout() == 6.0 +def test_resolve_llm_passthrough_timeout_stream_timeout_precedence(): + # streaming: stream_timeout wins at each level, then falls through to the non-stream keys + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": True, "stream_timeout": 1800, "timeout": 45}, + ) + == 1800.0 + ) + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": True}, + litellm_params={"stream_timeout": 1800, "timeout": 90}, + ) + == 1800.0 + ) + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": True}, + litellm_params={"timeout": 90}, + router_stream_timeout=1800, + ) + == 90.0 + ) + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": True}, + router_timeout=120, + router_stream_timeout=1800, + ) + == 1800.0 + ) + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": True}, + router_timeout=120, + ) + == 120.0 + ) + + # non-streaming: stream_timeout is ignored everywhere + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": False, "stream_timeout": 1800}, + litellm_params={"stream_timeout": 1800, "timeout": 90}, + router_stream_timeout=1800, + ) + == 90.0 + ) + with patch("litellm.proxy.proxy_server.general_settings", {}): + assert ( + resolve_llm_passthrough_timeout( + litellm_params={"stream_timeout": 1800}, + router_stream_timeout=1800, + ) + == DEFAULT_PASS_THROUGH_REQUEST_TIMEOUT_SECONDS + ) + + @pytest.mark.asyncio async def test_pass_through_request_uses_resolved_timeout(): with patch("litellm.proxy.proxy_server.proxy_logging_obj") as mock_proxy_logging: diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 3cabcd71627..70ce182c028 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -5480,6 +5480,64 @@ def test_update_kwargs_with_deployment_uses_pass_through_request_timeout(): assert kwargs["timeout"] == 6.0 +def test_update_kwargs_with_deployment_passthrough_honors_stream_timeout(): + """ + The SDK-native passthrough route (anthropic /v1/messages, bedrock /converse) resolves + its upstream timeout separately from the completion route. A streaming call must get + stream_timeout (deployment litellm_params first, then router_settings), while a + non-streaming call on the same deployment keeps the non-stream resolution. + """ + router = litellm.Router( + model_list=[ + { + "model_name": "anthropic-with-stream-timeout", + "litellm_params": { + "model": "anthropic/claude-sonnet-4-5", + "api_key": "fake-key", + "stream_timeout": 1800, + }, + }, + { + "model_name": "anthropic-router-default", + "litellm_params": { + "model": "anthropic/claude-sonnet-4-5", + "api_key": "fake-key", + }, + }, + ], + stream_timeout=900, + ) + per_deployment, router_default = router.model_list + + with patch( + "litellm.proxy.proxy_server.general_settings", + {"pass_through_request_timeout": 6}, + ): + kwargs: dict = {"stream": True} + router._update_kwargs_with_deployment( + deployment=per_deployment, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + assert kwargs["timeout"] == 1800.0 + + kwargs = {"stream": True} + router._update_kwargs_with_deployment( + deployment=router_default, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + assert kwargs["timeout"] == 900.0 + + kwargs = {"stream": False} + router._update_kwargs_with_deployment( + deployment=per_deployment, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + assert kwargs["timeout"] == 6.0 + + @pytest.mark.asyncio async def test_router_acompletion_with_unknown_model_and_default_fallback(): """ From efb2bcd87fae4c6a78bb562cbcde98778b967a79 Mon Sep 17 00:00:00 2001 From: clonylu Date: Tue, 15 Sep 2026 16:11:55 +0800 Subject: [PATCH 028/141] test(router): cover passthrough stream_timeout without patching proxy globals --- .../test_pass_through_endpoints.py | 14 ++--- tests/test_litellm/test_router.py | 56 ++++++++++--------- 2 files changed, 38 insertions(+), 32 deletions(-) diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index d697a114613..7f8663ea860 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -1167,14 +1167,14 @@ def test_resolve_llm_passthrough_timeout_stream_timeout_precedence(): ) == 90.0 ) - with patch("litellm.proxy.proxy_server.general_settings", {}): - assert ( - resolve_llm_passthrough_timeout( - litellm_params={"stream_timeout": 1800}, - router_stream_timeout=1800, - ) - == DEFAULT_PASS_THROUGH_REQUEST_TIMEOUT_SECONDS + assert ( + resolve_llm_passthrough_timeout( + litellm_params={"stream_timeout": 1800}, + router_timeout=120, + router_stream_timeout=1800, ) + == 120.0 + ) @pytest.mark.asyncio diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 70ce182c028..4c39f8ba4e4 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -5494,6 +5494,7 @@ def test_update_kwargs_with_deployment_passthrough_honors_stream_timeout(): "litellm_params": { "model": "anthropic/claude-sonnet-4-5", "api_key": "fake-key", + "timeout": 60, "stream_timeout": 1800, }, }, @@ -5505,37 +5506,42 @@ def test_update_kwargs_with_deployment_passthrough_honors_stream_timeout(): }, }, ], + timeout=120, stream_timeout=900, ) per_deployment, router_default = router.model_list - with patch( - "litellm.proxy.proxy_server.general_settings", - {"pass_through_request_timeout": 6}, - ): - kwargs: dict = {"stream": True} - router._update_kwargs_with_deployment( - deployment=per_deployment, - kwargs=kwargs, - function_name="_ageneric_api_call_with_fallbacks", - ) - assert kwargs["timeout"] == 1800.0 + kwargs: dict = {"stream": True} + router._update_kwargs_with_deployment( + deployment=per_deployment, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + assert kwargs["timeout"] == 1800.0 - kwargs = {"stream": True} - router._update_kwargs_with_deployment( - deployment=router_default, - kwargs=kwargs, - function_name="_ageneric_api_call_with_fallbacks", - ) - assert kwargs["timeout"] == 900.0 + kwargs = {"stream": True} + router._update_kwargs_with_deployment( + deployment=router_default, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + assert kwargs["timeout"] == 900.0 - kwargs = {"stream": False} - router._update_kwargs_with_deployment( - deployment=per_deployment, - kwargs=kwargs, - function_name="_ageneric_api_call_with_fallbacks", - ) - assert kwargs["timeout"] == 6.0 + kwargs = {"stream": False} + router._update_kwargs_with_deployment( + deployment=per_deployment, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + assert kwargs["timeout"] == 60.0 + + kwargs = {"stream": False} + router._update_kwargs_with_deployment( + deployment=router_default, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + assert kwargs["timeout"] == 120.0 @pytest.mark.asyncio From 911f66aff69fabb6666bde3f54db70960cb04b56 Mon Sep 17 00:00:00 2001 From: Emerson Gomes Date: Tue, 15 Sep 2026 12:12:57 -0500 Subject: [PATCH 029/141] feat(azure_ai): support FLUX.2 flex images --- litellm/images/main.py | 9 +- litellm/images/utils.py | 7 +- .../litellm_core_utils/llm_cost_calc/utils.py | 3 + .../image_edit/flux2_transformation.py | 64 ++++--- .../image_generation/cost_calculator.py | 34 +++- .../image_generation/flux_transformation.py | 91 +++++++-- ...odel_prices_and_context_window_backup.json | 19 ++ litellm/proxy/_lazy_openapi_snapshot.json | 2 +- litellm/types/llms/openai.py | 4 + model_prices_and_context_window.json | 19 ++ ...test_azure_ai_image_edit_transformation.py | 116 ++++++++++++ .../test_azure_ai_flux2_image_generation.py | 172 ++++++++++++++++++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 4 +- 13 files changed, 491 insertions(+), 53 deletions(-) create mode 100644 tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py diff --git a/litellm/images/main.py b/litellm/images/main.py index 6a94e7c8df2..81547a153c3 100644 --- a/litellm/images/main.py +++ b/litellm/images/main.py @@ -846,7 +846,12 @@ def image_edit( local_vars.update(kwargs) # Get ImageEditOptionalRequestParams with only valid parameters image_edit_optional_params: Final[ImageEditOptionalRequestParams] = ( - _get_ImageEditRequestUtils().get_requested_image_edit_optional_param(local_vars) + _get_ImageEditRequestUtils().get_requested_image_edit_optional_param( + local_vars, + provider_supported_params=frozenset( + image_edit_provider_config.get_supported_openai_params(model) + ).intersection(non_default_params), + ) ) # Get optional parameters for the responses API image_edit_request_params: Final[dict] = _get_ImageEditRequestUtils().get_optional_params_image_edit( @@ -857,7 +862,7 @@ def image_edit( additional_drop_params=kwargs.get("additional_drop_params"), ) - if ( + if image_edit_provider_config.use_multipart_form_data() and ( custom_llm_provider == "openai" or custom_llm_provider == "azure" or custom_llm_provider in litellm.openai_compatible_providers diff --git a/litellm/images/utils.py b/litellm/images/utils.py index 49b70870de6..24454954714 100644 --- a/litellm/images/utils.py +++ b/litellm/images/utils.py @@ -1,4 +1,4 @@ -from collections.abc import Mapping +from collections.abc import Collection, Mapping from io import BufferedReader, BytesIO from typing import Any, Final, cast, get_type_hints @@ -63,6 +63,7 @@ class ImageEditRequestUtils: @staticmethod def get_requested_image_edit_optional_param( params: Mapping[str, object], + provider_supported_params: Collection[str] = (), ) -> ImageEditOptionalRequestParams: """ Filter parameters to only include those defined in ImageEditOptionalRequestParams. @@ -73,7 +74,9 @@ class ImageEditRequestUtils: Returns: ImageEditOptionalRequestParams instance with only the valid parameters """ - valid_keys: Final = get_type_hints(ImageEditOptionalRequestParams).keys() + valid_keys: Final = frozenset(get_type_hints(ImageEditOptionalRequestParams)) | frozenset( + provider_supported_params + ) filtered_params: Final = {k: v for k, v in params.items() if k in valid_keys and v is not None} return cast(ImageEditOptionalRequestParams, filtered_params) diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py index baa9aab1087..0a0e92ff3a3 100644 --- a/litellm/litellm_core_utils/llm_cost_calc/utils.py +++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py @@ -1853,6 +1853,9 @@ class CostCalculatorUtils: return azure_ai_image_cost_calculator( model=model, image_response=completion_response, + size=resolved_size, + n=resolved_n, + optional_params=optional_params, ) elif custom_llm_provider == litellm.LlmProviders.FAL_AI.value: from litellm.llms.fal_ai.cost_calculator import ( diff --git a/litellm/llms/azure_ai/image_edit/flux2_transformation.py b/litellm/llms/azure_ai/image_edit/flux2_transformation.py index a09a80985b7..aa8905e5601 100644 --- a/litellm/llms/azure_ai/image_edit/flux2_transformation.py +++ b/litellm/llms/azure_ai/image_edit/flux2_transformation.py @@ -1,5 +1,7 @@ import base64 +from collections.abc import Mapping, Sequence from io import BufferedReader +from types import MappingProxyType from typing import Any, Final from httpx._types import RequestFiles @@ -24,7 +26,7 @@ class AzureFoundryFlux2ImageEditConfig(OpenAIImageEditConfig): Azure AI Foundry FLUX 2 image edit config Supports FLUX 2 models (e.g., flux.2-pro) for image editing. - Uses the same /providers/blackforestlabs/v1/flux-2-pro endpoint as image generation, + Uses the model-specific /providers/blackforestlabs/v1/flux-2-* endpoint as image generation, with the image passed as base64 in JSON body. """ @@ -33,11 +35,17 @@ class AzureFoundryFlux2ImageEditConfig(OpenAIImageEditConfig): FLUX 2 supports a subset of OpenAI image edit params """ return [ - "prompt", - "image", - "model", "n", "size", + "width", + "height", + "num_images", + "seed", + "safety_tolerance", + "output_format", + "aspect_ratio", + "guidance", + "steps", ] def map_openai_params( @@ -50,14 +58,14 @@ class AzureFoundryFlux2ImageEditConfig(OpenAIImageEditConfig): Map OpenAI params to FLUX 2 params. FLUX 2 uses the same param names as OpenAI for supported params. """ - mapped_params: Final[dict[str, Any]] = {} - supported_params: Final = self.get_supported_openai_params(model) - - for key, value in dict(image_edit_optional_params).items(): - if key in supported_params and value is not None: - mapped_params[key] = value - - return mapped_params + return AzureFoundryFluxImageGenerationConfig().map_openai_params( + non_default_params=MappingProxyType( + {key: value for key, value in image_edit_optional_params.items() if value is not None} + ), + optional_params=MappingProxyType({}), + model=model, + drop_params=drop_params, + ) def use_multipart_form_data(self) -> bool: """FLUX 2 uses JSON requests, not multipart/form-data.""" @@ -90,7 +98,7 @@ class AzureFoundryFlux2ImageEditConfig(OpenAIImageEditConfig): self, model: str, prompt: str | None, - image: FileTypes | None, + image: FileTypes | Sequence[FileTypes] | None, image_edit_optional_request_params: dict, litellm_params: GenericLiteLLMParams, headers: dict, @@ -107,29 +115,29 @@ class AzureFoundryFlux2ImageEditConfig(OpenAIImageEditConfig): if image is None: raise ValueError("FLUX 2 image edit requires an image.") - image_b64: Final = self._convert_image_to_base64(image) + images: Final = tuple(image) if isinstance(image, list) else (image,) + if not images: + raise ValueError("FLUX 2 image edit requires at least one image.") + max_reference_images: Final = 10 if "flex" in model.lower() else 8 + if len(images) > max_reference_images: + raise ValueError(f"{model} supports at most {max_reference_images} reference images.") - # Build request body with required params + reference_images: Final[Mapping[str, str]] = MappingProxyType( + { + "input_image" if index == 1 else f"input_image_{index}": self._convert_image_to_base64(reference_image) + for index, reference_image in enumerate(images, start=1) + } + ) request_body: Final[dict[str, Any]] = { "prompt": prompt, - "image": image_b64, "model": model, + **reference_images, + **image_edit_optional_request_params, } - - # Add mapped optional params (already filtered by map_openai_params) - request_body.update(image_edit_optional_request_params) - - # Return JSON body and empty files list (FLUX 2 doesn't use multipart) return request_body, [] def _convert_image_to_base64(self, image: Any) -> str: """Convert image file to base64 string""" - # Handle list of images (take first one) - if isinstance(image, list): - if len(image) == 0: - raise ValueError("Empty image list provided") - image = image[0] - if isinstance(image, BufferedReader): image_bytes = image.read() image.seek(0) # Reset file pointer for potential reuse @@ -151,7 +159,7 @@ class AzureFoundryFlux2ImageEditConfig(OpenAIImageEditConfig): """ Constructs a complete URL for Azure AI Foundry FLUX 2 image edits. - Uses the same /providers/blackforestlabs/v1/flux-2-pro endpoint as image generation. + Uses the same model-specific BFL provider endpoint as image generation. """ api_base = AzureFoundryModelInfo.get_api_base(api_base) diff --git a/litellm/llms/azure_ai/image_generation/cost_calculator.py b/litellm/llms/azure_ai/image_generation/cost_calculator.py index 106c7e42b83..086293f26c0 100644 --- a/litellm/llms/azure_ai/image_generation/cost_calculator.py +++ b/litellm/llms/azure_ai/image_generation/cost_calculator.py @@ -1,3 +1,4 @@ +from collections.abc import Mapping from typing import Any, Final import litellm @@ -10,6 +11,9 @@ from litellm.types.utils import ImageResponse def cost_calculator( model: str, image_response: Any, + size: str | None = None, + n: int | None = None, + optional_params: Mapping[str, object] | None = None, ) -> float: """ Azure AI image generation cost calculator @@ -28,10 +32,32 @@ def cost_calculator( if token_based_cost is not None: return token_based_cost + num_images: Final = n if n is not None else len(image_response.data or ()) output_cost_per_image: Final[float] = _model_info.get("output_cost_per_image") or 0.0 - num_images: int = 0 - if image_response.data: - num_images = len(image_response.data) - return output_cost_per_image * num_images + if output_cost_per_image: + return output_cost_per_image * num_images + + model_cost: Final = litellm.model_cost[_model_info["key"]] + input_cost_per_pixel: Final[float] = model_cost.get("input_cost_per_pixel") or 0.0 + if input_cost_per_pixel: + from litellm.cost_calculator import default_image_cost_calculator + + cost_model: Final = ( + model if model.startswith(f"{litellm.LlmProviders.AZURE_AI.value}/") else f"azure_ai/{model}" + ) + width: Final = optional_params.get("width") if optional_params else None + height: Final = optional_params.get("height") if optional_params else None + pixel_size: Final = ( + f"{width}x{height}" + if type(width) is int and type(height) is int and width > 0 and height > 0 + else size or image_response.size + ) + return default_image_cost_calculator( + model=cost_model, + custom_llm_provider=litellm.LlmProviders.AZURE_AI.value, + size=pixel_size, + n=num_images, + ) + return 0.0 raise ValueError(f"image_response must be of type ImageResponse got type={type(image_response)}") diff --git a/litellm/llms/azure_ai/image_generation/flux_transformation.py b/litellm/llms/azure_ai/image_generation/flux_transformation.py index 65b5a35af52..b10e8a6f35e 100644 --- a/litellm/llms/azure_ai/image_generation/flux_transformation.py +++ b/litellm/llms/azure_ai/image_generation/flux_transformation.py @@ -1,18 +1,13 @@ +from collections.abc import Mapping +from types import MappingProxyType from typing import Final from litellm.llms.openai.image_generation import GPTImageGenerationConfig +from litellm.types.llms.openai import OpenAIImageGenerationOptionalParams class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): - """ - Azure Foundry flux image generation config - - From manual testing it follows the gpt-image-1 image generation config - - (Azure Foundry does not have any docs on supported params at the time of writing) - - From our test suite - following GPTImageGenerationConfig is working for this model - """ + """Azure Foundry BFL API configuration for FLUX image generation.""" @staticmethod def get_flux2_image_generation_url( @@ -25,11 +20,11 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): FLUX 2 models on Azure AI use a different URL pattern than standard Azure OpenAI: - Standard: /openai/deployments/{model}/images/generations - - FLUX 2: /providers/blackforestlabs/v1/flux-2-pro + - FLUX 2: /providers/blackforestlabs/v1/{model-path} Args: api_base: Base URL (e.g., https://litellm-ci-cd-prod.services.ai.azure.com) - model: Model name (e.g., flux.2-pro) + model: Model name (e.g., FLUX.2-flex or FLUX.2-pro) api_version: API version (e.g., preview) Returns: @@ -47,9 +42,8 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): return api_base return f"{api_base}?api-version={api_version}" - # Construct the FLUX 2 provider path - # Model name flux.2-pro maps to endpoint flux-2-pro - return f"{api_base}/providers/blackforestlabs/v1/flux-2-pro?api-version={api_version}" + provider_model_path: Final = AzureFoundryFluxImageGenerationConfig.get_flux2_provider_model_path(model) + return f"{api_base}/providers/blackforestlabs/v1/{provider_model_path}?api-version={api_version}" @staticmethod def is_flux2_model(model: str) -> bool: @@ -64,3 +58,72 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): """ model_lower: Final = model.lower().replace(".", "-").replace("_", "-") return "flux-2" in model_lower or "flux2" in model_lower + + @staticmethod + def get_flux2_provider_model_path(model: str) -> str: + normalized_model: Final = model.lower().replace(".", "-").replace("_", "-") + return "flux-2-flex" if "flux-2-flex" in normalized_model else "flux-2-pro" + + def get_supported_openai_params( # mutable-ok: inherited config contract returns a list + self, model: str + ) -> list[OpenAIImageGenerationOptionalParams]: + if not self.is_flux2_model(model): + return super().get_supported_openai_params(model) + return [ # mutable-ok: BaseImageGenerationConfig requires a list + "n", + "size", + "output_format", + "seed", + "safety_tolerance", + "aspect_ratio", + "width", + "height", + "num_images", + "guidance", + "steps", + ] + + @staticmethod + def _map_parameter(name: str, value: object) -> tuple[tuple[str, object], ...]: + if name == "n": + return (("num_images", value),) + if name != "size": + return ((name, value),) + + try: + width, height = (int(dimension) for dimension in str(value).lower().split("x")) + except (TypeError, ValueError): + raise ValueError(f"Invalid size format '{value}'. Expected 'WxH', for example '1024x1024'.") + return (("width", width), ("height", height)) + + def map_openai_params( + self, + non_default_params: Mapping[str, object], + optional_params: Mapping[str, object], + model: str, + drop_params: bool, + ) -> dict: # mutable-ok: inherited config contract returns a dict + if not self.is_flux2_model(model): + return super().map_openai_params( + non_default_params=dict(non_default_params), + optional_params=dict(optional_params), + model=model, + drop_params=drop_params, + ) + supported_params: Final = self.get_supported_openai_params(model) + unsupported_params: Final = tuple(name for name in non_default_params if name not in supported_params) + if unsupported_params and not drop_params: + raise ValueError( + f"Parameters {unsupported_params} are not supported for model {model}. " + f"Supported parameters are {supported_params}. Set drop_params=True to drop unsupported parameters." + ) + + mapped_params: Final[Mapping[str, object]] = MappingProxyType( + { + mapped_name: mapped_value + for name, value in non_default_params.items() + if name in supported_params + for mapped_name, mapped_value in self._map_parameter(name, value) + } + ) + return {**optional_params, **mapped_params} # mutable-ok: inherited config contract returns a dict diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 9f91cf82f41..bcb2dfc53e8 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -9900,6 +9900,25 @@ "/v1/images/generations" ] }, + "azure_ai/FLUX.2-flex": { + "input_cost_per_pixel": 5e-08, + "litellm_provider": "azure_ai", + "max_input_tokens": 32000, + "max_tokens": 32000, + "mode": "image_generation", + "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/black-forest-labs/", + "supported_endpoints": [ + "/v1/images/generations", + "/v1/images/edits" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "image" + ] + }, "azure_ai/FW-DeepSeek-V3.2": { "deprecation_date": "2027-07-01", "cache_read_input_token_cost": 3.1e-07, diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index fb2f014e3d8..81889e4da0c 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -18974,7 +18974,7 @@ } } }, - "description": "\n Unified rate-limit error.\n\n Every rate-limit condition surfaced by litellm \u2014 whether it originated from\n an upstream LLM provider, a vendor batch endpoint, or one of litellm's own\n proxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\n max-iterations, etc.) \u2014 is raised as an instance of this class.\n\n The :attr:`category` attribute lets callers distinguish the source. See\n :class:`RateLimitErrorCategory` for the available values.\n " + "description": "\nUnified rate-limit error.\n\nEvery rate-limit condition surfaced by litellm \u2014 whether it originated from\nan upstream LLM provider, a vendor batch endpoint, or one of litellm's own\nproxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\nmax-iterations, etc.) \u2014 is raised as an instance of this class.\n\nThe :attr:`category` attribute lets callers distinguish the source. See\n:class:`RateLimitErrorCategory` for the available values.\n" }, "500": { "content": { diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py index e3eac9b9205..710b34116e5 100644 --- a/litellm/types/llms/openai.py +++ b/litellm/types/llms/openai.py @@ -1160,6 +1160,10 @@ OpenAIImageGenerationOptionalParams = Literal[ "image_url", "image_prompt_strength", "aspect_ratio", + "width", + "height", + "guidance", + "steps", "imageConfig", ] diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 9f91cf82f41..bcb2dfc53e8 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -9900,6 +9900,25 @@ "/v1/images/generations" ] }, + "azure_ai/FLUX.2-flex": { + "input_cost_per_pixel": 5e-08, + "litellm_provider": "azure_ai", + "max_input_tokens": 32000, + "max_tokens": 32000, + "mode": "image_generation", + "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/black-forest-labs/", + "supported_endpoints": [ + "/v1/images/generations", + "/v1/images/edits" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "image" + ] + }, "azure_ai/FW-DeepSeek-V3.2": { "deprecation_date": "2027-07-01", "cache_read_input_token_cost": 3.1e-07, diff --git a/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py b/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py index b6cb7ea9b54..94b23c5f1c2 100644 --- a/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py +++ b/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py @@ -1,12 +1,20 @@ +import base64 +import json +from collections.abc import Mapping +from typing import Final +import httpx +import pytest import litellm +from litellm.images.utils import ImageEditRequestUtils from litellm.llms.azure_ai.image_edit.flux2_transformation import ( AzureFoundryFlux2ImageEditConfig, ) from litellm.llms.azure_ai.image_edit.transformation import ( AzureFoundryFluxImageEditConfig, ) +from litellm.llms.custom_httpx.http_handler import HTTPHandler def test_azure_ai_validate_environment(): @@ -60,3 +68,111 @@ def test_flux2_validate_environment_with_entra_token(monkeypatch): assert headers["Authorization"] == "Bearer entra-token" assert headers["Content-Type"] == "application/json" + + +def test_flux2_image_edit_maps_openai_and_provider_parameters(): + config = AzureFoundryFlux2ImageEditConfig() + requested_params = ImageEditRequestUtils.get_requested_image_edit_optional_param( + { + "n": 2, + "size": "1536x1024", + "guidance": 4.5, + "steps": 32, + "unrelated": "discarded", + }, + provider_supported_params=config.get_supported_openai_params("FLUX.2-flex"), + ) + mapped_params = config.map_openai_params( + image_edit_optional_params=requested_params, + model="FLUX.2-flex", + drop_params=False, + ) + + assert mapped_params == { + "num_images": 2, + "width": 1536, + "height": 1024, + "guidance": 4.5, + "steps": 32, + } + + +@pytest.mark.parametrize( + ("model", "max_reference_images"), + [ + ("FLUX.2-flex", 10), + ("FLUX.2-pro", 8), + ], +) +def test_flux2_image_edit_uses_all_reference_fields(model: str, max_reference_images: int): + images = [f"image-{index}".encode() for index in range(1, max_reference_images + 1)] + request, files = AzureFoundryFlux2ImageEditConfig().transform_image_edit_request( + model=model, + prompt="Blend every reference", + image=images, + image_edit_optional_request_params={"guidance": 4.5, "steps": 20}, + litellm_params={}, + headers={}, + ) + + assert files == [] + assert request["input_image"] == base64.b64encode(images[0]).decode() + assert request[f"input_image_{max_reference_images}"] == base64.b64encode(images[-1]).decode() + assert "input_image_1" not in request + assert "image" not in request + assert len([key for key in request if key.startswith("input_image")]) == max_reference_images + assert request["guidance"] == 4.5 + assert request["steps"] == 20 + + +@pytest.mark.parametrize( + ("model", "reference_images"), + [ + ("FLUX.2-flex", 11), + ("FLUX.2-pro", 9), + ], +) +def test_flux2_image_edit_rejects_too_many_references(model: str, reference_images: int): + with pytest.raises(ValueError, match=f"at most {reference_images - 1} reference images"): + AzureFoundryFlux2ImageEditConfig().transform_image_edit_request( + model=model, + prompt="Blend every reference", + image=[b"image"] * reference_images, + image_edit_optional_request_params={}, + litellm_params={}, + headers={}, + ) + + +@pytest.mark.parametrize("dimensions", ({"size": "2048x1024"}, {"width": 2048, "height": 1024})) +@pytest.mark.usefixtures("local_model_cost_map") +def test_flux2_image_edit_preserves_controls_and_pixel_cost(dimensions: Mapping[str, int | str]): + def respond(request: httpx.Request) -> httpx.Response: + body: Final = json.loads(request.content) + assert body == { + "model": "FLUX.2-flex", + "prompt": "Add a hat", + "input_image": base64.b64encode(b"image").decode(), + "num_images": 2, + "width": 2048, + "height": 1024, + "guidance": 4.5, + "steps": 32, + } + return httpx.Response(200, json={"data": [{"b64_json": "aW1n"}, {"b64_json": "aW1n"}]}) + + client: Final = HTTPHandler(client=httpx.Client(transport=httpx.MockTransport(respond))) + response: Final = litellm.image_edit( + model="azure_ai/FLUX.2-flex", + image=b"image", + prompt="Add a hat", + api_key="test-key", + api_base="https://example.services.ai.azure.com", + client=client, + n=2, + guidance=4.5, + steps=32, + **dimensions, + ) + + assert response._hidden_params["response_cost"] == pytest.approx(5e-08 * 2048 * 1024 * 2) diff --git a/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py new file mode 100644 index 00000000000..07026cf4309 --- /dev/null +++ b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py @@ -0,0 +1,172 @@ +from collections.abc import Mapping +from typing import Final +from unittest.mock import MagicMock + +import httpx +import pytest + +import litellm +from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap +from litellm.litellm_core_utils.llm_cost_calc.utils import CostCalculatorUtils +from litellm.llms.azure.azure import AzureChatCompletion +from litellm.llms.azure.image_generation import get_azure_image_generation_config +from litellm.llms.azure.image_generation.http_utils import azure_deployment_image_generation_json_body +from litellm.llms.azure_ai.image_generation.flux_transformation import ( + AzureFoundryFluxImageGenerationConfig, +) +from litellm.types.utils import ImageObject, ImageResponse +from litellm.utils import _invalidate_model_cost_lowercase_map + + +@pytest.fixture(autouse=True) +def use_local_model_cost_map(monkeypatch): + monkeypatch.setattr(litellm, "model_cost", GetModelCostMap.load_local_model_cost_map()) + litellm.get_model_info.cache_clear() + _invalidate_model_cost_lowercase_map() + yield + litellm.get_model_info.cache_clear() + _invalidate_model_cost_lowercase_map() + + +@pytest.mark.parametrize( + ("model", "provider_path"), + [ + ("FLUX.2-flex", "flux-2-flex"), + ("FLUX.2-pro", "flux-2-pro"), + ], +) +def test_flux2_uses_model_specific_provider_url(model: str, provider_path: str): + url = AzureChatCompletion().create_azure_base_url( + azure_client_params={ + "azure_endpoint": "https://example.services.ai.azure.com/", + "api_version": "preview", + }, + model=model, + ) + + assert ( + url == f"https://example.services.ai.azure.com/providers/blackforestlabs/v1/{provider_path}?api-version=preview" + ) + + +def test_flux2_flex_maps_openai_and_provider_parameters(): + config = AzureFoundryFluxImageGenerationConfig() + mapped_params = config.map_openai_params( + non_default_params={ + "n": 2, + "size": "1536x1024", + "guidance": 4.5, + "steps": 32, + "output_format": "jpeg", + }, + optional_params={}, + model="FLUX.2-flex", + drop_params=False, + ) + url = config.get_flux2_image_generation_url( + api_base="https://example.services.ai.azure.com", + model="FLUX.2-flex", + api_version="preview", + ) + request = azure_deployment_image_generation_json_body( + api_base=url, + data={"model": "FLUX.2-flex", "prompt": "A red fox", **mapped_params}, + deployment_name="FLUX.2-flex", + ) + + assert request == { + "model": "FLUX.2-flex", + "prompt": "A red fox", + "num_images": 2, + "width": 1536, + "height": 1024, + "guidance": 4.5, + "steps": 32, + "output_format": "jpeg", + } + + +def test_flux2_flex_rejects_invalid_size(): + with pytest.raises(ValueError, match="Expected 'WxH'"): + AzureFoundryFluxImageGenerationConfig().map_openai_params( + non_default_params={"size": "large"}, + optional_params={}, + model="FLUX.2-flex", + drop_params=False, + ) + + +def test_flux2_flex_model_info(): + model_info = litellm.get_model_info( + model="FLUX.2-flex", + custom_llm_provider="azure_ai", + ) + catalog_info = litellm.model_cost["azure_ai/FLUX.2-flex"] + + assert model_info["mode"] == "image_generation" + assert model_info["max_input_tokens"] == 32000 + assert model_info["max_tokens"] == 32000 + assert model_info["supported_endpoints"] == ["/v1/images/generations", "/v1/images/edits"] + assert catalog_info["input_cost_per_pixel"] == 5e-08 + assert catalog_info["supported_modalities"] == ["text", "image"] + assert catalog_info["supported_output_modalities"] == ["image"] + + +def test_flux2_flex_cost_uses_generated_megapixels(): + response = ImageResponse( + data=[ + ImageObject(url="https://example.com/one.png"), + ImageObject(url="https://example.com/two.png"), + ] + ) + + cost = CostCalculatorUtils.route_image_generation_cost_calculator( + model="FLUX.2-flex", + completion_response=response, + custom_llm_provider="azure_ai", + size="2048x1024", + call_type="image_generation", + ) + + assert cost == pytest.approx(5e-08 * 2048 * 1024 * 2) + + +@pytest.mark.parametrize("model", ("FLUX-1.1-pro", "FLUX.1-Kontext-pro")) +def test_flux1_preserves_existing_openai_parameters(model: str): + params: Final = {"n": 2, "size": "1536x1024", "quality": "high", "user": "test-user"} + + mapped: Final = AzureFoundryFluxImageGenerationConfig().map_openai_params( + non_default_params=params, + optional_params={}, + model=model, + drop_params=False, + ) + + assert mapped == params + + +@pytest.mark.parametrize("dimensions", ({"size": "2048x1024"}, {"width": 2048, "height": 1024})) +def test_flux2_cost_uses_mapped_dimensions_after_response_transformation(dimensions: Mapping[str, int | str]): + params: Final = AzureFoundryFluxImageGenerationConfig().map_openai_params( + non_default_params={"n": 2, **dimensions}, + optional_params={}, + model="FLUX.2-flex", + drop_params=False, + ) + response: Final = get_azure_image_generation_config("FLUX.2-flex").transform_image_generation_response( + model="FLUX.2-flex", + raw_response=httpx.Response(200, json={"data": [{"b64_json": "aW1n"}, {"b64_json": "aW1n"}]}), + model_response=ImageResponse(), + logging_obj=MagicMock(), + request_data={"prompt": "A red fox", **params}, + optional_params=params, + litellm_params={}, + encoding=None, + ) + + assert litellm.completion_cost( + model="azure_ai/FLUX.2-flex", + completion_response=response, + optional_params=params, + call_type="image_generation", + ) == pytest.approx(5e-08 * 2048 * 1024 * 2) diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index b26f5e25b6f..6b7337ea8d1 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -35343,7 +35343,7 @@ export interface components { default_model?: string | null; /** * Deployment Affinity - * @description When True and a session_id is resolvable on the request, pin the deployment chosen inside each routed model group and reuse it whenever the session returns to that group, without pinning which group the session routes to. Independent of session_affinity, which pins the model group instead (and always carries this deployment pin with it): with session_affinity off, every turn is still classified on its own merits while a session that escalates to a stronger tier and comes back still lands on the deployment it used before, which is what keeps a provider prompt cache warm. Pins are held per model group, so switching tiers does not disturb the pin left behind in the previous group. On by default because re-shuffling a conversation across deployments of the same model discards that cache for no benefit; set False to keep every turn load-balanced across the group, which is what a deployment set with tight per-deployment rate limits wants. Inert when no session_id is resolvable, since there is nothing to key a pin on, and suppressed when plugins are configured, for the same reason session_affinity is. + * @description When True and a client session_id is resolvable, reuse the session's chosen model for each classified tier and its deployment within each model group. With session_affinity off, every turn is still classified: moving to another tier leaves the previous tier's model pin intact for a later return. Pins yield to current candidate, context, modality, and availability constraints. Adaptive selection chooses the initial model from its eligible pool, then reuses that choice per tier. This reduces avoidable provider prompt-cache misses; it does not guarantee cache hits. Set False to select models and load-balance deployments on every turn, unless session_affinity or user_turn classification requires a pin. Inert without a client session_id and suppressed when plugins are configured. * @default true */ deployment_affinity: boolean; @@ -35487,7 +35487,7 @@ export interface components { session_affinity: boolean; /** * Session Affinity Ttl Seconds - * @description TTL for the session affinity pin; refreshed on every cache hit. Bounds both the session_affinity model pin and the deployment_affinity deployment pin, so it measures idle time for the session's routing decisions rather than total session length + * @description TTL for the session affinity pin; refreshed on every cache hit. Bounds both the session_affinity model pin and the deployment_affinity per-tier model and deployment pins, so it measures idle time for the session's routing decisions rather than total session length * @default 3600 */ session_affinity_ttl_seconds: number; From ff1e2a02ba9b7bafee87896edcbf935749bf59c3 Mon Sep 17 00:00:00 2001 From: Emerson Gomes Date: Tue, 15 Sep 2026 12:33:58 -0500 Subject: [PATCH 030/141] fix(proxy): preserve CI-compatible OpenAPI snapshot formatting --- litellm/proxy/_lazy_openapi_snapshot.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index 81889e4da0c..fb2f014e3d8 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -18974,7 +18974,7 @@ } } }, - "description": "\nUnified rate-limit error.\n\nEvery rate-limit condition surfaced by litellm \u2014 whether it originated from\nan upstream LLM provider, a vendor batch endpoint, or one of litellm's own\nproxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\nmax-iterations, etc.) \u2014 is raised as an instance of this class.\n\nThe :attr:`category` attribute lets callers distinguish the source. See\n:class:`RateLimitErrorCategory` for the available values.\n" + "description": "\n Unified rate-limit error.\n\n Every rate-limit condition surfaced by litellm \u2014 whether it originated from\n an upstream LLM provider, a vendor batch endpoint, or one of litellm's own\n proxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\n max-iterations, etc.) \u2014 is raised as an instance of this class.\n\n The :attr:`category` attribute lets callers distinguish the source. See\n :class:`RateLimitErrorCategory` for the available values.\n " }, "500": { "content": { From 3821e5ace617f43122e757ec0d922c6f4a8c9b6a Mon Sep 17 00:00:00 2001 From: Emerson Gomes Date: Tue, 15 Sep 2026 12:43:24 -0500 Subject: [PATCH 031/141] fix(azure_ai): specify FLUX parameter mapping return type --- litellm/llms/azure_ai/image_generation/flux_transformation.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/litellm/llms/azure_ai/image_generation/flux_transformation.py b/litellm/llms/azure_ai/image_generation/flux_transformation.py index b10e8a6f35e..b9d5e11ff2a 100644 --- a/litellm/llms/azure_ai/image_generation/flux_transformation.py +++ b/litellm/llms/azure_ai/image_generation/flux_transformation.py @@ -102,7 +102,7 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): optional_params: Mapping[str, object], model: str, drop_params: bool, - ) -> dict: # mutable-ok: inherited config contract returns a dict + ) -> dict[str, object]: # mutable-ok: inherited config contract returns a dict if not self.is_flux2_model(model): return super().map_openai_params( non_default_params=dict(non_default_params), From 4595b4f62f209d3d71734e8f1f2692f9a726e9a1 Mon Sep 17 00:00:00 2001 From: Emerson Gomes Date: Tue, 15 Sep 2026 13:01:12 -0500 Subject: [PATCH 032/141] fix(azure-ai): coerce FLUX controls and preserve response dimensions --- .../image_generation/flux_transformation.py | 5 +++++ .../image_generation/gpt_transformation.py | 9 ++++++++- .../test_azure_ai_image_edit_transformation.py | 6 +++--- .../test_azure_ai_flux2_image_generation.py | 18 ++++++++++++++++++ 4 files changed, 34 insertions(+), 4 deletions(-) diff --git a/litellm/llms/azure_ai/image_generation/flux_transformation.py b/litellm/llms/azure_ai/image_generation/flux_transformation.py index b9d5e11ff2a..997205b1fc3 100644 --- a/litellm/llms/azure_ai/image_generation/flux_transformation.py +++ b/litellm/llms/azure_ai/image_generation/flux_transformation.py @@ -85,6 +85,11 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): @staticmethod def _map_parameter(name: str, value: object) -> tuple[tuple[str, object], ...]: + if isinstance(value, str): + if name in ("n", "num_images", "width", "height", "steps", "seed", "safety_tolerance"): + return (("num_images" if name == "n" else name, int(value)),) + if name == "guidance": + return ((name, float(value)),) if name == "n": return (("num_images", value),) if name != "size": diff --git a/litellm/llms/openai/image_generation/gpt_transformation.py b/litellm/llms/openai/image_generation/gpt_transformation.py index 090b2eba387..8dc4d8953ea 100644 --- a/litellm/llms/openai/image_generation/gpt_transformation.py +++ b/litellm/llms/openai/image_generation/gpt_transformation.py @@ -82,7 +82,14 @@ class GPTImageGenerationConfig(BaseImageGenerationConfig): ) # set optional params - image_response.size = image_response.size or optional_params.get("size", "1024x1024") + width: Final = optional_params.get("width") + height: Final = optional_params.get("height") + requested_size: Final = ( + f"{width}x{height}" + if isinstance(width, int) and isinstance(height, int) + else optional_params.get("size", "1024x1024") + ) + image_response.size = image_response.size or requested_size image_response.quality = image_response.quality or optional_params.get("quality", "high") image_response.output_format = image_response.output_format or optional_params.get("output_format", "png") diff --git a/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py b/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py index 94b23c5f1c2..d74afa88a6b 100644 --- a/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py +++ b/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py @@ -144,7 +144,7 @@ def test_flux2_image_edit_rejects_too_many_references(model: str, reference_imag ) -@pytest.mark.parametrize("dimensions", ({"size": "2048x1024"}, {"width": 2048, "height": 1024})) +@pytest.mark.parametrize("dimensions", ({"size": "2048x1024"}, {"width": 2048, "height": 1024}, {"width": "2048", "height": "1024"})) @pytest.mark.usefixtures("local_model_cost_map") def test_flux2_image_edit_preserves_controls_and_pixel_cost(dimensions: Mapping[str, int | str]): def respond(request: httpx.Request) -> httpx.Response: @@ -170,8 +170,8 @@ def test_flux2_image_edit_preserves_controls_and_pixel_cost(dimensions: Mapping[ api_base="https://example.services.ai.azure.com", client=client, n=2, - guidance=4.5, - steps=32, + guidance="4.5", + steps="32", **dimensions, ) diff --git a/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py index 07026cf4309..c1d46b7e919 100644 --- a/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py +++ b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py @@ -170,3 +170,21 @@ def test_flux2_cost_uses_mapped_dimensions_after_response_transformation(dimensi optional_params=params, call_type="image_generation", ) == pytest.approx(5e-08 * 2048 * 1024 * 2) + + +def test_flux2_response_preserves_mapped_dimensions(): + config = AzureFoundryFluxImageGenerationConfig() + params = config.map_openai_params( + non_default_params={"size": "2048x1024"}, optional_params={}, model="FLUX.2-flex", drop_params=False + ) + response = config.transform_image_generation_response( + model="FLUX.2-flex", + raw_response=httpx.Response(200, json={"data": [{"b64_json": "aW1n"}]}), + model_response=ImageResponse(), + logging_obj=MagicMock(), + request_data={"prompt": "A landscape"}, + optional_params=params, + litellm_params={}, + encoding=None, + ) + assert response.size == "2048x1024" From 92e55b3b2262c40e41178435453edf3802819229 Mon Sep 17 00:00:00 2001 From: yassin Date: Tue, 15 Sep 2026 21:17:00 +0000 Subject: [PATCH 033/141] perf(proxy): split aggregated usage query into key-free rollups and bounded top-N keys Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 4 + .../common_daily_activity.py | 241 +++++++----- .../common_daily_activity.py | 5 + .../test_common_daily_activity.py | 355 +++++++++++++++--- ui/litellm-dashboard/src/lib/http/schema.d.ts | 5 + 5 files changed, 479 insertions(+), 131 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index ba5ec73d435..36bc16a0ae3 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -2027,6 +2027,10 @@ MCP_SPEND_LOG_MODEL_PREFIX: Final[str] = "MCP: " PTU_SENTINEL_API_KEY: Final[str] = "__ptu_flat_cost__" PTU_ROLLUP_JOB_ID: Final[str] = "ptu_flat_cost_rollup_job" PTU_ROLLUP_LOCK_TTL_SECONDS: Final[int] = 900 +# Per-api_key rollups on the aggregated usage endpoint cover only the top N keys +# by spend so the result set stops growing with key count. Totals and the +# model/provider/endpoint rollups still cover every key. +USAGE_TOP_API_KEYS_LIMIT: Final[int] = 100 # Furthest back the catch-up pass looks for unpriced PTU days when a deployment # declares no ptu_effective_from, bounding the scan for an open-ended window. PTU_ROLLUP_MAX_BACKFILL_DAYS: Final[int] = 90 diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index 44ed0017e42..4aafd416263 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -9,7 +9,7 @@ from fastapi import HTTPException, status from typing_extensions import ReadOnly, TypedDict from litellm._logging import verbose_proxy_logger -from litellm.constants import PTU_SENTINEL_API_KEY +from litellm.constants import PTU_SENTINEL_API_KEY, USAGE_TOP_API_KEYS_LIMIT from litellm.proxy._types import CommonProxyErrors from litellm.proxy.spend_tracking.key_metadata_recovery import ( attach_user_emails, @@ -169,6 +169,32 @@ class _EntityRollupRow(_GroupingSetsRow): api_key_rolled: int +class _AggregatedQueryKwargs(TypedDict): + """Filter arguments shared by the three aggregated SQL builders.""" + + table_name: ReadOnly[str] + entity_id_field: ReadOnly[str] + entity_id: ReadOnly[str | list[str] | None] + start_date: ReadOnly[str] + end_date: ReadOnly[str] + model: ReadOnly[str | None] + api_key: ReadOnly[str | list[str] | None] + exclude_entity_ids: ReadOnly[list[str] | None] + timezone_offset_minutes: ReadOnly[int | None] + include_current_utc_day: ReadOnly[bool] + + +_SqlQuery = tuple[str, list[str]] + + +async def _query_raw_optional( + prisma_client: PrismaClient, query: _SqlQuery | None +) -> list[dict[str, object]] | None: # mutable-ok: prisma query_raw return shape + if query is None: + return None + return await prisma_client.db.query_raw(query[0], *query[1]) + + def _reported_flat_cost(record: DailySpendRecord | _GroupingSetsRow) -> float: """Flat cost a daily row reports, which is zero unless PTU cost attribution is enabled. @@ -689,6 +715,27 @@ def _ptu_flat_cost_select(table_name: str) -> str: return "0::float AS ptu_flat_cost" +def _rollup_metric_select(table_name: str) -> str: + return f""" + SUM(spend)::float AS spend, + {_ptu_flat_cost_select(table_name)}, + SUM(prompt_tokens)::bigint AS prompt_tokens, + SUM(completion_tokens)::bigint AS completion_tokens, + SUM(cache_read_input_tokens)::bigint AS cache_read_input_tokens, + SUM(cache_creation_input_tokens)::bigint AS cache_creation_input_tokens, + SUM(compression_saved_tokens)::bigint AS compression_saved_tokens, + SUM(compression_savings_spend)::float AS compression_savings_spend, + SUM(prompt_caching_savings_spend)::float AS prompt_caching_savings_spend, + SUM(gateway_injected_caching_savings_spend)::float AS gateway_injected_caching_savings_spend, + SUM(autorouter_savings_spend)::float AS autorouter_savings_spend, + SUM(api_requests)::bigint AS api_requests, + SUM(successful_requests)::bigint AS successful_requests, + SUM(failed_requests)::bigint AS failed_requests""" + + +_MODEL_GROUP_EXPR: Final = "COALESCE(NULLIF(model_group, ''), model)" + + def _build_aggregated_sql_query( *, table_name: str, @@ -702,12 +749,16 @@ def _build_aggregated_sql_query( timezone_offset_minutes: int | None = None, include_current_utc_day: bool = False, ) -> tuple[str, list[str]]: # mutable-ok: SQL text plus its ordered $N params - """Build a parameterized SQL GROUP BY query for aggregated daily activity. + """Build the key-free GROUPING SETS query for aggregated daily activity. - Groups by (date, api_key, model, model_group, custom_llm_provider, - mcp_namespaced_tool_name, endpoint) with SUMs on all metric columns. - The entity_id column is intentionally omitted from GROUP BY to collapse - rows across entities — this is where the biggest row reduction comes from. + Emits the grand total, per-date totals and the per-(date, model), model_group, + provider, mcp tool and endpoint rollups. api_key is never a grouping column here, + so the row count is bounded by dates x distinct models/providers/endpoints and + does not grow with the number of keys. Per-key rollups come from + _build_top_api_keys_sql_query. Both queries emit the same 7-bit group_level + bitmask (date, api_key, model, model_group, provider, mcp, endpoint); this one + hard-codes the api_key bit to "rolled up" so the dispatcher can consume the two + result sets as one stream. Returns: Tuple of (sql_query, params_list) ready for prisma_client.db.query_raw(). @@ -730,14 +781,6 @@ def _build_aggregated_sql_query( exclude_entity_ids=exclude_entity_ids, ) - # Postgres computes every rollup level the response needs — per-date - # totals, per-(date, model), per-(date, model, api_key), per-provider, - # etc. — in a single pass via GROUPING SETS. The GROUPING() bitmask - # encodes which level a row belongs to so Python can dispatch rows - # straight into their buckets without re-summing. The leaf grouping - # is omitted on purpose: nothing in the response shape needs it once - # all the rollups are present. - # # TODO: drop the successful_requests/failed_requests aggregates (and the # total_successful_requests metadata they feed) once the admin UI reads SGR # only from LiteLLM_DailyGatewayRequests. The remaining spend, token and @@ -745,44 +788,25 @@ def _build_aggregated_sql_query( sql_query: Final = f""" SELECT date, - api_key, + NULL::text AS api_key, model, - COALESCE(NULLIF(model_group, ''), model) AS model_group, + {_MODEL_GROUP_EXPR} AS model_group, custom_llm_provider, mcp_namespaced_tool_name, endpoint, - GROUPING(date, api_key, model, COALESCE(NULLIF(model_group, ''), model), - custom_llm_provider, mcp_namespaced_tool_name, - endpoint) AS group_level, - SUM(spend)::float AS spend, - {_ptu_flat_cost_select(table_name)}, - SUM(prompt_tokens)::bigint AS prompt_tokens, - SUM(completion_tokens)::bigint AS completion_tokens, - SUM(cache_read_input_tokens)::bigint AS cache_read_input_tokens, - SUM(cache_creation_input_tokens)::bigint AS cache_creation_input_tokens, - SUM(compression_saved_tokens)::bigint AS compression_saved_tokens, - SUM(compression_savings_spend)::float AS compression_savings_spend, - SUM(prompt_caching_savings_spend)::float AS prompt_caching_savings_spend, - SUM(gateway_injected_caching_savings_spend)::float AS gateway_injected_caching_savings_spend, - SUM(autorouter_savings_spend)::float AS autorouter_savings_spend, - SUM(api_requests)::bigint AS api_requests, - SUM(successful_requests)::bigint AS successful_requests, - SUM(failed_requests)::bigint AS failed_requests + (GROUPING(date) << 6) | {_API_KEY_ROLLED_UP_BIT} + | GROUPING(model, {_MODEL_GROUP_EXPR}, + custom_llm_provider, mcp_namespaced_tool_name, + endpoint) AS group_level,{_rollup_metric_select(table_name)} FROM "{pg_table}" WHERE {where_clause} GROUP BY GROUPING SETS ( (date), - (date, api_key), (date, model), - (date, model, api_key), - (date, COALESCE(NULLIF(model_group, ''), model)), - (date, COALESCE(NULLIF(model_group, ''), model), api_key), + (date, {_MODEL_GROUP_EXPR}), (date, custom_llm_provider), - (date, custom_llm_provider, api_key), (date, mcp_namespaced_tool_name), - (date, mcp_namespaced_tool_name, api_key), (date, endpoint), - (date, endpoint, api_key), () ) """ @@ -790,6 +814,80 @@ def _build_aggregated_sql_query( return sql_query, sql_params +def _build_top_api_keys_sql_query( + *, + table_name: str, + entity_id_field: str, + entity_id: str | list[str] | None, # mutable-ok: filter union shared with the paginated path + start_date: str, + end_date: str, + model: str | None, + api_key: str | list[str] | None, # mutable-ok: filter union shared with the paginated path + exclude_entity_ids: list[str] | None = None, # mutable-ok: filter union shared with the paginated path + timezone_offset_minutes: int | None = None, + include_current_utc_day: bool = False, +) -> tuple[str, list[str]]: # mutable-ok: SQL text plus its ordered $N params + """Per-key companion to _build_aggregated_sql_query. + + Ranks keys by spend over the same WHERE clause, keeps the top + USAGE_TOP_API_KEYS_LIMIT (ties broken by api_key so the set is stable across + refreshes) and emits the six (date, , api_key) rollups for those keys + only. The PTU flat-cost sentinel never ranks, so it cannot occupy a visible slot. + """ + pg_table: Final = _PRISMA_TO_PG_TABLE.get(table_name) + if pg_table is None: + raise ValueError(f"Unknown table name: {table_name}") + + adjusted_start, adjusted_end = _adjust_dates_for_timezone( + start_date, end_date, timezone_offset_minutes, include_current_utc_day + ) + + where_clause, where_params = _build_aggregated_where_clause( + entity_id_field=entity_id_field, + entity_id=entity_id, + adjusted_start=adjusted_start, + adjusted_end=adjusted_end, + model=model, + api_key=api_key, + exclude_entity_ids=exclude_entity_ids, + ) + sentinel_param: Final = f"${len(where_params) + 1}" + + sql_query: Final = f""" + WITH top_api_keys AS ( + SELECT api_key + FROM "{pg_table}" + WHERE {where_clause} AND api_key <> {sentinel_param} + GROUP BY api_key + ORDER BY SUM(spend) DESC, api_key + LIMIT {USAGE_TOP_API_KEYS_LIMIT} + ) + SELECT + date, + api_key, + model, + {_MODEL_GROUP_EXPR} AS model_group, + custom_llm_provider, + mcp_namespaced_tool_name, + endpoint, + GROUPING(date, api_key, model, {_MODEL_GROUP_EXPR}, + custom_llm_provider, mcp_namespaced_tool_name, + endpoint) AS group_level,{_rollup_metric_select(table_name)} + FROM "{pg_table}" + WHERE {where_clause} AND api_key IN (SELECT api_key FROM top_api_keys) + GROUP BY GROUPING SETS ( + (date, api_key), + (date, model, api_key), + (date, {_MODEL_GROUP_EXPR}, api_key), + (date, custom_llm_provider, api_key), + (date, mcp_namespaced_tool_name, api_key), + (date, endpoint, api_key) + ) + """ + + return sql_query, [*where_params, PTU_SENTINEL_API_KEY] + + def _build_entity_rollup_sql_query( *, table_name: str, @@ -832,21 +930,7 @@ def _build_entity_rollup_sql_query( "{entity_id_field}" AS entity_id, date, api_key, - GROUPING(api_key) AS api_key_rolled, - SUM(spend)::float AS spend, - {_ptu_flat_cost_select(table_name)}, - SUM(prompt_tokens)::bigint AS prompt_tokens, - SUM(completion_tokens)::bigint AS completion_tokens, - SUM(cache_read_input_tokens)::bigint AS cache_read_input_tokens, - SUM(cache_creation_input_tokens)::bigint AS cache_creation_input_tokens, - SUM(compression_saved_tokens)::bigint AS compression_saved_tokens, - SUM(compression_savings_spend)::float AS compression_savings_spend, - SUM(prompt_caching_savings_spend)::float AS prompt_caching_savings_spend, - SUM(gateway_injected_caching_savings_spend)::float AS gateway_injected_caching_savings_spend, - SUM(autorouter_savings_spend)::float AS autorouter_savings_spend, - SUM(api_requests)::bigint AS api_requests, - SUM(successful_requests)::bigint AS successful_requests, - SUM(failed_requests)::bigint AS failed_requests + GROUPING(api_key) AS api_key_rolled,{_rollup_metric_select(table_name)} FROM "{pg_table}" WHERE {where_clause} GROUP BY GROUPING SETS ( @@ -948,6 +1032,7 @@ async def _aggregate_spend_records( # current grouping set's key), 0 when the column is part of the key. _GROUP_GRAND_TOTAL: Final = 127 # 0b1111111 — all rolled up _GROUP_DATE: Final = 63 # 0b0111111 — only date kept +_API_KEY_ROLLED_UP_BIT: Final = 32 # 0b0100000 — api_key position in the 7-bit mask _GROUP_DATE_API_KEY: Final = 31 # 0b0011111 _GROUP_DATE_MODEL: Final = 47 # 0b0101111 _GROUP_DATE_MODEL_API_KEY: Final = 15 # 0b0001111 @@ -1311,9 +1396,11 @@ async def get_daily_activity_aggregated( ) -> SpendAnalyticsPaginatedResponse: """Aggregated variant that returns the full result set (no pagination). - Uses SQL GROUP BY to aggregate rows in the database rather than fetching - all individual rows into Python. This collapses rows across entities - (users/teams/orgs), reducing ~150k rows to ~2-3k grouped rows. + Runs two GROUPING SETS queries in parallel: a key-free one for totals and the + model/provider/mcp/endpoint rollups (row count independent of key cardinality) + and a bounded one for the per-key rollups of the top USAGE_TOP_API_KEYS_LIMIT + keys by spend. breakdown.api_keys and every api_key_breakdown therefore list at + most that many keys, while the totals and the key-free rollups cover every key. include_entity_breakdown runs a small companion rollup query and folds `breakdown.entities` onto the response, as entity-scoped views like Team Usage need. @@ -1333,7 +1420,7 @@ async def get_daily_activity_aggregated( ) try: - sql_query, sql_params = _build_aggregated_sql_query( + query_kwargs: Final = _AggregatedQueryKwargs( table_name=table_name, entity_id_field=entity_id_field, entity_id=entity_id, @@ -1345,36 +1432,17 @@ async def get_daily_activity_aggregated( timezone_offset_minutes=timezone_offset_minutes, include_current_utc_day=include_current_utc_day, ) + key_free_sql, key_free_params = _build_aggregated_sql_query(**query_kwargs) + top_keys_sql, top_keys_params = _build_top_api_keys_sql_query(**query_kwargs) + entity_query: Final = _build_entity_rollup_sql_query(**query_kwargs) if include_entity_breakdown else None - entity_query: Final = ( - _build_entity_rollup_sql_query( - table_name=table_name, - entity_id_field=entity_id_field, - entity_id=entity_id, - start_date=start_date, - end_date=end_date, - model=model, - api_key=api_key, - exclude_entity_ids=exclude_entity_ids, - timezone_offset_minutes=timezone_offset_minutes, - include_current_utc_day=include_current_utc_day, - ) - if include_entity_breakdown - else None + raw_key_free_rows, raw_top_key_rows, raw_entity_rows = await asyncio.gather( + prisma_client.db.query_raw(key_free_sql, *key_free_params), + prisma_client.db.query_raw(top_keys_sql, *top_keys_params), + _query_raw_optional(prisma_client, entity_query), ) - # Execute the GROUPING SETS query (one row per rollup level), alongside - # the per-entity companion rollup when the caller wants entities. - raw_rows, raw_entity_rows = ( - await asyncio.gather( - prisma_client.db.query_raw(sql_query, *sql_params), - prisma_client.db.query_raw(entity_query[0], *entity_query[1]), - ) - if entity_query is not None - else (await prisma_client.db.query_raw(sql_query, *sql_params), None) - ) - - records: Final = [_GroupingSetsRow(**row) for row in (raw_rows or [])] + records: Final = [_GroupingSetsRow(**row) for row in (*(raw_key_free_rows or ()), *(raw_top_key_rows or ()))] # The grouping-sets dispatcher places each row directly in its bucket # using the row's GROUPING() bitmask. No Python-side summing needed. @@ -1426,6 +1494,7 @@ async def get_daily_activity_aggregated( page=1, total_pages=1, has_more=False, + api_key_limit=USAGE_TOP_API_KEYS_LIMIT, ), ) diff --git a/litellm/types/proxy/management_endpoints/common_daily_activity.py b/litellm/types/proxy/management_endpoints/common_daily_activity.py index 090e5c42376..f58d40dfa9c 100644 --- a/litellm/types/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/types/proxy/management_endpoints/common_daily_activity.py @@ -96,6 +96,11 @@ class DailySpendMetadata(BaseModel): page: int = Field(default=1) total_pages: int = Field(default=1) has_more: bool = Field(default=False) + api_key_limit: int | None = Field( + default=None, + description="When set, api_keys and every api_key_breakdown list at most this many keys, " + "ranked by spend. Totals and the model, provider, mcp and endpoint rollups still cover every key.", + ) class SpendAnalyticsPaginatedResponse(BaseModel): diff --git a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py index 71896a18f48..dec98a1da27 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py +++ b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py @@ -1,17 +1,24 @@ +import re +from collections.abc import Sequence from datetime import datetime, timedelta, timezone from types import SimpleNamespace from typing import Final from unittest.mock import AsyncMock, MagicMock +import psycopg import pytest +from psycopg.rows import dict_row +from pytest_postgresql import factories from litellm.proxy.spend_tracking.ptu_feature_flag import PTU_COST_ATTRIBUTION_ENV_VAR +from litellm.constants import PTU_SENTINEL_API_KEY, USAGE_TOP_API_KEYS_LIMIT from litellm.proxy.management_endpoints.common_daily_activity import ( _adjust_dates_for_timezone, _build_aggregated_sql_query, _build_entity_rollup_sql_query, + _build_top_api_keys_sql_query, _is_user_agent_tag, _record_to_spend_metrics, get_api_key_metadata, @@ -159,7 +166,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "autorouter_savings_spend": 0.0, "failed_requests": 0, } - mock_rows = [ + key_free_rows = [ # (date, endpoint) — rolls up across api_keys and models { **base, @@ -185,31 +192,6 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "api_requests": 1, "successful_requests": 1, }, - # (date, endpoint, api_key) — populates the per-key sub-bucket - { - **base, - "date": "2024-01-01", - "endpoint": "/v1/chat/completions", - "api_key": "key-1", - "group_level": 30, - "spend": 15.0, - "prompt_tokens": 150, - "completion_tokens": 75, - "api_requests": 2, - "successful_requests": 2, - }, - { - **base, - "date": "2024-01-01", - "endpoint": "/v1/embeddings", - "api_key": "key-2", - "group_level": 30, - "spend": 3.0, - "prompt_tokens": 30, - "completion_tokens": 0, - "api_requests": 1, - "successful_requests": 1, - }, # (date) — per-date totals { **base, @@ -237,8 +219,35 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "successful_requests": 3, }, ] + top_key_rows = [ + # (date, endpoint, api_key) — populates the per-key sub-bucket + { + **base, + "date": "2024-01-01", + "endpoint": "/v1/chat/completions", + "api_key": "key-1", + "group_level": 30, + "spend": 15.0, + "prompt_tokens": 150, + "completion_tokens": 75, + "api_requests": 2, + "successful_requests": 2, + }, + { + **base, + "date": "2024-01-01", + "endpoint": "/v1/embeddings", + "api_key": "key-2", + "group_level": 30, + "spend": 3.0, + "prompt_tokens": 30, + "completion_tokens": 0, + "api_requests": 1, + "successful_requests": 1, + }, + ] - mock_prisma.db.query_raw = AsyncMock(return_value=mock_rows) + mock_prisma.db.query_raw = AsyncMock(side_effect=[key_free_rows, top_key_rows]) mock_prisma.db.litellm_verificationtoken = MagicMock() mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) @@ -284,8 +293,11 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): assert "key-2" in embeddings_endpoint.api_key_breakdown assert embeddings_endpoint.api_key_breakdown["key-2"].metrics.spend == 3.0 - # Verify query_raw was called (not find_many) - mock_prisma.db.query_raw.assert_called_once() + # One key-free rollup query plus one bounded per-key query, no find_many + assert mock_prisma.db.query_raw.call_count == 2 + key_free_sql, top_keys_sql = (call.args[0] for call in mock_prisma.db.query_raw.call_args_list) + assert "top_api_keys" not in key_free_sql + assert "WITH top_api_keys AS" in top_keys_sql @pytest.mark.asyncio @@ -812,7 +824,7 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): "autorouter_savings_spend": 0.0, "failed_requests": 0, } - mock_rows = [ + key_free_rows = [ { **base, "date": "2024-01-01", @@ -825,6 +837,8 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): "api_requests": 1, "successful_requests": 1, }, + ] + top_key_rows = [ { **base, "date": "2024-01-01", @@ -839,7 +853,7 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): }, ] - mock_prisma.db.query_raw = AsyncMock(return_value=mock_rows) + mock_prisma.db.query_raw = AsyncMock(side_effect=[key_free_rows, top_key_rows]) # Active table returns nothing for this key mock_prisma.db.litellm_verificationtoken = MagicMock() @@ -1240,17 +1254,61 @@ class TestBuildAggregatedSqlQuery: normalized = " ".join(sql.split()) fallback = "COALESCE(NULLIF(model_group, ''), model)" assert f"{fallback} AS model_group" in normalized - assert ( - f"GROUPING(date, api_key, model, {fallback}, " - "custom_llm_provider, mcp_namespaced_tool_name, endpoint) AS group_level" in normalized - ) - assert f"(date, {fallback}), (date, {fallback}, api_key)," in normalized + assert f"GROUPING(model, {fallback}, custom_llm_provider, mcp_namespaced_tool_name, endpoint)" in normalized + assert f"(date, {fallback})," in normalized assert "(date, model_group)" not in normalized assert "COALESCE(model_group, model)" not in normalized + def test_key_free_query_never_groups_by_api_key(self): + """The main rollup query must not emit one row per key, that is what blew up + the query engine at 3k+ keys. Every grouping set stays key-free and api_key + is projected as a NULL literal so the dispatcher's row shape is unchanged.""" + sql, _ = _build_aggregated_sql_query( + table_name="litellm_dailyuserspend", + entity_id_field="user_id", + entity_id=None, + start_date="2026-07-01", + end_date="2026-07-01", + model=None, + api_key=None, + ) + + normalized = " ".join(sql.split()) + grouping_block = normalized.split("GROUP BY GROUPING SETS (", 1)[1] + assert "api_key" not in grouping_block + assert "NULL::text AS api_key" in normalized + + def test_top_api_keys_query_ranks_keys_deterministically_and_shares_filters(self): + sql, params = _build_top_api_keys_sql_query( + table_name="litellm_dailyuserspend", + entity_id_field="user_id", + entity_id="user-1", + start_date="2026-05-29", + end_date="2026-06-02", + model="bedrock/global.anthropic.claude-opus-4-8", + api_key="sk-test", + timezone_offset_minutes=-330, + ) + + normalized = " ".join(sql.split()) + assert f"ORDER BY SUM(spend) DESC, api_key LIMIT {USAGE_TOP_API_KEYS_LIMIT}" in normalized + assert "api_key IN (SELECT api_key FROM top_api_keys)" in normalized + assert "api_key <> $6" in normalized + grouping_block = normalized.split("GROUP BY GROUPING SETS (", 1)[1] + assert grouping_block.count(", api_key)") == 6 + assert grouping_block.count("(date") == 6 + assert params == [ + "2026-05-29", + "2026-06-02", + "user-1", + "bedrock/global.anthropic.claude-opus-4-8", + "sk-test", + PTU_SENTINEL_API_KEY, + ] + class TestAggregatedEmptyEntityFilter: - _BUILDERS: Final = (_build_aggregated_sql_query, _build_entity_rollup_sql_query) + _BUILDERS: Final = (_build_aggregated_sql_query, _build_top_api_keys_sql_query, _build_entity_rollup_sql_query) @pytest.mark.parametrize("build", _BUILDERS) def test_empty_entity_list_emits_no_degenerate_in_clause(self, build): @@ -1267,7 +1325,8 @@ class TestAggregatedEmptyEntityFilter: normalized = " ".join(sql.split()) assert "IN ()" not in normalized assert '"team_id" IN' not in normalized - assert params == ["2026-08-01", "2026-08-19"] + sentinel_params = [PTU_SENTINEL_API_KEY] if build is _build_top_api_keys_sql_query else [] + assert params == ["2026-08-01", "2026-08-19", *sentinel_params] @pytest.mark.parametrize("build", _BUILDERS) def test_empty_entity_list_matches_nothing_rather_than_everything(self, build): @@ -1298,7 +1357,8 @@ class TestAggregatedEmptyEntityFilter: normalized = " ".join(sql.split()) assert '"team_id" IN ($3, $4)' in normalized assert "FALSE" not in normalized - assert params == ["2026-08-01", "2026-08-19", "team-alpha", "team-beta"] + sentinel_params = [PTU_SENTINEL_API_KEY] if build is _build_top_api_keys_sql_query else [] + assert params == ["2026-08-01", "2026-08-19", "team-alpha", "team-beta", *sentinel_params] @pytest.mark.asyncio @@ -1313,7 +1373,7 @@ async def test_get_daily_activity_aggregated_empty_result_set(): mock_prisma = MagicMock() mock_prisma.db = MagicMock() - mock_rows = [ + key_free_rows = [ { "date": None, "api_key": None, @@ -1338,7 +1398,7 @@ async def test_get_daily_activity_aggregated_empty_result_set(): "failed_requests": None, } ] - mock_prisma.db.query_raw = AsyncMock(return_value=mock_rows) + mock_prisma.db.query_raw = AsyncMock(side_effect=[key_free_rows, []]) result = await get_daily_activity_aggregated( prisma_client=mock_prisma, @@ -1365,6 +1425,211 @@ async def test_get_daily_activity_aggregated_empty_result_set(): assert result.metadata.total_compression_saved_tokens == 0 +_aggregated_postgresql_proc: Final = factories.postgresql_proc() +_aggregated_postgresql: Final = factories.postgresql("_aggregated_postgresql_proc") + +_DAILY_USER_SPEND_DDL: Final = """ + CREATE TABLE "LiteLLM_DailyUserSpend" ( + id TEXT PRIMARY KEY, + user_id TEXT, + date TEXT NOT NULL, + api_key TEXT NOT NULL, + model TEXT, + model_group TEXT, + custom_llm_provider TEXT, + mcp_namespaced_tool_name TEXT, + endpoint TEXT, + prompt_tokens BIGINT DEFAULT 0, + completion_tokens BIGINT DEFAULT 0, + cache_read_input_tokens BIGINT DEFAULT 0, + cache_creation_input_tokens BIGINT DEFAULT 0, + compression_saved_tokens BIGINT DEFAULT 0, + compression_savings_spend DOUBLE PRECISION DEFAULT 0, + prompt_caching_savings_spend DOUBLE PRECISION DEFAULT 0, + gateway_injected_caching_savings_spend DOUBLE PRECISION DEFAULT 0, + autorouter_savings_spend DOUBLE PRECISION DEFAULT 0, + spend DOUBLE PRECISION DEFAULT 0, + api_requests BIGINT DEFAULT 0, + successful_requests BIGINT DEFAULT 0, + failed_requests BIGINT DEFAULT 0 + ) +""" + + +def _seed_daily_user_spend(conn: psycopg.Connection, rows: Sequence[tuple[object, ...]]) -> None: + with conn.cursor() as cur: + cur.execute(_DAILY_USER_SPEND_DDL) + cur.executemany( + """ + INSERT INTO "LiteLLM_DailyUserSpend" + (id, user_id, date, api_key, model, model_group, custom_llm_provider, + endpoint, prompt_tokens, spend, api_requests, successful_requests) + VALUES (%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s) + """, + rows, + ) + conn.commit() + + +def _psycopg_query_raw(conn: psycopg.Connection, row_counts: list[int]): + """Run the proxy's $N-parameterized SQL through psycopg, recording each result size.""" + + async def query_raw(sql: str, *params: str) -> list[dict[str, object]]: + converted: Final = re.sub(r"\$(\d+)", r"%(p\1)s", sql) + with conn.cursor(row_factory=dict_row) as cur: + cur.execute( + converted, # pyright: ignore[reportArgumentType] # psycopg stubs want a literal-typed query + {f"p{i}": v for i, v in enumerate(params, start=1)}, + ) + rows: Final = cur.fetchall() + row_counts.append(len(rows)) + return rows + + return query_raw + + +@pytest.mark.asyncio +async def test_get_daily_activity_aggregated_bounds_api_key_rollups( + _aggregated_postgresql: psycopg.Connection, +): + """Run both GROUPING SETS queries against real Postgres with more keys than the cap. + + key-004 and key-005 tie on spend exactly at the USAGE_TOP_API_KEYS_LIMIT + cutoff; the api_key tiebreaker must keep key-004 and drop key-005. The PTU + sentinel outspends every key but must not take a slot. Excluded keys and the + sentinel still count toward the totals and the model rollup, which come from + the key-free query. + """ + n_keys: Final = USAGE_TOP_API_KEYS_LIMIT + 5 + key_rows: Final = [ + ( + f"row-{i:03d}", + f"user-{i:03d}", + "2026-06-01", + f"key-{i:03d}", + "gpt-5", + "", + "openai", + "/v1/chat/completions", + 10, + 6.0 if i == 4 else float(i + 1), + 1, + 1, + ) + for i in range(n_keys) + ] + sentinel_row: Final = ( + "row-ptu", + None, + "2026-06-01", + PTU_SENTINEL_API_KEY, + "gpt-5", + "", + "azure", + None, + 0, + 1000.0, + 0, + 0, + ) + _seed_daily_user_spend(_aggregated_postgresql, [*key_rows, sentinel_row]) + key_spend: Final = sum(6.0 if i == 4 else float(i + 1) for i in range(n_keys)) + + row_counts: Final[list[int]] = [] # mutable-ok: out-param for the query_raw shim + mock_prisma = MagicMock() + mock_prisma.db = MagicMock() + mock_prisma.db.query_raw = _psycopg_query_raw(_aggregated_postgresql, row_counts) + mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) + mock_prisma.db.litellm_deletedverificationtoken.find_many = AsyncMock(return_value=[]) + + result = await get_daily_activity_aggregated( + prisma_client=mock_prisma, + table_name="litellm_dailyuserspend", + entity_id_field="user_id", + entity_id=None, + entity_metadata_field=None, + start_date="2026-06-01", + end_date="2026-06-01", + model=None, + api_key=None, + ) + + # Key-free query: (), (date), (date, model), (date, model_group), two providers, + # one mcp NULL bucket, endpoint plus its NULL bucket = 9 rows regardless of key count. + # Top-key query: six per-key grouping sets, each capped at the limit. + assert row_counts == [9, 6 * USAGE_TOP_API_KEYS_LIMIT] + + assert result.metadata.total_spend == pytest.approx(key_spend + 1000.0) + assert result.metadata.total_api_requests == n_keys + assert result.metadata.api_key_limit == USAGE_TOP_API_KEYS_LIMIT + + expected_top: Final = {f"key-{i:03d}" for i in range(6, n_keys)} | {"key-004"} + day: Final = result.results[0] + assert day.metrics.spend == pytest.approx(key_spend + 1000.0) + assert set(day.breakdown.api_keys) == expected_top + assert day.breakdown.api_keys["key-004"].metrics.spend == 6.0 + assert "key-005" not in day.breakdown.api_keys + assert PTU_SENTINEL_API_KEY not in day.breakdown.api_keys + + assert day.breakdown.models["gpt-5"].metrics.spend == pytest.approx(key_spend + 1000.0) + assert set(day.breakdown.models["gpt-5"].api_key_breakdown) == expected_top + assert day.breakdown.providers["openai"].metrics.spend == pytest.approx(key_spend) + assert set(day.breakdown.providers["openai"].api_key_breakdown) == expected_top + assert day.breakdown.endpoints["/v1/chat/completions"].metrics.api_requests == n_keys + + +@pytest.mark.asyncio +async def test_get_daily_activity_aggregated_explicit_api_key_filter_scopes_both_queries( + _aggregated_postgresql: psycopg.Connection, +): + """An explicit api_key filter must scope the key-free totals and the per-key + rollups to that key alone, so the two result sets never disagree.""" + rows: Final = [ + ( + f"row-{i}", + f"user-{i}", + "2026-06-01", + f"key-{i}", + "gpt-5", + "", + "openai", + "/v1/chat/completions", + 10, + float(i + 1), + 1, + 1, + ) + for i in range(3) + ] + _seed_daily_user_spend(_aggregated_postgresql, rows) + + row_counts: Final[list[int]] = [] # mutable-ok: out-param for the query_raw shim + mock_prisma = MagicMock() + mock_prisma.db = MagicMock() + mock_prisma.db.query_raw = _psycopg_query_raw(_aggregated_postgresql, row_counts) + mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) + mock_prisma.db.litellm_deletedverificationtoken.find_many = AsyncMock(return_value=[]) + + result = await get_daily_activity_aggregated( + prisma_client=mock_prisma, + table_name="litellm_dailyuserspend", + entity_id_field="user_id", + entity_id=None, + entity_metadata_field=None, + start_date="2026-06-01", + end_date="2026-06-01", + model=None, + api_key="key-1", + ) + + assert result.metadata.total_spend == 2.0 + day: Final = result.results[0] + assert set(day.breakdown.api_keys) == {"key-1"} + assert day.breakdown.api_keys["key-1"].metrics.spend == 2.0 + assert day.breakdown.models["gpt-5"].metrics.spend == 2.0 + assert set(day.breakdown.models["gpt-5"].api_key_breakdown) == {"key-1"} + + def _no_spend_record(): """A rollup row for a key with no spend, where SUM() returns NULL (None).""" return SimpleNamespace( @@ -2128,8 +2393,8 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown(): {**base, "date": None, "group_level": 127, "spend": 18.0}, {**base, "date": "2024-01-01", "group_level": 63, "spend": 18.0}, {**base, "date": "2024-01-01", "model": "gpt-4o", "group_level": 47, "spend": 18.0}, - {**base, "date": "2024-01-01", "api_key": "key-1", "group_level": 31, "spend": 12.0}, ] + top_key_rows = [{**base, "date": "2024-01-01", "api_key": "key-1", "group_level": 31, "spend": 12.0}] entity_base = { key: value for key, value in base.items() @@ -2156,7 +2421,7 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown(): }, ] - mock_prisma.db.query_raw = AsyncMock(side_effect=[main_rows, entity_rows]) + mock_prisma.db.query_raw = AsyncMock(side_effect=[main_rows, top_key_rows, entity_rows]) mock_prisma.db.litellm_verificationtoken = MagicMock() mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) @@ -2173,9 +2438,9 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown(): include_entity_breakdown=True, ) - assert mock_prisma.db.query_raw.call_count == 2 + assert mock_prisma.db.query_raw.call_count == 3 main_sql = mock_prisma.db.query_raw.call_args_list[0][0][0] - entity_sql = mock_prisma.db.query_raw.call_args_list[1][0][0] + entity_sql = mock_prisma.db.query_raw.call_args_list[2][0][0] assert "entity_id" not in main_sql assert '"team_id" AS entity_id' in entity_sql assert '(date, "team_id"),' in entity_sql diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 4c16a8613eb..f9b234e0298 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -27430,6 +27430,11 @@ export interface components { }; /** DailySpendMetadata */ DailySpendMetadata: { + /** + * Api Key Limit + * @description When set, api_keys and every api_key_breakdown list at most this many keys, ranked by spend. Totals and the model, provider, mcp and endpoint rollups still cover every key. + */ + api_key_limit?: number | null; /** * Has More * @default false From a0311dddf773c9558b4da9732a4a15c9646e9336 Mon Sep 17 00:00:00 2001 From: yassin Date: Tue, 15 Sep 2026 21:28:09 +0000 Subject: [PATCH 034/141] chore(proxy): regenerate lazy OpenAPI snapshot for api_key_limit Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_lazy_openapi_snapshot.json | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index f3b579d22c7..a3f525f0773 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -3050,6 +3050,18 @@ }, "DailySpendMetadata": { "properties": { + "api_key_limit": { + "anyOf": [ + { + "type": "integer" + }, + { + "type": "null" + } + ], + "description": "When set, api_keys and every api_key_breakdown list at most this many keys, ranked by spend. Totals and the model, provider, mcp and endpoint rollups still cover every key.", + "title": "Api Key Limit" + }, "has_more": { "default": false, "title": "Has More", From f94a40f841c3dcf4498d4cb7acef7b482cf91dc5 Mon Sep 17 00:00:00 2001 From: yassin Date: Tue, 15 Sep 2026 21:52:37 +0000 Subject: [PATCH 035/141] perf(proxy): serve key-free rollups and top-N keys from one UNION ALL statement Both arms now run in a single query_raw call so totals and per-key breakdowns come from the same snapshot. USAGE_TOP_API_KEYS_LIMIT can be raised via env for deployments that need every key in the response. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 5 +- .../common_daily_activity.py | 134 ++++++------------ .../test_common_daily_activity.py | 98 ++++++------- 3 files changed, 87 insertions(+), 150 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index 36bc16a0ae3..565c6433c6e 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -2027,10 +2027,7 @@ MCP_SPEND_LOG_MODEL_PREFIX: Final[str] = "MCP: " PTU_SENTINEL_API_KEY: Final[str] = "__ptu_flat_cost__" PTU_ROLLUP_JOB_ID: Final[str] = "ptu_flat_cost_rollup_job" PTU_ROLLUP_LOCK_TTL_SECONDS: Final[int] = 900 -# Per-api_key rollups on the aggregated usage endpoint cover only the top N keys -# by spend so the result set stops growing with key count. Totals and the -# model/provider/endpoint rollups still cover every key. -USAGE_TOP_API_KEYS_LIMIT: Final[int] = 100 +USAGE_TOP_API_KEYS_LIMIT: Final[int] = int(os.getenv("USAGE_TOP_API_KEYS_LIMIT", "100")) # Furthest back the catch-up pass looks for unpriced PTU days when a deployment # declares no ptu_effective_from, bounding the scan for an open-ended window. PTU_ROLLUP_MAX_BACKFILL_DAYS: Final[int] = 90 diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index 4aafd416263..95a89460d90 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -170,8 +170,6 @@ class _EntityRollupRow(_GroupingSetsRow): class _AggregatedQueryKwargs(TypedDict): - """Filter arguments shared by the three aggregated SQL builders.""" - table_name: ReadOnly[str] entity_id_field: ReadOnly[str] entity_id: ReadOnly[str | list[str] | None] @@ -749,16 +747,14 @@ def _build_aggregated_sql_query( timezone_offset_minutes: int | None = None, include_current_utc_day: bool = False, ) -> tuple[str, list[str]]: # mutable-ok: SQL text plus its ordered $N params - """Build the key-free GROUPING SETS query for aggregated daily activity. + """Build the GROUPING SETS query for aggregated daily activity. - Emits the grand total, per-date totals and the per-(date, model), model_group, - provider, mcp tool and endpoint rollups. api_key is never a grouping column here, - so the row count is bounded by dates x distinct models/providers/endpoints and - does not grow with the number of keys. Per-key rollups come from - _build_top_api_keys_sql_query. Both queries emit the same 7-bit group_level - bitmask (date, api_key, model, model_group, provider, mcp, endpoint); this one - hard-codes the api_key bit to "rolled up" so the dispatcher can consume the two - result sets as one stream. + One statement, two UNION ALL arms over the same WHERE clause. The first arm is + key-free: grand total, per-date totals and the (date, model / model_group / + provider / mcp / endpoint) rollups, so its row count never grows with the number + of keys. The second arm emits the (date, , api_key) rollups for the + USAGE_TOP_API_KEYS_LIMIT highest-spend keys only. Both arms share the 7-bit + group_level bitmask (date, api_key, model, model_group, provider, mcp, endpoint). Returns: Tuple of (sql_query, params_list) ready for prisma_client.db.query_raw(). @@ -771,77 +767,6 @@ def _build_aggregated_sql_query( start_date, end_date, timezone_offset_minutes, include_current_utc_day ) - where_clause, sql_params = _build_aggregated_where_clause( - entity_id_field=entity_id_field, - entity_id=entity_id, - adjusted_start=adjusted_start, - adjusted_end=adjusted_end, - model=model, - api_key=api_key, - exclude_entity_ids=exclude_entity_ids, - ) - - # TODO: drop the successful_requests/failed_requests aggregates (and the - # total_successful_requests metadata they feed) once the admin UI reads SGR - # only from LiteLLM_DailyGatewayRequests. The remaining spend, token and - # api_requests rollups are still served from here. - sql_query: Final = f""" - SELECT - date, - NULL::text AS api_key, - model, - {_MODEL_GROUP_EXPR} AS model_group, - custom_llm_provider, - mcp_namespaced_tool_name, - endpoint, - (GROUPING(date) << 6) | {_API_KEY_ROLLED_UP_BIT} - | GROUPING(model, {_MODEL_GROUP_EXPR}, - custom_llm_provider, mcp_namespaced_tool_name, - endpoint) AS group_level,{_rollup_metric_select(table_name)} - FROM "{pg_table}" - WHERE {where_clause} - GROUP BY GROUPING SETS ( - (date), - (date, model), - (date, {_MODEL_GROUP_EXPR}), - (date, custom_llm_provider), - (date, mcp_namespaced_tool_name), - (date, endpoint), - () - ) - """ - - return sql_query, sql_params - - -def _build_top_api_keys_sql_query( - *, - table_name: str, - entity_id_field: str, - entity_id: str | list[str] | None, # mutable-ok: filter union shared with the paginated path - start_date: str, - end_date: str, - model: str | None, - api_key: str | list[str] | None, # mutable-ok: filter union shared with the paginated path - exclude_entity_ids: list[str] | None = None, # mutable-ok: filter union shared with the paginated path - timezone_offset_minutes: int | None = None, - include_current_utc_day: bool = False, -) -> tuple[str, list[str]]: # mutable-ok: SQL text plus its ordered $N params - """Per-key companion to _build_aggregated_sql_query. - - Ranks keys by spend over the same WHERE clause, keeps the top - USAGE_TOP_API_KEYS_LIMIT (ties broken by api_key so the set is stable across - refreshes) and emits the six (date, , api_key) rollups for those keys - only. The PTU flat-cost sentinel never ranks, so it cannot occupy a visible slot. - """ - pg_table: Final = _PRISMA_TO_PG_TABLE.get(table_name) - if pg_table is None: - raise ValueError(f"Unknown table name: {table_name}") - - adjusted_start, adjusted_end = _adjust_dates_for_timezone( - start_date, end_date, timezone_offset_minutes, include_current_utc_day - ) - where_clause, where_params = _build_aggregated_where_clause( entity_id_field=entity_id_field, entity_id=entity_id, @@ -852,9 +777,38 @@ def _build_top_api_keys_sql_query( exclude_entity_ids=exclude_entity_ids, ) sentinel_param: Final = f"${len(where_params) + 1}" + metric_select: Final = _rollup_metric_select(table_name) + # TODO: drop the successful_requests/failed_requests aggregates (and the + # total_successful_requests metadata they feed) once the admin UI reads SGR + # only from LiteLLM_DailyGatewayRequests. The remaining spend, token and + # api_requests rollups are still served from here. sql_query: Final = f""" - WITH top_api_keys AS ( + (SELECT + date, + NULL::text AS api_key, + model, + {_MODEL_GROUP_EXPR} AS model_group, + custom_llm_provider, + mcp_namespaced_tool_name, + endpoint, + (GROUPING(date) << 6) | {_API_KEY_ROLLED_UP_BIT} + | GROUPING(model, {_MODEL_GROUP_EXPR}, + custom_llm_provider, mcp_namespaced_tool_name, + endpoint) AS group_level,{metric_select} + FROM "{pg_table}" + WHERE {where_clause} + GROUP BY GROUPING SETS ( + (date), + (date, model), + (date, {_MODEL_GROUP_EXPR}), + (date, custom_llm_provider), + (date, mcp_namespaced_tool_name), + (date, endpoint), + () + )) + UNION ALL + (WITH top_api_keys AS ( SELECT api_key FROM "{pg_table}" WHERE {where_clause} AND api_key <> {sentinel_param} @@ -872,7 +826,7 @@ def _build_top_api_keys_sql_query( endpoint, GROUPING(date, api_key, model, {_MODEL_GROUP_EXPR}, custom_llm_provider, mcp_namespaced_tool_name, - endpoint) AS group_level,{_rollup_metric_select(table_name)} + endpoint) AS group_level,{metric_select} FROM "{pg_table}" WHERE {where_clause} AND api_key IN (SELECT api_key FROM top_api_keys) GROUP BY GROUPING SETS ( @@ -882,7 +836,7 @@ def _build_top_api_keys_sql_query( (date, custom_llm_provider, api_key), (date, mcp_namespaced_tool_name, api_key), (date, endpoint, api_key) - ) + )) """ return sql_query, [*where_params, PTU_SENTINEL_API_KEY] @@ -1432,17 +1386,15 @@ async def get_daily_activity_aggregated( timezone_offset_minutes=timezone_offset_minutes, include_current_utc_day=include_current_utc_day, ) - key_free_sql, key_free_params = _build_aggregated_sql_query(**query_kwargs) - top_keys_sql, top_keys_params = _build_top_api_keys_sql_query(**query_kwargs) + sql_query, sql_params = _build_aggregated_sql_query(**query_kwargs) entity_query: Final = _build_entity_rollup_sql_query(**query_kwargs) if include_entity_breakdown else None - raw_key_free_rows, raw_top_key_rows, raw_entity_rows = await asyncio.gather( - prisma_client.db.query_raw(key_free_sql, *key_free_params), - prisma_client.db.query_raw(top_keys_sql, *top_keys_params), + raw_rows, raw_entity_rows = await asyncio.gather( + prisma_client.db.query_raw(sql_query, *sql_params), _query_raw_optional(prisma_client, entity_query), ) - records: Final = [_GroupingSetsRow(**row) for row in (*(raw_key_free_rows or ()), *(raw_top_key_rows or ()))] + records: Final = [_GroupingSetsRow(**row) for row in (raw_rows or ())] # The grouping-sets dispatcher places each row directly in its bucket # using the row's GROUPING() bitmask. No Python-side summing needed. diff --git a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py index dec98a1da27..5ff3f89343b 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py +++ b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py @@ -18,7 +18,6 @@ from litellm.proxy.management_endpoints.common_daily_activity import ( _adjust_dates_for_timezone, _build_aggregated_sql_query, _build_entity_rollup_sql_query, - _build_top_api_keys_sql_query, _is_user_agent_tag, _record_to_spend_metrics, get_api_key_metadata, @@ -166,7 +165,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "autorouter_savings_spend": 0.0, "failed_requests": 0, } - key_free_rows = [ + mock_rows = [ # (date, endpoint) — rolls up across api_keys and models { **base, @@ -218,8 +217,6 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "api_requests": 3, "successful_requests": 3, }, - ] - top_key_rows = [ # (date, endpoint, api_key) — populates the per-key sub-bucket { **base, @@ -247,7 +244,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): }, ] - mock_prisma.db.query_raw = AsyncMock(side_effect=[key_free_rows, top_key_rows]) + mock_prisma.db.query_raw = AsyncMock(return_value=mock_rows) mock_prisma.db.litellm_verificationtoken = MagicMock() mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) @@ -293,11 +290,8 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): assert "key-2" in embeddings_endpoint.api_key_breakdown assert embeddings_endpoint.api_key_breakdown["key-2"].metrics.spend == 3.0 - # One key-free rollup query plus one bounded per-key query, no find_many - assert mock_prisma.db.query_raw.call_count == 2 - key_free_sql, top_keys_sql = (call.args[0] for call in mock_prisma.db.query_raw.call_args_list) - assert "top_api_keys" not in key_free_sql - assert "WITH top_api_keys AS" in top_keys_sql + # Verify query_raw was called (not find_many) + mock_prisma.db.query_raw.assert_called_once() @pytest.mark.asyncio @@ -484,9 +478,7 @@ async def test_get_api_key_metadata_recovers_double_hashed_key_via_reverse_hash( return_value=[SimpleNamespace(user_id="alice", user_email="alice@example.com")] ) mock_prisma.db.query_raw = AsyncMock( - return_value=[ - {"digest": double_hashed, "key_alias": "batch-worker", "team_id": "team-1", "user_id": "alice"} - ] + return_value=[{"digest": double_hashed, "key_alias": "batch-worker", "team_id": "team-1", "user_id": "alice"}] ) result = await get_api_key_metadata( @@ -824,7 +816,7 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): "autorouter_savings_spend": 0.0, "failed_requests": 0, } - key_free_rows = [ + mock_rows = [ { **base, "date": "2024-01-01", @@ -837,8 +829,6 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): "api_requests": 1, "successful_requests": 1, }, - ] - top_key_rows = [ { **base, "date": "2024-01-01", @@ -853,7 +843,7 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): }, ] - mock_prisma.db.query_raw = AsyncMock(side_effect=[key_free_rows, top_key_rows]) + mock_prisma.db.query_raw = AsyncMock(return_value=mock_rows) # Active table returns nothing for this key mock_prisma.db.litellm_verificationtoken = MagicMock() @@ -1226,6 +1216,7 @@ class TestBuildAggregatedSqlQuery: "user-1", "bedrock/global.anthropic.claude-opus-4-8", "sk-test", + PTU_SENTINEL_API_KEY, ] assert "model = $4" in sql assert "api_key = $5" in sql @@ -1259,9 +1250,9 @@ class TestBuildAggregatedSqlQuery: assert "(date, model_group)" not in normalized assert "COALESCE(model_group, model)" not in normalized - def test_key_free_query_never_groups_by_api_key(self): - """The main rollup query must not emit one row per key, that is what blew up - the query engine at 3k+ keys. Every grouping set stays key-free and api_key + def test_totals_arm_never_groups_by_api_key(self): + """The totals arm must not emit one row per key, that is what blew up the + query engine at 3k+ keys. Every grouping set there stays key-free and api_key is projected as a NULL literal so the dispatcher's row shape is unchanged.""" sql, _ = _build_aggregated_sql_query( table_name="litellm_dailyuserspend", @@ -1273,13 +1264,15 @@ class TestBuildAggregatedSqlQuery: api_key=None, ) - normalized = " ".join(sql.split()) - grouping_block = normalized.split("GROUP BY GROUPING SETS (", 1)[1] + totals_arm, _ = " ".join(sql.split()).split("UNION ALL") + grouping_block = totals_arm.split("GROUP BY GROUPING SETS (", 1)[1] assert "api_key" not in grouping_block - assert "NULL::text AS api_key" in normalized + assert "NULL::text AS api_key" in totals_arm - def test_top_api_keys_query_ranks_keys_deterministically_and_shares_filters(self): - sql, params = _build_top_api_keys_sql_query( + def test_per_key_arm_ranks_keys_deterministically_and_shares_filters(self): + """Both arms sit in one statement so totals and per-key rows come from the + same snapshot, and the per-key arm reuses the caller's filter params.""" + sql, params = _build_aggregated_sql_query( table_name="litellm_dailyuserspend", entity_id_field="user_id", entity_id="user-1", @@ -1290,25 +1283,20 @@ class TestBuildAggregatedSqlQuery: timezone_offset_minutes=-330, ) - normalized = " ".join(sql.split()) - assert f"ORDER BY SUM(spend) DESC, api_key LIMIT {USAGE_TOP_API_KEYS_LIMIT}" in normalized - assert "api_key IN (SELECT api_key FROM top_api_keys)" in normalized - assert "api_key <> $6" in normalized - grouping_block = normalized.split("GROUP BY GROUPING SETS (", 1)[1] + totals_arm, per_key_arm = " ".join(sql.split()).split("UNION ALL") + assert "top_api_keys" not in totals_arm + assert f"ORDER BY SUM(spend) DESC, api_key LIMIT {USAGE_TOP_API_KEYS_LIMIT}" in per_key_arm + assert "api_key IN (SELECT api_key FROM top_api_keys)" in per_key_arm + assert "api_key <> $6" in per_key_arm + assert per_key_arm.count("model = $4 AND api_key = $5") == 2 + grouping_block = per_key_arm.split("GROUP BY GROUPING SETS (", 1)[1] assert grouping_block.count(", api_key)") == 6 assert grouping_block.count("(date") == 6 - assert params == [ - "2026-05-29", - "2026-06-02", - "user-1", - "bedrock/global.anthropic.claude-opus-4-8", - "sk-test", - PTU_SENTINEL_API_KEY, - ] + assert params[-1] == PTU_SENTINEL_API_KEY class TestAggregatedEmptyEntityFilter: - _BUILDERS: Final = (_build_aggregated_sql_query, _build_top_api_keys_sql_query, _build_entity_rollup_sql_query) + _BUILDERS: Final = (_build_aggregated_sql_query, _build_entity_rollup_sql_query) @pytest.mark.parametrize("build", _BUILDERS) def test_empty_entity_list_emits_no_degenerate_in_clause(self, build): @@ -1325,7 +1313,7 @@ class TestAggregatedEmptyEntityFilter: normalized = " ".join(sql.split()) assert "IN ()" not in normalized assert '"team_id" IN' not in normalized - sentinel_params = [PTU_SENTINEL_API_KEY] if build is _build_top_api_keys_sql_query else [] + sentinel_params = [PTU_SENTINEL_API_KEY] if build is _build_aggregated_sql_query else [] assert params == ["2026-08-01", "2026-08-19", *sentinel_params] @pytest.mark.parametrize("build", _BUILDERS) @@ -1357,7 +1345,7 @@ class TestAggregatedEmptyEntityFilter: normalized = " ".join(sql.split()) assert '"team_id" IN ($3, $4)' in normalized assert "FALSE" not in normalized - sentinel_params = [PTU_SENTINEL_API_KEY] if build is _build_top_api_keys_sql_query else [] + sentinel_params = [PTU_SENTINEL_API_KEY] if build is _build_aggregated_sql_query else [] assert params == ["2026-08-01", "2026-08-19", "team-alpha", "team-beta", *sentinel_params] @@ -1373,7 +1361,7 @@ async def test_get_daily_activity_aggregated_empty_result_set(): mock_prisma = MagicMock() mock_prisma.db = MagicMock() - key_free_rows = [ + mock_rows = [ { "date": None, "api_key": None, @@ -1398,7 +1386,7 @@ async def test_get_daily_activity_aggregated_empty_result_set(): "failed_requests": None, } ] - mock_prisma.db.query_raw = AsyncMock(side_effect=[key_free_rows, []]) + mock_prisma.db.query_raw = AsyncMock(return_value=mock_rows) result = await get_daily_activity_aggregated( prisma_client=mock_prisma, @@ -1492,13 +1480,13 @@ def _psycopg_query_raw(conn: psycopg.Connection, row_counts: list[int]): async def test_get_daily_activity_aggregated_bounds_api_key_rollups( _aggregated_postgresql: psycopg.Connection, ): - """Run both GROUPING SETS queries against real Postgres with more keys than the cap. + """Run the GROUPING SETS statement against real Postgres with more keys than the cap. key-004 and key-005 tie on spend exactly at the USAGE_TOP_API_KEYS_LIMIT cutoff; the api_key tiebreaker must keep key-004 and drop key-005. The PTU sentinel outspends every key but must not take a slot. Excluded keys and the sentinel still count toward the totals and the model rollup, which come from - the key-free query. + the key-free arm. """ n_keys: Final = USAGE_TOP_API_KEYS_LIMIT + 5 key_rows: Final = [ @@ -1554,10 +1542,10 @@ async def test_get_daily_activity_aggregated_bounds_api_key_rollups( api_key=None, ) - # Key-free query: (), (date), (date, model), (date, model_group), two providers, + # Key-free arm: (), (date), (date, model), (date, model_group), two providers, # one mcp NULL bucket, endpoint plus its NULL bucket = 9 rows regardless of key count. - # Top-key query: six per-key grouping sets, each capped at the limit. - assert row_counts == [9, 6 * USAGE_TOP_API_KEYS_LIMIT] + # Per-key arm: six per-key grouping sets, each capped at the limit. + assert row_counts == [9 + 6 * USAGE_TOP_API_KEYS_LIMIT] assert result.metadata.total_spend == pytest.approx(key_spend + 1000.0) assert result.metadata.total_api_requests == n_keys @@ -1579,11 +1567,11 @@ async def test_get_daily_activity_aggregated_bounds_api_key_rollups( @pytest.mark.asyncio -async def test_get_daily_activity_aggregated_explicit_api_key_filter_scopes_both_queries( +async def test_get_daily_activity_aggregated_explicit_api_key_filter_scopes_both_arms( _aggregated_postgresql: psycopg.Connection, ): """An explicit api_key filter must scope the key-free totals and the per-key - rollups to that key alone, so the two result sets never disagree.""" + rollups to that key alone, so the two arms never disagree.""" rows: Final = [ ( f"row-{i}", @@ -2358,7 +2346,7 @@ def test_entity_rollup_sql_query_and_api_key_list_filter(): api_key=[], ) assert "FALSE" in empty_sql - assert empty_params == ["2024-01-01", "2024-01-31"] + assert empty_params == ["2024-01-01", "2024-01-31", PTU_SENTINEL_API_KEY] @pytest.mark.asyncio @@ -2393,8 +2381,8 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown(): {**base, "date": None, "group_level": 127, "spend": 18.0}, {**base, "date": "2024-01-01", "group_level": 63, "spend": 18.0}, {**base, "date": "2024-01-01", "model": "gpt-4o", "group_level": 47, "spend": 18.0}, + {**base, "date": "2024-01-01", "api_key": "key-1", "group_level": 31, "spend": 12.0}, ] - top_key_rows = [{**base, "date": "2024-01-01", "api_key": "key-1", "group_level": 31, "spend": 12.0}] entity_base = { key: value for key, value in base.items() @@ -2421,7 +2409,7 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown(): }, ] - mock_prisma.db.query_raw = AsyncMock(side_effect=[main_rows, top_key_rows, entity_rows]) + mock_prisma.db.query_raw = AsyncMock(side_effect=[main_rows, entity_rows]) mock_prisma.db.litellm_verificationtoken = MagicMock() mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) @@ -2438,9 +2426,9 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown(): include_entity_breakdown=True, ) - assert mock_prisma.db.query_raw.call_count == 3 + assert mock_prisma.db.query_raw.call_count == 2 main_sql = mock_prisma.db.query_raw.call_args_list[0][0][0] - entity_sql = mock_prisma.db.query_raw.call_args_list[2][0][0] + entity_sql = mock_prisma.db.query_raw.call_args_list[1][0][0] assert "entity_id" not in main_sql assert '"team_id" AS entity_id' in entity_sql assert '(date, "team_id"),' in entity_sql From c3e937b84544902b2de41a2748def7803b8f3368 Mon Sep 17 00:00:00 2001 From: yassin Date: Tue, 15 Sep 2026 22:11:20 +0000 Subject: [PATCH 036/141] docs(proxy): describe the single UNION ALL aggregate statement in the endpoint docstring Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../management_endpoints/common_daily_activity.py | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index 95a89460d90..8a3ba196ab2 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -1350,11 +1350,12 @@ async def get_daily_activity_aggregated( ) -> SpendAnalyticsPaginatedResponse: """Aggregated variant that returns the full result set (no pagination). - Runs two GROUPING SETS queries in parallel: a key-free one for totals and the - model/provider/mcp/endpoint rollups (row count independent of key cardinality) - and a bounded one for the per-key rollups of the top USAGE_TOP_API_KEYS_LIMIT - keys by spend. breakdown.api_keys and every api_key_breakdown therefore list at - most that many keys, while the totals and the key-free rollups cover every key. + Runs one GROUPING SETS statement with two UNION ALL arms: a key-free one for totals + and the model/provider/mcp/endpoint rollups (row count independent of key + cardinality) and a bounded one for the per-key rollups of the top + USAGE_TOP_API_KEYS_LIMIT keys by spend. breakdown.api_keys and every + api_key_breakdown therefore list at most that many keys, while the totals and the + key-free rollups cover every key. include_entity_breakdown runs a small companion rollup query and folds `breakdown.entities` onto the response, as entity-scoped views like Team Usage need. From aa7f1e16b8810a9321fd12539732ff738aeebd38 Mon Sep 17 00:00:00 2001 From: yucheng Date: Tue, 15 Sep 2026 23:53:14 +0000 Subject: [PATCH 037/141] feat(proxy): throttle failed Admin UI sign-ins per source and source/username Replace the username-global lockout with counters keyed by source address and by source/username pair. Each has a fixed counting window (60s) and a separate block TTL (300s). Blocks are soft: a correct password still signs in, wrong passwords from a blocked key take one of 5 held slots per worker and are held 30s before a 429. Once a pair is blocked its failures stop counting against the source. The source scope runs only when trusted_proxy_ranges is set, IPv6 is grouped by /64, and per-source limits accept IP and CIDR overrides with longest-prefix matching. Redis is authoritative through one Lua script per failure, with bounded per-worker fallback when Redis raises. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/caching/redis_cache.py | 4 +- litellm/proxy/_types.py | 23 +- litellm/proxy/auth/login_throttle.py | 578 +++++---- litellm/proxy/auth/login_utils.py | 26 +- litellm/proxy/proxy_server.py | 12 +- .../proxy/auth/test_login_utils.py | 1072 +++++++++-------- .../proxy/proxy_server/conftest.py | 31 +- .../proxy_server/test_routes_login_sso.py | 121 +- tests/test_litellm/proxy/test_proxy_server.py | 34 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 26 +- 10 files changed, 1043 insertions(+), 884 deletions(-) diff --git a/litellm/caching/redis_cache.py b/litellm/caching/redis_cache.py index 3f1bac12563..b4b2b1a334c 100644 --- a/litellm/caching/redis_cache.py +++ b/litellm/caching/redis_cache.py @@ -874,7 +874,7 @@ class RedisCache(BaseCache): return _LUA_COUNT.validate_python(count) @_redis_circuit_breaker_guard_sync - def batch_get_counts(self, key_list: list[str] | tuple[str, ...]) -> tuple[int | None, ...]: + def batch_get_counts(self, key_list: list[str]) -> tuple[int | None, ...]: """Read integer counters for ``key_list``, in order, raising when Redis cannot answer. ``batch_get_cache`` swallows every failure and returns an empty dict, which the caller @@ -885,7 +885,7 @@ class RedisCache(BaseCache): return _decoded_counts(self._run_redis_mget_operation(keys=namespaced_keys)) @_redis_circuit_breaker_guard - async def async_batch_get_counts(self, key_list: list[str] | tuple[str, ...]) -> tuple[int | None, ...]: + async def async_batch_get_counts(self, key_list: list[str]) -> tuple[int | None, ...]: """Async twin of ``batch_get_counts``, raising on failure the same way.""" namespaced_keys: Final = [self.check_and_fix_namespace(key=key) for key in key_list] return _decoded_counts(await self._async_run_redis_mget_operation(keys=namespaced_keys)) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index ddcc7b2dece..2537555f316 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2736,20 +2736,29 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): description="sends alerts if requests hang for 5min+", ) ui_access_mode: Literal["admin_only", "all"] | None = Field("all", description="Control access to the Proxy UI") - max_failed_login_attempts: int | None = Field( - None, - ge=1, - description="Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Set under `general_settings` in config.yaml. Defaults to 50", - ) max_failed_login_attempts_per_source: int | None = Field( None, ge=1, - description="Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Set under `general_settings` in config.yaml. Defaults to 250", + description="Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set, since otherwise every client behind an ingress shares one address. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10", + ) + max_failed_login_attempts_per_source_overrides: dict[str, int] | None = Field( + None, + description="Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins. Set under `general_settings` in config.yaml", + ) + max_failed_login_attempts_per_user: int | None = Field( + None, + ge=1, + description="Failed Admin UI sign-in attempts allowed from one source address for one username within `failed_login_window_seconds`. One more blocks that address for that username for `failed_login_block_seconds`, and its further failures stop counting against `max_failed_login_attempts_per_source`, so a script stuck on one account does not block everyone behind the same address. Set under `general_settings` in config.yaml. Defaults to 5", ) failed_login_window_seconds: int | None = Field( None, ge=1, - description="Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Set under `general_settings` in config.yaml. Defaults to 900", + description="Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Set under `general_settings` in config.yaml. Defaults to 60", + ) + failed_login_block_seconds: int | None = Field( + None, + ge=1, + description="How long a blocked source address, or source address and username, stays blocked. The block is soft: a correct password still signs in, but each attempt from a blocked key takes one of 5 held slots per worker and a wrong password is held for 30 seconds before it is refused with 429. Set under `general_settings` in config.yaml. Defaults to 300", ) allowed_routes: list | None = Field(None, description="Proxy API Endpoints you want users to be able to access") reject_clientside_metadata_tags: bool | None = Field( diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 50333ffad05..fb708ecf872 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -1,142 +1,231 @@ """Failed-login accounting for the Admin UI sign-in path. -Counts failed credential checks over a fixed window against two independent keys, the -username on its own and the source address on its own, so that one username attacked from -many sources and one source spraying many usernames are both counted. Repeated failures -are answered slowly, doubling from one second, and refused with 429 once either counter -reaches its limit. Built per request by ``LoginThrottle.from_request`` because it carries -that request's resolved source address, and because the coordination cache is assigned at -startup and can be reassigned later. +Wrong passwords are counted over a short window per source address and per source-and-username +pair; too many in one window blocks that key for a fixed time. Blocks are soft: a correct password +still signs in, while a wrong one from a blocked key is held open before its 429 and only a few can +be held at once, which bounds how many guesses a blocked key gets checked. A blocked pair stops +counting against its source, so one script stuck on one account does not block the whole office. """ +from __future__ import annotations + import asyncio import hashlib -from collections.abc import Awaitable, Callable, Mapping +import ipaddress +import math +import time +from collections.abc import AsyncGenerator, Mapping +from contextlib import asynccontextmanager from dataclasses import dataclass from functools import cache from types import MappingProxyType -from typing import Final, NamedTuple, NoReturn +from typing import Final, Literal, NamedTuple, NoReturn -from fastapi import Request +from fastapi import Request, status +from pydantic import TypeAdapter, ValidationError from litellm._logging import verbose_proxy_logger -from litellm.caching.dual_cache import DualCache from litellm.caching.in_memory_cache import InMemoryCache from litellm.caching.redis_cache import RedisCache from litellm.proxy._types import ProxyErrorTypes, ProxyException from litellm.proxy.auth.network import TrustedProxyConfig, normalize_cidr_ranges, resolve_client_ip -from litellm.proxy.auth.trusted_proxy_utils import TRUSTED_PROXY_RANGES_KEY from litellm.secret_managers.main import get_secret_bool -DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS: Final = 50 -DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE: Final = 250 -DEFAULT_FAILED_LOGIN_WINDOW_SECONDS: Final = 900 +DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE: Final = 10 +DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_USER: Final = 5 +DEFAULT_FAILED_LOGIN_WINDOW_SECONDS: Final = 60 +DEFAULT_FAILED_LOGIN_BLOCK_SECONDS: Final = 300 -USERNAME_DELAY_ONSET: Final = 3 -SOURCE_DELAY_ONSET: Final = 25 -FIRST_DELAY_SECONDS: Final = 1.0 -MAX_DELAY_SECONDS: Final = 30.0 -MAX_CONCURRENT_DELAYS_PER_SOURCE: Final = 5 +BLOCKED_ATTEMPT_HOLD_SECONDS: Final = 30 +MAX_HELD_ATTEMPTS_PER_KEY: Final = 5 +IPV6_SOURCE_PREFIX_LENGTH: Final = 64 -_MAX_DELAY_DOUBLINGS: Final = 16 +SOURCE_LIMIT_KEY: Final = "max_failed_login_attempts_per_source" +SOURCE_LIMIT_OVERRIDES_KEY: Final = "max_failed_login_attempts_per_source_overrides" +USER_LIMIT_KEY: Final = "max_failed_login_attempts_per_user" +WINDOW_KEY: Final = "failed_login_window_seconds" +BLOCK_KEY: Final = "failed_login_block_seconds" +TRUSTED_PROXY_RANGES_KEY: Final = "trusted_proxy_ranges" _CACHE_KEY_PREFIX: Final = "login_fail" _UNKNOWN_SOURCE: Final = "unknown" -_MAX_LOGGED_USERNAME_CHARS: Final = 128 +_MAX_TRACKED_COUNTERS: Final = 20_000 +_MAX_TRACKED_BLOCKS: Final = 10_000 +_NO_SETTINGS: Final[Mapping[str, object]] = MappingProxyType({}) +_NOT_BLOCKED: Final = (0, 0) +_LOCAL_BLOCK_EXPIRY: Final = TypeAdapter[float | None](float | None) +_SOURCE_LIMIT_OVERRIDES: Final = TypeAdapter[Mapping[str, object]](Mapping[str, object]) -_MAX_TRACKED_LOGIN_USERNAMES: Final = 10_000 -_MAX_TRACKED_LOGIN_SOURCES: Final = 10_000 +Scope = Literal["user", "source"] + +_BlockTtls = tuple[int, int] +_LUA_BLOCK_TTLS: Final = TypeAdapter[_BlockTtls](_BlockTtls) +_Network = ipaddress.IPv4Network | ipaddress.IPv6Network + +# KEYS: pair counter, pair block, source counter, source block (one cluster slot via the source hash tag) +# ARGV: pair limit, source limit (0 = source scope off), window seconds, block seconds +# Both scripts return {pair block TTL, source block TTL}; 0 or below means not blocked +_BLOCK_TTLS_LUA: Final = "return {redis.call('TTL', KEYS[2]), redis.call('TTL', KEYS[4])}" +_RECORD_FAILURE_LUA: Final = ( + "local function bump(count_key, block_key, limit) " + "local blocked = redis.call('TTL', block_key) " + "if blocked > 0 then return blocked end " + "local count = redis.call('INCR', count_key) " + "if redis.call('TTL', count_key) < 0 then redis.call('EXPIRE', count_key, ARGV[3]) end " + "if count > limit then redis.call('SET', block_key, '1', 'EX', ARGV[4]) return tonumber(ARGV[4]) end " + "return 0 end " + "local user_block = bump(KEYS[1], KEYS[2], tonumber(ARGV[1])) " + "local source_block = 0 " + "if tonumber(ARGV[2]) > 0 and user_block == 0 then " + "source_block = bump(KEYS[3], KEYS[4], tonumber(ARGV[2])) end " + "return {user_block, source_block}" +) + +_COUNTERS: Final = InMemoryCache( + max_size_in_memory=_MAX_TRACKED_COUNTERS, default_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS +) +_BLOCKS: Final = InMemoryCache(max_size_in_memory=_MAX_TRACKED_BLOCKS, default_ttl=DEFAULT_FAILED_LOGIN_BLOCK_SECONDS) +_HELD_ATTEMPTS: Final[dict[str, int]] = {} -def _bounded_store(max_entries: int) -> DualCache: - return DualCache( - in_memory_cache=InMemoryCache(max_size_in_memory=max_entries), - default_in_memory_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS, - ) +async def _sleep(seconds: float) -> None: + await asyncio.sleep(seconds) -_FAILED_LOGIN_USERNAME_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_USERNAMES) -_FAILED_LOGIN_SOURCE_CACHE: Final = _bounded_store(_MAX_TRACKED_LOGIN_SOURCES) -_NO_SETTINGS: Final = MappingProxyType({}) -_UNAVAILABLE: Final = object() - -_DELAYS_IN_FLIGHT: Final[dict[str, int]] = {} # mutable-ok: per-source slots taken and released around each held delay +@cache +def _rate_limit_disabled() -> bool: + return get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT", default_value=False) is True @cache def warn_login_counters_are_per_worker(num_workers: str) -> None: - """Warn once per process that failed sign-in counters are not shared across workers.""" verbose_proxy_logger.warning( - "Running %s workers but Redis is not configured for LiteLLM caching. " - "Failed Admin UI sign-in attempts are counted per worker, so an attacker " - "gets max_failed_login_attempts guesses per worker instead of overall. " - "Configure Redis via the 'cache' section in your proxy config.", + "Running %s workers but Redis is not configured. Failed Admin UI sign-in attempts are counted " + "per worker, so the effective limits are %s times the configured values. Configure Redis " + "to share one count across workers.", + num_workers, num_workers, ) @cache -def _rate_limit_disabled() -> bool: - """Resolved once per process so an unauthenticated flood never reaches the secret manager.""" - return bool(get_secret_bool("LITELLM_DISABLE_LOGIN_RATE_LIMIT", False)) +def warn_source_login_limit_is_off() -> None: + verbose_proxy_logger.warning( + "%s is not set, so failed Admin UI sign-in attempts are limited per source address and username " + "only. Set it to the address ranges of the proxies in front of LiteLLM to also limit each " + "source address across usernames.", + TRUSTED_PROXY_RANGES_KEY, + ) -async def _sleep(seconds: float) -> None: - """The wait a rejected sign-in is held for. Replaced in tests so the suite pays no wall clock.""" - await asyncio.sleep(seconds) - - -class FailureCounts(NamedTuple): - """Failures recorded so far in this window against each of the two keys.""" - - username: int - source: int - - -def _parse_int_setting(value: object) -> object: - if not isinstance(value, str): - return value - try: - return int(value.strip()) - except ValueError: - return value - - -def _int_setting(name: str, value: object, default: int, minimum: int) -> int: - if value is None: +def _positive_int(raw: object, key: str, default: int) -> int: + if raw is None: return default - parsed: Final = _parse_int_setting(value) - if isinstance(parsed, bool) or not isinstance(parsed, int) or parsed < minimum: + try: + value: Final = int(str(raw)) + except (TypeError, ValueError): + verbose_proxy_logger.warning("Invalid %s value %r; using %s", key, raw, default) + return default + if value < 1: + verbose_proxy_logger.warning("Invalid %s value %s (must be >= 1); using %s", key, value, default) + return default + return value + + +def _int_setting(settings: Mapping[str, object], key: str, default: int) -> int: + return _positive_int(settings.get(key), key, default) + + +def _parse_address(client_ip: str) -> ipaddress.IPv4Address | ipaddress.IPv6Address | None: + try: + return ipaddress.ip_address(client_ip) + except ValueError: + return None + + +def _parse_network(raw_range: str) -> _Network | None: + try: + return ipaddress.ip_network(raw_range.strip(), strict=False) + except ValueError: verbose_proxy_logger.warning( - "general_settings.%s=%r is not an integer >= %s; using the default of %s", name, value, minimum, default + "Invalid address or range %r in %s; skipping", raw_range, SOURCE_LIMIT_OVERRIDES_KEY + ) + return None + + +def _source_limit(settings: Mapping[str, object], client_ip: str) -> int: + """Failure allowance for this address: the most specific configured range containing it, else the default.""" + default: Final = _int_setting(settings, SOURCE_LIMIT_KEY, DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE) + raw_overrides: Final = settings.get(SOURCE_LIMIT_OVERRIDES_KEY) + if raw_overrides is None: + return default + try: + overrides: Final = _SOURCE_LIMIT_OVERRIDES.validate_python(raw_overrides) + except ValidationError: + verbose_proxy_logger.warning( + "Invalid %s value; expected a mapping of address or range to limit", SOURCE_LIMIT_OVERRIDES_KEY ) return default - return parsed + address: Final = _parse_address(client_ip) + if address is None: + return default + matches: Final = sorted( + (network.prefixlen, _positive_int(raw_limit, SOURCE_LIMIT_OVERRIDES_KEY, default)) + for raw_range, raw_limit in overrides.items() + if (network := _parse_network(raw_range)) is not None and address in network + ) + return matches[-1][1] if matches else default -def _as_count(cached: object) -> int: - return int(cached) if isinstance(cached, int | float) and not isinstance(cached, bool) else 0 +def source_group(client_ip: str) -> str: + """The bucket an address is counted in: IPv4 as is, IPv6 by its /64, so one prefix holder cannot rotate.""" + address: Final = _parse_address(client_ip) + if address is None: + return client_ip + if isinstance(address, ipaddress.IPv6Address): + mapped: Final = address.ipv4_mapped + if mapped is not None: + return str(mapped) + return str(ipaddress.ip_network((address, IPV6_SOURCE_PREFIX_LENGTH), strict=False)) + return str(address) + + +class _Keys(NamedTuple): + pair_counter: str + pair_block: str + source_counter: str + source_block: str + + +@dataclass(frozen=True, slots=True) +class Block: + scope: Scope + retry_after: int @dataclass(frozen=True, slots=True) class LoginThrottle: - """Fixed-window failed-login accounting for one request's username and source address.""" + """Failed-login limits for one request's source address; ``source_limit`` is None when the + source scope is off because ``trusted_proxy_ranges`` is unset and the peer address is the ingress.""" client_ip: str - max_attempts: int - max_attempts_per_source: int + source_limit: int | None + user_limit: int window_seconds: int - username_cache: DualCache - source_cache: DualCache + block_seconds: int + counters: InMemoryCache + blocks: InMemoryCache redis_cache: RedisCache | None = None enabled: bool = True @classmethod def from_request( - cls, request: Request, general_settings: Mapping[str, object] | None, redis_cache: RedisCache | None - ) -> "LoginThrottle": - """Build the throttle for this request from the proxy's general_settings and shared Redis cache.""" - settings: Final = general_settings or _NO_SETTINGS + cls, + request: Request, + general_settings: Mapping[str, object] | None, + redis_cache: RedisCache | None, + ) -> LoginThrottle: + settings: Final = general_settings if general_settings is not None else _NO_SETTINGS cidrs: Final = normalize_cidr_ranges( settings.get(TRUSTED_PROXY_RANGES_KEY), setting_name=TRUSTED_PROXY_RANGES_KEY ) @@ -145,199 +234,166 @@ class LoginThrottle: ) return cls( client_ip=resolved or _UNKNOWN_SOURCE, - max_attempts=_int_setting( - "max_failed_login_attempts", - settings.get("max_failed_login_attempts"), - DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS, - 1, - ), - max_attempts_per_source=_int_setting( - "max_failed_login_attempts_per_source", - settings.get("max_failed_login_attempts_per_source"), - DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE, - 1, - ), - window_seconds=_int_setting( - "failed_login_window_seconds", - settings.get("failed_login_window_seconds"), - DEFAULT_FAILED_LOGIN_WINDOW_SECONDS, - 1, - ), - username_cache=_FAILED_LOGIN_USERNAME_CACHE, - source_cache=_FAILED_LOGIN_SOURCE_CACHE, + source_limit=_source_limit(settings, resolved) if cidrs and resolved is not None else None, + user_limit=_int_setting(settings, USER_LIMIT_KEY, DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_USER), + window_seconds=_int_setting(settings, WINDOW_KEY, DEFAULT_FAILED_LOGIN_WINDOW_SECONDS), + block_seconds=_int_setting(settings, BLOCK_KEY, DEFAULT_FAILED_LOGIN_BLOCK_SECONDS), + counters=_COUNTERS, + blocks=_BLOCKS, redis_cache=redis_cache, enabled=not _rate_limit_disabled(), ) - @staticmethod - def _loggable(username: str) -> str: - """The username with anything that could forge a log line removed.""" - return "".join(c for c in username if c.isprintable())[:_MAX_LOGGED_USERNAME_CHARS] + def _keys(self, username: str) -> _Keys: + group: Final = source_group(self.client_ip) + user: Final = hashlib.sha256(username.casefold().encode("utf-8")).hexdigest() + return _Keys( + pair_counter=f"{_CACHE_KEY_PREFIX}:{{{group}}}:user:{user}", + pair_block=f"{_CACHE_KEY_PREFIX}:{{{group}}}:block:user:{user}", + source_counter=f"{_CACHE_KEY_PREFIX}:{{{group}}}:source", + source_block=f"{_CACHE_KEY_PREFIX}:{{{group}}}:block:source", + ) - @staticmethod - def _username_key(username: str) -> str: - identity: Final = hashlib.sha256(username.casefold().encode("utf-8")).hexdigest() - return f"{_CACHE_KEY_PREFIX}:user:{identity}" - - def _source_key(self) -> str: - return f"{_CACHE_KEY_PREFIX}:source:{self.client_ip}" - - async def _outcome(self, work: Awaitable[object]) -> object: + @asynccontextmanager + async def attempt(self, username: str, *, exempt: bool = False) -> AsyncGenerator[LoginAttempt]: + if not self.enabled or exempt: + yield LoginAttempt(throttle=self, username=username, block=None) + return + keys: Final = self._keys(username) + block: Final = await self._active_block(keys) + if block is None: + yield LoginAttempt(throttle=self, username=username, block=None) + return + slot: Final = keys.pair_block if block.scope == "user" else keys.source_block + held: Final = _HELD_ATTEMPTS.get(slot, 0) + if held >= MAX_HELD_ATTEMPTS_PER_KEY: + verbose_proxy_logger.warning( + "Admin UI sign-in refused: %s attempts already held for a blocked %s; username=%r source=%s", + held, + block.scope, + username, + self.client_ip, + ) + self.refuse(BLOCKED_ATTEMPT_HOLD_SECONDS) + _HELD_ATTEMPTS[slot] = held + 1 try: - return await work - except Exception as exc: # noqa: BLE001 # an unreachable cache must never deny a valid credential - verbose_proxy_logger.warning("login attempt accounting unavailable: %s", exc) - return _UNAVAILABLE + yield LoginAttempt(throttle=self, username=username, block=block) + finally: + remaining: Final = _HELD_ATTEMPTS.get(slot, 1) - 1 + if remaining > 0: + _HELD_ATTEMPTS[slot] = remaining + else: + _HELD_ATTEMPTS.pop(slot, None) - async def _shared(self, work: Callable[[RedisCache], Awaitable[object]]) -> object: - """The Redis result, or ``_UNAVAILABLE`` when Redis is not configured or the call raised.""" - redis_cache: Final = self.redis_cache - if redis_cache is None: - return _UNAVAILABLE - return await self._outcome(work(redis_cache)) + async def _active_block(self, keys: _Keys) -> Block | None: + local: Final = self._local_block_ttls(keys) + shared: Final = await self._shared_block_ttls(keys) + user_ttl: Final = max(local[0], shared[0]) + source_ttl: Final = max(local[1], shared[1]) + if user_ttl > 0: + return Block(scope="user", retry_after=user_ttl) + if self.source_limit is not None and source_ttl > 0: + return Block(scope="source", retry_after=source_ttl) + return None - async def _failures(self, store: DualCache, key: str) -> int: - """The shared count plus this worker's own. - - A failure is written to exactly one of the two: Redis, or this worker's store when Redis - refused it. So the local store is empty while Redis is healthy, and once Redis answers - again the guesses it missed still count. Read through ``async_batch_get_counts`` because - ``async_get_cache`` turns a failed GET into ``None``, which would pass as an empty counter. - """ - local: Final = _as_count(await self._outcome(store.async_get_cache(key=key))) - shared: Final = await self._shared(lambda redis_cache: redis_cache.async_batch_get_counts((key,))) - if not isinstance(shared, tuple): - return local - return _as_count(shared[0]) + local - - async def _remaining_window(self, key: str) -> int: - """Seconds until this counter expires. - - Counters are only ever written together with their expiry, so a counter without one - was stripped out of band (PERSIST, a restore). It is given the full window again, - since nothing increments a key once the limit is reached. - """ + async def _shared_block_ttls(self, keys: _Keys) -> _BlockTtls: if self.redis_cache is None: - return self.window_seconds - ttl: Final = await self._shared(lambda redis_cache: redis_cache.async_get_ttl(key)) - if isinstance(ttl, int) and ttl > 0: - return min(ttl, self.window_seconds) - await self._shared(lambda redis_cache: redis_cache.async_increment_with_floor(key, 0, self.window_seconds)) - return self.window_seconds + return _NOT_BLOCKED + try: + return _LUA_BLOCK_TTLS.validate_python( + await self.redis_cache.async_register_script(_BLOCK_TTLS_LUA)(list(keys), []) + ) + except Exception as err: + self._warn_redis(err) + return _NOT_BLOCKED - def _refused(self, retry_after: int, param: str) -> ProxyException: - return ProxyException( + def _local_block_ttls(self, keys: _Keys) -> _BlockTtls: + return self._local_block_ttl(keys.pair_block), self._local_block_ttl(keys.source_block) + + def _local_block_ttl(self, block_key: str) -> int: + expires_at: Final = _LOCAL_BLOCK_EXPIRY.validate_python(self.blocks.get_cache(block_key)) + if expires_at is None: + return 0 + return max(math.ceil(expires_at - time.time()), 0) + + async def record_failure(self, username: str) -> _BlockTtls: + keys: Final = self._keys(username) + source_limit: Final = self.source_limit or 0 + if self.redis_cache is not None: + try: + return _LUA_BLOCK_TTLS.validate_python( + await self.redis_cache.async_register_script(_RECORD_FAILURE_LUA)( + list(keys), [self.user_limit, source_limit, self.window_seconds, self.block_seconds] + ) + ) + except Exception as err: + self._warn_redis(err) + user_block: Final = self._local_bump(keys.pair_counter, keys.pair_block, self.user_limit) + if source_limit == 0 or user_block > 0: + return user_block, 0 + return user_block, self._local_bump(keys.source_counter, keys.source_block, source_limit) + + def _local_bump(self, count_key: str, block_key: str, limit: int) -> int: + blocked: Final = self._local_block_ttl(block_key) + if blocked > 0: + return blocked + count: Final = int(self.counters.increment_cache(count_key, 1, ttl=self.window_seconds)) + if count <= limit: + return 0 + self.blocks.set_cache(block_key, time.time() + self.block_seconds, ttl=self.block_seconds) + return self.block_seconds + + async def clear_pair(self, username: str) -> None: + pair_counter: Final = self._keys(username).pair_counter + if self.redis_cache is not None: + try: + await self.redis_cache.async_delete_cache(pair_counter) + except Exception as err: + self._warn_redis(err) + self.counters.delete_cache(pair_counter) + + def _warn_redis(self, err: Exception) -> None: + verbose_proxy_logger.warning( + "Redis failed while counting Admin UI sign-in attempts; using this worker's own counters " + "until it recovers: %s", + err, + ) + + @staticmethod + def refuse(retry_after: int) -> NoReturn: + raise ProxyException( message="Too many failed sign-in attempts. Try again later.", type=ProxyErrorTypes.auth_error, - param=param, - code=429, - headers={"Retry-After": str(retry_after)}, # mutable-ok: ProxyException coerces header values + param="username", + code=status.HTTP_429_TOO_MANY_REQUESTS, + headers={"Retry-After": str(retry_after)}, ) - async def _refuse(self, key: str, scope: str, param: str, username: str, failures: int, limit: int) -> NoReturn: - retry_after: Final = await self._remaining_window(key) + +@dataclass(frozen=True, slots=True) +class LoginAttempt: + throttle: LoginThrottle + username: str + block: Block | None + + async def succeeded(self) -> None: + if not self.throttle.enabled: + return + await self.throttle.clear_pair(self.username) + + async def failed(self) -> None: + if not self.throttle.enabled: + return + if self.block is not None: + await _sleep(BLOCKED_ATTEMPT_HOLD_SECONDS) + self.throttle.refuse(max(self.block.retry_after - BLOCKED_ATTEMPT_HOLD_SECONDS, 1)) + user_block, source_block = await self.throttle.record_failure(self.username) + if user_block == 0 and source_block == 0: + return verbose_proxy_logger.warning( - "Admin UI sign-in attempts exhausted for %s; username=%s source=%s failures=%s limit=%s window=%ss", - scope, - self._loggable(username), - self.client_ip, - failures, - limit, - self.window_seconds, + "Admin UI sign-in blocked for %s seconds after too many failures; scope=%s username=%r source=%s", + user_block or source_block, + "user" if user_block else "source", + self.username, + self.throttle.client_ip, ) - raise self._refused(retry_after, param) - - async def raise_if_blocked(self, username: str) -> None: - """Refuse before the database lookup and before the invite-link password hash.""" - if not self.enabled: - return - username_key: Final = self._username_key(username) - source_key: Final = self._source_key() - username_failures: Final = await self._failures(self.username_cache, username_key) - if username_failures >= self.max_attempts: - await self._refuse( - username_key, "username", "max_failed_login_attempts", username, username_failures, self.max_attempts - ) - source_failures: Final = await self._failures(self.source_cache, source_key) - if source_failures >= self.max_attempts_per_source: - await self._refuse( - source_key, - "source address", - "max_failed_login_attempts_per_source", - username, - source_failures, - self.max_attempts_per_source, - ) - - async def _bump(self, store: DualCache, key: str) -> int: - shared: Final = await self._shared( - lambda redis_cache: redis_cache.async_increment_with_floor(key, 1, self.window_seconds) - ) - if shared is _UNAVAILABLE: - return _as_count( - await self._outcome(store.async_increment_cache(key=key, value=1, ttl=self.window_seconds)) - ) - return _as_count(shared) + _as_count(await self._outcome(store.async_get_cache(key=key))) - - async def record_failure(self, username: str) -> FailureCounts: - """Count one rejected credential guess against this username and against this source.""" - if not self.enabled: - return FailureCounts(username=0, source=0) - return FailureCounts( - username=await self._bump(self.username_cache, self._username_key(username)), - source=await self._bump(self.source_cache, self._source_key()), - ) - - @staticmethod - def delay_seconds(counts: FailureCounts) -> float: - """Seconds to hold a rejected attempt for, doubling per failure past whichever onset is further along.""" - steps: Final = min( - max(counts.username - USERNAME_DELAY_ONSET, counts.source - SOURCE_DELAY_ONSET), - _MAX_DELAY_DOUBLINGS, - ) - if steps < 0: - return 0.0 - return min(FIRST_DELAY_SECONDS * float(2**steps), MAX_DELAY_SECONDS) - - async def delay_for(self, username: str, counts: FailureCounts) -> None: - """Hold this rejected attempt open before answering it, so guessing costs wall-clock time. - - Only ever reached once the credentials are known to be wrong, so a valid password is - never delayed. Sources are capped at ``MAX_CONCURRENT_DELAYS_PER_SOURCE`` held - connections; over that, the attempt is refused immediately instead of parking a socket. - """ - if not self.enabled: - return - delay: Final = self.delay_seconds(counts) - if delay <= 0: - return - in_flight: Final = _DELAYS_IN_FLIGHT.get(self.client_ip, 0) - if in_flight >= MAX_CONCURRENT_DELAYS_PER_SOURCE: - verbose_proxy_logger.warning( - "Admin UI sign-in attempts held concurrently exhausted; username=%s source=%s in_flight=%s", - self._loggable(username), - self.client_ip, - in_flight, - ) - raise self._refused(int(MAX_DELAY_SECONDS), "concurrent_failed_logins") - _DELAYS_IN_FLIGHT[self.client_ip] = in_flight + 1 - try: - await _sleep(delay) - finally: - remaining: Final = _DELAYS_IN_FLIGHT.get(self.client_ip, 1) - 1 - if remaining > 0: - _DELAYS_IN_FLIGHT[self.client_ip] = remaining - else: - _DELAYS_IN_FLIGHT.pop(self.client_ip, None) - - async def clear(self, username: str) -> None: - """Drop the username counter after a successful sign-in. - - The source counter is left alone. It is shared by every account behind that address, - so one success there says nothing about the other attempts it is counting. - """ - if not self.enabled: - return - key: Final = self._username_key(username) - await self._shared(lambda redis_cache: redis_cache.async_delete_cache(key)) - await self._outcome(self.username_cache.async_delete_cache(key=key)) diff --git a/litellm/proxy/auth/login_utils.py b/litellm/proxy/auth/login_utils.py index b88a7d14ad8..6f52babb255 100644 --- a/litellm/proxy/auth/login_utils.py +++ b/litellm/proxy/auth/login_utils.py @@ -27,7 +27,7 @@ from litellm.proxy._types import ( UserAPIKeyAuth, ) from litellm.proxy.auth.auth_utils import is_sso_provider_fully_configured -from litellm.proxy.auth.login_throttle import LoginThrottle +from litellm.proxy.auth.login_throttle import LoginAttempt, LoginThrottle from litellm.proxy.management_endpoints.internal_user_endpoints import user_update from litellm.proxy.management_endpoints.key_management_endpoints import ( generate_key_helper_fn, @@ -219,9 +219,21 @@ async def authenticate_user( admin_credentials_match: Final = _admin_credentials_match(username, password, master_key, general_settings) - if not admin_credentials_match: - await throttle.raise_if_blocked(username) + async with throttle.attempt(username, exempt=admin_credentials_match) as attempt: + return await _sign_in( + username, password, master_key, prisma_client, attempt, general_settings, admin_credentials_match + ) + +async def _sign_in( + username: str, + password: str, + master_key: str, + prisma_client: PrismaClient | None, + attempt: LoginAttempt, + general_settings: Mapping[str, object], + admin_credentials_match: bool, +) -> LoginResult: # Check if we can find the `username` in the db. On the UI, users can enter username=their email _user_row: LiteLLM_UserTable | None = None user_role: ( @@ -315,7 +327,7 @@ async def authenticate_user( key = ExperimentalUIJWTToken.get_experimental_ui_login_jwt_auth_token(user_info) - await throttle.clear(username) + await attempt.succeeded() return LoginResult( user_id=user_id, @@ -372,7 +384,7 @@ async def authenticate_user( key = response["token"] - await throttle.clear(username) + await attempt.succeeded() return LoginResult( user_id=user_id, @@ -382,7 +394,7 @@ async def authenticate_user( login_method="username_password", ) else: - await throttle.delay_for(username, await throttle.record_failure(username)) + await attempt.failed() raise ProxyException( message=_invalid_credentials_message(general_settings), type=ProxyErrorTypes.auth_error, @@ -390,7 +402,7 @@ async def authenticate_user( code=401, ) else: - await throttle.delay_for(username, await throttle.record_failure(username)) + await attempt.failed() raise ProxyException( message=_invalid_credentials_message(general_settings), type=ProxyErrorTypes.auth_error, diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 43013e85e05..5b89beb38d2 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -325,7 +325,12 @@ from litellm.proxy.auth.auth_utils import ( from litellm.proxy.auth.fallback_model_access import router_fallback_access_check from litellm.proxy.auth.handle_jwt import JWTHandler from litellm.proxy.auth.litellm_license import AUTO_ROUTER_LICENSE_REMEDY, LicenseCheck -from litellm.proxy.auth.login_throttle import LoginThrottle, warn_login_counters_are_per_worker +from litellm.proxy.auth.login_throttle import ( + TRUSTED_PROXY_RANGES_KEY, + LoginThrottle, + warn_login_counters_are_per_worker, + warn_source_login_limit_is_off, +) from litellm.proxy.auth.model_checks import ( expand_wildcard_deployments_for_model_info, get_all_fallbacks, @@ -5803,11 +5808,10 @@ class ProxyConfig: if general_settings is None: general_settings = {} - ### FAILED-LOGIN ACCOUNTING MULTI-INSTANCE PREREQUISITE CHECK ### - # Failed Admin UI sign-in counters live in redis_usage_cache when available so a - # brute-force run is counted once across workers instead of once per worker. if os.getenv("NUM_WORKERS", "1") != "1" and redis_usage_cache is None: warn_login_counters_are_per_worker(os.getenv("NUM_WORKERS", "1")) + if not general_settings.get(TRUSTED_PROXY_RANGES_KEY): + warn_source_login_limit_is_off() _bg_hc_model_groups: Final = parse_background_health_check_model_groups(general_settings) _enable_hc_routing = False diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index c53476f5148..22fcedaa19d 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -25,33 +25,32 @@ class _RecordedSleeps: @pytest.fixture(autouse=True) def login_delays(monkeypatch): - """Replace the failed-login wait, so the suite pays no wall clock and can read it back.""" + """Replace the hold on a blocked wrong password, so the suite pays no wall clock and can read it back.""" from litellm.proxy.auth import login_throttle recorded = _RecordedSleeps() monkeypatch.setattr(login_throttle, "_sleep", recorded) - login_throttle._DELAYS_IN_FLIGHT.clear() + login_throttle._HELD_ATTEMPTS.clear() yield recorded - login_throttle._DELAYS_IN_FLIGHT.clear() + login_throttle._HELD_ATTEMPTS.clear() def _unlimited_throttle(): - """A throttle wired to a real in-memory store with a limit no test can reach.""" - from litellm.caching.dual_cache import DualCache + """A throttle wired to real in-memory stores with limits no test can reach.""" + from litellm.caching.in_memory_cache import InMemoryCache from litellm.proxy.auth.login_throttle import LoginThrottle - store: Final = DualCache() return LoginThrottle( client_ip="1.2.3.4", - max_attempts=10_000, - max_attempts_per_source=10_000, - window_seconds=900, - username_cache=store, - source_cache=store, + source_limit=None, + user_limit=10_000, + window_seconds=60, + block_seconds=300, + counters=InMemoryCache(), + blocks=InMemoryCache(), ) - from litellm.constants import LITELLM_PROXY_ADMIN_NAME from litellm.proxy._types import ( LiteLLM_UserTable, @@ -658,29 +657,37 @@ class TestEncodeUiSessionJwt: def _throttle( - max_attempts: int = 3, - window_seconds: int = 900, + user_limit: int = 2, + source_limit: int | None = None, + window_seconds: int = 60, + block_seconds: int = 300, client_ip: str = "1.2.3.4", - cache=None, + stores=None, redis_cache=None, - max_attempts_per_source: int = 10_000, ): - """A throttle over a real in-memory store, so the tests exercise the true counters.""" - from litellm.caching.dual_cache import DualCache + """A throttle over real in-memory stores, so the tests exercise the true counters and blocks.""" + from litellm.caching.in_memory_cache import InMemoryCache from litellm.proxy.auth.login_throttle import LoginThrottle - store: Final = cache if cache is not None else DualCache() + counters, blocks = stores if stores is not None else (InMemoryCache(), InMemoryCache()) return LoginThrottle( client_ip=client_ip, - max_attempts=max_attempts, - max_attempts_per_source=max_attempts_per_source, + source_limit=source_limit, + user_limit=user_limit, window_seconds=window_seconds, - username_cache=store, - source_cache=store, + block_seconds=block_seconds, + counters=counters, + blocks=blocks, redis_cache=redis_cache, ) +def _stores(): + from litellm.caching.in_memory_cache import InMemoryCache + + return InMemoryCache(), InMemoryCache() + + async def _guess(throttle, username: str = "admin", password: str = "wrong"): from litellm.proxy.auth.login_utils import authenticate_user @@ -693,97 +700,376 @@ async def _guess(throttle, username: str = "admin", password: str = "wrong"): ) +async def _fail(throttle, username: str = "admin") -> str: + """One wrong guess; returns the status code it was answered with.""" + from litellm.proxy._types import ProxyException + + with pytest.raises(ProxyException) as exc: + await _guess(throttle, username=username) + return exc.value.code + + +def _known_user(email: str = "known@example.com"): + user = MagicMock() + user.user_id = "u-1" + user.user_email = email + user.user_role = "internal_user" + user.password = "scrypt:stored" + repo = MagicMock() + repo.return_value.table.find_first = AsyncMock(return_value=user) + return repo + + +async def _db_login(throttle, username: str, password: str, *, correct: bool): + """A database user's sign-in with the stored hash faked, so no database or scrypt is needed.""" + from litellm.proxy.auth.login_utils import authenticate_user + + with ( + patch("litellm.proxy.auth.login_utils.UserRepository", _known_user(username)), + patch( # test-quality-ok: reaches the known-DB-user branch without a database + "litellm.proxy.auth.login_utils.verify_password", return_value=correct + ), + patch("litellm.proxy.auth.login_utils._rehash_password_if_needed", new=AsyncMock()), + patch( # test-quality-ok: success mints a UI key; faked so no DB is needed + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ), + ): + return await authenticate_user( + username=username, password=password, master_key="sk-master", prisma_client=MagicMock(), throttle=throttle + ) + + +def _local_count(throttle, key: str) -> int: + return int(throttle.counters.get_cache(key) or 0) + + @pytest.mark.asyncio -async def test_attempts_are_refused_once_the_limit_is_reached(monkeypatch): - """The limit denies further attempts for the window, and the denial carries Retry-After.""" +async def test_too_many_failures_for_one_username_block_that_pair_and_carry_retry_after(monkeypatch): + """One failure past the pair limit blocks the source for that username; the next wrong guess is held + and answered 429 with the block's remaining time, and the counter is not touched by blocked guesses.""" from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=3, window_seconds=77) + throttle = _throttle(user_limit=2, block_seconds=77) + keys = throttle._keys("admin") - for _ in range(3): - with pytest.raises(ProxyException) as first: - await _guess(throttle) - assert first.value.code == "401" + assert [await _fail(throttle) for _ in range(3)] == ["401", "401", "401"], "the limit itself is a plain 401" + assert throttle._local_block_ttl(keys.pair_block) == 77 with pytest.raises(ProxyException) as blocked: await _guess(throttle) assert blocked.value.code == "429" - assert blocked.value.headers.get("Retry-After") == "77" + assert blocked.value.headers.get("Retry-After") == "47", "the 30s hold is taken off the remaining block" + assert _local_count(throttle, keys.pair_counter) == 3, "a blocked guess is not counted again" @pytest.mark.asyncio -async def test_a_correct_admin_password_is_accepted_while_blocked(monkeypatch): - """The configured admin credentials are compared before the gate, so the operator gets in. - - A throttle that refuses a valid password hands anyone who can reach the login form a - denial of service against the one account that can fix it. - """ - from litellm.proxy._types import ProxyException +async def test_a_wrong_password_from_a_blocked_key_is_held_before_it_is_refused(monkeypatch, login_delays): + """The hold is the rate cap: a blocked key gets one verified guess per held slot per 30 seconds.""" + from litellm.proxy.auth.login_throttle import BLOCKED_ATTEMPT_HOLD_SECONDS + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(user_limit=1) + + assert [await _fail(throttle) for _ in range(2)] == ["401", "401"] + assert login_delays.seconds == [], "an unblocked wrong password is answered at once" + + assert await _fail(throttle) == "429" + assert login_delays.seconds == [BLOCKED_ATTEMPT_HOLD_SECONDS] + + +@pytest.mark.asyncio +async def test_a_correct_password_signs_in_while_its_pair_is_blocked(monkeypatch): + """The block is soft: the real user is still verified and gets in, so nobody can be locked out by + guessing at their account.""" monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") monkeypatch.setenv("DATABASE_URL", "postgresql://stub") - throttle = _throttle(max_attempts=2) + throttle = _throttle(user_limit=1) - for _ in range(2): - with pytest.raises(ProxyException): - await _guess(throttle) + assert [await _fail(throttle, username="user@corp.com") for _ in range(3)] == ["401", "401", "429"] - with pytest.raises(ProxyException) as still_blocked: - await _guess(throttle) - assert still_blocked.value.code == "429", "a wrong password is still refused" - - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed - "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) - ): - result = await _guess(throttle, password="right") + result = await _db_login(throttle, "user@corp.com", "right", correct=True) assert result.key == "sk-ui" @pytest.mark.asyncio -async def test_a_blocked_attempt_does_not_extend_the_window(monkeypatch): - """Hammering while blocked must not push the counter or refresh its TTL.""" - from litellm.proxy._types import ProxyException - +async def test_a_correct_password_signs_in_while_its_source_is_blocked(monkeypatch): + """Same for the source-wide block: it slows guessing from that address, it does not refuse a user.""" monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=2) - key = throttle._username_key("admin") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + throttle = _throttle(user_limit=100, source_limit=2) - for _ in range(2): - with pytest.raises(ProxyException): - await _guess(throttle) - counted_at_limit = await throttle._failures(throttle.username_cache, key) + for i in range(3): + assert await _fail(throttle, username=f"other-{i}@corp.com") == "401" + assert await _fail(throttle, username="other-9@corp.com") == "429", "the source is blocked for everyone" - for _ in range(5): - with pytest.raises(ProxyException): - await _guess(throttle) - - assert await throttle._failures(throttle.username_cache, key) == counted_at_limit == 2 + result = await _db_login(throttle, "user@corp.com", "right", correct=True) + assert result.key == "sk-ui" @pytest.mark.asyncio -async def test_a_successful_sign_in_clears_the_bucket(monkeypatch): - """Success resets the budget rather than leaving the operator near the limit.""" - from litellm.proxy._types import ProxyException +async def test_a_successful_sign_in_clears_the_pair_counter_but_not_the_source_counter(monkeypatch): + """One account's success says nothing about the other guesses the address is making.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + throttle = _throttle(user_limit=5, source_limit=50) + keys = throttle._keys("user@corp.com") + + for _ in range(2): + assert await _fail(throttle, username="user@corp.com") == "401" + assert _local_count(throttle, keys.pair_counter) == 2 + assert _local_count(throttle, keys.source_counter) == 2 + + await _db_login(throttle, "user@corp.com", "right", correct=True) + + assert _local_count(throttle, keys.pair_counter) == 0 + assert _local_count(throttle, keys.source_counter) == 2 + + +@pytest.mark.asyncio +async def test_once_a_pair_is_blocked_its_failures_stop_counting_against_the_source(monkeypatch): + """A script stuck on one account trips the pair block and then leaves the office's shared address alone.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(user_limit=2, source_limit=4) + keys = throttle._keys("stuck-script@corp.com") + + assert [await _fail(throttle, username="stuck-script@corp.com") for _ in range(3)] == ["401"] * 3 + assert _local_count(throttle, keys.source_counter) == 2, "failures before the pair block count for the source" + + for _ in range(5): + assert await _fail(throttle, username="stuck-script@corp.com") == "429" + assert _local_count(throttle, keys.source_counter) == 2, "blocked-pair failures must not reach the source" + + assert await _fail(throttle, username="colleague@corp.com") == "401", "a colleague still signs in normally" + assert throttle._local_block_ttl(keys.source_block) == 0 + + +@pytest.mark.asyncio +async def test_the_blocking_failure_itself_does_not_count_against_the_source(monkeypatch): + """The guess that installs the pair block is the first one that stops counting, so a pair limit of B + costs the source exactly B, not B plus one.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(user_limit=2, source_limit=2) + keys = throttle._keys("stuck@corp.com") + + assert [await _fail(throttle, username="stuck@corp.com") for _ in range(3)] == ["401", "401", "401"] + + assert _local_count(throttle, keys.source_counter) == 2 + assert throttle._local_block_ttl(keys.source_block) == 0, "the third guess blocked the pair, not the source" + + +@pytest.mark.asyncio +async def test_too_many_failures_across_usernames_block_the_whole_source(monkeypatch): + """A spray of one guess per username never trips a pair; the source counter is what stops it.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(user_limit=5, source_limit=3, block_seconds=200) + + assert [await _fail(throttle, username=f"sprayed-{i}@corp.com") for i in range(4)] == ["401"] * 4 + + assert await _fail(throttle, username="sprayed-99@corp.com") == "429" + assert throttle._local_block_ttl(throttle._keys("x").source_block) == 200 + + +@pytest.mark.asyncio +async def test_without_trusted_proxy_ranges_the_source_scope_is_off(monkeypatch): + """Behind an ingress every client shares the peer address, so a source-wide block would block them all. + The pair scope still applies.""" + from litellm.proxy.auth.login_throttle import LoginThrottle + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + request = MagicMock() + request.headers = {"x-forwarded-for": "203.0.113.9"} + request.client = MagicMock() + request.client.host = "10.0.0.1" + throttle = LoginThrottle.from_request( + request, general_settings={"max_failed_login_attempts_per_source": 1}, redis_cache=None + ) + + assert throttle.source_limit is None + assert throttle.client_ip == "10.0.0.1", "the header is not trusted without a configured proxy range" + assert [await _fail(throttle, username=f"user-{i}@corp.com") for i in range(6)] == ["401"] * 6 + + +@pytest.mark.asyncio +async def test_with_trusted_proxy_ranges_the_source_is_the_forwarded_client(monkeypatch): + """The header is walked right to left past the trusted hops, so a forged left-most entry cannot pick the bucket.""" + from litellm.proxy.auth.login_throttle import LoginThrottle + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + settings = {"trusted_proxy_ranges": ["10.0.0.0/8"], "max_failed_login_attempts_per_source": 2} + + def _from(peer: str, forwarded: str): + request = MagicMock() + request.headers = {"x-forwarded-for": forwarded} + request.client = MagicMock() + request.client.host = peer + return LoginThrottle.from_request(request, general_settings=settings, redis_cache=None) + + via_proxy = _from("10.0.0.1", "1.1.1.1, 203.0.113.9, 10.0.0.2") + assert via_proxy.client_ip == "203.0.113.9" + assert via_proxy.source_limit == 2 + + direct = _from("198.51.100.7", "203.0.113.9") + assert direct.client_ip == "198.51.100.7", "a peer outside the trusted ranges cannot forward anything" + + +def test_source_overrides_pick_the_most_specific_matching_range(): + """An exact address beats a /16 beats a /8; an address in none of them keeps the default.""" + from litellm.proxy.auth.login_throttle import LoginThrottle + + settings = { + "trusted_proxy_ranges": ["10.0.0.0/8"], + "max_failed_login_attempts_per_source": 7, + "max_failed_login_attempts_per_source_overrides": { + "203.0.0.0/8": 100, + "203.0.113.0/24": 200, + "203.0.113.9": 300, + "not-an-address": 999, + "198.51.100.0/24": "not-a-number", + }, + } + + def _limit(client: str) -> int | None: + request = MagicMock() + request.headers = {"x-forwarded-for": client} + request.client = MagicMock() + request.client.host = "10.0.0.1" + return LoginThrottle.from_request(request, general_settings=settings, redis_cache=None).source_limit + + assert _limit("203.0.113.9") == 300 + assert _limit("203.0.113.10") == 200 + assert _limit("203.0.1.1") == 100 + assert _limit("192.0.2.1") == 7 + assert _limit("198.51.100.1") == 7, "a garbage limit falls back to the default rather than a huge or zero budget" + + +def test_ipv6_sources_are_grouped_by_their_64_bit_prefix(): + """A /64 holder has 2^64 addresses; counting each one separately would hand them unlimited fresh buckets.""" + from litellm.proxy.auth.login_throttle import source_group + + assert source_group("2001:db8:1:2::1") == source_group("2001:db8:1:2:ffff:ffff:ffff:ffff") == "2001:db8:1:2::/64" + assert source_group("2001:db8:1:3::1") != source_group("2001:db8:1:2::1") + assert source_group("::ffff:203.0.113.9") == source_group("203.0.113.9") == "203.0.113.9" + assert source_group("unknown") == "unknown" + + +@pytest.mark.asyncio +async def test_two_ipv6_addresses_in_one_64_share_the_source_budget(monkeypatch): + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + stores = _stores() + first = _throttle(user_limit=50, source_limit=2, client_ip="2001:db8:1:2::1", stores=stores) + second = _throttle(user_limit=50, source_limit=2, client_ip="2001:db8:1:2::2", stores=stores) + + assert [await _fail(first, username=f"a-{i}@corp.com") for i in range(3)] == ["401"] * 3 + assert await _fail(second, username="b@corp.com") == "429" + + +@pytest.mark.asyncio +async def test_one_source_being_blocked_does_not_touch_another(monkeypatch): + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + stores = _stores() + attacker = _throttle(user_limit=50, source_limit=2, client_ip="203.0.113.9", stores=stores) + neighbour = _throttle(user_limit=50, source_limit=2, client_ip="198.51.100.7", stores=stores) + + assert [await _fail(attacker, username=f"t-{i}@corp.com") for i in range(3)] == ["401"] * 3 + assert await _fail(attacker, username="t-9@corp.com") == "429" + assert await _fail(neighbour, username="t-9@corp.com") == "401" + + +@pytest.mark.asyncio +async def test_the_same_username_from_another_source_has_its_own_budget(monkeypatch): + """The pair carries the address on purpose: an attacker elsewhere cannot lock a user out of their own office.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + stores = _stores() + attacker = _throttle(user_limit=1, client_ip="203.0.113.9", stores=stores) + office = _throttle(user_limit=1, client_ip="198.51.100.7", stores=stores) + + assert [await _fail(attacker, username="victim@corp.com") for _ in range(3)] == ["401", "401", "429"] + assert await _fail(office, username="victim@corp.com") == "401" + + +@pytest.mark.asyncio +async def test_the_counting_window_is_anchored_at_the_first_failure(monkeypatch): + """Later failures must not push the expiry out, or a slow guesser keeps their own count alive forever.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(user_limit=50, window_seconds=60) + key = throttle._keys("admin").pair_counter + + await _fail(throttle) + first_expiry = throttle.counters.ttl_dict[key] + for _ in range(3): + await _fail(throttle) + + assert throttle.counters.ttl_dict[key] == first_expiry + + +@pytest.mark.asyncio +async def test_the_block_outlives_the_counting_window(monkeypatch): + """Counters expire after the window and blocks after the block time; the two are separate keys.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(user_limit=1, window_seconds=10, block_seconds=300) + keys = throttle._keys("admin") + + assert [await _fail(throttle) for _ in range(2)] == ["401", "401"] + + throttle.counters.delete_cache(keys.pair_counter) + + assert await _fail(throttle) == "429", "an expired counter must not lift an active block" + assert 290 <= throttle._local_block_ttl(keys.pair_block) <= 300 + + +@pytest.mark.asyncio +async def test_the_block_time_is_fixed_and_not_refreshed_by_blocked_guesses(monkeypatch): + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + throttle = _throttle(user_limit=1, block_seconds=300) + key = throttle._keys("admin").pair_block + + assert [await _fail(throttle) for _ in range(2)] == ["401", "401"] + installed_at = throttle.blocks.ttl_dict[key] + + for _ in range(4): + assert await _fail(throttle) == "429" + + assert throttle.blocks.ttl_dict[key] == installed_at + + +@pytest.mark.asyncio +async def test_the_configured_admin_credentials_bypass_the_throttle(monkeypatch): + """The only account that can fix a misconfiguration is exempt: no hold, no slot, even while blocked.""" + from litellm.proxy.auth import login_throttle as lt monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") monkeypatch.setenv("DATABASE_URL", "postgresql://stub") - throttle = _throttle(max_attempts=3) + throttle = _throttle(user_limit=1) - for _ in range(2): - with pytest.raises(ProxyException): - await _guess(throttle) + assert [await _fail(throttle) for _ in range(3)] == ["401", "401", "429"] - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed - "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + with ( + patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), + patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ), ): - await _guess(throttle, password="right") - - assert await throttle._failures(throttle.username_cache, throttle._username_key("admin")) == 0 + result = await _guess(throttle, password="right") + assert result.key == "sk-ui" + assert lt._HELD_ATTEMPTS == {} @pytest.mark.asyncio @@ -791,7 +1077,7 @@ async def test_a_configuration_error_never_counts(monkeypatch): """A 500 from an unset master key is not a guess and must not consume the budget.""" from litellm.proxy._types import ProxyException - throttle = _throttle(max_attempts=2) + throttle = _throttle(user_limit=2) for _ in range(5): with pytest.raises(ProxyException) as exc: await authenticate_user( @@ -799,44 +1085,20 @@ async def test_a_configuration_error_never_counts(monkeypatch): ) assert exc.value.code == "500" - assert await throttle._failures(throttle.username_cache, throttle._username_key("admin")) == 0 + assert _local_count(throttle, throttle._keys("admin").pair_counter) == 0 @pytest.mark.asyncio -async def test_the_username_is_case_folded_into_one_bucket(monkeypatch): +async def test_the_username_is_case_folded_into_one_pair(monkeypatch): """The DB lookup is case-insensitive, so casing must not multiply the budget.""" - from litellm.proxy._types import ProxyException - monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=4) + throttle = _throttle(user_limit=4) - for name in ("admin@corp.com", "ADMIN@corp.com", "Admin@corp.com", "aDmIn@corp.com"): - with pytest.raises(ProxyException) as exc: - await _guess(throttle, username=name) - assert exc.value.code == "401" + for name in ("admin@corp.com", "ADMIN@corp.com", "Admin@corp.com", "aDmIn@corp.com", "admin@CORP.com"): + assert await _fail(throttle, username=name) == "401" - with pytest.raises(ProxyException) as blocked: - await _guess(throttle, username="admin@CORP.com") - assert blocked.value.code == "429" - - -@pytest.mark.asyncio -async def test_a_different_username_from_the_same_source_is_unaffected(monkeypatch): - """The counters are independent, so one username's failures do not exhaust another's.""" - from litellm.proxy._types import ProxyException - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=2) - - for _ in range(3): - with pytest.raises(ProxyException): - await _guess(throttle, username="admin") - - with pytest.raises(ProxyException) as other: - await _guess(throttle, username="someone-else@example.com") - assert other.value.code == "401", "a second username must still reach the credential check" + assert await _fail(throttle, username="admin@Corp.com") == "429" @pytest.mark.asyncio @@ -848,26 +1110,9 @@ async def test_both_credential_rejections_are_indistinguishable(monkeypatch): monkeypatch.setenv("UI_PASSWORD", "right") with pytest.raises(ProxyException) as unknown: - await _guess(_throttle(max_attempts=99), username="nobody@example.com") - - fake_user = MagicMock() - fake_user.user_id = "u-1" - fake_user.user_email = "known@example.com" - fake_user.user_role = "internal_user" - fake_user.password = "scrypt:fake" - repo = MagicMock() - repo.return_value.table.find_first = AsyncMock(return_value=fake_user) - with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( # test-quality-ok: reaches the known-DB-user branch without a database - "litellm.proxy.auth.login_utils.verify_password", return_value=False - ): - with pytest.raises(ProxyException) as known: - await authenticate_user( - username="known@example.com", - password="wrong", - master_key="sk-master", - prisma_client=MagicMock(), - throttle=_throttle(max_attempts=99), - ) + await _guess(_throttle(user_limit=99), username="nobody@example.com") + with pytest.raises(ProxyException) as known: + await _db_login(_throttle(user_limit=99), "known@example.com", "wrong", correct=False) assert unknown.value.message == known.value.message assert "known@example.com" not in unknown.value.message + known.value.message @@ -875,13 +1120,12 @@ async def test_both_credential_rejections_are_indistinguishable(monkeypatch): @pytest.mark.asyncio async def test_a_user_with_no_password_set_does_not_consume_the_budget(monkeypatch): - """That 401 is deterministic and guards no secret, so counting it would only let - someone burn a passwordless account's bucket.""" + """That 401 is deterministic and guards no secret, so counting it would only let someone burn the pair.""" from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=2) + throttle = _throttle(user_limit=2) passwordless = MagicMock() passwordless.user_id = "u-2" @@ -891,7 +1135,9 @@ async def test_a_user_with_no_password_set_does_not_consume_the_budget(monkeypat repo = MagicMock() repo.return_value.table.find_first = AsyncMock(return_value=passwordless) - with patch("litellm.proxy.auth.login_utils.UserRepository", repo): # test-quality-ok: reaches the passwordless-DB-user branch without a database + with patch( + "litellm.proxy.auth.login_utils.UserRepository", repo + ): # test-quality-ok: reaches the passwordless-DB-user branch without a database for _ in range(5): with pytest.raises(ProxyException) as exc: await authenticate_user( @@ -903,185 +1149,36 @@ async def test_a_user_with_no_password_set_does_not_consume_the_budget(monkeypat ) assert exc.value.code == "401" - assert await throttle._failures(throttle.username_cache, throttle._username_key("nopass@example.com")) == 0 + assert _local_count(throttle, throttle._keys("nopass@example.com").pair_counter) == 0 @pytest.mark.asyncio async def test_a_wrong_password_for_a_known_user_also_counts(monkeypatch): - """The database-user branch must charge the bucket too, not just the unknown-user branch.""" + """The database-user branch must charge the pair too, not just the unknown-user branch.""" from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=3) + throttle = _throttle(user_limit=2) - known = MagicMock() - known.user_id = "u-1" - known.user_email = "known@example.com" - known.user_role = "internal_user" - known.password = "scrypt:stored" - repo = MagicMock() - repo.return_value.table.find_first = AsyncMock(return_value=known) - - async def _attempt(): - return await authenticate_user( - username="known@example.com", - password="wrong", - master_key="sk-master", - prisma_client=MagicMock(), - throttle=throttle, - ) - - with patch("litellm.proxy.auth.login_utils.UserRepository", repo), patch( # test-quality-ok: reaches the known-DB-user branch without a database - "litellm.proxy.auth.login_utils.verify_password", return_value=False - ): - for _ in range(3): - with pytest.raises(ProxyException) as rejected: - await _attempt() - assert rejected.value.code == "401" - - with pytest.raises(ProxyException) as blocked: - await _attempt() - assert blocked.value.code == "429" - - -@pytest.mark.asyncio -async def test_one_source_exhausting_its_own_budget_does_not_refuse_another_source(monkeypatch): - """The source counter is per address, so a noisy office does not take its neighbour down.""" - from litellm.caching.dual_cache import DualCache - from litellm.proxy._types import ProxyException - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - shared_store = DualCache() - attacker = _throttle( - max_attempts=10_000, max_attempts_per_source=2, client_ip="203.0.113.9", cache=shared_store - ) - operator = _throttle( - max_attempts=10_000, max_attempts_per_source=2, client_ip="198.51.100.7", cache=shared_store - ) - - for i in range(2): - with pytest.raises(ProxyException): - await _guess(attacker, username=f"target-{i}@corp.com") - - with pytest.raises(ProxyException) as blocked: - await _guess(attacker, username="target-2@corp.com") - assert blocked.value.code == "429" - - with pytest.raises(ProxyException) as unaffected: - await _guess(operator, username="target-3@corp.com") - assert unaffected.value.code == "401", "the other address must still reach the credential check" - - -@pytest.mark.asyncio -async def test_a_username_exhausted_from_one_source_is_refused_from_another(monkeypatch): - """The username counter carries no address, so spreading the guesses buys nothing. - - The pair key this replaced reset the budget for every new address, which is exactly the - shape of a credential-stuffing run from a proxy pool. - """ - from litellm.caching.dual_cache import DualCache - from litellm.proxy._types import ProxyException - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - shared_store = DualCache() - first_hop = _throttle(max_attempts=2, client_ip="203.0.113.9", cache=shared_store) - second_hop = _throttle(max_attempts=2, client_ip="198.51.100.7", cache=shared_store) - - for _ in range(2): - with pytest.raises(ProxyException): - await _guess(first_hop, username="victim@corp.com") - - with pytest.raises(ProxyException) as rotated: - await _guess(second_hop, username="victim@corp.com") - assert rotated.value.code == "429" - - -@pytest.mark.asyncio -async def test_a_source_wide_spray_is_counted_even_though_each_username_is_fresh(monkeypatch): - """One guess against each of many usernames never trips a username counter, only the source one.""" - from litellm.proxy._types import ProxyException - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=10_000, max_attempts_per_source=6, client_ip="203.0.113.11") - - for i in range(6): + for _ in range(3): with pytest.raises(ProxyException) as rejected: - await _guess(throttle, username=f"sprayed-{i}@corp.com") + await _db_login(throttle, "known@example.com", "wrong", correct=False) assert rejected.value.code == "401" with pytest.raises(ProxyException) as blocked: - await _guess(throttle, username="sprayed-7@corp.com") + await _db_login(throttle, "known@example.com", "wrong", correct=False) assert blocked.value.code == "429" - assert await throttle._failures(throttle.username_cache, throttle._username_key("sprayed-7@corp.com")) == 0 @pytest.mark.asyncio -async def test_a_successful_sign_in_leaves_the_source_counter_alone(monkeypatch): - """One account's success says nothing about the other attempts the address is making.""" - from litellm.proxy._types import ProxyException - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - monkeypatch.setenv("DATABASE_URL", "postgresql://stub") - throttle = _throttle(max_attempts=10) - - for _ in range(2): - with pytest.raises(ProxyException): - await _guess(throttle) - - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed - "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) - ): - await _guess(throttle, password="right") - - assert await throttle._failures(throttle.username_cache, throttle._username_key("admin")) == 0 - assert await throttle._failures(throttle.source_cache, throttle._source_key()) == 2 - - -@pytest.mark.asyncio -async def test_the_delay_doubles_from_one_second_and_is_capped(monkeypatch, login_delays): - """Guessing has to cost wall clock, and the cost has to stop short of an unbounded hang.""" - from litellm.proxy._types import ProxyException - from litellm.proxy.auth.login_throttle import MAX_DELAY_SECONDS - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=10_000) - - for _ in range(9): - with pytest.raises(ProxyException): - await _guess(throttle) - - assert login_delays.seconds == [1.0, 2.0, 4.0, 8.0, 16.0, 30.0, 30.0], ( - "the first two failures answer immediately, then the wait doubles up to the cap" - ) - assert max(login_delays.seconds) == MAX_DELAY_SECONDS - - -@pytest.mark.asyncio -async def test_the_delay_tracks_whichever_counter_is_further_past_its_onset(monkeypatch): - """A source deep into a spray must not be answered instantly just because the username is fresh.""" - from litellm.proxy.auth.login_throttle import FailureCounts, LoginThrottle - - assert LoginThrottle.delay_seconds(FailureCounts(username=1, source=1)) == 0.0 - assert LoginThrottle.delay_seconds(FailureCounts(username=2, source=24)) == 0.0 - assert LoginThrottle.delay_seconds(FailureCounts(username=3, source=1)) == 1.0 - assert LoginThrottle.delay_seconds(FailureCounts(username=1, source=25)) == 1.0 - assert LoginThrottle.delay_seconds(FailureCounts(username=4, source=28)) == 8.0 - - -@pytest.mark.asyncio -async def test_held_attempts_from_one_source_are_capped(monkeypatch): - """Holding a rejected attempt open must not let one address park unlimited sockets.""" +async def test_held_attempts_from_one_blocked_key_are_capped(monkeypatch): + """Holding a wrong guess open must not let one blocked key park unlimited sockets in password checks.""" import asyncio from litellm.proxy._types import ProxyException from litellm.proxy.auth import login_throttle as lt - from litellm.proxy.auth.login_throttle import MAX_CONCURRENT_DELAYS_PER_SOURCE + from litellm.proxy.auth.login_throttle import MAX_HELD_ATTEMPTS_PER_KEY monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") @@ -1091,134 +1188,102 @@ async def test_held_attempts_from_one_source_are_capped(monkeypatch): await release.wait() monkeypatch.setattr(lt, "_sleep", _park) - throttle = _throttle(max_attempts=10_000, client_ip="203.0.113.44") - await throttle.record_failure("admin") - await throttle.record_failure("admin") + throttle = _throttle(user_limit=1, client_ip="203.0.113.44") + slot = throttle._keys("admin").pair_block + assert [await _fail(throttle) for _ in range(2)] == ["401", "401"] - held = [asyncio.create_task(_guess(throttle)) for _ in range(MAX_CONCURRENT_DELAYS_PER_SOURCE)] + held = [asyncio.create_task(_guess(throttle)) for _ in range(MAX_HELD_ATTEMPTS_PER_KEY)] for _ in range(1000): - if lt._DELAYS_IN_FLIGHT.get("203.0.113.44") == MAX_CONCURRENT_DELAYS_PER_SOURCE: + if lt._HELD_ATTEMPTS.get(slot) == MAX_HELD_ATTEMPTS_PER_KEY: break await asyncio.sleep(0) - assert lt._DELAYS_IN_FLIGHT.get("203.0.113.44") == MAX_CONCURRENT_DELAYS_PER_SOURCE + assert lt._HELD_ATTEMPTS.get(slot) == MAX_HELD_ATTEMPTS_PER_KEY try: with pytest.raises(ProxyException) as over_cap: await _guess(throttle) assert over_cap.value.code == "429" assert over_cap.value.headers.get("Retry-After") == "30" + assert await _fail(throttle, username="someone-else@corp.com") == "401", "other keys are not affected" finally: release.set() for task in held: with pytest.raises(ProxyException): await task - with pytest.raises(ProxyException) as after_drain: - await _guess(throttle) - assert after_drain.value.code == "401", "the cap must release once the held attempts answer" + assert lt._HELD_ATTEMPTS.get(slot) is None, "the slots are released once the held attempts answer" @pytest.mark.asyncio -async def test_disabling_the_control_removes_the_delay_as_well(monkeypatch, login_delays): +async def test_disabling_the_control_removes_the_hold_as_well(monkeypatch, login_delays): """The escape hatch has to turn off the whole control, not only the refusal.""" import dataclasses - from litellm.proxy._types import ProxyException - monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = dataclasses.replace(_throttle(max_attempts=2), enabled=False) - - for _ in range(6): - with pytest.raises(ProxyException) as rejected: - await _guess(throttle) - assert rejected.value.code == "401" + throttle = dataclasses.replace(_throttle(user_limit=1), enabled=False) + assert [await _fail(throttle) for _ in range(6)] == ["401"] * 6 assert login_delays.seconds == [] class _FakeRedis: - """Redis whose only counter write is the atomic INCRBY-plus-EXPIRE Lua call. + """Redis whose only writes are the throttle's two scripts, run atomically as one call each. - `async_increment` is deliberately absent: a two-step increment would fail the test - with AttributeError, because Redis could then commit a count without its expiry. + Mirrors the Lua: a blocked key returns its remaining block time and is not counted; a counter + is expired on first write; one over the limit installs the block; a blocked pair stops the + source from being counted. The real scripts are exercised against a live Redis in the PR's + proof, this fake only has to be faithful enough for the worker-sharing tests. """ def __init__(self): self.values: dict = {} self.ttls: dict = {} + self.scripts: list[str] = [] - async def async_get_cache(self, key, **kwargs): - return self.values.get(key) + def async_register_script(self, script: str): + from litellm.proxy.auth import login_throttle as lt - async def async_batch_get_counts(self, key_list): - return tuple(self.values.get(key) for key in key_list) + async def _run(keys, args): + self.scripts.append(script) + if script == lt._BLOCK_TTLS_LUA: + return [self._ttl(keys[1]), self._ttl(keys[3])] + assert script == lt._RECORD_FAILURE_LUA + user_limit, source_limit, window, block = (int(a) for a in args) + user_block = self._bump(keys[0], keys[1], user_limit, window, block) + if source_limit > 0 and user_block == 0: + return [user_block, self._bump(keys[2], keys[3], source_limit, window, block)] + return [user_block, 0] - async def async_increment_with_floor(self, key, value, ttl): - self.values[key] = self.values.get(key, 0) + value - self.ttls.setdefault(key, ttl) - return self.values[key] + return _run - async def async_get_ttl(self, key): - return self.ttls.get(key) + def _ttl(self, key: str) -> int: + return self.ttls.get(key, -2) if key in self.values else -2 + + def _bump(self, count_key: str, block_key: str, limit: int, window: int, block: int) -> int: + if self._ttl(block_key) > 0: + return self._ttl(block_key) + self.values[count_key] = self.values.get(count_key, 0) + 1 + self.ttls.setdefault(count_key, window) + if self.values[count_key] > limit: + self.values[block_key] = 1 + self.ttls[block_key] = block + return block + return 0 async def async_delete_cache(self, key): self.values.pop(key, None) self.ttls.pop(key, None) - def persist(self): - self.ttls.clear() - - -@pytest.mark.asyncio -async def test_counters_are_written_with_their_expiry_and_re_armed_if_stripped(monkeypatch): - """Regression: a counter with no TTL would refuse the pair forever. - - Nothing increments a key once the limit is reached, so a counter that ever exists - without an expiry stays refused with no way back. Every write must therefore carry the - expiry, and a refusal that finds it stripped (PERSIST) must put the window back. - """ - from litellm.proxy._types import ProxyException - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - redis = _FakeRedis() - throttle = _throttle(max_attempts=2, window_seconds=77, redis_cache=redis) - - for _ in range(2): - with pytest.raises(ProxyException): - await _guess(throttle) - - assert redis.values, "failures must land in the shared counter" - assert set(redis.ttls) == set(redis.values), "no counter may exist without its expiry" - assert set(redis.ttls.values()) == {77} - - redis.persist() - with pytest.raises(ProxyException) as blocked: - await _guess(throttle) - assert blocked.value.code == "429" - assert blocked.value.headers.get("Retry-After") == "77" - assert set(redis.ttls) >= {k for k in redis.values if ":user:" in k}, "the refusal must re-arm a stripped expiry" - class _DownRedis(_FakeRedis): - """Redis whose every call fails, as during an outage or an open circuit breaker. + """Redis whose every call fails, as during an outage or an open circuit breaker.""" - `async_get_cache` returns None rather than raising, as the real one does: it swallows the - error, so a failed GET is indistinguishable from an empty key to anyone reading through it. - """ + def async_register_script(self, script: str): + async def _run(keys, args): + raise ConnectionError("redis is down") - async def async_get_cache(self, key, **kwargs): - return None - - async def async_batch_get_counts(self, key_list): - raise ConnectionError("redis is down") - - async def async_increment_with_floor(self, key, value, ttl): - raise ConnectionError("redis is down") - - async def async_get_ttl(self, key): - raise ConnectionError("redis is down") + return _run async def async_delete_cache(self, key): raise ConnectionError("redis is down") @@ -1226,124 +1291,84 @@ class _DownRedis(_FakeRedis): @pytest.mark.asyncio async def test_redis_is_the_only_counter_while_it_answers(monkeypatch): - """Regression: every worker must spend the same budget, and a success must clear it for all. - - Counting in this worker's memory as well as in Redis let the two drift apart: a worker - whose Redis write failed kept its own count while the others gave the attacker fresh - guesses, and a stale local count outlived the shared clear after a correct password. - """ - from litellm.caching.dual_cache import DualCache - from litellm.proxy._types import ProxyException + """Every worker must spend the same budget, see the same block, and a success must clear the pair for all.""" from litellm.proxy.auth.login_throttle import _CACHE_KEY_PREFIX monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") redis = _FakeRedis() - first_worker_store = DualCache() - second_worker_store = DualCache() - first_worker = _throttle(max_attempts=2, cache=first_worker_store, redis_cache=redis) - second_worker = _throttle(max_attempts=2, cache=second_worker_store, redis_cache=redis) + first_worker = _throttle(user_limit=2, stores=_stores(), redis_cache=redis) + second_worker = _throttle(user_limit=2, stores=_stores(), redis_cache=redis) - for _ in range(2): - with pytest.raises(ProxyException, match="Invalid credentials"): - await _guess(first_worker) - - assert not [k for k in first_worker_store.in_memory_cache.cache_dict if str(k).startswith(_CACHE_KEY_PREFIX)], ( + assert [await _fail(first_worker, username="user@corp.com") for _ in range(3)] == ["401"] * 3 + assert not [k for k in first_worker.counters.cache_dict if str(k).startswith(_CACHE_KEY_PREFIX)], ( "with Redis answering, no worker may keep a counter of its own" ) - with pytest.raises(ProxyException) as blocked: - await _guess(second_worker) - assert blocked.value.code == "429", "the second worker must see the budget the first one spent" + assert not first_worker.blocks.cache_dict - monkeypatch.setenv("DATABASE_URL", "postgresql://stub") - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed - "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) - ): - await _guess(second_worker, password="right") + assert await _fail(second_worker, username="user@corp.com") == "429", "the second worker sees the block" - assert not [k for k in redis.values if ":user:" in k], "a success must clear the shared username counter" - with pytest.raises(ProxyException, match="Invalid credentials"): - await _guess(first_worker) + await _db_login(second_worker, "user@corp.com", "right", correct=True) + + assert not [k for k in redis.values if ":user:" in k and ":block:" not in k], ( + "success clears the shared pair counter" + ) + assert [k for k in redis.values if ":block:user:" in k], "an active block is not lifted by one success" @pytest.mark.asyncio async def test_a_redis_outage_falls_back_to_this_workers_own_counter(monkeypatch): - """With Redis raising, guesses are still counted and refused, per worker, instead of unbounded.""" + """With Redis raising, guesses are still counted and blocked per worker, with a warning, instead of unbounded.""" + import logging + + from litellm._logging import verbose_proxy_logger from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=2, redis_cache=_DownRedis()) + throttle = _throttle(user_limit=2, block_seconds=300, redis_cache=_DownRedis()) - for _ in range(2): - with pytest.raises(ProxyException, match="Invalid credentials"): + records: list[logging.LogRecord] = [] + handler = logging.Handler() + handler.emit = records.append + verbose_proxy_logger.addHandler(handler) + try: + assert [await _fail(throttle) for _ in range(3)] == ["401"] * 3 + with pytest.raises(ProxyException) as blocked: await _guess(throttle) + finally: + verbose_proxy_logger.removeHandler(handler) - with pytest.raises(ProxyException) as blocked: - await _guess(throttle) assert blocked.value.code == "429" - assert blocked.value.headers.get("Retry-After") == "900" - - -class _WriteRefusingRedis(_FakeRedis): - """Redis that answers reads but raises on writes until `recover()` is called.""" - - def __init__(self): - super().__init__() - self.writable = False - - def recover(self): - self.writable = True - - async def async_increment_with_floor(self, key, value, ttl): - if not self.writable: - raise ConnectionError("redis write failed") - return await super().async_increment_with_floor(key, value, ttl) + assert blocked.value.headers.get("Retry-After") == "270" + assert any("Redis failed while counting Admin UI sign-in attempts" in r.getMessage() for r in records) @pytest.mark.asyncio -async def test_failures_redis_refused_still_count_once_redis_recovers(monkeypatch): - """Regression: a guess Redis could not record must not be forgotten when Redis comes back. - - Such a guess lands in this worker's own store. Reading only Redis afterwards handed the - attacker that guess again, so the budget was the limit plus however many writes failed. - """ - from litellm.proxy._types import ProxyException - +async def test_a_failed_redis_delete_still_clears_this_workers_counter(monkeypatch): + """The fail-open tradeoff: when Redis cannot clear the pair, the worker clears what it holds and moves on.""" monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - redis = _WriteRefusingRedis() - throttle = _throttle(max_attempts=2, redis_cache=redis) + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + throttle = _throttle(user_limit=5, redis_cache=_DownRedis()) + key = throttle._keys("user@corp.com").pair_counter - with pytest.raises(ProxyException, match="Invalid credentials"): - await _guess(throttle) - assert not redis.values, "the refused write must not have reached Redis" + assert [await _fail(throttle, username="user@corp.com") for _ in range(2)] == ["401", "401"] + assert _local_count(throttle, key) == 2 - redis.recover() - with pytest.raises(ProxyException, match="Invalid credentials"): - await _guess(throttle) - assert [v for k, v in redis.values.items() if ":user:" in k] == [1], "only the recorded guess is in Redis" - - with pytest.raises(ProxyException) as blocked: - await _guess(throttle) - assert blocked.value.code == "429", "the guess Redis missed and the one it took must add up to the limit" + await _db_login(throttle, "user@corp.com", "right", correct=True) + assert _local_count(throttle, key) == 0 @pytest.mark.asyncio async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): - """Regression: throttle entries must not evict cached credentials. - - user_api_key_cache holds at most 200 in-memory entries and evicts the soonest to - expire first, so parking 900s sign-in counters there let a stream of made-up usernames - push out the much shorter lived credential entries, sending every ordinary API request - back to the database. - """ + """Regression: throttle entries must not evict cached credentials from user_api_key_cache.""" from litellm.proxy import proxy_server as ps from litellm.proxy.auth.login_throttle import _CACHE_KEY_PREFIX, LoginThrottle monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - auth_cache_keys_before = set(ps.user_api_key_cache.in_memory_cache.cache_dict) request = MagicMock() @@ -1353,53 +1378,66 @@ async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): throttle = LoginThrottle.from_request(request, general_settings={}, redis_cache=None) for i in range(25): - with pytest.raises(ProxyException, match="Invalid credentials"): - await _guess(throttle, username=f"made-up-{i}@example.com") + assert await _fail(throttle, username=f"made-up-{i}@example.com") == "401" added = set(ps.user_api_key_cache.in_memory_cache.cache_dict) - auth_cache_keys_before - assert not [k for k in added if str(k).startswith(_CACHE_KEY_PREFIX)], ( - "sign-in counters must live in their own cache, not the key-authentication cache" - ) + assert not [k for k in added if str(k).startswith(_CACHE_KEY_PREFIX)] def test_settings_that_arrive_as_environment_strings_are_honored(): - """An `os.environ/VAR` reference in general_settings resolves to a string, not an int. - - Regression: a digit string fell back to the default with only a log line, so an operator - tightening the limits through environment substitution silently kept the stock ceilings. - """ + """An `os.environ/VAR` reference in general_settings resolves to a string, not an int.""" from litellm.proxy.auth.login_throttle import LoginThrottle request = MagicMock() - request.headers = {} + request.headers = {"x-forwarded-for": "203.0.113.9"} request.client = MagicMock() - request.client.host = "1.2.3.4" + request.client.host = "10.0.0.1" throttle = LoginThrottle.from_request( request, general_settings={ - "max_failed_login_attempts": "7", + "trusted_proxy_ranges": "10.0.0.0/8", "max_failed_login_attempts_per_source": " 70 ", + "max_failed_login_attempts_per_user": "7", "failed_login_window_seconds": "not-a-number", + "failed_login_block_seconds": "-5", }, redis_cache=None, ) - assert throttle.max_attempts == 7 - assert throttle.max_attempts_per_source == 70 - assert throttle.window_seconds == 900, "garbage still falls back to the default" + assert throttle.source_limit == 70 + assert throttle.user_limit == 7 + assert throttle.window_seconds == 60, "garbage falls back to the default" + assert throttle.block_seconds == 300, "a value below one would block nothing or forever" + + +def test_the_defaults_are_the_agreed_ones(): + from litellm.proxy.auth.login_throttle import LoginThrottle + + request = MagicMock() + request.headers = {"x-forwarded-for": "203.0.113.9"} + request.client = MagicMock() + request.client.host = "10.0.0.1" + throttle = LoginThrottle.from_request( + request, general_settings={"trusted_proxy_ranges": ["10.0.0.0/8"]}, redis_cache=None + ) + + assert (throttle.source_limit, throttle.user_limit, throttle.window_seconds, throttle.block_seconds) == ( + 10, + 5, + 60, + 300, + ) def test_the_disable_flag_is_read_once_not_per_login_attempt(monkeypatch): - """Regression: the kill switch was read through the secret manager on every unauthenticated request. - - With a hosted secret manager in read mode that is a synchronous network call per guess, so a - flood of wrong passwords could exhaust the secret manager even after the source was refused. - """ + """Regression: the kill switch was read through the secret manager on every unauthenticated request.""" from litellm.proxy.auth import login_throttle reads: Final[list[str]] = [] # mutable-ok: test-only call recorder - monkeypatch.setattr(login_throttle, "get_secret_bool", lambda name, default: reads.append(name) or default) + monkeypatch.setattr( + login_throttle, "get_secret_bool", lambda name, default_value: reads.append(name) or default_value + ) login_throttle._rate_limit_disabled.cache_clear() request = MagicMock() request.headers = {} @@ -1413,105 +1451,71 @@ def test_the_disable_flag_is_read_once_not_per_login_attempt(monkeypatch): assert reads == ["LITELLM_DISABLE_LOGIN_RATE_LIMIT"] -def test_a_negative_or_boolean_setting_falls_back_to_the_default(): - """A limit below one would refuse everyone; a bool is a typo, not a count.""" - from litellm.proxy.auth.login_throttle import LoginThrottle - - request = MagicMock() - request.headers = {} - request.client = MagicMock() - request.client.host = "1.2.3.4" - - throttle = LoginThrottle.from_request( - request, - general_settings={"max_failed_login_attempts": "-7", "max_failed_login_attempts_per_source": True}, - redis_cache=None, - ) - - assert throttle.max_attempts == 50 - assert throttle.max_attempts_per_source == 250 - - @pytest.mark.asyncio -async def test_a_refused_username_cannot_forge_log_lines(monkeypatch): +async def test_a_blocked_username_cannot_forge_log_lines(monkeypatch): """The username reaches a warning log, so it must not carry newlines or control bytes.""" import logging from litellm._logging import verbose_proxy_logger - from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(max_attempts=1) + throttle = _throttle(user_limit=1) forged = "victim@example.com\nWARNING: sign-in succeeded for attacker\x00" - with pytest.raises(ProxyException): - await _guess(throttle, username=forged) + assert await _fail(throttle, username=forged) == "401" records: list[logging.LogRecord] = [] handler = logging.Handler() handler.emit = records.append verbose_proxy_logger.addHandler(handler) try: - with pytest.raises(ProxyException) as blocked: - await _guess(throttle, username=forged) + assert await _fail(throttle, username=forged) == "401" finally: verbose_proxy_logger.removeHandler(handler) - assert blocked.value.code == "429" - emitted = [r.getMessage() for r in records if "sign-in attempts exhausted" in r.getMessage()] - assert emitted, "the refusal must be logged" + emitted = [r.getMessage() for r in records if "Admin UI sign-in blocked" in r.getMessage()] + assert emitted, "installing the block must be logged" assert "\n" not in emitted[0] and "\x00" not in emitted[0] assert "victim@example.com" in emitted[0] @pytest.mark.asyncio -async def test_a_username_spray_cannot_evict_an_existing_counter(monkeypatch): - """Regression: the in-memory tier must hold more counters than a spray can create. - - The default in-memory cache keeps 200 entries and evicts the soonest to expire, and - every counter shares one window, so eviction was effectively oldest-first. A few - hundred made-up usernames therefore pushed out the attacker's own counter and handed - back a fresh allowance against the real account. Username and source counters must also - live in separate stores, or the same spray evicts the source counter meant to stop it. - """ - from litellm.proxy._types import ProxyException +async def test_a_username_spray_cannot_evict_an_active_block(monkeypatch): + """Counters and blocks live in separate bounded stores, so a flood of made-up pairs fills the counter + store while the blocks it already earned stay in force.""" + from litellm.caching.in_memory_cache import InMemoryCache from litellm.proxy.auth.login_throttle import ( + _BLOCKS, + _COUNTERS, + _MAX_TRACKED_BLOCKS, + _MAX_TRACKED_COUNTERS, LoginThrottle, - _FAILED_LOGIN_SOURCE_CACHE, - _FAILED_LOGIN_USERNAME_CACHE, - _MAX_TRACKED_LOGIN_SOURCES, - _MAX_TRACKED_LOGIN_USERNAMES, ) monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - assert _MAX_TRACKED_LOGIN_SOURCES >= 10_000 - assert _MAX_TRACKED_LOGIN_USERNAMES >= 10_000 - assert _FAILED_LOGIN_SOURCE_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_SOURCES - assert _FAILED_LOGIN_USERNAME_CACHE.in_memory_cache.max_size_in_memory == _MAX_TRACKED_LOGIN_USERNAMES - assert _FAILED_LOGIN_SOURCE_CACHE.in_memory_cache is not _FAILED_LOGIN_USERNAME_CACHE.in_memory_cache - + assert _MAX_TRACKED_COUNTERS >= 10_000 and _MAX_TRACKED_BLOCKS >= 10_000 + assert _COUNTERS is not _BLOCKS + counters, blocks = InMemoryCache(max_size_in_memory=50), InMemoryCache(max_size_in_memory=50) throttle = LoginThrottle( client_ip="10.9.9.9", - max_attempts=3, - max_attempts_per_source=10_000, - window_seconds=900, - username_cache=_FAILED_LOGIN_USERNAME_CACHE, - source_cache=_FAILED_LOGIN_SOURCE_CACHE, + source_limit=None, + user_limit=1, + window_seconds=60, + block_seconds=300, + counters=counters, + blocks=blocks, ) victim = "spray-victim@corp.com" - for _ in range(3): - with pytest.raises(ProxyException): - await _guess(throttle, username=victim) + assert [await _fail(throttle, username=victim) for _ in range(2)] == ["401", "401"] - for i in range(500): + for i in range(200): await throttle.record_failure(f"spray-filler-{i}@corp.com") - assert await throttle._failures(throttle.username_cache, throttle._username_key(victim)) == 3, "the counter must survive a spray" - with pytest.raises(ProxyException) as blocked: - await _guess(throttle, username=victim) - assert blocked.value.code == "429" + assert len(counters.cache_dict) <= 50, "the counter store is bounded" + assert counters.get_cache(throttle._keys(victim).pair_counter) is None, "the victim's counter was evicted" + assert await _fail(throttle, username=victim) == "429", "the block survived the spray" def _patch_sso_configured(stack: ExitStack, *, configured: bool) -> None: diff --git a/tests/test_litellm/proxy/proxy_server/conftest.py b/tests/test_litellm/proxy/proxy_server/conftest.py index 56aa87f1e50..a349d378985 100644 --- a/tests/test_litellm/proxy/proxy_server/conftest.py +++ b/tests/test_litellm/proxy/proxy_server/conftest.py @@ -517,38 +517,25 @@ def make_key( def reset_login_throttle(monkeypatch): """Clear the Admin UI failed-login counters between tests. - `client` is session scoped and the counters live in shared module stores with a 900s - window, so without this a failed sign-in test could return 429 in unrelated tests later. + `client` is session scoped and the counters live in shared module stores with a 300s block + window, so without this a failed sign-in test could block unrelated tests later. Only the throttle's own keys are removed, so other cache entries remain untouched. """ from litellm.proxy import proxy_server as ps from litellm.proxy.auth import login_throttle - from litellm.proxy.auth.login_throttle import ( - _CACHE_KEY_PREFIX, - _FAILED_LOGIN_SOURCE_CACHE, - _FAILED_LOGIN_USERNAME_CACHE, - ) + from litellm.proxy.auth.login_throttle import _BLOCKS, _CACHE_KEY_PREFIX, _COUNTERS async def _no_delay(_seconds: float) -> None: - """The escalating wait on a rejected sign-in, replaced so the route tests stay fast.""" + """The hold on a rejected sign-in from a blocked key, replaced so the route tests stay fast.""" monkeypatch.setattr(login_throttle, "_sleep", _no_delay) def _drop_throttle_keys() -> None: - login_throttle._DELAYS_IN_FLIGHT.clear() - for cache in (_FAILED_LOGIN_USERNAME_CACHE, _FAILED_LOGIN_SOURCE_CACHE): - in_memory = getattr(cache, "in_memory_cache", None) - if in_memory is None: - continue - tracked = tuple( - key - for store in (getattr(in_memory, "cache_dict", None), getattr(in_memory, "ttl_dict", None)) - if isinstance(store, dict) - for key in tuple(store) - if str(key).startswith(_CACHE_KEY_PREFIX) - ) - for key in tracked: - in_memory.delete_cache(key) + login_throttle._HELD_ATTEMPTS.clear() + for store in (_COUNTERS, _BLOCKS): + for key in tuple(store.cache_dict) + tuple(store.ttl_dict): + if key.startswith(_CACHE_KEY_PREFIX): + store.delete_cache(key) monkeypatch.setattr(ps, "redis_usage_cache", None) _drop_throttle_keys() diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py index f8382c64e6a..7399e64c421 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py @@ -10,11 +10,8 @@ Routes covered: from __future__ import annotations -from concurrent.futures import ThreadPoolExecutor from unittest.mock import AsyncMock, MagicMock -import pytest - from .conftest import normalize # --------------------------------------------------------------------------- @@ -495,15 +492,38 @@ def _install_real_auth(monkeypatch, **settings): def _form_login(client, username="admin", password="wrong"): - return client.post( - "/login", data={"username": username, "password": password}, follow_redirects=False - ).status_code + return client.post("/login", data={"username": username, "password": password}, follow_redirects=False).status_code def _json_login(client, path, username="admin", password="wrong"): return client.post(path, json={"username": username, "password": password}).status_code +def _db_user(monkeypatch, email: str): + """A database user with a stored hash, faked so the route reaches the known-user branch without Postgres.""" + from unittest.mock import AsyncMock, MagicMock + + from litellm.proxy import proxy_server as ps + + user = MagicMock() + user.user_id = "u-1" + user.user_email = email + user.user_role = "internal_user" + user.password = "scrypt:stored" + repo = MagicMock() + repo.return_value.table.find_first = AsyncMock(return_value=user) + monkeypatch.setattr(ps, "prisma_client", MagicMock()) + monkeypatch.setattr("litellm.proxy.auth.login_utils.UserRepository", repo) + monkeypatch.setattr("litellm.proxy.auth.login_utils._rehash_password_if_needed", AsyncMock()) + monkeypatch.setattr( + "litellm.proxy.auth.login_utils.verify_password", lambda given, stored: given == "right-db-password" + ) + monkeypatch.setattr( + "litellm.proxy.auth.login_utils.generate_key_helper_fn", AsyncMock(return_value={"token": "sk-ui"}) + ) + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + + def test_budget_is_shared_across_every_login_endpoint(client, monkeypatch, reset_login_throttle): """The endpoint is not part of the key, so spending the budget on one route blocks the rest. @@ -511,92 +531,117 @@ def test_budget_is_shared_across_every_login_endpoint(client, monkeypatch, reset """ _install_real_auth( monkeypatch, - max_failed_login_attempts=10, + max_failed_login_attempts_per_user=10, control_plane_url="https://cp.example.com", ) assert [_form_login(client) for _ in range(5)] == [401] * 5 assert [_json_login(client, "/v2/login") for _ in range(5)] == [401] * 5 - assert _json_login(client, "/v3/login") == 429, "the eleventh attempt must be refused on a third route" + assert _json_login(client, "/v3/login") == 401, "the eleventh failure crosses the limit and installs the block" + assert _json_login(client, "/v3/login") == 429, "the twelfth attempt must be refused on a third route" def test_budget_is_shared_across_username_casing(client, monkeypatch, reset_login_throttle): """The database lookup is case-insensitive, so casing must not partition the counter.""" - _install_real_auth(monkeypatch, max_failed_login_attempts=10) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=3) - assert [_json_login(client, "/v2/login", username="admin@corp.com") for _ in range(5)] == [401] * 5 - assert [_json_login(client, "/v2/login", username="ADMIN@corp.com") for _ in range(5)] == [401] * 5 + assert [_json_login(client, "/v2/login", username="admin@corp.com") for _ in range(2)] == [401] * 2 + assert [_json_login(client, "/v2/login", username="ADMIN@corp.com") for _ in range(2)] == [401] * 2 assert _json_login(client, "/v2/login", username="Admin@corp.com") == 429 def test_a_refused_attempt_carries_retry_after(client, monkeypatch, reset_login_throttle): - """The 429 tells the caller how long the window has left.""" - _install_real_auth(monkeypatch, max_failed_login_attempts=2, failed_login_window_seconds=77) + """The 429 tells the caller how long the block has left, after the 30 seconds it was already held.""" + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1, failed_login_block_seconds=77) assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] refused = client.post("/v2/login", json={"username": "admin", "password": "wrong"}) assert refused.status_code == 429 - assert refused.headers.get("retry-after") == "77" + assert refused.headers.get("retry-after") == "47" def test_the_form_returns_a_human_readable_lockout_page(client, monkeypatch, reset_login_throttle): """The no-JavaScript form must render a wait page when its POST is throttled.""" - _install_real_auth(monkeypatch, max_failed_login_attempts=2, failed_login_window_seconds=77) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1, failed_login_block_seconds=77) assert [_form_login(client) for _ in range(2)] == [401, 401] refused = client.post("/login", data={"username": "admin", "password": "wrong"}) assert refused.status_code == 429 assert refused.headers.get("content-type", "").startswith("text/html") - assert "Try again in about 77 seconds" in refused.text - assert refused.headers.get("retry-after") == "77" + assert "Try again in about 47 seconds" in refused.text + assert refused.headers.get("retry-after") == "47" def test_a_second_username_from_the_same_source_still_gets_through(client, monkeypatch, reset_login_throttle): - """The username counter carries no address, so one account exhausting it cannot block another.""" - _install_real_auth(monkeypatch, max_failed_login_attempts=2) + """The pair block is per username, so one account's block cannot take the office down with it.""" + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) - for _ in range(3): - _json_login(client, "/v2/login", username="admin") + assert [_json_login(client, "/v2/login", username="admin") for _ in range(3)] == [401, 401, 429] assert _json_login(client, "/v2/login", username="someone-else@example.com") == 401 -def test_a_spray_across_usernames_is_refused_on_the_source_counter(client, monkeypatch, reset_login_throttle): - """A fresh username per guess keeps every username counter at one, so the address is what stops it.""" - _install_real_auth(monkeypatch, max_failed_login_attempts=100, max_failed_login_attempts_per_source=4) +def test_a_spray_across_usernames_is_blocked_on_the_source_when_the_source_is_attributable( + client, monkeypatch, reset_login_throttle +): + """A fresh username per guess keeps every pair at one, so the address is what stops it.""" + _install_real_auth(monkeypatch, trusted_proxy_ranges=["10.0.0.0/8"], max_failed_login_attempts_per_source=4) - sprayed = [_json_login(client, "/v2/login", username=f"sprayed-{i}@corp.com") for i in range(4)] - assert sprayed == [401] * 4 + sprayed = [_json_login(client, "/v2/login", username=f"sprayed-{i}@corp.com") for i in range(5)] + assert sprayed == [401] * 5 - assert _json_login(client, "/v2/login", username="sprayed-5@corp.com") == 429 + assert _json_login(client, "/v2/login", username="sprayed-6@corp.com") == 429 -def test_the_configured_admin_password_still_signs_in_while_refused(client, monkeypatch, reset_login_throttle): +def test_a_spray_across_usernames_is_not_blocked_without_trusted_proxy_ranges( + client, monkeypatch, reset_login_throttle +): + """Without a configured proxy range the peer address is whoever fronts the proxy, shared by every + client, so a source-wide block would block them all and the source scope stays off.""" + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=4) + + sprayed = [_json_login(client, "/v2/login", username=f"sprayed-{i}@corp.com") for i in range(8)] + assert sprayed == [401] * 8 + + +def test_the_configured_admin_password_still_signs_in_while_blocked(client, monkeypatch, reset_login_throttle): """The operator must never be locked out of the console by traffic aimed at it.""" from unittest.mock import AsyncMock, patch - _install_real_auth(monkeypatch, max_failed_login_attempts=2) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) monkeypatch.setenv("DATABASE_URL", "postgresql://stub") - assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] - assert _json_login(client, "/v2/login") == 429 + assert [_json_login(client, "/v2/login") for _ in range(3)] == [401, 401, 429] - with patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed - "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + with ( + patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), + patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed + "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) + ), ): assert _json_login(client, "/v2/login", password="right-password") == 200 -def test_sign_in_succeeds_again_once_the_budget_is_restored(client, monkeypatch, reset_login_throttle): - """A cleared bucket lets the same username straight back in.""" - _install_real_auth(monkeypatch, max_failed_login_attempts=2) +def test_a_database_users_correct_password_signs_in_while_blocked(client, monkeypatch, reset_login_throttle): + """The block is soft: guessing at an account slows the guesser down, it does not lock the owner out.""" + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) + _db_user(monkeypatch, "user@corp.com") - assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] - assert _json_login(client, "/v2/login") == 429 + assert [_json_login(client, "/v2/login", username="user@corp.com") for _ in range(3)] == [401, 401, 429] + + assert _json_login(client, "/v2/login", username="user@corp.com", password="right-db-password") == 200 + assert _json_login(client, "/v2/login", username="user@corp.com") == 429, "the block itself is still in force" + + +def test_sign_in_succeeds_again_once_the_block_is_cleared(client, monkeypatch, reset_login_throttle): + """A cleared store lets the same username straight back to a plain credential check.""" + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) + + assert [_json_login(client, "/v2/login") for _ in range(3)] == [401, 401, 429] reset_login_throttle() assert _json_login(client, "/v2/login") == 401 diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 2833686f115..07696f23da7 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -3438,6 +3438,34 @@ async def test_load_config_warns_per_worker_login_counters_without_general_setti assert "Running 4 workers but Redis is not configured" in caplog.text +@pytest.mark.asyncio +async def test_load_config_warns_that_the_source_login_limit_is_off_without_trusted_proxy_ranges( + tmp_path, monkeypatch, caplog +): + """The per-source failed-login limit is skipped when the source cannot be attributed, and the + operator must be told so at startup; a configured range silences it.""" + import logging + + from litellm.proxy.auth.login_throttle import warn_source_login_limit_is_off + from litellm.proxy.proxy_server import ProxyConfig + + monkeypatch.setenv("NUM_WORKERS", "1") + warn_source_login_limit_is_off.cache_clear() + config_file = tmp_path / "config.yaml" + config_file.write_text("model_list: []\n") + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + await ProxyConfig().load_config(router=MagicMock(), config_file_path=str(config_file)) + assert "trusted_proxy_ranges is not set" in caplog.text + + caplog.clear() + warn_source_login_limit_is_off.cache_clear() + config_file.write_text("model_list: []\ngeneral_settings:\n trusted_proxy_ranges: ['10.0.0.0/8']\n") + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + await ProxyConfig().load_config(router=MagicMock(), config_file_path=str(config_file)) + assert "trusted_proxy_ranges is not set" not in caplog.text + + @pytest.mark.asyncio async def test_load_environment_variables_direct_and_os_environ(): """ @@ -13331,14 +13359,16 @@ async def test_login_throttle_settings_are_not_hot_applied_from_the_database(): ps.general_settings.clear() await ProxyConfig()._update_general_settings( db_general_settings={ - "max_failed_login_attempts": 999, + "max_failed_login_attempts_per_user": 999, "max_failed_login_attempts_per_source": 999, "failed_login_window_seconds": 1, + "failed_login_block_seconds": 1, } ) - assert "max_failed_login_attempts" not in ps.general_settings + assert "max_failed_login_attempts_per_user" not in ps.general_settings assert "max_failed_login_attempts_per_source" not in ps.general_settings assert "failed_login_window_seconds" not in ps.general_settings + assert "failed_login_block_seconds" not in ps.general_settings finally: ps.general_settings.clear() ps.general_settings.update(original) diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 099ce397ffe..6203e79bccd 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -26447,9 +26447,14 @@ export interface components { * @description If True, router fallbacks configured in router_settings are only attempted when the calling key (and its team and project) is allowed to call the fallback model; unauthorized fallback targets are skipped and the primary model's error is returned. Default is False. */ enforce_fallback_model_access?: boolean | null; + /** + * Failed Login Block Seconds + * @description How long a blocked source address, or source address and username, stays blocked. The block is soft: a correct password still signs in, but each attempt from a blocked key takes one of 5 held slots per worker and a wrong password is held for 30 seconds before it is refused with 429. Set under `general_settings` in config.yaml. Defaults to 300 + */ + failed_login_block_seconds?: number | null; /** * Failed Login Window Seconds - * @description Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Set under `general_settings` in config.yaml. Defaults to 900 + * @description Fixed window in seconds over which failed Admin UI sign-in attempts are counted. The window starts at the first failure and is not extended by later ones. Set under `general_settings` in config.yaml. Defaults to 60 */ failed_login_window_seconds?: number | null; /** @@ -26496,16 +26501,23 @@ export interface components { * @description max batch input file size in MB for /v1/files uploads with purpose=batch, if a file is larger than this size it will be rejected before being forwarded to the provider */ max_batch_file_size_mb?: number | null; - /** - * Max Failed Login Attempts - * @description Number of failed Admin UI sign-in attempts allowed for one username, from any source address, within `failed_login_window_seconds`, before further attempts for that username are refused with 429. Attempts are answered with a doubling delay well before this ceiling. Set under `general_settings` in config.yaml. Defaults to 50 - */ - max_failed_login_attempts?: number | null; /** * Max Failed Login Attempts Per Source - * @description Number of failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`, before further attempts from that address are refused with 429. Counted independently of `max_failed_login_attempts`. Set under `general_settings` in config.yaml. Defaults to 250 + * @description Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set, since otherwise every client behind an ingress shares one address. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10 */ max_failed_login_attempts_per_source?: number | null; + /** + * Max Failed Login Attempts Per Source Overrides + * @description Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins. Set under `general_settings` in config.yaml + */ + max_failed_login_attempts_per_source_overrides?: { + [key: string]: number; + } | null; + /** + * Max Failed Login Attempts Per User + * @description Failed Admin UI sign-in attempts allowed from one source address for one username within `failed_login_window_seconds`. One more blocks that address for that username for `failed_login_block_seconds`, and its further failures stop counting against `max_failed_login_attempts_per_source`, so a script stuck on one account does not block everyone behind the same address. Set under `general_settings` in config.yaml. Defaults to 5 + */ + max_failed_login_attempts_per_user?: number | null; /** * Max File Size Mb * @description max file size in MB for /v1/files uploads, for any purpose, if a file is larger than this size it will be rejected before being forwarded to the provider From 7c2c59da903f72d13829c9b2f486b35bef1cb6a6 Mon Sep 17 00:00:00 2001 From: yucheng Date: Wed, 16 Sep 2026 00:13:51 +0000 Subject: [PATCH 038/141] test(proxy): explain the internal patches in the login throttle tests Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../test_litellm/proxy/auth/test_login_utils.py | 16 +++++++++++----- .../proxy/proxy_server/test_routes_login_sso.py | 4 +++- 2 files changed, 14 insertions(+), 6 deletions(-) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 22fcedaa19d..ac40b5364f9 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -725,11 +725,15 @@ async def _db_login(throttle, username: str, password: str, *, correct: bool): from litellm.proxy.auth.login_utils import authenticate_user with ( - patch("litellm.proxy.auth.login_utils.UserRepository", _known_user(username)), + patch( # test-quality-ok: the user lookup is the database boundary; faked so no DB is needed + "litellm.proxy.auth.login_utils.UserRepository", _known_user(username) + ), patch( # test-quality-ok: reaches the known-DB-user branch without a database "litellm.proxy.auth.login_utils.verify_password", return_value=correct ), - patch("litellm.proxy.auth.login_utils._rehash_password_if_needed", new=AsyncMock()), + patch( # test-quality-ok: the rehash writes to the database; faked so no DB is needed + "litellm.proxy.auth.login_utils._rehash_password_if_needed", new=AsyncMock() + ), patch( # test-quality-ok: success mints a UI key; faked so no DB is needed "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ), @@ -1062,7 +1066,9 @@ async def test_the_configured_admin_credentials_bypass_the_throttle(monkeypatch) assert [await _fail(throttle) for _ in range(3)] == ["401", "401", "429"] with ( - patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), + patch( # test-quality-ok: the admin sign-in upserts the admin row; faked so no DB is needed + "litellm.proxy.auth.login_utils.user_update", new=AsyncMock() + ), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ), @@ -1135,9 +1141,9 @@ async def test_a_user_with_no_password_set_does_not_consume_the_budget(monkeypat repo = MagicMock() repo.return_value.table.find_first = AsyncMock(return_value=passwordless) - with patch( + with patch( # test-quality-ok: reaches the passwordless-DB-user branch without a database "litellm.proxy.auth.login_utils.UserRepository", repo - ): # test-quality-ok: reaches the passwordless-DB-user branch without a database + ): for _ in range(5): with pytest.raises(ProxyException) as exc: await authenticate_user( diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py index 7399e64c421..82e86086518 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py @@ -618,7 +618,9 @@ def test_the_configured_admin_password_still_signs_in_while_blocked(client, monk assert [_json_login(client, "/v2/login") for _ in range(3)] == [401, 401, 429] with ( - patch("litellm.proxy.auth.login_utils.user_update", new=AsyncMock()), + patch( # test-quality-ok: the admin sign-in upserts the admin row; faked so no DB is needed + "litellm.proxy.auth.login_utils.user_update", new=AsyncMock() + ), patch( # test-quality-ok: success mints a UI key and persists the user; faked so no DB is needed "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ), From bc60b49e98f17bbe09a457ba96c6b2d3d0650836 Mon Sep 17 00:00:00 2001 From: yucheng Date: Wed, 16 Sep 2026 00:55:11 +0000 Subject: [PATCH 039/141] fix(proxy): key held sign-in attempts on the source while the source is blocked An active source block now takes precedence over a pair block, so every blocked username behind one blocked source shares the source's five held slots instead of getting five each Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 4 +- .../proxy/auth/test_login_utils.py | 46 +++++++++++++++++++ 2 files changed, 48 insertions(+), 2 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index fb708ecf872..393411d0670 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -290,10 +290,10 @@ class LoginThrottle: shared: Final = await self._shared_block_ttls(keys) user_ttl: Final = max(local[0], shared[0]) source_ttl: Final = max(local[1], shared[1]) - if user_ttl > 0: - return Block(scope="user", retry_after=user_ttl) if self.source_limit is not None and source_ttl > 0: return Block(scope="source", retry_after=source_ttl) + if user_ttl > 0: + return Block(scope="user", retry_after=user_ttl) return None async def _shared_block_ttls(self, keys: _Keys) -> _BlockTtls: diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index ac40b5364f9..65c388240e5 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1220,6 +1220,52 @@ async def test_held_attempts_from_one_blocked_key_are_capped(monkeypatch): assert lt._HELD_ATTEMPTS.get(slot) is None, "the slots are released once the held attempts answer" +@pytest.mark.asyncio +async def test_a_blocked_source_shares_one_held_slot_pool_across_its_blocked_usernames(monkeypatch): + """Once the source is blocked, a pair block for a username must not hand that username its own five slots.""" + import asyncio + + from litellm.proxy._types import ProxyException + from litellm.proxy.auth import login_throttle as lt + from litellm.proxy.auth.login_throttle import MAX_HELD_ATTEMPTS_PER_KEY + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + release = asyncio.Event() + + async def _park(_seconds: float) -> None: + await release.wait() + + monkeypatch.setattr(lt, "_sleep", _park) + throttle = _throttle(user_limit=1, source_limit=3, client_ip="203.0.113.45") + assert [await _fail(throttle) for _ in range(2)] == ["401", "401"], "the admin pair is now blocked" + assert [await _fail(throttle, username=f"spray-{i}@corp.com") for i in range(3)] == ["401"] * 3 + source_slot = throttle._keys("admin").source_block + assert throttle._local_block_ttl(source_slot) > 0, "the source is now blocked as well" + + usernames = ["admin", *(f"fresh-{i}@corp.com" for i in range(MAX_HELD_ATTEMPTS_PER_KEY - 1))] + held = [asyncio.create_task(_guess(throttle, username=name)) for name in usernames] + for _ in range(1000): + if lt._HELD_ATTEMPTS.get(source_slot) == MAX_HELD_ATTEMPTS_PER_KEY: + break + await asyncio.sleep(0) + assert lt._HELD_ATTEMPTS == {source_slot: MAX_HELD_ATTEMPTS_PER_KEY} + + try: + for name in ("admin", "fresh-0@corp.com", "never-seen@corp.com"): + with pytest.raises(ProxyException) as over_cap: + await _guess(throttle, username=name) + assert over_cap.value.code == "429" + assert over_cap.value.headers.get("Retry-After") == "30" + finally: + release.set() + for task in held: + with pytest.raises(ProxyException): + await task + + assert lt._HELD_ATTEMPTS == {} + + @pytest.mark.asyncio async def test_disabling_the_control_removes_the_hold_as_well(monkeypatch, login_delays): """The escape hatch has to turn off the whole control, not only the refusal.""" From 3c972cb31f006e13d9e1fbb14054785cc15a6c7d Mon Sep 17 00:00:00 2001 From: yucheng Date: Wed, 16 Sep 2026 01:33:54 +0000 Subject: [PATCH 040/141] fix(proxy): apply source overrides to IPv4-mapped IPv6 sign-in sources Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 9 +++++---- tests/test_litellm/proxy/auth/test_login_utils.py | 2 ++ 2 files changed, 7 insertions(+), 4 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 393411d0670..4ab57ca461f 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -137,10 +137,14 @@ def _int_setting(settings: Mapping[str, object], key: str, default: int) -> int: def _parse_address(client_ip: str) -> ipaddress.IPv4Address | ipaddress.IPv6Address | None: + """The address as it is limited and counted: an IPv4-mapped IPv6 address is its IPv4 address.""" try: - return ipaddress.ip_address(client_ip) + address: Final = ipaddress.ip_address(client_ip) except ValueError: return None + if isinstance(address, ipaddress.IPv6Address) and address.ipv4_mapped is not None: + return address.ipv4_mapped + return address def _parse_network(raw_range: str) -> _Network | None: @@ -183,9 +187,6 @@ def source_group(client_ip: str) -> str: if address is None: return client_ip if isinstance(address, ipaddress.IPv6Address): - mapped: Final = address.ipv4_mapped - if mapped is not None: - return str(mapped) return str(ipaddress.ip_network((address, IPV6_SOURCE_PREFIX_LENGTH), strict=False)) return str(address) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 65c388240e5..df5bd29f8ff 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -955,6 +955,8 @@ def test_source_overrides_pick_the_most_specific_matching_range(): assert _limit("203.0.1.1") == 100 assert _limit("192.0.2.1") == 7 assert _limit("198.51.100.1") == 7, "a garbage limit falls back to the default rather than a huge or zero budget" + assert _limit("::ffff:203.0.113.9") == 300, "a mapped address gets the limit of the IPv4 bucket it is counted in" + assert _limit("::ffff:203.0.113.10") == 200 def test_ipv6_sources_are_grouped_by_their_64_bit_prefix(): From c03a42a9c8144f8c2346ba8cd9dd342e1fe71149 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 16 Sep 2026 07:17:27 +0000 Subject: [PATCH 041/141] fix(azure): keep api-version query after vector store search path Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../openai/vector_stores/transformation.py | 3 +- .../llms/azure/vector_stores/__init__.py | 0 ...test_azure_vector_stores_transformation.py | 20 ++++++++++++ ...est_openai_vector_stores_transformation.py | 32 +++++++++++-------- 4 files changed, 40 insertions(+), 15 deletions(-) create mode 100644 tests/test_litellm/llms/azure/vector_stores/__init__.py create mode 100644 tests/test_litellm/llms/azure/vector_stores/test_azure_vector_stores_transformation.py diff --git a/litellm/llms/openai/vector_stores/transformation.py b/litellm/llms/openai/vector_stores/transformation.py index 125e5168c69..57fe5b04838 100644 --- a/litellm/llms/openai/vector_stores/transformation.py +++ b/litellm/llms/openai/vector_stores/transformation.py @@ -101,7 +101,8 @@ class OpenAIVectorStoreConfig(BaseVectorStoreConfig): extra_body: dict[str, object] | None = None, ) -> tuple[str, dict]: encoded_vector_store_id: Final = encode_url_path_segment(vector_store_id, field_name="vector_store_id") - url: Final = f"{api_base}/{encoded_vector_store_id}/search" + base_url, query_separator, query_string = api_base.partition("?") + url: Final = f"{base_url}/{encoded_vector_store_id}/search{query_separator}{query_string}" typed_request_body: Final = VectorStoreSearchRequest( query=query, filters=vector_store_search_optional_params.get("filters", None), diff --git a/tests/test_litellm/llms/azure/vector_stores/__init__.py b/tests/test_litellm/llms/azure/vector_stores/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/test_litellm/llms/azure/vector_stores/test_azure_vector_stores_transformation.py b/tests/test_litellm/llms/azure/vector_stores/test_azure_vector_stores_transformation.py new file mode 100644 index 00000000000..59bec08fca6 --- /dev/null +++ b/tests/test_litellm/llms/azure/vector_stores/test_azure_vector_stores_transformation.py @@ -0,0 +1,20 @@ +from litellm.llms.azure.vector_stores.transformation import AzureOpenAIVectorStoreConfig + + +def test_transform_search_vector_store_request_preserves_azure_query_string(): + config = AzureOpenAIVectorStoreConfig() + api_base = config.get_complete_url( + api_base="https://x.openai.azure.com", + litellm_params={"api_version": "2024-10-21"}, + ) + + url, _ = config.transform_search_vector_store_request( + vector_store_id="vs_1", + query="hello", + vector_store_search_optional_params={}, + api_base=api_base, + litellm_logging_obj=None, + litellm_params={"api_version": "2024-10-21"}, + ) + + assert url == "https://x.openai.azure.com/openai/vector_stores/vs_1/search?api-version=2024-10-21" diff --git a/tests/test_litellm/llms/openai/vector_stores/test_openai_vector_stores_transformation.py b/tests/test_litellm/llms/openai/vector_stores/test_openai_vector_stores_transformation.py index e7b1aab45b4..ea1f9e87ed8 100644 --- a/tests/test_litellm/llms/openai/vector_stores/test_openai_vector_stores_transformation.py +++ b/tests/test_litellm/llms/openai/vector_stores/test_openai_vector_stores_transformation.py @@ -7,11 +7,8 @@ from litellm.types.vector_stores import ( class TestOpenAIVectorStoreAPIConfig: - @pytest.mark.parametrize("metadata", [{}, None]) - def test_transform_create_vector_store_request_with_metadata_empty_or_none( - self, metadata - ): + def test_transform_create_vector_store_request_with_metadata_empty_or_none(self, metadata): """ Test transform_create_vector_store_request when metadata is None or empty dict. """ @@ -24,9 +21,7 @@ class TestOpenAIVectorStoreAPIConfig: "metadata": metadata, } - url, request_body = config.transform_create_vector_store_request( - vector_store_create_params, api_base - ) + url, request_body = config.transform_create_vector_store_request(vector_store_create_params, api_base) assert url == api_base assert request_body["name"] == "test-vector-store" @@ -50,9 +45,7 @@ class TestOpenAIVectorStoreAPIConfig: "metadata": large_metadata, } - url, request_body = config.transform_create_vector_store_request( - vector_store_create_params, api_base - ) + url, request_body = config.transform_create_vector_store_request(vector_store_create_params, api_base) assert url == api_base assert request_body["name"] == "test-vector-store" @@ -77,8 +70,19 @@ class TestOpenAIVectorStoreAPIConfig: litellm_params={}, ) - assert ( - url - == "https://api.openai.com/v1/vector_stores/..%2F..%2Ffiles%3Fx%3D1%23frag/search" - ) + assert url == "https://api.openai.com/v1/vector_stores/..%2F..%2Ffiles%3Fx%3D1%23frag/search" assert request_body["query"] == "hello" + + def test_transform_search_vector_store_request_preserves_query_string(self): + config = OpenAIVectorStoreConfig() + + url, _ = config.transform_search_vector_store_request( + vector_store_id="vs_1", + query="hello", + vector_store_search_optional_params={}, + api_base="https://x.openai.azure.com/openai/vector_stores?api-version=2024-10-21", + litellm_logging_obj=None, + litellm_params={}, + ) + + assert url == "https://x.openai.azure.com/openai/vector_stores/vs_1/search?api-version=2024-10-21" From cfd83548b30cf0afdf9a5c9f77a573582dfa7256 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 09:04:37 +0000 Subject: [PATCH 042/141] refactor(proxy): drop narrating docstrings from the aggregated usage query path Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../management_endpoints/common_daily_activity.py | 14 -------------- 1 file changed, 14 deletions(-) diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index cb81d57ac77..5f63f639a11 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -761,13 +761,6 @@ def _build_aggregated_sql_query( ) -> tuple[str, list[str]]: # mutable-ok: SQL text plus its ordered $N params """Build the GROUPING SETS query for aggregated daily activity. - One statement, two UNION ALL arms over the same WHERE clause. The first arm is - key-free: grand total, per-date totals and the (date, model / model_group / - provider / mcp / endpoint) rollups, so its row count never grows with the number - of keys. The second arm emits the (date, , api_key) rollups for the - USAGE_TOP_API_KEYS_LIMIT highest-spend keys only. Both arms share the 7-bit - group_level bitmask (date, api_key, model, model_group, provider, mcp, endpoint). - Returns: Tuple of (sql_query, params_list) ready for prisma_client.db.query_raw(). """ @@ -1366,13 +1359,6 @@ async def get_daily_activity_aggregated( ) -> SpendAnalyticsPaginatedResponse: """Aggregated variant that returns the full result set (no pagination). - Runs one GROUPING SETS statement with two UNION ALL arms: a key-free one for totals - and the model/provider/mcp/endpoint rollups (row count independent of key - cardinality) and a bounded one for the per-key rollups of the top - USAGE_TOP_API_KEYS_LIMIT keys by spend. breakdown.api_keys and every - api_key_breakdown therefore list at most that many keys, while the totals and the - key-free rollups cover every key. - include_entity_breakdown runs a small companion rollup query and folds `breakdown.entities` onto the response, as entity-scoped views like Team Usage need. From f863e746123bea7a6d0d1c730e68db35d8936854 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 09:22:20 +0000 Subject: [PATCH 043/141] test(proxy): assert aggregated usage behavior against Postgres instead of SQL text Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../test_common_daily_activity.py | 114 +++++++----------- 1 file changed, 41 insertions(+), 73 deletions(-) diff --git a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py index 471cabbbb59..e1a4d0d02a8 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py +++ b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py @@ -1239,79 +1239,6 @@ class TestBuildAggregatedSqlQuery: assert "model = $4" in sql assert "api_key = $5" in sql - def test_model_group_rollups_fall_back_to_model_name(self): - """Aggregated model_groups rollups must fall back to model for group-less rows. - - The (date, model_group) grouping level cannot recover the model column - after the fact (it is rolled up), so the fallback has to happen in SQL; - without it, group-less rows silently vanish from the model_groups - breakdown that the usage UI now renders by default. Group-less rows are - stored as empty strings, not NULL (spend_tracking_utils defaults - model_group to ""), so a plain COALESCE is not enough: the fallback must - be NULLIF-wrapped to catch both - """ - sql, _ = _build_aggregated_sql_query( - table_name="litellm_dailyuserspend", - entity_id_field="user_id", - entity_id=None, - start_date="2026-07-01", - end_date="2026-07-01", - model=None, - api_key=None, - ) - - normalized = " ".join(sql.split()) - fallback = "COALESCE(NULLIF(model_group, ''), model)" - assert f"{fallback} AS model_group" in normalized - assert f"GROUPING(model, {fallback}, custom_llm_provider, mcp_namespaced_tool_name, endpoint)" in normalized - assert f"(date, {fallback})," in normalized - assert "(date, model_group)" not in normalized - assert "COALESCE(model_group, model)" not in normalized - - def test_totals_arm_never_groups_by_api_key(self): - """The totals arm must not emit one row per key, that is what blew up the - query engine at 3k+ keys. Every grouping set there stays key-free and api_key - is projected as a NULL literal so the dispatcher's row shape is unchanged.""" - sql, _ = _build_aggregated_sql_query( - table_name="litellm_dailyuserspend", - entity_id_field="user_id", - entity_id=None, - start_date="2026-07-01", - end_date="2026-07-01", - model=None, - api_key=None, - ) - - totals_arm, _ = " ".join(sql.split()).split("UNION ALL") - grouping_block = totals_arm.split("GROUP BY GROUPING SETS (", 1)[1] - assert "api_key" not in grouping_block - assert "NULL::text AS api_key" in totals_arm - - def test_per_key_arm_ranks_keys_deterministically_and_shares_filters(self): - """Both arms sit in one statement so totals and per-key rows come from the - same snapshot, and the per-key arm reuses the caller's filter params.""" - sql, params = _build_aggregated_sql_query( - table_name="litellm_dailyuserspend", - entity_id_field="user_id", - entity_id="user-1", - start_date="2026-05-29", - end_date="2026-06-02", - model="bedrock/global.anthropic.claude-opus-4-8", - api_key="sk-test", - timezone_offset_minutes=-330, - ) - - totals_arm, per_key_arm = " ".join(sql.split()).split("UNION ALL") - assert "top_api_keys" not in totals_arm - assert f"ORDER BY SUM(spend) DESC, api_key LIMIT {USAGE_TOP_API_KEYS_LIMIT}" in per_key_arm - assert "api_key IN (SELECT api_key FROM top_api_keys)" in per_key_arm - assert "api_key <> $6" in per_key_arm - assert per_key_arm.count("model = $4 AND api_key = $5") == 2 - grouping_block = per_key_arm.split("GROUP BY GROUPING SETS (", 1)[1] - assert grouping_block.count(", api_key)") == 6 - assert grouping_block.count("(date") == 6 - assert params[-1] == PTU_SENTINEL_API_KEY - class TestAggregatedEmptyEntityFilter: _BUILDERS: Final = (_build_aggregated_sql_query, _build_entity_rollup_sql_query) @@ -1640,6 +1567,47 @@ async def test_get_daily_activity_aggregated_explicit_api_key_filter_scopes_both assert set(day.breakdown.models["gpt-5"].api_key_breakdown) == {"key-1"} +@pytest.mark.asyncio +async def test_get_daily_activity_aggregated_model_group_rollups_fall_back_to_model_name( + _aggregated_postgresql: psycopg.Connection, +): + """Rows stored with an empty or NULL model_group must land in the model_groups + breakdown under their model name instead of vanishing from the usage UI.""" + rows: Final = [ + ("row-0", "user-0", "2026-06-01", "key-0", "gpt-5", "gpt-5-eu", "openai", "/v1/chat/completions", 10, 7.0, 1, 1), + ("row-1", "user-1", "2026-06-01", "key-1", "gpt-5", "", "openai", "/v1/chat/completions", 10, 3.0, 1, 1), + ("row-2", "user-2", "2026-06-01", "key-2", "claude-x", None, "anthropic", "/v1/messages", 10, 2.0, 1, 1), + ] + _seed_daily_user_spend(_aggregated_postgresql, rows) + + mock_prisma = MagicMock() + mock_prisma.db = MagicMock() + mock_prisma.db.query_raw = _psycopg_query_raw(_aggregated_postgresql, []) + mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) + mock_prisma.db.litellm_deletedverificationtoken.find_many = AsyncMock(return_value=[]) + + result = await get_daily_activity_aggregated( + prisma_client=mock_prisma, + table_name="litellm_dailyuserspend", + entity_id_field="user_id", + entity_id=None, + entity_metadata_field=None, + start_date="2026-06-01", + end_date="2026-06-01", + model=None, + api_key=None, + ) + + breakdown: Final = result.results[0].breakdown + assert set(breakdown.model_groups) == {"gpt-5-eu", "gpt-5", "claude-x"} + assert breakdown.model_groups["gpt-5-eu"].metrics.spend == 7.0 + assert breakdown.model_groups["gpt-5"].metrics.spend == 3.0 + assert breakdown.model_groups["claude-x"].metrics.spend == 2.0 + assert set(breakdown.model_groups["gpt-5"].api_key_breakdown) == {"key-1"} + assert set(breakdown.models) == {"gpt-5", "claude-x"} + assert breakdown.models["gpt-5"].metrics.spend == 10.0 + + def _no_spend_record(): """A rollup row for a key with no spend, where SUM() returns NULL (None).""" return SimpleNamespace( From 61b3611b8c6667c8eae5dadc6036a6d95e897d04 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 09:22:20 +0000 Subject: [PATCH 044/141] fix(ui): block the global usage export when the aggregated key cap is reached Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../components/EntityUsage/EntityUsage.tsx | 2 +- .../_components/components/UsagePageView.tsx | 3 +- .../exportBlockedReason.test.ts | 37 ++++++++++++++++++- .../EntityUsageExport/exportBlockedReason.ts | 18 ++++++++- 4 files changed, 56 insertions(+), 4 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx index 6687bd4df03..27a4163dff7 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx @@ -664,7 +664,7 @@ const EntityUsage: React.FC = ({ { key: "endpoints", label: "Endpoint Activity", content: }, ]; - const spendFetchState = { coversRange, cancelled, failed }; + const spendFetchState = { coversRange, cancelled, failed, apiKeyLimitReached: undefined }; return (
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx index 691c5dc839a..48322477fb4 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx @@ -30,7 +30,7 @@ import { ActivityMetrics, processActivityData } from "@/components/activity_metr import CloudZeroExportModal from "@/components/cloudzero_export_modal"; import UserDropdown from "@/components/common_components/UserDropdown"; import EntityUsageExportModal from "@/components/EntityUsageExport"; -import { getExportBlockedReason } from "@/components/EntityUsageExport/exportBlockedReason"; +import { getApiKeyLimitReached, getExportBlockedReason } from "@/components/EntityUsageExport/exportBlockedReason"; import KeyActivityPanel from "@/components/UsagePage/components/KeyActivityPanel"; import { Team } from "@/components/key_team_helpers/key_list"; import { @@ -256,6 +256,7 @@ const UsagePage: React.FC = ({ teams, organizations }) => { coversRange: activeAggregated !== null || paginatedResult.coversRange, cancelled: paginatedResult.cancelled, failed: paginatedResult.failed, + apiKeyLimitReached: getApiKeyLimitReached(userSpendData.results, userSpendData.metadata?.api_key_limit), }; const exportBlockedReason = getExportBlockedReason(spendFetchState); diff --git a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts index e39b01a5dea..86783e2e4a0 100644 --- a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts +++ b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts @@ -1,14 +1,24 @@ import { describe, expect, it } from "vitest"; -import { getExportBlockedReason, type UsageFetchState } from "./exportBlockedReason"; +import type { DailyData } from "@/components/UsagePage/types"; + +import { getApiKeyLimitReached, getExportBlockedReason, type UsageFetchState } from "./exportBlockedReason"; const state = (overrides: Partial = {}): UsageFetchState => ({ coversRange: true, cancelled: false, failed: false, + apiKeyLimitReached: undefined, ...overrides, }); +const dayWithKeys = (date: string, ...keys: string[]): DailyData => + ({ + date, + metrics: {}, + breakdown: { api_keys: Object.fromEntries(keys.map((k) => [k, { metrics: {}, metadata: {} }])) }, + }) as unknown as DailyData; + describe("getExportBlockedReason", () => { it("lets the export through once the data on screen covers the range", () => { expect(getExportBlockedReason(state())).toBeUndefined(); @@ -31,4 +41,29 @@ describe("getExportBlockedReason", () => { expect(reason).toMatch(/failed to load/i); expect(reason).not.toMatch(/stopped/i); }); + + it("blocks when the aggregated endpoint hit its key cap, since a per-team CSV would miss keys", () => { + const reason = getExportBlockedReason(state({ apiKeyLimitReached: 100 })); + + expect(reason).toMatch(/100 highest-spend keys/); + expect(reason).toMatch(/USAGE_TOP_API_KEYS_LIMIT/); + }); +}); + +describe("getApiKeyLimitReached", () => { + it("reports the cap once the distinct keys across every day reach it", () => { + const results = [dayWithKeys("2026-06-01", "key-1", "key-2"), dayWithKeys("2026-06-02", "key-2", "key-3")]; + + expect(getApiKeyLimitReached(results, 3)).toBe(3); + }); + + it("stays quiet while fewer keys than the cap came back, which means every key is on screen", () => { + const results = [dayWithKeys("2026-06-01", "key-1", "key-2"), dayWithKeys("2026-06-02", "key-2")]; + + expect(getApiKeyLimitReached(results, 3)).toBeUndefined(); + }); + + it("stays quiet when the response carries no cap, as the paginated fallback does", () => { + expect(getApiKeyLimitReached([dayWithKeys("2026-06-01", "key-1")], undefined)).toBeUndefined(); + }); }); diff --git a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts index 71408ba8f3f..ca756090351 100644 --- a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts +++ b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts @@ -1,13 +1,29 @@ +import type { DailyData } from "@/components/UsagePage/types"; + export interface UsageFetchState { coversRange: boolean; cancelled: boolean; failed: boolean; + apiKeyLimitReached: number | undefined; } -export const getExportBlockedReason = ({ coversRange, cancelled, failed }: UsageFetchState): string | undefined => { +export const getApiKeyLimitReached = (results: DailyData[], apiKeyLimit: unknown): number | undefined => { + if (typeof apiKeyLimit !== "number") return undefined; + const keys = new Set(results.flatMap((day) => Object.keys(day.breakdown.api_keys ?? {}))); + return keys.size >= apiKeyLimit ? apiKeyLimit : undefined; +}; + +export const getExportBlockedReason = ({ + coversRange, + cancelled, + failed, + apiKeyLimitReached, +}: UsageFetchState): string | undefined => { if (failed) return "Some spend data failed to load, so an export would under-report. Reload the page to try again."; if (cancelled) return "Loading was stopped before the whole range arrived, so an export would under-report. Reload the page to load it all."; if (!coversRange) return "Spend data is still loading, so an export would under-report. Wait for it to finish."; + if (apiKeyLimitReached !== undefined) + return `Only the ${apiKeyLimitReached} highest-spend keys were loaded, so a per-team export would under-report. Raise USAGE_TOP_API_KEYS_LIMIT on the proxy to load more keys.`; return undefined; }; From aec36a2bac12222a681f6552103b6899657abc04 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 09:36:13 +0000 Subject: [PATCH 045/141] style(proxy): drop em dash from api_key bit comment Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/management_endpoints/common_daily_activity.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index 5f63f639a11..93e6e60c77c 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -991,7 +991,7 @@ async def _aggregate_spend_records( # current grouping set's key), 0 when the column is part of the key. _GROUP_GRAND_TOTAL: Final = 127 # 0b1111111 — all rolled up _GROUP_DATE: Final = 63 # 0b0111111 — only date kept -_API_KEY_ROLLED_UP_BIT: Final = 32 # 0b0100000 — api_key position in the 7-bit mask +_API_KEY_ROLLED_UP_BIT: Final = 32 # 0b0100000 _GROUP_DATE_API_KEY: Final = 31 # 0b0011111 _GROUP_DATE_MODEL: Final = 47 # 0b0101111 _GROUP_DATE_MODEL_API_KEY: Final = 15 # 0b0001111 From 87894f2e7f343a75cc6d4f0977deceb0d564dde9 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 10:02:11 +0000 Subject: [PATCH 046/141] fix(proxy): report total_api_keys so exact-limit key sets are not treated as truncated Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_lazy_openapi_snapshot.json | 12 ++++ .../common_daily_activity.py | 15 ++-- .../common_daily_activity.py | 5 ++ .../test_common_daily_activity.py | 68 +++++++++++++++++-- .../components/EntityUsage/EntityUsage.tsx | 2 +- .../_components/components/UsagePageView.tsx | 7 +- .../exportBlockedReason.test.ts | 37 ++++------ .../EntityUsageExport/exportBlockedReason.ts | 20 +++--- ui/litellm-dashboard/src/lib/http/schema.d.ts | 5 ++ 9 files changed, 126 insertions(+), 45 deletions(-) diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index 5ff35747779..fb82a540761 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -3072,6 +3072,18 @@ "title": "Page", "type": "integer" }, + "total_api_keys": { + "anyOf": [ + { + "type": "integer" + }, + { + "type": "null" + } + ], + "description": "Distinct API keys matching the filters. When this exceeds api_key_limit, the per-key lists are truncated to the highest-spend keys.", + "title": "Total Api Keys" + }, "total_api_requests": { "default": 0, "title": "Total Api Requests", diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index 93e6e60c77c..a36b6052981 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -155,6 +155,7 @@ class _GroupingSetsRow(SimpleNamespace): mcp_namespaced_tool_name: str | None endpoint: str | None group_level: int + distinct_api_keys: int | None spend: float | None prompt_tokens: int | None completion_tokens: int | None @@ -800,7 +801,8 @@ def _build_aggregated_sql_query( (GROUPING(date) << 6) | {_API_KEY_ROLLED_UP_BIT} | GROUPING(model, {_MODEL_GROUP_EXPR}, custom_llm_provider, mcp_namespaced_tool_name, - endpoint) AS group_level,{metric_select} + endpoint) AS group_level, + NULL::bigint AS distinct_api_keys,{metric_select} FROM "{pg_table}" WHERE {where_clause} GROUP BY GROUPING SETS ( @@ -814,7 +816,7 @@ def _build_aggregated_sql_query( )) UNION ALL (WITH top_api_keys AS ( - SELECT api_key + SELECT api_key, COUNT(*) OVER () AS distinct_api_keys FROM "{pg_table}" WHERE {where_clause} AND api_key <> {sentinel_param} GROUP BY api_key @@ -831,9 +833,10 @@ def _build_aggregated_sql_query( endpoint, GROUPING(date, api_key, model, {_MODEL_GROUP_EXPR}, custom_llm_provider, mcp_namespaced_tool_name, - endpoint) AS group_level,{metric_select} - FROM "{pg_table}" - WHERE {where_clause} AND api_key IN (SELECT api_key FROM top_api_keys) + endpoint) AS group_level, + MAX(top_api_keys.distinct_api_keys) AS distinct_api_keys,{metric_select} + FROM "{pg_table}" JOIN top_api_keys USING (api_key) + WHERE {where_clause} GROUP BY GROUPING SETS ( (date, api_key), (date, model, api_key), @@ -1398,6 +1401,7 @@ async def get_daily_activity_aggregated( ) records: Final = [_GroupingSetsRow(**row) for row in (raw_rows or ())] + total_api_keys: Final = next((r.distinct_api_keys for r in records if r.distinct_api_keys is not None), 0) # The grouping-sets dispatcher places each row directly in its bucket # using the row's GROUPING() bitmask. No Python-side summing needed. @@ -1452,6 +1456,7 @@ async def get_daily_activity_aggregated( total_pages=1, has_more=False, api_key_limit=USAGE_TOP_API_KEYS_LIMIT, + total_api_keys=total_api_keys, ), ) diff --git a/litellm/types/proxy/management_endpoints/common_daily_activity.py b/litellm/types/proxy/management_endpoints/common_daily_activity.py index 2893eb30b3c..5d42b1230a0 100644 --- a/litellm/types/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/types/proxy/management_endpoints/common_daily_activity.py @@ -105,6 +105,11 @@ class DailySpendMetadata(BaseModel): description="When set, api_keys and every api_key_breakdown list at most this many keys, " "ranked by spend. Totals and the model, provider, mcp and endpoint rollups still cover every key.", ) + total_api_keys: int | None = Field( + default=None, + description="Distinct API keys matching the filters. When this exceeds api_key_limit, the per-key " + "lists are truncated to the highest-spend keys.", + ) class SpendAnalyticsPaginatedResponse(BaseModel): diff --git a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py index e1a4d0d02a8..e7f8bcc7e4f 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py +++ b/tests/test_litellm/proxy/management_endpoints/test_common_daily_activity.py @@ -175,6 +175,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "endpoint": "/v1/chat/completions", "api_key": None, "group_level": 62, + "distinct_api_keys": None, "spend": 15.0, "prompt_tokens": 150, "completion_tokens": 75, @@ -187,6 +188,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "endpoint": "/v1/embeddings", "api_key": None, "group_level": 62, + "distinct_api_keys": None, "spend": 3.0, "prompt_tokens": 30, "completion_tokens": 0, @@ -200,6 +202,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "endpoint": None, "api_key": None, "group_level": 63, + "distinct_api_keys": None, "spend": 18.0, "prompt_tokens": 180, "completion_tokens": 75, @@ -213,6 +216,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "endpoint": None, "api_key": None, "group_level": 127, + "distinct_api_keys": None, "spend": 18.0, "prompt_tokens": 180, "completion_tokens": 75, @@ -226,6 +230,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "endpoint": "/v1/chat/completions", "api_key": "key-1", "group_level": 30, + "distinct_api_keys": 2, "spend": 15.0, "prompt_tokens": 150, "completion_tokens": 75, @@ -238,6 +243,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown(): "endpoint": "/v1/embeddings", "api_key": "key-2", "group_level": 30, + "distinct_api_keys": 2, "spend": 3.0, "prompt_tokens": 30, "completion_tokens": 0, @@ -839,6 +845,7 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): "endpoint": "/v1/chat/completions", "api_key": None, "group_level": 62, + "distinct_api_keys": None, "spend": 10.0, "prompt_tokens": 100, "completion_tokens": 50, @@ -851,6 +858,7 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys(): "endpoint": "/v1/chat/completions", "api_key": "deleted-key-hash", "group_level": 30, + "distinct_api_keys": 1, "spend": 10.0, "prompt_tokens": 100, "completion_tokens": 50, @@ -1316,6 +1324,7 @@ async def test_get_daily_activity_aggregated_empty_result_set(): "mcp_namespaced_tool_name": None, "endpoint": None, "group_level": 127, + "distinct_api_keys": None, "spend": None, "prompt_tokens": None, "completion_tokens": None, @@ -1499,6 +1508,7 @@ async def test_get_daily_activity_aggregated_bounds_api_key_rollups( assert result.metadata.total_spend == pytest.approx(key_spend + 1000.0) assert result.metadata.total_api_requests == n_keys assert result.metadata.api_key_limit == USAGE_TOP_API_KEYS_LIMIT + assert result.metadata.total_api_keys == n_keys expected_top: Final = {f"key-{i:03d}" for i in range(6, n_keys)} | {"key-004"} day: Final = result.results[0] @@ -1560,6 +1570,7 @@ async def test_get_daily_activity_aggregated_explicit_api_key_filter_scopes_both ) assert result.metadata.total_spend == 2.0 + assert result.metadata.total_api_keys == 1 day: Final = result.results[0] assert set(day.breakdown.api_keys) == {"key-1"} assert day.breakdown.api_keys["key-1"].metrics.spend == 2.0 @@ -1567,6 +1578,55 @@ async def test_get_daily_activity_aggregated_explicit_api_key_filter_scopes_both assert set(day.breakdown.models["gpt-5"].api_key_breakdown) == {"key-1"} +@pytest.mark.asyncio +async def test_get_daily_activity_aggregated_reports_exact_limit_key_count_as_complete( + _aggregated_postgresql: psycopg.Connection, +): + """With exactly USAGE_TOP_API_KEYS_LIMIT keys nothing is dropped, and the + response must say so: total_api_keys equals the limit rather than exceeding it.""" + rows: Final = [ + ( + f"row-{i:03d}", + f"user-{i:03d}", + "2026-06-01", + f"key-{i:03d}", + "gpt-5", + "", + "openai", + "/v1/chat/completions", + 10, + float(i + 1), + 1, + 1, + ) + for i in range(USAGE_TOP_API_KEYS_LIMIT) + ] + _seed_daily_user_spend(_aggregated_postgresql, rows) + + row_counts: Final[list[int]] = [] # mutable-ok: out-param for the query_raw shim + mock_prisma = MagicMock() + mock_prisma.db = MagicMock() + mock_prisma.db.query_raw = _psycopg_query_raw(_aggregated_postgresql, row_counts) + mock_prisma.db.litellm_verificationtoken.find_many = AsyncMock(return_value=[]) + mock_prisma.db.litellm_deletedverificationtoken.find_many = AsyncMock(return_value=[]) + + result = await get_daily_activity_aggregated( + prisma_client=mock_prisma, + table_name="litellm_dailyuserspend", + entity_id_field="user_id", + entity_id=None, + entity_metadata_field=None, + start_date="2026-06-01", + end_date="2026-06-01", + model=None, + api_key=None, + ) + + assert result.metadata.total_api_keys == USAGE_TOP_API_KEYS_LIMIT + assert result.metadata.api_key_limit == USAGE_TOP_API_KEYS_LIMIT + assert len(result.results[0].breakdown.api_keys) == USAGE_TOP_API_KEYS_LIMIT + + @pytest.mark.asyncio async def test_get_daily_activity_aggregated_model_group_rollups_fall_back_to_model_name( _aggregated_postgresql: psycopg.Connection, @@ -2427,10 +2487,10 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown(): "successful_requests": 0, } main_rows = [ - {**base, "date": None, "group_level": 127, "spend": 18.0}, - {**base, "date": "2024-01-01", "group_level": 63, "spend": 18.0}, - {**base, "date": "2024-01-01", "model": "gpt-4o", "group_level": 47, "spend": 18.0}, - {**base, "date": "2024-01-01", "api_key": "key-1", "group_level": 31, "spend": 12.0}, + {**base, "date": None, "group_level": 127, "distinct_api_keys": None, "spend": 18.0}, + {**base, "date": "2024-01-01", "group_level": 63, "distinct_api_keys": None, "spend": 18.0}, + {**base, "date": "2024-01-01", "model": "gpt-4o", "group_level": 47, "distinct_api_keys": None, "spend": 18.0}, + {**base, "date": "2024-01-01", "api_key": "key-1", "group_level": 31, "distinct_api_keys": 1, "spend": 12.0}, ] entity_base = { key: value diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx index 27a4163dff7..001fee4b7bc 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx @@ -664,7 +664,7 @@ const EntityUsage: React.FC = ({ { key: "endpoints", label: "Endpoint Activity", content: }, ]; - const spendFetchState = { coversRange, cancelled, failed, apiKeyLimitReached: undefined }; + const spendFetchState = { coversRange, cancelled, failed, apiKeyTruncation: undefined }; return (
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx index 48322477fb4..3a97e54edfc 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx @@ -30,7 +30,7 @@ import { ActivityMetrics, processActivityData } from "@/components/activity_metr import CloudZeroExportModal from "@/components/cloudzero_export_modal"; import UserDropdown from "@/components/common_components/UserDropdown"; import EntityUsageExportModal from "@/components/EntityUsageExport"; -import { getApiKeyLimitReached, getExportBlockedReason } from "@/components/EntityUsageExport/exportBlockedReason"; +import { getApiKeyTruncation, getExportBlockedReason } from "@/components/EntityUsageExport/exportBlockedReason"; import KeyActivityPanel from "@/components/UsagePage/components/KeyActivityPanel"; import { Team } from "@/components/key_team_helpers/key_list"; import { @@ -256,7 +256,10 @@ const UsagePage: React.FC = ({ teams, organizations }) => { coversRange: activeAggregated !== null || paginatedResult.coversRange, cancelled: paginatedResult.cancelled, failed: paginatedResult.failed, - apiKeyLimitReached: getApiKeyLimitReached(userSpendData.results, userSpendData.metadata?.api_key_limit), + apiKeyTruncation: getApiKeyTruncation( + userSpendData.metadata?.api_key_limit, + userSpendData.metadata?.total_api_keys, + ), }; const exportBlockedReason = getExportBlockedReason(spendFetchState); diff --git a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts index 86783e2e4a0..8491b31f5f9 100644 --- a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts +++ b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.test.ts @@ -1,24 +1,15 @@ import { describe, expect, it } from "vitest"; -import type { DailyData } from "@/components/UsagePage/types"; - -import { getApiKeyLimitReached, getExportBlockedReason, type UsageFetchState } from "./exportBlockedReason"; +import { getApiKeyTruncation, getExportBlockedReason, type UsageFetchState } from "./exportBlockedReason"; const state = (overrides: Partial = {}): UsageFetchState => ({ coversRange: true, cancelled: false, failed: false, - apiKeyLimitReached: undefined, + apiKeyTruncation: undefined, ...overrides, }); -const dayWithKeys = (date: string, ...keys: string[]): DailyData => - ({ - date, - metrics: {}, - breakdown: { api_keys: Object.fromEntries(keys.map((k) => [k, { metrics: {}, metadata: {} }])) }, - }) as unknown as DailyData; - describe("getExportBlockedReason", () => { it("lets the export through once the data on screen covers the range", () => { expect(getExportBlockedReason(state())).toBeUndefined(); @@ -42,28 +33,26 @@ describe("getExportBlockedReason", () => { expect(reason).not.toMatch(/stopped/i); }); - it("blocks when the aggregated endpoint hit its key cap, since a per-team CSV would miss keys", () => { - const reason = getExportBlockedReason(state({ apiKeyLimitReached: 100 })); + it("blocks when the aggregated endpoint dropped keys, since a per-team CSV would miss them", () => { + const reason = getExportBlockedReason(state({ apiKeyTruncation: { limit: 100, total: 3000 } })); - expect(reason).toMatch(/100 highest-spend keys/); + expect(reason).toMatch(/100 highest-spend keys of 3000/); expect(reason).toMatch(/USAGE_TOP_API_KEYS_LIMIT/); }); }); -describe("getApiKeyLimitReached", () => { - it("reports the cap once the distinct keys across every day reach it", () => { - const results = [dayWithKeys("2026-06-01", "key-1", "key-2"), dayWithKeys("2026-06-02", "key-2", "key-3")]; - - expect(getApiKeyLimitReached(results, 3)).toBe(3); +describe("getApiKeyTruncation", () => { + it("reports truncation once the proxy saw more keys than it returned", () => { + expect(getApiKeyTruncation(100, 101)).toEqual({ limit: 100, total: 101 }); }); - it("stays quiet while fewer keys than the cap came back, which means every key is on screen", () => { - const results = [dayWithKeys("2026-06-01", "key-1", "key-2"), dayWithKeys("2026-06-02", "key-2")]; - - expect(getApiKeyLimitReached(results, 3)).toBeUndefined(); + it("stays quiet when exactly the cap exists, since every key is on screen", () => { + expect(getApiKeyTruncation(100, 100)).toBeUndefined(); + expect(getApiKeyTruncation(100, 7)).toBeUndefined(); }); it("stays quiet when the response carries no cap, as the paginated fallback does", () => { - expect(getApiKeyLimitReached([dayWithKeys("2026-06-01", "key-1")], undefined)).toBeUndefined(); + expect(getApiKeyTruncation(undefined, undefined)).toBeUndefined(); + expect(getApiKeyTruncation(100, null)).toBeUndefined(); }); }); diff --git a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts index ca756090351..6c5a5f83231 100644 --- a/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts +++ b/ui/litellm-dashboard/src/components/EntityUsageExport/exportBlockedReason.ts @@ -1,29 +1,31 @@ -import type { DailyData } from "@/components/UsagePage/types"; +export interface ApiKeyTruncation { + limit: number; + total: number; +} export interface UsageFetchState { coversRange: boolean; cancelled: boolean; failed: boolean; - apiKeyLimitReached: number | undefined; + apiKeyTruncation: ApiKeyTruncation | undefined; } -export const getApiKeyLimitReached = (results: DailyData[], apiKeyLimit: unknown): number | undefined => { - if (typeof apiKeyLimit !== "number") return undefined; - const keys = new Set(results.flatMap((day) => Object.keys(day.breakdown.api_keys ?? {}))); - return keys.size >= apiKeyLimit ? apiKeyLimit : undefined; +export const getApiKeyTruncation = (apiKeyLimit: unknown, totalApiKeys: unknown): ApiKeyTruncation | undefined => { + if (typeof apiKeyLimit !== "number" || typeof totalApiKeys !== "number") return undefined; + return totalApiKeys > apiKeyLimit ? { limit: apiKeyLimit, total: totalApiKeys } : undefined; }; export const getExportBlockedReason = ({ coversRange, cancelled, failed, - apiKeyLimitReached, + apiKeyTruncation, }: UsageFetchState): string | undefined => { if (failed) return "Some spend data failed to load, so an export would under-report. Reload the page to try again."; if (cancelled) return "Loading was stopped before the whole range arrived, so an export would under-report. Reload the page to load it all."; if (!coversRange) return "Spend data is still loading, so an export would under-report. Wait for it to finish."; - if (apiKeyLimitReached !== undefined) - return `Only the ${apiKeyLimitReached} highest-spend keys were loaded, so a per-team export would under-report. Raise USAGE_TOP_API_KEYS_LIMIT on the proxy to load more keys.`; + if (apiKeyTruncation !== undefined) + return `Only the ${apiKeyTruncation.limit} highest-spend keys of ${apiKeyTruncation.total} were loaded, so a per-team export would under-report. Raise USAGE_TOP_API_KEYS_LIMIT on the proxy to load more keys.`; return undefined; }; diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index c9e9c550afa..84656e7eee1 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -27503,6 +27503,11 @@ export interface components { * @default 1 */ page: number; + /** + * Total Api Keys + * @description Distinct API keys matching the filters. When this exceeds api_key_limit, the per-key lists are truncated to the highest-spend keys. + */ + total_api_keys?: number | null; /** * Total Api Requests * @default 0 From 38c1139377279b061b808d12b9b0f8b22c05f339 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 10:13:05 +0000 Subject: [PATCH 047/141] feat(ui): note on the Key Activity tab when only the top-spend keys were loaded Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../usage/_components/components/UsagePageView.tsx | 2 +- .../UsagePage/components/KeyActivityPanel.test.tsx | 10 ++++++++++ .../UsagePage/components/KeyActivityPanel.tsx | 14 +++++++++++++- 3 files changed, 24 insertions(+), 2 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx index 3a97e54edfc..a9ab0f17f40 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx @@ -908,7 +908,7 @@ const UsagePage: React.FC = ({ teams, organizations }) => { - + diff --git a/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.test.tsx b/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.test.tsx index 693ac20a360..830139143e9 100644 --- a/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.test.tsx +++ b/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.test.tsx @@ -68,4 +68,14 @@ describe("KeyActivityPanel", () => { expect(screen.getByLabelText("Search keys")).toHaveValue(""); expect(screen.getByTestId("rendered-keys")).toHaveTextContent("hash-alicehash-bob"); }); + + it("says how many keys the proxy left out when only the top spenders were loaded", () => { + render(); + expect(screen.getByRole("note")).toHaveTextContent("Only the 2 highest-spend keys of 3,000 are loaded"); + }); + + it("shows no truncation note when every key is loaded", () => { + render(); + expect(screen.queryByRole("note")).not.toBeInTheDocument(); + }); }); diff --git a/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.tsx b/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.tsx index 8287a04d0c7..8b2141f8528 100644 --- a/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.tsx +++ b/ui/litellm-dashboard/src/components/UsagePage/components/KeyActivityPanel.tsx @@ -2,6 +2,7 @@ import { Search, X } from "lucide-react"; import React, { useMemo, useState } from "react"; import { ActivityMetrics } from "@/components/activity_metrics"; +import type { ApiKeyTruncation } from "@/components/EntityUsageExport/exportBlockedReason"; import { InputGroup, InputGroupAddon, InputGroupButton, InputGroupInput } from "@/components/ui/input-group"; import { filterKeyActivity } from "../keyActivityFilter"; @@ -10,9 +11,14 @@ import type { ModelActivityData } from "../types"; interface KeyActivityPanelProps { keyMetrics: Record; hidePromptCachingMetrics?: boolean; + apiKeyTruncation?: ApiKeyTruncation; } -const KeyActivityPanel: React.FC = ({ keyMetrics, hidePromptCachingMetrics = false }) => { +const KeyActivityPanel: React.FC = ({ + keyMetrics, + hidePromptCachingMetrics = false, + apiKeyTruncation, +}) => { const [query, setQuery] = useState(""); const filtered = useMemo(() => filterKeyActivity(keyMetrics, query), [keyMetrics, query]); const totalKeys = Object.keys(keyMetrics).length; @@ -43,6 +49,12 @@ const KeyActivityPanel: React.FC = ({ keyMetrics, hidePro Showing {shownKeys.toLocaleString()} of {totalKeys.toLocaleString()} keys + {apiKeyTruncation !== undefined && ( + + Only the {apiKeyTruncation.limit.toLocaleString()} highest-spend keys of{" "} + {apiKeyTruncation.total.toLocaleString()} are loaded + + )}
{isFiltering && totalKeys > 0 && shownKeys === 0 ? (

From 09fa0833b43820c039f45027ce26663de5a5dc54 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 10:17:46 +0000 Subject: [PATCH 048/141] feat(ui): surface top-key truncation on the team usage view Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../components/EntityUsage/EntityUsage.test.tsx | 17 +++++++++++++++++ .../components/EntityUsage/EntityUsage.tsx | 15 ++++++++++++--- 2 files changed, 29 insertions(+), 3 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx index 2a6c2ede478..5846a63bc70 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx @@ -569,6 +569,23 @@ describe("EntityUsage", () => { expect(screen.getAllByText("Activity Metrics")[1]).toBeInTheDocument(); }); + it("tells the team view how many keys the proxy left out of the per-key lists", async () => { + mockTeamDailyActivityAggregatedCall.mockResolvedValue({ + ...mockSpendData, + metadata: { ...mockSpendData.metadata, api_key_limit: 100, total_api_keys: 3000 }, + }); + render(); + + await waitFor(() => { + expect(mockTeamDailyActivityAggregatedCall).toHaveBeenCalled(); + }); + act(() => { + fireEvent.click(screen.getByText("Key Activity")); + }); + + expect(await screen.findByRole("note")).toHaveTextContent("Only the 100 highest-spend keys of 3,000 are loaded"); + }); + // An inactive tab panel is marked aria-selected="false" by one tab library and hidden by the // other, so treat either as "not on screen" and the assertion holds whichever one is rendering. const isShowing = (element: HTMLElement): boolean => { diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx index 001fee4b7bc..27460b21108 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.tsx @@ -25,7 +25,7 @@ import TeamMultiSelect from "@/components/common_components/team_multi_select"; import UserDropdown from "@/components/common_components/UserDropdown"; import { ActivityMetrics, processActivityData } from "@/components/activity_metrics"; import { UsageExportHeader } from "@/components/EntityUsageExport"; -import { getExportBlockedReason } from "@/components/EntityUsageExport/exportBlockedReason"; +import { getApiKeyTruncation, getExportBlockedReason } from "@/components/EntityUsageExport/exportBlockedReason"; import type { EntityType } from "@/components/EntityUsageExport/types"; import { agentDailyActivityCall, @@ -71,6 +71,8 @@ interface EntitySpendData { total_successful_requests: number; total_failed_requests: number; total_tokens: number; + api_key_limit?: number | null; + total_api_keys?: number | null; }; } @@ -160,6 +162,7 @@ const EntityUsage: React.FC = ({ }); const spendData = spendDataRaw as unknown as EntitySpendData; + const apiKeyTruncation = getApiKeyTruncation(spendData.metadata?.api_key_limit, spendData.metadata?.total_api_keys); const { data: agentSpendDataRaw, @@ -659,12 +662,18 @@ const EntityUsage: React.FC = ({ { key: "keys", label: "Key Activity", - content: , + content: ( + + ), }, { key: "endpoints", label: "Endpoint Activity", content: }, ]; - const spendFetchState = { coversRange, cancelled, failed, apiKeyTruncation: undefined }; + const spendFetchState = { coversRange, cancelled, failed, apiKeyTruncation }; return (

From 303a9058cc9db4ef7251c8e7ecbb90a63e17c36e Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 10:21:33 +0000 Subject: [PATCH 049/141] refactor(proxy): type entity rollup rows by their own projection Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../common_daily_activity.py | 25 +++++++++++-------- 1 file changed, 14 insertions(+), 11 deletions(-) diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index a36b6052981..a5e292b177d 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -146,16 +146,9 @@ class _AggregatedSpendData(TypedDict): totals: SpendMetrics -class _GroupingSetsRow(SimpleNamespace): +class _RollupMetricsRow(SimpleNamespace): date: str api_key: str | None - model: str | None - model_group: str | None - custom_llm_provider: str | None - mcp_namespaced_tool_name: str | None - endpoint: str | None - group_level: int - distinct_api_keys: int | None spend: float | None prompt_tokens: int | None completion_tokens: int | None @@ -173,7 +166,17 @@ class _GroupingSetsRow(SimpleNamespace): timed_requests: int | None -class _EntityRollupRow(_GroupingSetsRow): +class _GroupingSetsRow(_RollupMetricsRow): + model: str | None + model_group: str | None + custom_llm_provider: str | None + mcp_namespaced_tool_name: str | None + endpoint: str | None + group_level: int + distinct_api_keys: int | None + + +class _EntityRollupRow(_RollupMetricsRow): entity_id: str | None api_key_rolled: int @@ -202,7 +205,7 @@ async def _query_raw_optional( return await prisma_client.db.query_raw(query[0], *query[1]) -def _reported_flat_cost(record: DailySpendRecord | _GroupingSetsRow) -> float: +def _reported_flat_cost(record: DailySpendRecord | _RollupMetricsRow) -> float: """Flat cost a daily row reports, which is zero unless PTU cost attribution is enabled. Both read paths funnel through here: the paginated path reads the ``ptu_flat_cost`` @@ -1008,7 +1011,7 @@ _GROUP_DATE_ENDPOINT: Final = 62 # 0b0111110 _GROUP_DATE_ENDPOINT_API_KEY: Final = 30 # 0b0011110 -def _record_to_spend_metrics(record: _GroupingSetsRow) -> SpendMetrics: +def _record_to_spend_metrics(record: _RollupMetricsRow) -> SpendMetrics: """Build a SpendMetrics directly from one already-aggregated rollup row. SUM() over zero rows is SQL NULL, so rollup rows (notably the grand-total From b74733d4e6971984d3af6e78c565ff37c47cffe2 Mon Sep 17 00:00:00 2001 From: yassin Date: Wed, 16 Sep 2026 10:53:04 +0000 Subject: [PATCH 050/141] feat(ui): note on the cache leakage card when only the top-spend keys were loaded Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../_components/CacheLeakageCard.test.tsx | 24 +++++++++++++++++++ .../_components/CacheLeakageCard.tsx | 9 ++++++- .../useDailyActivityRange.test.tsx | 17 ++++++++++++- .../_components/useDailyActivityRange.ts | 3 +++ 4 files changed, 51 insertions(+), 2 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.test.tsx index 54af13d8a90..8cb18344734 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.test.tsx @@ -178,4 +178,28 @@ describe("CacheLeakageCard", () => { screen.queryByText("Data is still loading; rows and totals will update as the rest of the range arrives."), ).not.toBeInTheDocument(); }); + + it("says which keys are missing from the key ranking when the proxy capped the per-key lists", () => { + const day = dayWithKeys("2026-07-12", { + "hash-leaky": key("leaky-key", { prompt_tokens: 10000, cache_read_input_tokens: 0 }), + }); + renderWith([day], { apiKeyTruncation: { limit: 100, total: 3000 } }); + + expect(screen.getByRole("note")).toHaveTextContent( + "Only the 100 highest-spend keys of 3,000 are loaded, so a lower-spend key that leaks more is not listed here.", + ); + + fireEvent.click(screen.getByRole("tab", { name: "By model" })); + + expect(screen.queryByRole("note")).not.toBeInTheDocument(); + }); + + it("keeps the key ranking note off when every key was loaded", () => { + const day = dayWithKeys("2026-07-12", { + "hash-leaky": key("leaky-key", { prompt_tokens: 10000, cache_read_input_tokens: 0 }), + }); + renderWith([day]); + + expect(screen.queryByRole("note")).not.toBeInTheDocument(); + }); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.tsx index 3f27449ebe1..a0877b04648 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.tsx @@ -81,7 +81,7 @@ const SortableHead = ({ }; const CacheLeakageCard: React.FC = ({ activity }) => { - const { dateValue, onDateChange, results, loading, isFetchingMore } = activity; + const { dateValue, onDateChange, results, loading, isFetchingMore, apiKeyTruncation } = activity; const [dimension, setDimension] = useState("key"); const [sort, setSort] = useState({ column: "potentialSavings", dir: "desc" }); const leakage = useMemo(() => computeCacheLeakage(results, dimension), [results, dimension]); @@ -123,6 +123,13 @@ const CacheLeakageCard: React.FC = ({ activity }) => { + {dimension === "key" && apiKeyTruncation !== undefined && ( +

+ Only the {apiKeyTruncation.limit.toLocaleString()} highest-spend keys of{" "} + {apiKeyTruncation.total.toLocaleString()} are loaded, so a lower-spend key that leaks more is not listed + here. Raise USAGE_TOP_API_KEYS_LIMIT on the proxy to load more keys. +

+ )} {rows.length > 0 && isFetchingMore && (

Data is still loading; rows and totals will update as the rest of the range arrives. diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx index 00902aa9fdd..4059303d5a5 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx @@ -4,12 +4,13 @@ import { describe, expect, it, vi } from "vitest"; const mockUsePaginatedDailyActivity = vi.fn(); const mockCancel = vi.fn(); +let mockMetadata: Record = {}; vi.mock("@/app/(dashboard)/usage/_components/hooks/usePaginatedDailyActivity", () => ({ usePaginatedDailyActivity: (args: unknown) => { mockUsePaginatedDailyActivity(args); return { - data: { results: [] }, + data: { results: [], metadata: mockMetadata }, loading: false, isFetchingMore: false, progress: { currentPage: 4, totalPages: 9 }, @@ -80,4 +81,18 @@ describe("useDailyActivityRange", () => { expect(mockUsePaginatedDailyActivity).toHaveBeenLastCalledWith(expect.objectContaining({ enabled: false })); }); + + it("reports how many keys the proxy left out of the per-key lists", () => { + mockMetadata = { api_key_limit: 100, total_api_keys: 3000 }; + const { result } = renderHook(() => useDailyActivityRange("test-token", "u1", "proxy_admin")); + + expect(result.current.apiKeyTruncation).toEqual({ limit: 100, total: 3000 }); + }); + + it("reports no key truncation when every key fit under the proxy limit", () => { + mockMetadata = { api_key_limit: 100, total_api_keys: 100 }; + const { result } = renderHook(() => useDailyActivityRange("test-token", "u1", "proxy_admin")); + + expect(result.current.apiKeyTruncation).toBeUndefined(); + }); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts index 9f793a68bf5..92dd24b8d6d 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts @@ -1,6 +1,7 @@ import { useMemo, useState } from "react"; import { userDailyActivityAggregatedCall, userDailyActivityCall } from "@/components/networking"; +import { ApiKeyTruncation, getApiKeyTruncation } from "@/components/EntityUsageExport/exportBlockedReason"; import { DailyData } from "@/components/UsagePage/types"; import { spendScopeUserId } from "@/utils/roles"; import { usePaginatedDailyActivity } from "@/app/(dashboard)/usage/_components/hooks/usePaginatedDailyActivity"; @@ -22,6 +23,7 @@ export interface DailyActivityRange { cancelled: boolean; failed: boolean; cancel: () => void; + apiKeyTruncation?: ApiKeyTruncation; } /** @@ -78,6 +80,7 @@ export const useScopedDailyActivityRange = ( cancelled, failed, cancel, + apiKeyTruncation: getApiKeyTruncation(data.metadata?.api_key_limit, data.metadata?.total_api_keys), }; }; From 4148bf283c917423b30ba8bf6c5c79b0fc1983e5 Mon Sep 17 00:00:00 2001 From: yucheng Date: Wed, 16 Sep 2026 23:08:03 +0000 Subject: [PATCH 051/141] fix(proxy): catch only Redis failures when falling back to local login counters Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 4ab57ca461f..59d91c12d6e 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -23,10 +23,11 @@ from typing import Final, Literal, NamedTuple, NoReturn from fastapi import Request, status from pydantic import TypeAdapter, ValidationError +from redis.exceptions import RedisError from litellm._logging import verbose_proxy_logger from litellm.caching.in_memory_cache import InMemoryCache -from litellm.caching.redis_cache import RedisCache +from litellm.caching.redis_cache import RedisCache, RedisCircuitBreakerOpenError from litellm.proxy._types import ProxyErrorTypes, ProxyException from litellm.proxy.auth.network import TrustedProxyConfig, normalize_cidr_ranges, resolve_client_ip from litellm.secret_managers.main import get_secret_bool @@ -53,6 +54,7 @@ _MAX_TRACKED_COUNTERS: Final = 20_000 _MAX_TRACKED_BLOCKS: Final = 10_000 _NO_SETTINGS: Final[Mapping[str, object]] = MappingProxyType({}) _NOT_BLOCKED: Final = (0, 0) +_REDIS_FAILURES: Final = (RedisError, RedisCircuitBreakerOpenError, OSError, asyncio.TimeoutError) _LOCAL_BLOCK_EXPIRY: Final = TypeAdapter[float | None](float | None) _SOURCE_LIMIT_OVERRIDES: Final = TypeAdapter[Mapping[str, object]](Mapping[str, object]) @@ -304,7 +306,7 @@ class LoginThrottle: return _LUA_BLOCK_TTLS.validate_python( await self.redis_cache.async_register_script(_BLOCK_TTLS_LUA)(list(keys), []) ) - except Exception as err: + except _REDIS_FAILURES as err: self._warn_redis(err) return _NOT_BLOCKED @@ -327,7 +329,7 @@ class LoginThrottle: list(keys), [self.user_limit, source_limit, self.window_seconds, self.block_seconds] ) ) - except Exception as err: + except _REDIS_FAILURES as err: self._warn_redis(err) user_block: Final = self._local_bump(keys.pair_counter, keys.pair_block, self.user_limit) if source_limit == 0 or user_block > 0: @@ -349,7 +351,7 @@ class LoginThrottle: if self.redis_cache is not None: try: await self.redis_cache.async_delete_cache(pair_counter) - except Exception as err: + except _REDIS_FAILURES as err: self._warn_redis(err) self.counters.delete_cache(pair_counter) From 438b4e6a3f34249b61a83c1d6d959daa64f56e21 Mon Sep 17 00:00:00 2001 From: yucheng Date: Wed, 16 Sep 2026 23:15:35 +0000 Subject: [PATCH 052/141] fix(proxy): type the login throttle's local store and pass frozen Redis script arguments Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 33 +++++++++++++++++++--------- 1 file changed, 23 insertions(+), 10 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 59d91c12d6e..d17e3fd0d53 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -19,7 +19,7 @@ from contextlib import asynccontextmanager from dataclasses import dataclass from functools import cache from types import MappingProxyType -from typing import Final, Literal, NamedTuple, NoReturn +from typing import Final, Literal, NamedTuple, NoReturn, Protocol, TypeAlias from fastapi import Request, status from pydantic import TypeAdapter, ValidationError @@ -58,11 +58,24 @@ _REDIS_FAILURES: Final = (RedisError, RedisCircuitBreakerOpenError, OSError, asy _LOCAL_BLOCK_EXPIRY: Final = TypeAdapter[float | None](float | None) _SOURCE_LIMIT_OVERRIDES: Final = TypeAdapter[Mapping[str, object]](Mapping[str, object]) -Scope = Literal["user", "source"] +Scope: TypeAlias = Literal["user", "source"] -_BlockTtls = tuple[int, int] +_BlockTtls: TypeAlias = tuple[int, int] _LUA_BLOCK_TTLS: Final = TypeAdapter[_BlockTtls](_BlockTtls) -_Network = ipaddress.IPv4Network | ipaddress.IPv6Network +_Network: TypeAlias = ipaddress.IPv4Network | ipaddress.IPv6Network + + +class LocalStore(Protocol): + """The per-worker store behind the counters and blocks; ``InMemoryCache`` satisfies it.""" + + def get_cache(self, key: str) -> object: ... + + def set_cache(self, key: str, value: float, *, ttl: int) -> None: ... + + def increment_cache(self, key: str, value: float, *, ttl: int) -> float: ... + + def delete_cache(self, key: str) -> None: ... + # KEYS: pair counter, pair block, source counter, source block (one cluster slot via the source hash tag) # ARGV: pair limit, source limit (0 = source scope off), window seconds, block seconds @@ -87,7 +100,7 @@ _COUNTERS: Final = InMemoryCache( max_size_in_memory=_MAX_TRACKED_COUNTERS, default_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS ) _BLOCKS: Final = InMemoryCache(max_size_in_memory=_MAX_TRACKED_BLOCKS, default_ttl=DEFAULT_FAILED_LOGIN_BLOCK_SECONDS) -_HELD_ATTEMPTS: Final[dict[str, int]] = {} +_HELD_ATTEMPTS: Final[dict[str, int]] = {} # mutable-ok: in-flight hold counts rise on entry and fall on exit async def _sleep(seconds: float) -> None: @@ -216,8 +229,8 @@ class LoginThrottle: user_limit: int window_seconds: int block_seconds: int - counters: InMemoryCache - blocks: InMemoryCache + counters: LocalStore + blocks: LocalStore redis_cache: RedisCache | None = None enabled: bool = True @@ -304,7 +317,7 @@ class LoginThrottle: return _NOT_BLOCKED try: return _LUA_BLOCK_TTLS.validate_python( - await self.redis_cache.async_register_script(_BLOCK_TTLS_LUA)(list(keys), []) + await self.redis_cache.async_register_script(_BLOCK_TTLS_LUA)(keys, ()) ) except _REDIS_FAILURES as err: self._warn_redis(err) @@ -326,7 +339,7 @@ class LoginThrottle: try: return _LUA_BLOCK_TTLS.validate_python( await self.redis_cache.async_register_script(_RECORD_FAILURE_LUA)( - list(keys), [self.user_limit, source_limit, self.window_seconds, self.block_seconds] + keys, (self.user_limit, source_limit, self.window_seconds, self.block_seconds) ) ) except _REDIS_FAILURES as err: @@ -369,7 +382,7 @@ class LoginThrottle: type=ProxyErrorTypes.auth_error, param="username", code=status.HTTP_429_TOO_MANY_REQUESTS, - headers={"Retry-After": str(retry_after)}, + headers={"Retry-After": str(retry_after)}, # mutable-ok: ProxyException writes into its headers dict ) From 0a8423d77b7fe99e572d8ea5e923fcd05c6985a2 Mon Sep 17 00:00:00 2001 From: yucheng Date: Thu, 17 Sep 2026 00:28:39 +0000 Subject: [PATCH 053/141] fix(proxy): keep the sign-in hold pool from refusing a correct password The held-attempt cap ran before the password check, so five parked wrong guesses from a blocked source turned the soft block into a lockout for the real user. The slot is now taken only after a wrong password, and the pool-full refusal carries the block's remaining time as Retry-After Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 49 ++++++++++--------- .../proxy/auth/test_login_utils.py | 44 ++++++++++++++++- 2 files changed, 69 insertions(+), 24 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index d17e3fd0d53..a8899b9c675 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -281,25 +281,7 @@ class LoginThrottle: yield LoginAttempt(throttle=self, username=username, block=None) return slot: Final = keys.pair_block if block.scope == "user" else keys.source_block - held: Final = _HELD_ATTEMPTS.get(slot, 0) - if held >= MAX_HELD_ATTEMPTS_PER_KEY: - verbose_proxy_logger.warning( - "Admin UI sign-in refused: %s attempts already held for a blocked %s; username=%r source=%s", - held, - block.scope, - username, - self.client_ip, - ) - self.refuse(BLOCKED_ATTEMPT_HOLD_SECONDS) - _HELD_ATTEMPTS[slot] = held + 1 - try: - yield LoginAttempt(throttle=self, username=username, block=block) - finally: - remaining: Final = _HELD_ATTEMPTS.get(slot, 1) - 1 - if remaining > 0: - _HELD_ATTEMPTS[slot] = remaining - else: - _HELD_ATTEMPTS.pop(slot, None) + yield LoginAttempt(throttle=self, username=username, block=block, slot=slot) async def _active_block(self, keys: _Keys) -> Block | None: local: Final = self._local_block_ttls(keys) @@ -391,6 +373,7 @@ class LoginAttempt: throttle: LoginThrottle username: str block: Block | None + slot: str | None = None async def succeeded(self) -> None: if not self.throttle.enabled: @@ -400,9 +383,8 @@ class LoginAttempt: async def failed(self) -> None: if not self.throttle.enabled: return - if self.block is not None: - await _sleep(BLOCKED_ATTEMPT_HOLD_SECONDS) - self.throttle.refuse(max(self.block.retry_after - BLOCKED_ATTEMPT_HOLD_SECONDS, 1)) + if self.block is not None and self.slot is not None: + await self._hold_then_refuse(self.block, self.slot) user_block, source_block = await self.throttle.record_failure(self.username) if user_block == 0 and source_block == 0: return @@ -413,3 +395,26 @@ class LoginAttempt: self.username, self.throttle.client_ip, ) + + async def _hold_then_refuse(self, block: Block, slot: str) -> NoReturn: + held: Final = _HELD_ATTEMPTS.get(slot, 0) + if held >= MAX_HELD_ATTEMPTS_PER_KEY: + verbose_proxy_logger.warning( + "Admin UI sign-in refused at once: %s wrong attempts already held for a blocked %s; " + "username=%r source=%s", + held, + block.scope, + self.username, + self.throttle.client_ip, + ) + self.throttle.refuse(block.retry_after) + _HELD_ATTEMPTS[slot] = held + 1 + try: + await _sleep(BLOCKED_ATTEMPT_HOLD_SECONDS) + finally: + remaining: Final = _HELD_ATTEMPTS.get(slot, 1) - 1 + if remaining > 0: + _HELD_ATTEMPTS[slot] = remaining + else: + _HELD_ATTEMPTS.pop(slot, None) + self.throttle.refuse(max(block.retry_after - BLOCKED_ATTEMPT_HOLD_SECONDS, 1)) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index df5bd29f8ff..df35e916aa0 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1211,7 +1211,7 @@ async def test_held_attempts_from_one_blocked_key_are_capped(monkeypatch): with pytest.raises(ProxyException) as over_cap: await _guess(throttle) assert over_cap.value.code == "429" - assert over_cap.value.headers.get("Retry-After") == "30" + assert over_cap.value.headers.get("Retry-After") == "300", "refused at once, for the whole block" assert await _fail(throttle, username="someone-else@corp.com") == "401", "other keys are not affected" finally: release.set() @@ -1222,6 +1222,46 @@ async def test_held_attempts_from_one_blocked_key_are_capped(monkeypatch): assert lt._HELD_ATTEMPTS.get(slot) is None, "the slots are released once the held attempts answer" +@pytest.mark.asyncio +async def test_a_full_hold_pool_still_lets_the_right_password_in(monkeypatch): + """Five parked wrong guesses from the office must not turn the soft block into a lockout for the real user.""" + import asyncio + + from litellm.proxy.auth import login_throttle as lt + from litellm.proxy.auth.login_throttle import MAX_HELD_ATTEMPTS_PER_KEY + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + release = asyncio.Event() + + async def _park(_seconds: float) -> None: + await release.wait() + + monkeypatch.setattr(lt, "_sleep", _park) + throttle = _throttle(user_limit=1, source_limit=3, client_ip="203.0.113.46") + assert [await _fail(throttle, username=f"spray-{i}@corp.com") for i in range(4)] == ["401"] * 4 + source_slot = throttle._keys("known@example.com").source_block + assert throttle._local_block_ttl(source_slot) > 0, "the source is blocked" + + held = [ + asyncio.create_task(_guess(throttle, username="known@example.com")) for _ in range(MAX_HELD_ATTEMPTS_PER_KEY) + ] + for _ in range(1000): + if lt._HELD_ATTEMPTS.get(source_slot) == MAX_HELD_ATTEMPTS_PER_KEY: + break + await asyncio.sleep(0) + assert lt._HELD_ATTEMPTS == {source_slot: MAX_HELD_ATTEMPTS_PER_KEY} + + try: + signed_in = await _db_login(throttle, "known@example.com", "right", correct=True) + assert signed_in.user_id == "u-1" + finally: + release.set() + for task in held: + with pytest.raises(ProxyException): + await task + + @pytest.mark.asyncio async def test_a_blocked_source_shares_one_held_slot_pool_across_its_blocked_usernames(monkeypatch): """Once the source is blocked, a pair block for a username must not hand that username its own five slots.""" @@ -1258,7 +1298,7 @@ async def test_a_blocked_source_shares_one_held_slot_pool_across_its_blocked_use with pytest.raises(ProxyException) as over_cap: await _guess(throttle, username=name) assert over_cap.value.code == "429" - assert over_cap.value.headers.get("Retry-After") == "30" + assert over_cap.value.headers.get("Retry-After") == "300" finally: release.set() for task in held: From 8a645bcc00dc07e9d4b14b2b735596623e8b8947 Mon Sep 17 00:00:00 2001 From: yucheng Date: Thu, 17 Sep 2026 00:36:12 +0000 Subject: [PATCH 054/141] test(proxy): stub DATABASE_URL in the hold-pool regression test so it passes off the dev box Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/proxy/auth/test_login_utils.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index df35e916aa0..d1fdb2d6a70 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1232,6 +1232,7 @@ async def test_a_full_hold_pool_still_lets_the_right_password_in(monkeypatch): monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") release = asyncio.Event() async def _park(_seconds: float) -> None: From ed18edbbdd39ce326ed0448250b3ea767c94c4d1 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 17 Sep 2026 01:56:05 +0000 Subject: [PATCH 055/141] chore(proxy): drop a comment that restated the NUM_WORKERS assignment Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/proxy_cli.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/litellm/proxy/proxy_cli.py b/litellm/proxy/proxy_cli.py index 1ade0652855..9f2e4c9802e 100644 --- a/litellm/proxy/proxy_cli.py +++ b/litellm/proxy/proxy_cli.py @@ -1411,8 +1411,6 @@ def run_server( # DO NOT DELETE - enables global variables to work across files from litellm.proxy.proxy_server import app - # Write the resolved --num_workers back to its env var so worker processes can read - # the fleet size at startup (the failed-login accounting warning keys off it) os.environ["NUM_WORKERS"] = str(num_workers) # Auto-create PROMETHEUS_MULTIPROC_DIR for multi-worker setups From 2e8dc0a627b552d408910d84628692eca5452be3 Mon Sep 17 00:00:00 2001 From: yassin Date: Thu, 17 Sep 2026 03:03:19 +0000 Subject: [PATCH 056/141] feat(proxy): add Azure AI Speech pass-through route Adds /azure_speech/{endpoint:path}, an authenticated pass-through for the Azure AI Speech REST APIs: short-audio recognition on .stt.speech.microsoft.com and batch transcription on .api.cognitive.microsoft.com. The proxy resolves the subscription key through PassthroughEndpointRouter (AZURE_SPEECH_API_KEY or an Admin UI credential), picks the host from AZURE_SPEECH_REGION or AZURE_SPEECH_API_BASE, injects Ocp-Apim-Subscription-Key, strips the caller's Authorization and subscription-key headers, forwards the raw audio body byte for byte, and records a zero-cost SpendLogs row tagged azure_speech since the price map has no Azure Speech STT entry Resolves LIT-7939 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- gateway/routes/allowlist.py | 1 + helm/litellm/templates/ingress.yaml | 2 +- litellm/constants.py | 10 + litellm/passthrough/utils.py | 1 + litellm/proxy/_lazy_features.py | 1 + litellm/proxy/_lazy_openapi_snapshot.json | 222 +++++++++++++ litellm/proxy/_types.py | 1 + litellm/proxy/auth/user_api_key_auth.py | 7 + .../proxy/common_utils/http_parsing_utils.py | 13 +- .../llm_passthrough_endpoints.py | 118 +++++++ ...zure_speech_passthrough_logging_handler.py | 84 +++++ .../pass_through_endpoints.py | 3 +- .../pass_through_endpoints/success_handler.py | 24 +- .../provider_create_fields.json | 18 ++ litellm/types/utils.py | 1 + terraform/litellm/aws/locals.tf | 2 +- terraform/litellm/gcp/locals.tf | 2 +- ...est_billable_request_metrics_middleware.py | 5 + ...zure_speech_passthrough_logging_handler.py | 123 ++++++++ .../test_llm_pass_through_endpoints.py | 294 ++++++++++++++++++ .../test_passthrough_endpoint_router.py | 16 + .../src/components/provider_info_helpers.tsx | 4 + ui/litellm-dashboard/src/lib/http/schema.d.ts | 231 ++++++++++++++ 23 files changed, 1177 insertions(+), 6 deletions(-) create mode 100644 litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py create mode 100644 tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py diff --git a/gateway/routes/allowlist.py b/gateway/routes/allowlist.py index 099c6d5179f..5b8c44809fe 100644 --- a/gateway/routes/allowlist.py +++ b/gateway/routes/allowlist.py @@ -82,6 +82,7 @@ GATEWAY_PATH_PREFIXES: tuple[str, ...] = ( "/anthropic/", "/azure/", "/azure_ai/", + "/azure_speech/", "/aws/", "/bedrock/", "/comprehendmedical", diff --git a/helm/litellm/templates/ingress.yaml b/helm/litellm/templates/ingress.yaml index d42558b9396..94ac4d2d8b0 100644 --- a/helm/litellm/templates/ingress.yaml +++ b/helm/litellm/templates/ingress.yaml @@ -66,7 +66,7 @@ "/v1/video" "/v1/videos" "/video" "/videos" "/v1/search" "/search" "/v1/containers" "/containers" "/v1/evals" "/v1/memory" "/queue/chat" "/v1beta" "/interactions" - "/anthropic" "/azure" "/azure_ai" "/aws" "/bedrock" "/comprehendmedical" "/cohere" "/gemini" "/google" + "/anthropic" "/azure" "/azure_ai" "/azure_speech" "/aws" "/bedrock" "/comprehendmedical" "/cohere" "/gemini" "/google" "/vertex_ai" "/vertex-ai" "/assemblyai" "/eu.assemblyai" "/langfuse" "/vllm" "/mistral" "/groq" "/voyage" "/cursor" "/milvus" "/openai_passthrough" "/toolset" diff --git a/litellm/constants.py b/litellm/constants.py index 8409a161800..69de889a326 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1570,6 +1570,16 @@ ALLOWED_VERTEX_AI_PASSTHROUGH_HEADERS: Final = { # Works for all LLM pass-through endpoints (Vertex AI, Anthropic, Bedrock, etc.) PASS_THROUGH_HEADER_PREFIX: Final = "x-pass-" +AZURE_SPEECH_CUSTOM_LLM_PROVIDER: Final = "azure_speech" +AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX: Final = "/azure_speech" +AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX: Final = "/speech/" +AZURE_SPEECH_BATCH_PATH_PREFIX: Final = "/speechtotext/" +AZURE_SPEECH_STT_DOMAIN: Final = "stt.speech.microsoft.com" +AZURE_SPEECH_COGNITIVE_SERVICES_DOMAIN: Final = "api.cognitive.microsoft.com" +AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER: Final = "Ocp-Apim-Subscription-Key" +AZURE_SPEECH_SHORT_AUDIO_MODEL: Final = "short-audio" +AZURE_SPEECH_BATCH_MODEL: Final = "batch-transcription" + BASE_MCP_ROUTE: Final = "/mcp" BATCH_STATUS_POLL_INTERVAL_SECONDS: Final = int(os.getenv("BATCH_STATUS_POLL_INTERVAL_SECONDS", 3600)) # 1 hour diff --git a/litellm/passthrough/utils.py b/litellm/passthrough/utils.py index 7eb14fcc118..452c9c7de9d 100644 --- a/litellm/passthrough/utils.py +++ b/litellm/passthrough/utils.py @@ -17,6 +17,7 @@ _PASS_THROUGH_PROTECTED_HEADERS: Final[frozenset] = frozenset( "api-key", "x-api-key", "x-goog-api-key", + "ocp-apim-subscription-key", "host", "content-length", "accept-encoding", diff --git a/litellm/proxy/_lazy_features.py b/litellm/proxy/_lazy_features.py index faf95397fa5..92cc014967d 100644 --- a/litellm/proxy/_lazy_features.py +++ b/litellm/proxy/_lazy_features.py @@ -196,6 +196,7 @@ LAZY_FEATURES: Final[tuple[LazyFeature, ...]] = ( "/assemblyai/", "/azure/", "/azure_ai/", + "/azure_speech/", "/bedrock/", "/cohere/", "/comprehendmedical", diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index 74f38b3ca6d..dbfbc317d24 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -17133,6 +17133,228 @@ ] } }, + "/azure_speech/{endpoint}": { + "delete": { + "description": "Pass-through for the Azure AI Speech REST APIs (speech to text), e.g.\n`POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US`\nwith the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`.\n\nThe body is forwarded byte for byte and the proxy injects its own\n`Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key\nand is never forwarded.\n\n[Docs](https://docs.litellm.ai/docs/pass_through/azure_speech)", + "operationId": "azure_speech_proxy_route_azure_speech__endpoint__delete", + "parameters": [ + { + "in": "path", + "name": "endpoint", + "required": true, + "schema": { + "title": "Endpoint", + "type": "string" + } + } + ], + "responses": { + "200": { + "content": { + "application/json": { + "schema": {} + } + }, + "description": "Successful Response" + }, + "422": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/HTTPValidationError" + } + } + }, + "description": "Validation Error" + } + }, + "security": [ + { + "APIKeyHeader": [] + } + ], + "summary": "Azure Speech Proxy Route", + "tags": [ + "llm_passthrough" + ] + }, + "get": { + "description": "Pass-through for the Azure AI Speech REST APIs (speech to text), e.g.\n`POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US`\nwith the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`.\n\nThe body is forwarded byte for byte and the proxy injects its own\n`Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key\nand is never forwarded.\n\n[Docs](https://docs.litellm.ai/docs/pass_through/azure_speech)", + "operationId": "azure_speech_proxy_route_azure_speech__endpoint__get", + "parameters": [ + { + "in": "path", + "name": "endpoint", + "required": true, + "schema": { + "title": "Endpoint", + "type": "string" + } + } + ], + "responses": { + "200": { + "content": { + "application/json": { + "schema": {} + } + }, + "description": "Successful Response" + }, + "422": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/HTTPValidationError" + } + } + }, + "description": "Validation Error" + } + }, + "security": [ + { + "APIKeyHeader": [] + } + ], + "summary": "Azure Speech Proxy Route", + "tags": [ + "llm_passthrough" + ] + }, + "patch": { + "description": "Pass-through for the Azure AI Speech REST APIs (speech to text), e.g.\n`POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US`\nwith the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`.\n\nThe body is forwarded byte for byte and the proxy injects its own\n`Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key\nand is never forwarded.\n\n[Docs](https://docs.litellm.ai/docs/pass_through/azure_speech)", + "operationId": "azure_speech_proxy_route_azure_speech__endpoint__patch", + "parameters": [ + { + "in": "path", + "name": "endpoint", + "required": true, + "schema": { + "title": "Endpoint", + "type": "string" + } + } + ], + "responses": { + "200": { + "content": { + "application/json": { + "schema": {} + } + }, + "description": "Successful Response" + }, + "422": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/HTTPValidationError" + } + } + }, + "description": "Validation Error" + } + }, + "security": [ + { + "APIKeyHeader": [] + } + ], + "summary": "Azure Speech Proxy Route", + "tags": [ + "llm_passthrough" + ] + }, + "post": { + "description": "Pass-through for the Azure AI Speech REST APIs (speech to text), e.g.\n`POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US`\nwith the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`.\n\nThe body is forwarded byte for byte and the proxy injects its own\n`Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key\nand is never forwarded.\n\n[Docs](https://docs.litellm.ai/docs/pass_through/azure_speech)", + "operationId": "azure_speech_proxy_route_azure_speech__endpoint__post", + "parameters": [ + { + "in": "path", + "name": "endpoint", + "required": true, + "schema": { + "title": "Endpoint", + "type": "string" + } + } + ], + "responses": { + "200": { + "content": { + "application/json": { + "schema": {} + } + }, + "description": "Successful Response" + }, + "422": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/HTTPValidationError" + } + } + }, + "description": "Validation Error" + } + }, + "security": [ + { + "APIKeyHeader": [] + } + ], + "summary": "Azure Speech Proxy Route", + "tags": [ + "llm_passthrough" + ] + }, + "put": { + "description": "Pass-through for the Azure AI Speech REST APIs (speech to text), e.g.\n`POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US`\nwith the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`.\n\nThe body is forwarded byte for byte and the proxy injects its own\n`Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key\nand is never forwarded.\n\n[Docs](https://docs.litellm.ai/docs/pass_through/azure_speech)", + "operationId": "azure_speech_proxy_route_azure_speech__endpoint__put", + "parameters": [ + { + "in": "path", + "name": "endpoint", + "required": true, + "schema": { + "title": "Endpoint", + "type": "string" + } + } + ], + "responses": { + "200": { + "content": { + "application/json": { + "schema": {} + } + }, + "description": "Successful Response" + }, + "422": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/HTTPValidationError" + } + } + }, + "description": "Validation Error" + } + }, + "security": [ + { + "APIKeyHeader": [] + } + ], + "summary": "Azure Speech Proxy Route", + "tags": [ + "llm_passthrough" + ] + } + }, "/bedrock/{endpoint}": { "delete": { "description": "This is the v1 passthrough for Bedrock.\nV2 is handled by the `/bedrock/v2` endpoint.\n[Docs](https://docs.litellm.ai/docs/pass_through/bedrock)", diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 680c63393e8..489186a8f69 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -468,6 +468,7 @@ class LiteLLMRoutes(enum.Enum): mapped_pass_through_routes = [ "/bedrock", "/comprehendmedical", + "/azure_speech", "/vertex-ai", "/vertex_ai", "/cohere", diff --git a/litellm/proxy/auth/user_api_key_auth.py b/litellm/proxy/auth/user_api_key_auth.py index 4cbd4213463..5f2e0a3c1a8 100644 --- a/litellm/proxy/auth/user_api_key_auth.py +++ b/litellm/proxy/auth/user_api_key_auth.py @@ -105,6 +105,7 @@ from litellm.proxy.common_utils.http_parsing_utils import ( _safe_get_request_headers, _safe_get_request_query_params, _safe_set_request_parsed_body, + is_opaque_audio_pass_through_request, populate_request_with_path_params, read_raw_json_body, rewrite_request_model, @@ -1354,6 +1355,12 @@ async def _read_request_body_deferring_parse_failure( must run (resolving identity onto the request's trace) before the 400 goes out; the caller re-raises the returned exception once identity is seeded. """ + if is_opaque_audio_pass_through_request( + route=get_request_route(request=request), + content_type=_safe_get_request_headers(request=request).get("content-type", ""), + ): + _safe_set_request_parsed_body(request=request, parsed_body={}) # mutable-ok: the body cache stores a plain dict + return {}, None # mutable-ok: request_data is a plain dict across the whole auth path try: parsed_body: Final = await _read_request_body(request=request) except ProxyException as parse_exception: diff --git a/litellm/proxy/common_utils/http_parsing_utils.py b/litellm/proxy/common_utils/http_parsing_utils.py index f5b6a0a766d..29dc36f3dba 100644 --- a/litellm/proxy/common_utils/http_parsing_utils.py +++ b/litellm/proxy/common_utils/http_parsing_utils.py @@ -9,7 +9,12 @@ from fastapi import Request, UploadFile, status from typing_extensions import NotRequired, ReadOnly, Required from litellm._logging import verbose_proxy_logger -from litellm.constants import CLIENT_REQUESTED_MODEL_SCOPE_KEY, MAX_REQUEST_BODY_SIZE_TO_REPAIR_MB +from litellm.constants import ( + AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX, + AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, + CLIENT_REQUESTED_MODEL_SCOPE_KEY, + MAX_REQUEST_BODY_SIZE_TO_REPAIR_MB, +) from litellm.proxy._types import ProxyException from litellm.proxy.common_utils.callback_utils import ( get_metadata_variable_name_from_kwargs, @@ -214,6 +219,12 @@ async def _read_request_body(request: Request | None) -> dict: return {} +def is_opaque_audio_pass_through_request(route: str, content_type: str) -> bool: + return route.startswith( + f"{AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX}{AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX}" + ) and _normalize_media_type(content_type).startswith("audio/") + + async def read_raw_json_body(request: Request | None) -> bytes | None: if request is None or _safe_get_request_parsed_body(request=request) is None: return None diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index b9b8cb3a22b..e64eac87a7f 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -30,6 +30,13 @@ from litellm import get_llm_provider from litellm._logging import verbose_proxy_logger from litellm.constants import ( ALLOWED_VERTEX_AI_PASSTHROUGH_HEADERS, + AZURE_SPEECH_BATCH_PATH_PREFIX, + AZURE_SPEECH_COGNITIVE_SERVICES_DOMAIN, + AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX, + AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, + AZURE_SPEECH_STT_DOMAIN, + AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER, BEDROCK_AGENT_RUNTIME_PASS_THROUGH_ROUTES, ) from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix @@ -1316,6 +1323,117 @@ async def comprehend_medical_sdk_proxy_route( ) +AZURE_SPEECH_FORWARDED_REQUEST_HEADERS: Final = ("content-type", "accept") +AZURE_SPEECH_ENDPOINT_FAMILY_DOMAINS: Final = MappingProxyType( + { + AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX: AZURE_SPEECH_STT_DOMAIN, + AZURE_SPEECH_BATCH_PATH_PREFIX: AZURE_SPEECH_COGNITIVE_SERVICES_DOMAIN, + } +) + + +def resolve_azure_speech_base_url(endpoint_path: str, api_base: str | None, region: str | None) -> httpx.URL | None: + """ + Azure AI Speech serves the two REST families from different regional hosts: short-audio + recognition under ``{region}.stt.speech.microsoft.com`` and batch transcription under + ``{region}.api.cognitive.microsoft.com``. An operator-configured ``api_base`` (custom + domain or private endpoint) serves both and wins over the region. Returns ``None`` when + the path is outside both families so the operator key is never sent for an unknown API. + """ + domain: Final = next( + ( + family_domain + for family_prefix, family_domain in AZURE_SPEECH_ENDPOINT_FAMILY_DOMAINS.items() + if endpoint_path.startswith(family_prefix) + ), + None, + ) + if domain is None: + return None + if api_base: + return httpx.URL(api_base) + if not region: + return None + return httpx.URL(f"https://{region}.{domain}") + + +@router.api_route( + f"{AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX}/{{endpoint:path}}", + methods=["GET", "POST", "PUT", "DELETE", "PATCH"], # mutable-ok: fastapi route methods must be a list + tags=["Azure AI Speech Pass-through", "pass-through"], # mutable-ok: fastapi route tags must be a list +) +async def azure_speech_proxy_route( + endpoint: str, + request: Request, + fastapi_response: Response, + user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)], +): + """ + Pass-through for the Azure AI Speech REST APIs (speech to text), e.g. + `POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US` + with the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`. + + The body is forwarded byte for byte and the proxy injects its own + `Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key + and is never forwarded. + + [Docs](https://docs.litellm.ai/docs/pass_through/azure_speech) + """ + endpoint_path: Final = httpx.URL(endpoint).path + normalized_endpoint_path: Final = endpoint_path if endpoint_path.startswith("/") else f"/{endpoint_path}" + base_url: Final = resolve_azure_speech_base_url( + endpoint_path=normalized_endpoint_path, + api_base=get_secret_str(secret_name="AZURE_SPEECH_API_BASE"), + region=get_secret_str(secret_name="AZURE_SPEECH_REGION"), + ) + if base_url is None: + raise HTTPException( + status_code=400, + detail=( + f"Unsupported Azure Speech path: {normalized_endpoint_path}. Supported prefixes are " + f"{AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX} and {AZURE_SPEECH_BATCH_PATH_PREFIX}; set " + "AZURE_SPEECH_REGION or AZURE_SPEECH_API_BASE in the proxy environment." + ), + ) + azure_speech_api_key: Final = passthrough_endpoint_router.get_credentials( + custom_llm_provider=AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + region_name=None, + ) + if azure_speech_api_key is None: + raise HTTPException( + status_code=400, + detail="Azure Speech credentials not found. Set AZURE_SPEECH_API_KEY in the proxy environment.", + ) + + target_url: Final = base_url.copy_with( + path=HttpPassThroughEndpointHelpers.join_base_and_endpoint_path(base_url, normalized_endpoint_path) + ) + request_headers: Final = _safe_get_request_headers(request) + upstream_headers: Final = MappingProxyType( + { + header_name: header_value + for header_name, header_value in ( + *( + (header_name, request_headers[header_name]) + for header_name in AZURE_SPEECH_FORWARDED_REQUEST_HEADERS + if header_name in request_headers + ), + (AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER, azure_speech_api_key), + ) + } + ) + raw_body: Final = await request.body() + + endpoint_func: Final = create_pass_through_route( + endpoint=endpoint, + target=str(target_url), + custom_headers=upstream_headers, + custom_llm_provider=AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + ) + setattr(request.state, LITELLM_PASS_THROUGH_RAW_BODY_STATE_KEY, raw_body) + return await endpoint_func(request, fastapi_response, user_api_key_dict) + + def _resolve_vertex_model_from_router( model_id: str, llm_router: litellm.Router | None, diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py new file mode 100644 index 00000000000..a7084a9545e --- /dev/null +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py @@ -0,0 +1,84 @@ +from collections.abc import Mapping +from datetime import datetime +from typing import Final +from urllib.parse import urlparse + +import httpx + +from litellm._logging import verbose_proxy_logger +from litellm.constants import ( + AZURE_SPEECH_BATCH_MODEL, + AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + AZURE_SPEECH_SHORT_AUDIO_MODEL, + AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, +) +from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.litellm_core_utils.litellm_logging import ( + get_standard_logging_object_payload, +) +from litellm.proxy._types import PassThroughEndpointLoggingTypedDict +from litellm.types.utils import StandardPassThroughResponseObject + + +class AzureSpeechPassthroughLoggingHandler: + @staticmethod + def _model_from_url_route(url_route: str) -> str: + path: Final = urlparse(url_route).path + if path.startswith(AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX): + return f"{AZURE_SPEECH_CUSTOM_LLM_PROVIDER}/{AZURE_SPEECH_SHORT_AUDIO_MODEL}" + return f"{AZURE_SPEECH_CUSTOM_LLM_PROVIDER}/{AZURE_SPEECH_BATCH_MODEL}" + + @staticmethod + def azure_speech_passthrough_handler( + httpx_response: httpx.Response, + logging_obj: LiteLLMLoggingObj, + url_route: str, + result: str, + start_time: datetime, + end_time: datetime, + cache_hit: bool, + request_body: Mapping[str, object], + **kwargs: object, # kwargs-ok: the passthrough logging dispatch forwards shared logging kwargs to every handler + ) -> PassThroughEndpointLoggingTypedDict: + """ + Records model and provider for an Azure AI Speech REST call. Azure bills per audio + hour after the fact and neither the short-audio response nor the batch job carries + a billable duration this path can trust, so response_cost is recorded as 0.0 rather + than estimated. + """ + try: + model_name: Final = AzureSpeechPassthroughLoggingHandler._model_from_url_route(url_route) + + updated_kwargs: Final = { # mutable-ok: the logging pipeline requires a plain kwargs dict + **kwargs, + "model": model_name, + "custom_llm_provider": AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + "response_cost": 0.0, + } + logging_obj.model_call_details.update( + model=model_name, + custom_llm_provider=AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + response_cost=0.0, + ) + + standard_logging_object: Final = get_standard_logging_object_payload( + kwargs=updated_kwargs, + init_response_obj=StandardPassThroughResponseObject(response=result), + start_time=start_time, + end_time=end_time, + logging_obj=logging_obj, + status="success", + ) + + handler_payload: Final[PassThroughEndpointLoggingTypedDict] = { + "result": StandardPassThroughResponseObject(response=result), + "kwargs": {**updated_kwargs, "standard_logging_object": standard_logging_object}, + } + except Exception as e: # noqa: BLE001 # logging must never fail the forwarded request + verbose_proxy_logger.exception("Error in Azure Speech passthrough logging handler: %s", e) + fallback_payload: Final[PassThroughEndpointLoggingTypedDict] = { + "result": StandardPassThroughResponseObject(response=result), + "kwargs": kwargs, + } + return fallback_payload + return handler_payload diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py index 685c19062bb..81268a7cf6e 100644 --- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py @@ -52,6 +52,7 @@ from litellm.litellm_core_utils.core_helpers import ( from litellm.litellm_core_utils.initialize_dynamic_callback_params import validate_no_callback_env_reference from litellm.litellm_core_utils.internal_call_metadata import MODEL_ACCESS_GROUP_METADATA_KEY from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.litellm_core_utils.litellm_logging import _get_masked_values from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from litellm.litellm_core_utils.safe_json_dumps import safe_dumps from litellm.llms.base_llm.managed_resources.utils import ( @@ -1023,7 +1024,7 @@ async def pass_through_request( verbose_proxy_logger.debug( "Pass through endpoint sending request to \nURL %s\nheaders: %s\nbody: %s\n", url, - upstream_headers, + _get_masked_values(upstream_headers), _parsed_body, ) diff --git a/litellm/proxy/pass_through_endpoints/success_handler.py b/litellm/proxy/pass_through_endpoints/success_handler.py index 76a471302f4..919de5c1088 100644 --- a/litellm/proxy/pass_through_endpoints/success_handler.py +++ b/litellm/proxy/pass_through_endpoints/success_handler.py @@ -5,6 +5,7 @@ from urllib.parse import urlparse import httpx +from litellm.constants import AZURE_SPEECH_CUSTOM_LLM_PROVIDER from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.proxy._types import PassThroughEndpointLoggingResultValues from litellm.types.passthrough_endpoints.pass_through_endpoints import ( @@ -256,6 +257,24 @@ class PassThroughEndpointLogging: ) standard_logging_response_object = comprehend_medical_handler_result["result"] # rebind-ok: elif-chain kwargs = comprehend_medical_handler_result["kwargs"] # rebind-ok: elif-chain contract + elif self.is_azure_speech_route(custom_llm_provider): + from .llm_provider_handlers.azure_speech_passthrough_logging_handler import ( + AzureSpeechPassthroughLoggingHandler, + ) + + azure_speech_handler_result: Final = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=httpx_response, + logging_obj=logging_obj, + url_route=url_route, + result=result, + start_time=start_time, + end_time=end_time, + cache_hit=cache_hit, + request_body=request_body, + **kwargs, + ) + standard_logging_response_object = azure_speech_handler_result["result"] # rebind-ok: elif-chain + kwargs = azure_speech_handler_result["kwargs"] # rebind-ok: elif-chain contract elif self.is_vertex_ai_live_route(url_route): from .llm_provider_handlers.vertex_ai_live_passthrough_logging_handler import ( VertexAILivePassthroughLoggingHandler, @@ -300,7 +319,7 @@ class PassThroughEndpointLogging: ): standard_logging_response_object: PassThroughEndpointLoggingResultValues | None = None logging_obj.model_call_details["passthrough_logging_payload"] = passthrough_logging_payload - if self.is_assemblyai_route(url_route): + if self.is_assemblyai_route(url_route) and not self.is_azure_speech_route(custom_llm_provider): if AssemblyAIPassthroughLoggingHandler._should_log_request(httpx_response.request.method) is not True: return self.assemblyai_passthrough_logging_handler.assemblyai_passthrough_logging_handler( @@ -389,6 +408,9 @@ class PassThroughEndpointLogging: def is_comprehend_medical_route(self, custom_llm_provider: str | None) -> bool: return custom_llm_provider == "comprehendmedical" + def is_azure_speech_route(self, custom_llm_provider: str | None) -> bool: + return custom_llm_provider == AZURE_SPEECH_CUSTOM_LLM_PROVIDER + def is_langfuse_route(self, url_route: str): parsed_url: Final = urlparse(url_route) for route in self.TRACKED_LANGFUSE_ROUTES: diff --git a/litellm/proxy/public_endpoints/provider_create_fields.json b/litellm/proxy/public_endpoints/provider_create_fields.json index cd781abee26..e6673ec99aa 100644 --- a/litellm/proxy/public_endpoints/provider_create_fields.json +++ b/litellm/proxy/public_endpoints/provider_create_fields.json @@ -586,6 +586,24 @@ ], "default_model_placeholder": "azure_ai/command-r-plus" }, + { + "provider": "Azure_Speech", + "provider_display_name": "Azure AI Speech", + "litellm_provider": "azure_speech", + "credential_fields": [ + { + "key": "api_key", + "label": "Azure AI Speech Subscription Key", + "placeholder": null, + "tooltip": "The Ocp-Apim-Subscription-Key for your Azure AI Speech resource. The proxy injects it on every /azure_speech/* pass-through request. Region and API base come from AZURE_SPEECH_REGION / AZURE_SPEECH_API_BASE", + "required": true, + "field_type": "password", + "options": null, + "default_value": null + } + ], + "default_model_placeholder": "azure_speech/short-audio" + }, { "provider": "AZURE_TEXT", "provider_display_name": "Azure Text", diff --git a/litellm/types/utils.py b/litellm/types/utils.py index aaa16fd2d44..3c2c549e89e 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -4060,6 +4060,7 @@ class LlmProviders(str, Enum): TOPAZ = "topaz" SAP_GENERATIVE_AI_HUB = "sap" ASSEMBLYAI = "assemblyai" + AZURE_SPEECH = "azure_speech" CHARITY_ENGINE = "charity_engine" GITHUB_COPILOT = "github_copilot" SNOWFLAKE = "snowflake" diff --git a/terraform/litellm/aws/locals.tf b/terraform/litellm/aws/locals.tf index bd5b97b0f50..fcf1f7b905f 100644 --- a/terraform/litellm/aws/locals.tf +++ b/terraform/litellm/aws/locals.tf @@ -86,7 +86,7 @@ locals { "/queue/chat/*", "/v1beta/*", "/interactions/*", - "/anthropic/*", "/azure/*", "/azure_ai/*", "/aws/*", "/bedrock/*", "/comprehendmedical*", + "/anthropic/*", "/azure/*", "/azure_ai/*", "/azure_speech/*", "/aws/*", "/bedrock/*", "/comprehendmedical*", "/cohere/*", "/gemini/*", "/google/*", "/vertex_ai/*", "/vertex-ai/*", "/assemblyai/*", "/eu.assemblyai/*", diff --git a/terraform/litellm/gcp/locals.tf b/terraform/litellm/gcp/locals.tf index 3861413d496..dca7b05f1c9 100644 --- a/terraform/litellm/gcp/locals.tf +++ b/terraform/litellm/gcp/locals.tf @@ -55,7 +55,7 @@ locals { "/queue/chat/*", "/v1beta/*", "/interactions/*", - "/anthropic/*", "/azure/*", "/azure_ai/*", "/aws/*", "/bedrock/*", "/comprehendmedical*", + "/anthropic/*", "/azure/*", "/azure_ai/*", "/azure_speech/*", "/aws/*", "/bedrock/*", "/comprehendmedical*", "/cohere/*", "/gemini/*", "/google/*", "/vertex_ai/*", "/vertex-ai/*", "/assemblyai/*", "/eu.assemblyai/*", diff --git a/tests/test_litellm/proxy/middleware/test_billable_request_metrics_middleware.py b/tests/test_litellm/proxy/middleware/test_billable_request_metrics_middleware.py index 9c61412bd6e..5ce7aa858c1 100644 --- a/tests/test_litellm/proxy/middleware/test_billable_request_metrics_middleware.py +++ b/tests/test_litellm/proxy/middleware/test_billable_request_metrics_middleware.py @@ -116,6 +116,11 @@ def test_is_pure_asgi_not_base_http_middleware(): # Bare AWS-SDK-shaped route carries the operation in X-Amz-Target and writes SpendLogs ("/comprehendmedical", (BillableCategory.LLM, "/comprehendmedical")), ("/comprehendmedical/DetectEntitiesV2", (BillableCategory.LLM, "/comprehendmedical")), + ( + "/azure_speech/speech/recognition/conversation/cognitiveservices/v1", + (BillableCategory.LLM, "/azure_speech"), + ), + ("/azure_speech/speechtotext/v3.2/transcriptions", (BillableCategory.LLM, "/azure_speech")), ("/mcp", (BillableCategory.MCP, "/mcp")), ("/mcp/", (BillableCategory.MCP, "/mcp")), ("/mcp/tools/list", (BillableCategory.MCP, "/mcp")), diff --git a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py new file mode 100644 index 00000000000..91f7bf94281 --- /dev/null +++ b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py @@ -0,0 +1,123 @@ +from datetime import datetime +from unittest.mock import MagicMock + +import httpx +import pytest + +from litellm.proxy.pass_through_endpoints.llm_provider_handlers.azure_speech_passthrough_logging_handler import ( + AzureSpeechPassthroughLoggingHandler, +) +from litellm.proxy.pass_through_endpoints.success_handler import ( + PassThroughEndpointLogging, +) + +SHORT_AUDIO_URL = "https://eastus.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US" +BATCH_URL = "https://eastus.api.cognitive.microsoft.com/speechtotext/v3.2/transcriptions" +TRANSCRIPT = '{"RecognitionStatus":"Success","DisplayText":"Hello world."}' + + +def _make_response(url: str) -> httpx.Response: + request = httpx.Request("POST", url, headers={"Ocp-Apim-Subscription-Key": "server-secret"}) + return httpx.Response(200, request=request, text=TRANSCRIPT) + + +def _make_logging_obj() -> MagicMock: + logging_obj = MagicMock() + logging_obj.litellm_call_id = "test-call-id" + logging_obj.model_call_details = {} + return logging_obj + + +class TestAzureSpeechPassthroughHandler: + @pytest.mark.parametrize( + "url_route,expected_model", + [ + (SHORT_AUDIO_URL, "azure_speech/short-audio"), + (BATCH_URL, "azure_speech/batch-transcription"), + (f"{BATCH_URL}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab/files", "azure_speech/batch-transcription"), + ], + ) + def test_records_model_provider_and_zero_cost(self, url_route: str, expected_model: str): + logging_obj = _make_logging_obj() + + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(url_route), + logging_obj=logging_obj, + url_route=url_route, + result=TRANSCRIPT, + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert handler_result["result"] == {"response": TRANSCRIPT} + assert handler_result["kwargs"]["model"] == expected_model + assert handler_result["kwargs"]["custom_llm_provider"] == "azure_speech" + assert handler_result["kwargs"]["response_cost"] == 0.0 + assert handler_result["kwargs"]["standard_logging_object"]["response_cost"] == 0.0 + assert handler_result["kwargs"]["standard_logging_object"]["model"] == expected_model + assert logging_obj.model_call_details["model"] == expected_model + assert logging_obj.model_call_details["custom_llm_provider"] == "azure_speech" + assert logging_obj.model_call_details["response_cost"] == 0.0 + + def test_subscription_key_never_reaches_the_logging_payload(self): + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(SHORT_AUDIO_URL), + logging_obj=_make_logging_obj(), + url_route=SHORT_AUDIO_URL, + result=TRANSCRIPT, + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert "server-secret" not in repr(handler_result) + + +class TestIsAzureSpeechRoute: + def test_matches_by_provider_tag(self): + assert PassThroughEndpointLogging().is_azure_speech_route("azure_speech") + + @pytest.mark.parametrize("provider", ["azure", "azure_ai", "comprehendmedical", None]) + def test_does_not_match_other_providers(self, provider: str | None): + assert not PassThroughEndpointLogging().is_azure_speech_route(provider) + + def test_config_driven_passthrough_to_azure_speech_host_is_not_claimed(self): + normalized = PassThroughEndpointLogging().normalize_llm_passthrough_logging_payload( + httpx_response=_make_response(SHORT_AUDIO_URL), + response_body={"RecognitionStatus": "Success"}, + request_body={}, + logging_obj=_make_logging_obj(), + url_route=SHORT_AUDIO_URL, + result=TRANSCRIPT, + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + custom_llm_provider=None, + ) + + assert normalized["kwargs"].get("model") != "azure_speech/short-audio" + assert "response_cost" not in normalized["kwargs"] + + +class TestNormalizeDispatch: + def test_normalize_routes_to_azure_speech_handler(self): + normalized = PassThroughEndpointLogging().normalize_llm_passthrough_logging_payload( + httpx_response=_make_response(SHORT_AUDIO_URL), + response_body={"RecognitionStatus": "Success"}, + request_body={}, + logging_obj=_make_logging_obj(), + url_route=SHORT_AUDIO_URL, + result=TRANSCRIPT, + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + custom_llm_provider="azure_speech", + ) + + assert normalized["standard_logging_response_object"] == {"response": TRANSCRIPT} + assert normalized["kwargs"]["model"] == "azure_speech/short-audio" + assert normalized["kwargs"]["custom_llm_provider"] == "azure_speech" + assert normalized["kwargs"]["response_cost"] == 0.0 diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index 6e82c90514d..9fe7f5b6ee9 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -2,6 +2,7 @@ import asyncio import base64 import contextlib import json +import logging import os import traceback from collections.abc import Iterator, Mapping @@ -6136,3 +6137,296 @@ class TestAzureRelayDeploymentSegment: ) assert [call["model"] for call in captured] == ["gpt", "gpt"] + + +AZURE_SPEECH_SHORT_AUDIO_ENDPOINT: Final = "/speech/recognition/conversation/cognitiveservices/v1" +AZURE_SPEECH_BATCH_ENDPOINT: Final = "/speechtotext/v3.2/transcriptions" +AZURE_SPEECH_PCM16_HEADER: Final = ( + b"RIFF\x24\x0c\x00\x00WAVEfmt \x10\x00\x00\x00\x01\x00\x01\x00\x80\x3e\x00\x00\x00\x7d\x00\x00\x02\x00\x10\x00data\x00\x0c\x00\x00" +) +AZURE_SPEECH_WAV_BYTES: Final = AZURE_SPEECH_PCM16_HEADER + b"\x00" * 3072 +AZURE_SPEECH_NON_UTF8_WAV_BYTES: Final = AZURE_SPEECH_PCM16_HEADER + bytes(range(256)) * 12 +AZURE_SPEECH_TRANSCRIPT: Final = {"RecognitionStatus": "Success", "DisplayText": "The eagle has landed."} + + +@pytest.fixture +def azure_speech_client(monkeypatch: pytest.MonkeyPatch) -> Iterator[TestClient]: + from litellm.proxy.proxy_server import app + + monkeypatch.setenv("AZURE_SPEECH_API_KEY", "server-subscription-key") + monkeypatch.setenv("AZURE_SPEECH_REGION", "eastus") + monkeypatch.delenv("AZURE_SPEECH_API_BASE", raising=False) + monkeypatch.delenv("SERVER_ROOT_PATH", raising=False) + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + litellm.in_memory_llm_clients_cache.flush_cache() + monkeypatch.setitem(app.dependency_overrides, user_api_key_auth, lambda: UserAPIKeyAuth(api_key="sk-virtual")) + yield TestClient(app) + + +class TestAzureSpeechProxyRoute: + """Drives the real FastAPI route with respx standing in for the Azure hosts only.""" + + def test_short_audio_forwards_raw_wav_bytes_with_server_key(self, azure_speech_client: TestClient) -> None: + with respx.mock(assert_all_called=True) as upstream: + route = upstream.post( + f"https://eastus.stt.speech.microsoft.com{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}" + ).mock(return_value=httpx.Response(200, json=AZURE_SPEECH_TRANSCRIPT)) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", + params={"language": "en-US", "format": "detailed"}, + content=AZURE_SPEECH_WAV_BYTES, + headers={ + "Content-Type": "audio/wav; codecs=audio/pcm; samplerate=16000", + "Authorization": "Bearer sk-virtual", + "Ocp-Apim-Subscription-Key": "caller-supplied-key", + "x-pass-ocp-apim-subscription-key": "caller-supplied-key", + }, + ) + + assert (response.status_code, response.json()) == (200, AZURE_SPEECH_TRANSCRIPT) + sent = route.calls.last.request + assert sent.content == AZURE_SPEECH_WAV_BYTES + assert dict(sent.url.params) == {"language": "en-US", "format": "detailed"} + assert sent.headers["ocp-apim-subscription-key"] == "server-subscription-key" + assert sent.headers["content-type"] == "audio/wav; codecs=audio/pcm; samplerate=16000" + assert "authorization" not in sent.headers + assert "caller-supplied-key" not in repr(sent.headers) + + def test_batch_json_goes_to_the_cognitive_services_host(self, azure_speech_client: TestClient) -> None: + body: Final = {"contentUrls": ["https://example.com/a.wav"], "locale": "en-US", "displayName": "job"} + with respx.mock(assert_all_called=True) as upstream: + route = upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( + return_value=httpx.Response(201, json={"self": "https://eastus.api.cognitive.microsoft.com/x"}) + ) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", + json=body, + headers={"Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 201 + sent = route.calls.last.request + assert json.loads(sent.content) == body + assert sent.headers["ocp-apim-subscription-key"] == "server-subscription-key" + assert "authorization" not in sent.headers + + def test_batch_multipart_upload_is_forwarded_byte_for_byte(self, azure_speech_client: TestClient) -> None: + with respx.mock(assert_all_called=True) as upstream: + route = upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( + return_value=httpx.Response(201, json={"status": "NotStarted"}) + ) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", + files={"audio": ("eagle.wav", AZURE_SPEECH_NON_UTF8_WAV_BYTES, "audio/wav")}, + data={"definition": json.dumps({"locales": ["en-US"]})}, + headers={"Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 201 + sent = route.calls.last.request + assert sent.headers["content-type"].startswith("multipart/form-data; boundary=") + assert AZURE_SPEECH_NON_UTF8_WAV_BYTES in sent.content + assert b'name="definition"' in sent.content + assert sent.headers["ocp-apim-subscription-key"] == "server-subscription-key" + assert "authorization" not in sent.headers + + def test_batch_get_is_forwarded_with_the_job_id_path(self, azure_speech_client: TestClient) -> None: + job_path: Final = f"{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab/files" + with respx.mock(assert_all_called=True) as upstream: + route = upstream.get(f"https://eastus.api.cognitive.microsoft.com{job_path}").mock( + return_value=httpx.Response(200, json={"values": []}) + ) + + response = azure_speech_client.get(f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-virtual"}) + + assert (response.status_code, response.json()) == (200, {"values": []}) + assert route.calls.last.request.headers["ocp-apim-subscription-key"] == "server-subscription-key" + + @pytest.mark.parametrize("method", ["GET", "POST"]) + def test_batch_requests_are_logged_as_azure_speech_not_assemblyai( + self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch, method: str + ) -> None: + from litellm.integrations.custom_logger import CustomLogger + + class _Recorder(CustomLogger): + def __init__(self) -> None: + super().__init__() + self.payloads: list[dict[str, object]] = [] # mutable-ok: test recorder accumulates callback payloads + + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time) -> None: + self.payloads.append(kwargs["standard_logging_object"]) + + recorder: Final = _Recorder() + monkeypatch.setattr(litellm, "_async_success_callback", [*litellm._async_success_callback, recorder]) + with respx.mock(assert_all_called=True) as upstream: + upstream.request(method, f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"values": []}) + ) + + response = azure_speech_client.request( + method, + f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", + json={"locale": "en-US"} if method == "POST" else None, + headers={"Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 200 + assert [(p["model"], p["custom_llm_provider"], p["response_cost"]) for p in recorder.payloads] == [ + ("azure_speech/batch-transcription", "azure_speech", 0.0) + ] + + def test_api_base_wins_over_region_for_both_families( + self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch + ) -> None: + monkeypatch.setenv("AZURE_SPEECH_API_BASE", "https://my-speech.cognitiveservices.azure.com") + with respx.mock(assert_all_called=True) as upstream: + short_audio = upstream.post( + f"https://my-speech.cognitiveservices.azure.com{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}" + ).mock(return_value=httpx.Response(200, json=AZURE_SPEECH_TRANSCRIPT)) + batch = upstream.get(f"https://my-speech.cognitiveservices.azure.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"values": []}) + ) + + azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", + content=AZURE_SPEECH_WAV_BYTES, + headers={"Content-Type": "audio/wav", "Authorization": "Bearer sk-virtual"}, + ) + azure_speech_client.get(f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", headers={"Authorization": "Bearer x"}) + + assert short_audio.called and batch.called + + @pytest.mark.parametrize("endpoint", ["openai/deployments/whisper/audio/transcriptions", "speech", "speechtotext"]) + def test_unknown_path_family_is_rejected_before_any_upstream_call( + self, azure_speech_client: TestClient, endpoint: str + ) -> None: + with respx.mock(assert_all_called=False) as upstream: + catch_all = upstream.route().mock(return_value=httpx.Response(200)) + + response = azure_speech_client.post( + f"/azure_speech/{endpoint}", content=b"x", headers={"Authorization": "Bearer sk-virtual"} + ) + + assert response.status_code == 400 + assert not catch_all.called + + def test_missing_region_and_base_is_rejected_before_any_upstream_call( + self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch + ) -> None: + monkeypatch.delenv("AZURE_SPEECH_REGION") + with respx.mock(assert_all_called=False) as upstream: + catch_all = upstream.route().mock(return_value=httpx.Response(200)) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", + content=AZURE_SPEECH_WAV_BYTES, + headers={"Content-Type": "audio/wav", "Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 400 + assert "AZURE_SPEECH_REGION" in response.text + assert not catch_all.called + + def test_missing_api_key_is_rejected_before_any_upstream_call( + self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch + ) -> None: + monkeypatch.delenv("AZURE_SPEECH_API_KEY") + with respx.mock(assert_all_called=False) as upstream: + catch_all = upstream.route().mock(return_value=httpx.Response(200)) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", + content=AZURE_SPEECH_WAV_BYTES, + headers={"Content-Type": "audio/wav", "Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 400 + assert "AZURE_SPEECH_API_KEY" in response.text + assert not catch_all.called + + def test_azure_speech_is_a_mapped_pass_through_route(self) -> None: + from litellm.proxy._types import LiteLLMRoutes + + assert "/azure_speech" in LiteLLMRoutes.mapped_pass_through_routes.value + + +def _azure_speech_real_auth_attrs() -> dict[str, object]: + from litellm.caching.caching import DualCache + from litellm.proxy.utils import ProxyLogging + + user_api_key_cache: Final = DualCache() + return { + "prisma_client": None, + "user_api_key_cache": user_api_key_cache, + "proxy_logging_obj": ProxyLogging(user_api_key_cache=user_api_key_cache), + "master_key": "sk-master-key", + "general_settings": {}, + "llm_model_list": [], + "llm_router": None, + "open_telemetry_logger": None, + "user_custom_auth": None, + "jwt_handler": None, + } + + +class TestAzureSpeechRawBodyThroughRealAuth: + """user_api_key_auth reads the body before the route runs; raw audio must not be parsed as JSON.""" + + def _post_wav( + self, monkeypatch: pytest.MonkeyPatch, path: str, api_key: str, body: bytes = AZURE_SPEECH_WAV_BYTES + ) -> httpx.Response: + from litellm.proxy.proxy_server import app + + monkeypatch.setenv("AZURE_SPEECH_API_KEY", "server-subscription-key") + monkeypatch.setenv("AZURE_SPEECH_REGION", "eastus") + monkeypatch.delenv("AZURE_SPEECH_API_BASE", raising=False) + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + litellm.in_memory_llm_clients_cache.flush_cache() + with patch.multiple( # test-quality-ok: the real user_api_key_auth reads proxy_server module globals (master_key, caches) that have no injection seam + "litellm.proxy.proxy_server", **_azure_speech_real_auth_attrs() + ): + client = TestClient(app) + return client.post( + path, + params={"language": "en-US"}, + content=body, + headers={"Content-Type": "audio/wav", "Authorization": f"Bearer {api_key}"}, + ) + + @pytest.mark.parametrize("body", [AZURE_SPEECH_WAV_BYTES, AZURE_SPEECH_NON_UTF8_WAV_BYTES], ids=["ascii", "binary"]) + def test_master_key_with_raw_wav_body_reaches_azure_without_a_parse_attempt( + self, monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture, body: bytes + ) -> None: + with respx.mock(assert_all_called=True) as upstream, caplog.at_level(logging.ERROR, logger="LiteLLM Proxy"): + route = upstream.post( + f"https://eastus.stt.speech.microsoft.com{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}" + ).mock(return_value=httpx.Response(200, json=AZURE_SPEECH_TRANSCRIPT)) + + response = self._post_wav( + monkeypatch, f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", "sk-master-key", body=body + ) + + assert (response.status_code, response.json()) == (200, AZURE_SPEECH_TRANSCRIPT) + assert route.calls.last.request.content == body + assert [record.message for record in caplog.records if "request body" in record.message] == [] + + def test_wrong_litellm_key_with_raw_wav_body_is_rejected(self, monkeypatch: pytest.MonkeyPatch) -> None: + with respx.mock(assert_all_called=False) as upstream: + catch_all = upstream.route().mock(return_value=httpx.Response(200)) + + response = self._post_wav(monkeypatch, f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", "sk-wrong") + + assert response.status_code in (400, 401), response.text + assert not catch_all.called + + @pytest.mark.parametrize("path", ["/v1/chat/completions", f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}"]) + def test_audio_content_type_off_the_short_audio_route_is_still_parsed_as_json( + self, monkeypatch: pytest.MonkeyPatch, path: str + ) -> None: + response = self._post_wav(monkeypatch, path, "sk-master-key", body=b'{}{"model": "gpt-4o"}') + + assert response.status_code == 400 + assert "Invalid JSON payload" in response.text diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py index e3cbc2d507f..7a272a49853 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py @@ -159,6 +159,22 @@ def test_assemblyai_region_matching(): assert passthrough_router.get_credentials(custom_llm_provider="assemblyai", region_name=None) == "sk-us" +def test_azure_speech_dashboard_credential_resolves_through_flagged_deployment(monkeypatch): + monkeypatch.delenv("AZURE_SPEECH_API_KEY", raising=False) + CredentialAccessor.upsert_credentials([_credential("azure-speech-prod", "azure-subscription-key")]) + llm_router = litellm.Router( + model_list=[ + _flagged_deployment("azure_speech/short-audio", litellm_credential_name="azure-speech-prod"), + ] + ) + passthrough_router = _passthrough_router(llm_router) + + assert ( + passthrough_router.get_credentials(custom_llm_provider="azure_speech", region_name=None) + == "azure-subscription-key" + ) + + def test_env_fallback_when_no_router(monkeypatch): passthrough_router = _passthrough_router(None) monkeypatch.setenv("OPENAI_API_KEY", "sk-from-env") diff --git a/ui/litellm-dashboard/src/components/provider_info_helpers.tsx b/ui/litellm-dashboard/src/components/provider_info_helpers.tsx index de83cd00790..1a1a2fd73aa 100644 --- a/ui/litellm-dashboard/src/components/provider_info_helpers.tsx +++ b/ui/litellm-dashboard/src/components/provider_info_helpers.tsx @@ -80,6 +80,7 @@ export enum Providers { SageMaker = "AWS SageMaker", Azure = "Azure", Azure_AI_Studio = "Azure AI Foundry (Studio)", + Azure_Speech = "Azure AI Speech", AZURE_TEXT = "Azure Text", BASETEN = "Baseten", BYTEZ = "Bytez", @@ -193,6 +194,7 @@ export const provider_map: Record = { AUTO_ROUTER: "auto_router", Azure: "azure", Azure_AI_Studio: "azure_ai", + Azure_Speech: "azure_speech", AZURE_TEXT: "azure_text", BASETEN: "baseten", Bedrock: "bedrock", @@ -310,6 +312,7 @@ export const providerLogoMap: Partial> = { [Providers.AssemblyAI]: assemblyaiSmallLogo.src, [Providers.Azure]: microsoftAzureLogo.src, [Providers.Azure_AI_Studio]: microsoftAzureLogo.src, + [Providers.Azure_Speech]: microsoftAzureLogo.src, [Providers.AZURE_TEXT]: microsoftAzureLogo.src, [Providers.BASETEN]: basetenLogo.src, [Providers.Bedrock]: bedrockLogo.src, @@ -427,6 +430,7 @@ const providerPlaceholderMap: Partial> = { [Providers.Anthropic]: "claude-3-opus", [Providers.Azure]: "my-deployment", [Providers.Azure_AI_Studio]: "azure_ai/command-r-plus", + [Providers.Azure_Speech]: "azure_speech/short-audio", [Providers.Bedrock]: "claude-3-opus", [Providers.CHATGPT]: "chatgpt/gpt-5.4", [Providers.Cognition]: "cognition/swe-1.7", diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 872875cc535..62b9921302a 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -1612,6 +1612,82 @@ export interface paths { patch: operations["azure_proxy_route_azure_ai__endpoint__patch"]; trace?: never; }; + "/azure_speech/{endpoint}": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + /** + * Azure Speech Proxy Route + * @description Pass-through for the Azure AI Speech REST APIs (speech to text), e.g. + * `POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US` + * with the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`. + * + * The body is forwarded byte for byte and the proxy injects its own + * `Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key + * and is never forwarded. + * + * [Docs](https://docs.litellm.ai/docs/pass_through/azure_speech) + */ + get: operations["azure_speech_proxy_route_azure_speech__endpoint__get"]; + /** + * Azure Speech Proxy Route + * @description Pass-through for the Azure AI Speech REST APIs (speech to text), e.g. + * `POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US` + * with the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`. + * + * The body is forwarded byte for byte and the proxy injects its own + * `Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key + * and is never forwarded. + * + * [Docs](https://docs.litellm.ai/docs/pass_through/azure_speech) + */ + put: operations["azure_speech_proxy_route_azure_speech__endpoint__put"]; + /** + * Azure Speech Proxy Route + * @description Pass-through for the Azure AI Speech REST APIs (speech to text), e.g. + * `POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US` + * with the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`. + * + * The body is forwarded byte for byte and the proxy injects its own + * `Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key + * and is never forwarded. + * + * [Docs](https://docs.litellm.ai/docs/pass_through/azure_speech) + */ + post: operations["azure_speech_proxy_route_azure_speech__endpoint__post"]; + /** + * Azure Speech Proxy Route + * @description Pass-through for the Azure AI Speech REST APIs (speech to text), e.g. + * `POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US` + * with the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`. + * + * The body is forwarded byte for byte and the proxy injects its own + * `Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key + * and is never forwarded. + * + * [Docs](https://docs.litellm.ai/docs/pass_through/azure_speech) + */ + delete: operations["azure_speech_proxy_route_azure_speech__endpoint__delete"]; + options?: never; + head?: never; + /** + * Azure Speech Proxy Route + * @description Pass-through for the Azure AI Speech REST APIs (speech to text), e.g. + * `POST /azure_speech/speech/recognition/conversation/cognitiveservices/v1?language=en-US` + * with the raw audio as the body, or `POST /azure_speech/speechtotext/v3.2/transcriptions`. + * + * The body is forwarded byte for byte and the proxy injects its own + * `Ocp-Apim-Subscription-Key`; the caller's `Authorization` header is the LiteLLM key + * and is never forwarded. + * + * [Docs](https://docs.litellm.ai/docs/pass_through/azure_speech) + */ + patch: operations["azure_speech_proxy_route_azure_speech__endpoint__patch"]; + trace?: never; + }; "/batches": { parameters: { query?: never; @@ -43394,6 +43470,161 @@ export interface operations { }; }; }; + azure_speech_proxy_route_azure_speech__endpoint__get: { + parameters: { + query?: never; + header?: never; + path: { + endpoint: string; + }; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": unknown; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; + azure_speech_proxy_route_azure_speech__endpoint__put: { + parameters: { + query?: never; + header?: never; + path: { + endpoint: string; + }; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": unknown; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; + azure_speech_proxy_route_azure_speech__endpoint__post: { + parameters: { + query?: never; + header?: never; + path: { + endpoint: string; + }; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": unknown; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; + azure_speech_proxy_route_azure_speech__endpoint__delete: { + parameters: { + query?: never; + header?: never; + path: { + endpoint: string; + }; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": unknown; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; + azure_speech_proxy_route_azure_speech__endpoint__patch: { + parameters: { + query?: never; + header?: never; + path: { + endpoint: string; + }; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": unknown; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; list_batches_batches_get: { parameters: { query?: { From f2305879d06073dfa2fa9f8001659c48ae61e7d1 Mon Sep 17 00:00:00 2001 From: yassin Date: Thu, 17 Sep 2026 03:31:29 +0000 Subject: [PATCH 057/141] feat(proxy): price Azure Speech short audio pass-through from the recognized duration Short audio responses carry Offset and Duration in 100ns ticks; convert their sum to seconds and price it with the existing azure/speech/azure-stt entry through transcription_cost. Batch calls and responses without an integer duration stay at zero cost Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 2 + ...zure_speech_passthrough_logging_handler.py | 54 ++++++++--- .../pass_through_endpoints/success_handler.py | 1 + ...zure_speech_passthrough_logging_handler.py | 95 ++++++++++++++++--- .../test_llm_pass_through_endpoints.py | 43 +++++++++ 5 files changed, 171 insertions(+), 24 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index 69de889a326..15a1d054e26 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1579,6 +1579,8 @@ AZURE_SPEECH_COGNITIVE_SERVICES_DOMAIN: Final = "api.cognitive.microsoft.com" AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER: Final = "Ocp-Apim-Subscription-Key" AZURE_SPEECH_SHORT_AUDIO_MODEL: Final = "short-audio" AZURE_SPEECH_BATCH_MODEL: Final = "batch-transcription" +AZURE_SPEECH_PRICING_MODEL: Final = "azure/speech/azure-stt" +AZURE_SPEECH_TICKS_PER_SECOND: Final = 10_000_000 BASE_MCP_ROUTE: Final = "/mcp" diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py index a7084a9545e..74587acd453 100644 --- a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py @@ -1,4 +1,4 @@ -from collections.abc import Mapping +from collections.abc import Mapping, Sequence from datetime import datetime from typing import Final from urllib.parse import urlparse @@ -9,9 +9,12 @@ from litellm._logging import verbose_proxy_logger from litellm.constants import ( AZURE_SPEECH_BATCH_MODEL, AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + AZURE_SPEECH_PRICING_MODEL, AZURE_SPEECH_SHORT_AUDIO_MODEL, AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, + AZURE_SPEECH_TICKS_PER_SECOND, ) +from litellm.cost_calculator import transcription_cost from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.litellm_logging import ( get_standard_logging_object_payload, @@ -21,16 +24,50 @@ from litellm.types.utils import StandardPassThroughResponseObject class AzureSpeechPassthroughLoggingHandler: + @staticmethod + def _is_short_audio_route(url_route: str) -> bool: + return urlparse(url_route).path.startswith(AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX) + @staticmethod def _model_from_url_route(url_route: str) -> str: - path: Final = urlparse(url_route).path - if path.startswith(AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX): + if AzureSpeechPassthroughLoggingHandler._is_short_audio_route(url_route): return f"{AZURE_SPEECH_CUSTOM_LLM_PROVIDER}/{AZURE_SPEECH_SHORT_AUDIO_MODEL}" return f"{AZURE_SPEECH_CUSTOM_LLM_PROVIDER}/{AZURE_SPEECH_BATCH_MODEL}" + @staticmethod + def _recognized_audio_seconds(response_body: Mapping[str, object] | Sequence[object] | None) -> float: + if not isinstance(response_body, Mapping): + return 0.0 + offset: Final = response_body.get("Offset") + duration: Final = response_body.get("Duration") + if not isinstance(offset, int) or not isinstance(duration, int): + return 0.0 + return (offset + duration) / AZURE_SPEECH_TICKS_PER_SECOND + + @staticmethod + def _response_cost(url_route: str, response_body: Mapping[str, object] | Sequence[object] | None) -> float: + if not AzureSpeechPassthroughLoggingHandler._is_short_audio_route(url_route): + return 0.0 + audio_seconds: Final = AzureSpeechPassthroughLoggingHandler._recognized_audio_seconds(response_body) + if audio_seconds <= 0.0: + return 0.0 + try: + prompt_cost, completion_cost = transcription_cost( + model=AZURE_SPEECH_PRICING_MODEL, + custom_llm_provider="azure", + duration=audio_seconds, + ) + except Exception as e: # noqa: BLE001 # a missing price entry must not drop the spend log row + verbose_proxy_logger.warning( + "No price for %s, logging Azure Speech call at zero cost: %s", AZURE_SPEECH_PRICING_MODEL, e + ) + return 0.0 + return prompt_cost + completion_cost + @staticmethod def azure_speech_passthrough_handler( httpx_response: httpx.Response, + response_body: Mapping[str, object] | Sequence[object] | None, logging_obj: LiteLLMLoggingObj, url_route: str, result: str, @@ -40,25 +77,20 @@ class AzureSpeechPassthroughLoggingHandler: request_body: Mapping[str, object], **kwargs: object, # kwargs-ok: the passthrough logging dispatch forwards shared logging kwargs to every handler ) -> PassThroughEndpointLoggingTypedDict: - """ - Records model and provider for an Azure AI Speech REST call. Azure bills per audio - hour after the fact and neither the short-audio response nor the batch job carries - a billable duration this path can trust, so response_cost is recorded as 0.0 rather - than estimated. - """ try: model_name: Final = AzureSpeechPassthroughLoggingHandler._model_from_url_route(url_route) + response_cost: Final = AzureSpeechPassthroughLoggingHandler._response_cost(url_route, response_body) updated_kwargs: Final = { # mutable-ok: the logging pipeline requires a plain kwargs dict **kwargs, "model": model_name, "custom_llm_provider": AZURE_SPEECH_CUSTOM_LLM_PROVIDER, - "response_cost": 0.0, + "response_cost": response_cost, } logging_obj.model_call_details.update( model=model_name, custom_llm_provider=AZURE_SPEECH_CUSTOM_LLM_PROVIDER, - response_cost=0.0, + response_cost=response_cost, ) standard_logging_object: Final = get_standard_logging_object_payload( diff --git a/litellm/proxy/pass_through_endpoints/success_handler.py b/litellm/proxy/pass_through_endpoints/success_handler.py index 919de5c1088..2ccb8ad525d 100644 --- a/litellm/proxy/pass_through_endpoints/success_handler.py +++ b/litellm/proxy/pass_through_endpoints/success_handler.py @@ -264,6 +264,7 @@ class PassThroughEndpointLogging: azure_speech_handler_result: Final = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( httpx_response=httpx_response, + response_body=response_body, logging_obj=logging_obj, url_route=url_route, result=result, diff --git a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py index 91f7bf94281..5ffb1f7785d 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py @@ -1,9 +1,11 @@ +import json from datetime import datetime from unittest.mock import MagicMock import httpx import pytest +import litellm from litellm.proxy.pass_through_endpoints.llm_provider_handlers.azure_speech_passthrough_logging_handler import ( AzureSpeechPassthroughLoggingHandler, ) @@ -13,7 +15,24 @@ from litellm.proxy.pass_through_endpoints.success_handler import ( SHORT_AUDIO_URL = "https://eastus.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US" BATCH_URL = "https://eastus.api.cognitive.microsoft.com/speechtotext/v3.2/transcriptions" -TRANSCRIPT = '{"RecognitionStatus":"Success","DisplayText":"Hello world."}' +TRANSCRIPT_BODY = {"RecognitionStatus": "Success", "Offset": 5000000, "Duration": 25000000, "DisplayText": "Hello world."} +TRANSCRIPT = json.dumps(TRANSCRIPT_BODY) +TRANSCRIPT_AUDIO_SECONDS = 3.0 +PRICE_PER_SECOND = 0.5 + + +@pytest.fixture(autouse=True) +def azure_stt_price(monkeypatch: pytest.MonkeyPatch): + monkeypatch.setitem( + litellm.model_cost, + "azure/speech/azure-stt", + { + "litellm_provider": "azure", + "mode": "audio_transcription", + "input_cost_per_second": PRICE_PER_SECOND, + "output_cost_per_second": 0.0, + }, + ) def _make_response(url: str) -> httpx.Response: @@ -30,18 +49,19 @@ def _make_logging_obj() -> MagicMock: class TestAzureSpeechPassthroughHandler: @pytest.mark.parametrize( - "url_route,expected_model", + "url_route,expected_model,expected_cost", [ - (SHORT_AUDIO_URL, "azure_speech/short-audio"), - (BATCH_URL, "azure_speech/batch-transcription"), - (f"{BATCH_URL}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab/files", "azure_speech/batch-transcription"), + (SHORT_AUDIO_URL, "azure_speech/short-audio", TRANSCRIPT_AUDIO_SECONDS * PRICE_PER_SECOND), + (BATCH_URL, "azure_speech/batch-transcription", 0.0), + (f"{BATCH_URL}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab/files", "azure_speech/batch-transcription", 0.0), ], ) - def test_records_model_provider_and_zero_cost(self, url_route: str, expected_model: str): + def test_records_model_provider_and_cost(self, url_route: str, expected_model: str, expected_cost: float): logging_obj = _make_logging_obj() handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( httpx_response=_make_response(url_route), + response_body=TRANSCRIPT_BODY, logging_obj=logging_obj, url_route=url_route, result=TRANSCRIPT, @@ -54,16 +74,65 @@ class TestAzureSpeechPassthroughHandler: assert handler_result["result"] == {"response": TRANSCRIPT} assert handler_result["kwargs"]["model"] == expected_model assert handler_result["kwargs"]["custom_llm_provider"] == "azure_speech" - assert handler_result["kwargs"]["response_cost"] == 0.0 - assert handler_result["kwargs"]["standard_logging_object"]["response_cost"] == 0.0 + assert handler_result["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert handler_result["kwargs"]["standard_logging_object"]["response_cost"] == pytest.approx(expected_cost) assert handler_result["kwargs"]["standard_logging_object"]["model"] == expected_model assert logging_obj.model_call_details["model"] == expected_model assert logging_obj.model_call_details["custom_llm_provider"] == "azure_speech" - assert logging_obj.model_call_details["response_cost"] == 0.0 + assert logging_obj.model_call_details["response_cost"] == pytest.approx(expected_cost) + + @pytest.mark.parametrize( + "response_body", + [ + {"RecognitionStatus": "NoMatch", "Offset": 0, "Duration": 0}, + {"RecognitionStatus": "InitialSilenceTimeout"}, + {"Offset": "5000000", "Duration": "25000000"}, + {}, + [], + None, + ], + ) + def test_short_audio_without_recognized_duration_logs_zero_cost( + self, response_body: dict[str, object] | list[dict[str, object]] | None + ): + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(SHORT_AUDIO_URL), + response_body=response_body, + logging_obj=_make_logging_obj(), + url_route=SHORT_AUDIO_URL, + result="", + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert handler_result["kwargs"]["model"] == "azure_speech/short-audio" + assert handler_result["kwargs"]["response_cost"] == 0.0 + + def test_missing_price_entry_still_logs_the_row_at_zero_cost(self, monkeypatch: pytest.MonkeyPatch): + monkeypatch.delitem(litellm.model_cost, "azure/speech/azure-stt") + + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(SHORT_AUDIO_URL), + response_body=TRANSCRIPT_BODY, + logging_obj=_make_logging_obj(), + url_route=SHORT_AUDIO_URL, + result=TRANSCRIPT, + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert handler_result["kwargs"]["model"] == "azure_speech/short-audio" + assert handler_result["kwargs"]["custom_llm_provider"] == "azure_speech" + assert handler_result["kwargs"]["response_cost"] == 0.0 def test_subscription_key_never_reaches_the_logging_payload(self): handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( httpx_response=_make_response(SHORT_AUDIO_URL), + response_body=TRANSCRIPT_BODY, logging_obj=_make_logging_obj(), url_route=SHORT_AUDIO_URL, result=TRANSCRIPT, @@ -106,18 +175,18 @@ class TestNormalizeDispatch: def test_normalize_routes_to_azure_speech_handler(self): normalized = PassThroughEndpointLogging().normalize_llm_passthrough_logging_payload( httpx_response=_make_response(SHORT_AUDIO_URL), - response_body={"RecognitionStatus": "Success"}, + response_body=TRANSCRIPT_BODY, request_body={}, logging_obj=_make_logging_obj(), url_route=SHORT_AUDIO_URL, - result=TRANSCRIPT, + result="", start_time=datetime.now(), end_time=datetime.now(), cache_hit=False, custom_llm_provider="azure_speech", ) - assert normalized["standard_logging_response_object"] == {"response": TRANSCRIPT} + assert normalized["standard_logging_response_object"] == {"response": ""} assert normalized["kwargs"]["model"] == "azure_speech/short-audio" assert normalized["kwargs"]["custom_llm_provider"] == "azure_speech" - assert normalized["kwargs"]["response_cost"] == 0.0 + assert normalized["kwargs"]["response_cost"] == pytest.approx(TRANSCRIPT_AUDIO_SECONDS * PRICE_PER_SECOND) diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index 9fe7f5b6ee9..a1102543ae8 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -6278,6 +6278,49 @@ class TestAzureSpeechProxyRoute: ("azure_speech/batch-transcription", "azure_speech", 0.0) ] + def test_short_audio_spend_is_priced_from_the_recognized_duration( + self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch + ) -> None: + from litellm.integrations.custom_logger import CustomLogger + + class _Recorder(CustomLogger): + def __init__(self) -> None: + super().__init__() + self.payloads: list[dict[str, object]] = [] # mutable-ok: test recorder accumulates callback payloads + + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time) -> None: + self.payloads.append(kwargs["standard_logging_object"]) + + recorder: Final = _Recorder() + monkeypatch.setattr(litellm, "_async_success_callback", [*litellm._async_success_callback, recorder]) + monkeypatch.setitem( + litellm.model_cost, + "azure/speech/azure-stt", + { + "litellm_provider": "azure", + "mode": "audio_transcription", + "input_cost_per_second": 0.25, + "output_cost_per_second": 0.0, + }, + ) + transcript: Final = {**AZURE_SPEECH_TRANSCRIPT, "Offset": 10_000_000, "Duration": 30_000_000} + with respx.mock(assert_all_called=True) as upstream: + upstream.post(f"https://eastus.stt.speech.microsoft.com{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}").mock( + return_value=httpx.Response(200, json=transcript) + ) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", + content=AZURE_SPEECH_WAV_BYTES, + headers={"Content-Type": "audio/wav", "Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 200 + assert [(p["model"], p["custom_llm_provider"]) for p in recorder.payloads] == [ + ("azure_speech/short-audio", "azure_speech") + ] + assert recorder.payloads[0]["response_cost"] == pytest.approx(4.0 * 0.25) + def test_api_base_wins_over_region_for_both_families( self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch ) -> None: From 6e56ba86c5ed3851f8ef4b4b309c1e85949606f9 Mon Sep 17 00:00:00 2001 From: yassin Date: Thu, 17 Sep 2026 04:03:22 +0000 Subject: [PATCH 058/141] test(proxy): clear leaked auth dependency override before Azure Speech real-auth tests Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../pass_through_endpoints/test_llm_pass_through_endpoints.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index a1102543ae8..05ef44b89b6 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -6423,6 +6423,7 @@ class TestAzureSpeechRawBodyThroughRealAuth: ) -> httpx.Response: from litellm.proxy.proxy_server import app + monkeypatch.delitem(app.dependency_overrides, user_api_key_auth, raising=False) monkeypatch.setenv("AZURE_SPEECH_API_KEY", "server-subscription-key") monkeypatch.setenv("AZURE_SPEECH_REGION", "eastus") monkeypatch.delenv("AZURE_SPEECH_API_BASE", raising=False) From 1d8f19e4fde7615dc1828ebf7b4bf1e0d510af85 Mon Sep 17 00:00:00 2001 From: yassin Date: Thu, 17 Sep 2026 04:14:26 +0000 Subject: [PATCH 059/141] fix(proxy): keep Azure Speech multipart bodies intact through auth user_api_key_auth called request.form() on multipart Azure Speech batch uploads, consuming the Starlette stream before the pass-through handler could read the raw bytes. The opaque body predicate now covers multipart on the whole /azure_speech prefix so auth caches an empty parsed body and the upload is forwarded byte for byte Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../proxy/common_utils/http_parsing_utils.py | 9 +-- .../test_llm_pass_through_endpoints.py | 63 ++++++++++++++++--- 2 files changed, 61 insertions(+), 11 deletions(-) diff --git a/litellm/proxy/common_utils/http_parsing_utils.py b/litellm/proxy/common_utils/http_parsing_utils.py index 29dc36f3dba..1c17c46e5af 100644 --- a/litellm/proxy/common_utils/http_parsing_utils.py +++ b/litellm/proxy/common_utils/http_parsing_utils.py @@ -11,7 +11,6 @@ from typing_extensions import NotRequired, ReadOnly, Required from litellm._logging import verbose_proxy_logger from litellm.constants import ( AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX, - AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, CLIENT_REQUESTED_MODEL_SCOPE_KEY, MAX_REQUEST_BODY_SIZE_TO_REPAIR_MB, ) @@ -220,9 +219,11 @@ async def _read_request_body(request: Request | None) -> dict: def is_opaque_audio_pass_through_request(route: str, content_type: str) -> bool: - return route.startswith( - f"{AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX}{AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX}" - ) and _normalize_media_type(content_type).startswith("audio/") + """Azure Speech bodies (raw audio, multipart uploads) are forwarded byte for byte, so auth must not consume them.""" + media_type: Final = _normalize_media_type(content_type) + return route.startswith(f"{AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX}/") and ( + media_type.startswith("audio/") or media_type == "multipart/form-data" + ) async def read_raw_json_body(request: Request | None) -> bytes | None: diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index 05ef44b89b6..43fc28c34c4 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -6418,8 +6418,8 @@ def _azure_speech_real_auth_attrs() -> dict[str, object]: class TestAzureSpeechRawBodyThroughRealAuth: """user_api_key_auth reads the body before the route runs; raw audio must not be parsed as JSON.""" - def _post_wav( - self, monkeypatch: pytest.MonkeyPatch, path: str, api_key: str, body: bytes = AZURE_SPEECH_WAV_BYTES + def _post( + self, monkeypatch: pytest.MonkeyPatch, path: str, api_key: str, content_type: str, body: bytes ) -> httpx.Response: from litellm.proxy.proxy_server import app @@ -6437,9 +6437,14 @@ class TestAzureSpeechRawBodyThroughRealAuth: path, params={"language": "en-US"}, content=body, - headers={"Content-Type": "audio/wav", "Authorization": f"Bearer {api_key}"}, + headers={"Content-Type": content_type, "Authorization": f"Bearer {api_key}"}, ) + def _post_wav( + self, monkeypatch: pytest.MonkeyPatch, path: str, api_key: str, body: bytes = AZURE_SPEECH_WAV_BYTES + ) -> httpx.Response: + return self._post(monkeypatch, path, api_key, "audio/wav", body) + @pytest.mark.parametrize("body", [AZURE_SPEECH_WAV_BYTES, AZURE_SPEECH_NON_UTF8_WAV_BYTES], ids=["ascii", "binary"]) def test_master_key_with_raw_wav_body_reaches_azure_without_a_parse_attempt( self, monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture, body: bytes @@ -6466,11 +6471,55 @@ class TestAzureSpeechRawBodyThroughRealAuth: assert response.status_code in (400, 401), response.text assert not catch_all.called - @pytest.mark.parametrize("path", ["/v1/chat/completions", f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}"]) - def test_audio_content_type_off_the_short_audio_route_is_still_parsed_as_json( - self, monkeypatch: pytest.MonkeyPatch, path: str + def test_master_key_with_multipart_batch_upload_is_forwarded_byte_for_byte( + self, monkeypatch: pytest.MonkeyPatch ) -> None: - response = self._post_wav(monkeypatch, path, "sk-master-key", body=b'{}{"model": "gpt-4o"}') + boundary: Final = "lit7939boundary" + multipart_body: Final = ( + f"--{boundary}\r\nContent-Disposition: form-data; name=\"definition\"\r\n\r\n".encode() + + json.dumps({"locales": ["en-US"]}).encode() + + f"\r\n--{boundary}\r\nContent-Disposition: form-data; name=\"audio\"; filename=\"eagle.wav\"\r\n" + "Content-Type: audio/wav\r\n\r\n".encode() + + AZURE_SPEECH_NON_UTF8_WAV_BYTES + + f"\r\n--{boundary}--\r\n".encode() + ) + with respx.mock(assert_all_called=True) as upstream: + route = upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( + return_value=httpx.Response(201, json={"status": "NotStarted"}) + ) + + response = self._post( + monkeypatch, + f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", + "sk-master-key", + f"multipart/form-data; boundary={boundary}", + multipart_body, + ) + + assert (response.status_code, response.json()) == (201, {"status": "NotStarted"}) + sent = route.calls.last.request + assert sent.content == multipart_body + assert sent.headers["content-type"] == f"multipart/form-data; boundary={boundary}" + assert sent.headers["ocp-apim-subscription-key"] == "server-subscription-key" + + @pytest.mark.parametrize("content_type", ["audio/wav", "multipart/form-data; boundary=x"]) + def test_wrong_litellm_key_with_multipart_batch_upload_is_rejected( + self, monkeypatch: pytest.MonkeyPatch, content_type: str + ) -> None: + with respx.mock(assert_all_called=False) as upstream: + catch_all = upstream.route().mock(return_value=httpx.Response(200)) + + response = self._post( + monkeypatch, f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", "sk-wrong", content_type, b"--x--\r\n" + ) + + assert response.status_code in (400, 401), response.text + assert not catch_all.called + + def test_audio_content_type_off_the_azure_speech_route_is_still_parsed_as_json( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + response = self._post_wav(monkeypatch, "/v1/chat/completions", "sk-master-key", body=b'{}{"model": "gpt-4o"}') assert response.status_code == 400 assert "Invalid JSON payload" in response.text From 0986f404f8bc189854a9a7d88dfd4af376c84566 Mon Sep 17 00:00:00 2001 From: yucheng Date: Thu, 17 Sep 2026 06:08:12 +0000 Subject: [PATCH 060/141] fix(proxy): match IPv4-mapped IPv6 peers against IPv4 trusted proxy ranges Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/network.py | 8 +++++++- tests/test_litellm/proxy/auth/test_network.py | 7 +++++++ 2 files changed, 14 insertions(+), 1 deletion(-) diff --git a/litellm/proxy/auth/network.py b/litellm/proxy/auth/network.py index 32ad18d4deb..28466156100 100644 --- a/litellm/proxy/auth/network.py +++ b/litellm/proxy/auth/network.py @@ -49,11 +49,17 @@ def parse_trusted_proxy_ranges( return networks +def _unmapped(addr: ipaddress.IPv4Address | ipaddress.IPv6Address) -> ipaddress.IPv4Address | ipaddress.IPv6Address: + if isinstance(addr, ipaddress.IPv6Address) and addr.ipv4_mapped is not None: + return addr.ipv4_mapped + return addr + + def ip_in_networks(client_ip: str | None, networks: list[TrustedProxyNetwork]) -> bool: if not client_ip or not networks: return False try: - addr: Final = ipaddress.ip_address(client_ip.strip()) + addr: Final = _unmapped(ipaddress.ip_address(client_ip.strip())) except ValueError: return False return any(addr in network for network in networks) diff --git a/tests/test_litellm/proxy/auth/test_network.py b/tests/test_litellm/proxy/auth/test_network.py index b67723305e4..83233dc370d 100644 --- a/tests/test_litellm/proxy/auth/test_network.py +++ b/tests/test_litellm/proxy/auth/test_network.py @@ -57,6 +57,13 @@ def test_xff_honored_from_trusted_peer(): assert via_proxy is True +def test_ipv4_mapped_peer_and_hop_match_ipv4_trusted_ranges(): + request = make_request(headers={"x-forwarded-for": "203.0.113.9, ::ffff:10.0.0.5"}, client=("::ffff:10.0.0.1", 1)) + ip, via_proxy = resolve_client_ip(request, TRUSTED) + assert ip == "203.0.113.9" + assert via_proxy is True + + def test_spoofed_xff_from_untrusted_peer_is_ignored(): request = make_request( headers={"x-forwarded-for": "203.0.113.9"}, client=("8.8.8.8", 1) From c05095373d101f5192a4703788c940b5925bc2aa Mon Sep 17 00:00:00 2001 From: yucheng Date: Thu, 17 Sep 2026 06:41:23 +0000 Subject: [PATCH 061/141] fix(proxy): keep mapped-notation trusted proxy ranges matching mapped peers Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/network.py | 5 +++-- tests/test_litellm/proxy/auth/test_network.py | 8 ++++++++ 2 files changed, 11 insertions(+), 2 deletions(-) diff --git a/litellm/proxy/auth/network.py b/litellm/proxy/auth/network.py index 28466156100..8a20e207113 100644 --- a/litellm/proxy/auth/network.py +++ b/litellm/proxy/auth/network.py @@ -59,10 +59,11 @@ def ip_in_networks(client_ip: str | None, networks: list[TrustedProxyNetwork]) - if not client_ip or not networks: return False try: - addr: Final = _unmapped(ipaddress.ip_address(client_ip.strip())) + addr: Final = ipaddress.ip_address(client_ip.strip()) except ValueError: return False - return any(addr in network for network in networks) + candidates: Final = (addr, _unmapped(addr)) + return any(candidate in network for candidate in candidates for network in networks) def _is_valid_ip(value: str) -> bool: diff --git a/tests/test_litellm/proxy/auth/test_network.py b/tests/test_litellm/proxy/auth/test_network.py index 83233dc370d..e743ce8cd23 100644 --- a/tests/test_litellm/proxy/auth/test_network.py +++ b/tests/test_litellm/proxy/auth/test_network.py @@ -64,6 +64,14 @@ def test_ipv4_mapped_peer_and_hop_match_ipv4_trusted_ranges(): assert via_proxy is True +def test_ipv4_mapped_peer_still_matches_mapped_notation_trusted_range(): + config = TrustedProxyConfig(use_forwarded_for=True, trusted_proxy_cidrs=["::ffff:10.0.0.0/104"]) + request = make_request(headers={"x-forwarded-for": "203.0.113.9"}, client=("::ffff:10.0.0.1", 1)) + ip, via_proxy = resolve_client_ip(request, config) + assert ip == "203.0.113.9" + assert via_proxy is True + + def test_spoofed_xff_from_untrusted_peer_is_ignored(): request = make_request( headers={"x-forwarded-for": "203.0.113.9"}, client=("8.8.8.8", 1) From 9a365d2021a0c6a24be4989b4aca4bb898f60e45 Mon Sep 17 00:00:00 2001 From: yucheng Date: Thu, 17 Sep 2026 16:10:23 +0000 Subject: [PATCH 062/141] feat(proxy): hard-block throttled Admin UI sign-ins with no credential bypass A blocked source, or source and username pair, is now refused with 429 before the database lookup and password check, in place of the soft block that held wrong guesses for 30 seconds and let a correct password through. The env admin credentials and the master key typed into the login form are refused like any other credential while blocked; recovery is the master key as an API bearer token, which never goes through the sign-in path trusted_proxy_ranges: [] now means clients connect directly, so the peer address is the source and the per-source limit stays on. Only an unset or malformed value leaves the topology unknown, warns at startup and turns the per-source limit off Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 6 +- litellm/proxy/auth/login_throttle.py | 100 +++--- litellm/proxy/auth/login_utils.py | 18 +- litellm/proxy/auth/network.py | 3 +- litellm/proxy/proxy_server.py | 4 +- .../proxy/auth/test_login_utils.py | 333 ++++++++---------- .../proxy/proxy_server/conftest.py | 7 - .../proxy_server/test_routes_login_sso.py | 59 +++- tests/test_litellm/proxy/test_proxy_server.py | 16 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 6 +- 10 files changed, 273 insertions(+), 279 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 08380d8b537..80147863304 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2755,7 +2755,7 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): max_failed_login_attempts_per_source: int | None = Field( None, ge=1, - description="Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set, since otherwise every client behind an ingress shares one address. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10", + description="Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and this limit is off. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10", ) max_failed_login_attempts_per_source_overrides: dict[str, int] | None = Field( None, @@ -2774,7 +2774,7 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): failed_login_block_seconds: int | None = Field( None, ge=1, - description="How long a blocked source address, or source address and username, stays blocked. The block is soft: a correct password still signs in, but each attempt from a blocked key takes one of 5 held slots per worker and a wrong password is held for 30 seconds before it is refused with 429. Set under `general_settings` in config.yaml. Defaults to 300", + description="How long a blocked source address, or source address and username, stays blocked. Every attempt from a blocked key, right or wrong, is refused with 429 before the password is checked; the block is not extended by refused attempts. Set under `general_settings` in config.yaml. Defaults to 300", ) allowed_routes: list | None = Field(None, description="Proxy API Endpoints you want users to be able to access") reject_clientside_metadata_tags: bool | None = Field( @@ -2885,7 +2885,7 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): ) trusted_proxy_ranges: list[str] | None = Field( None, - description="CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler.", + description="CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler, and whose X-Forwarded-For is used to attribute Admin UI sign-in attempts to a source address. Set it to an empty list when clients connect directly, so the peer address is the source. Left unset, the per-source sign-in limit is off.", ) store_model_in_db: bool | None = Field( None, diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index a8899b9c675..0106d58e426 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -1,10 +1,10 @@ """Failed-login accounting for the Admin UI sign-in path. Wrong passwords are counted over a short window per source address and per source-and-username -pair; too many in one window blocks that key for a fixed time. Blocks are soft: a correct password -still signs in, while a wrong one from a blocked key is held open before its 429 and only a few can -be held at once, which bounds how many guesses a blocked key gets checked. A blocked pair stops +pair; too many in one window blocks that key for a fixed time. While a key is blocked every attempt +from it, right or wrong, is refused with 429 before the password is checked. A blocked pair stops counting against its source, so one script stuck on one account does not block the whole office. +Recovery is the master key over the API, which never passes through here, or waiting out the block. """ from __future__ import annotations @@ -14,8 +14,7 @@ import hashlib import ipaddress import math import time -from collections.abc import AsyncGenerator, Mapping -from contextlib import asynccontextmanager +from collections.abc import Mapping from dataclasses import dataclass from functools import cache from types import MappingProxyType @@ -37,8 +36,6 @@ DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_USER: Final = 5 DEFAULT_FAILED_LOGIN_WINDOW_SECONDS: Final = 60 DEFAULT_FAILED_LOGIN_BLOCK_SECONDS: Final = 300 -BLOCKED_ATTEMPT_HOLD_SECONDS: Final = 30 -MAX_HELD_ATTEMPTS_PER_KEY: Final = 5 IPV6_SOURCE_PREFIX_LENGTH: Final = 64 SOURCE_LIMIT_KEY: Final = "max_failed_login_attempts_per_source" @@ -100,11 +97,6 @@ _COUNTERS: Final = InMemoryCache( max_size_in_memory=_MAX_TRACKED_COUNTERS, default_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS ) _BLOCKS: Final = InMemoryCache(max_size_in_memory=_MAX_TRACKED_BLOCKS, default_ttl=DEFAULT_FAILED_LOGIN_BLOCK_SECONDS) -_HELD_ATTEMPTS: Final[dict[str, int]] = {} # mutable-ok: in-flight hold counts rise on entry and fall on exit - - -async def _sleep(seconds: float) -> None: - await asyncio.sleep(seconds) @cache @@ -127,12 +119,25 @@ def warn_login_counters_are_per_worker(num_workers: str) -> None: def warn_source_login_limit_is_off() -> None: verbose_proxy_logger.warning( "%s is not set, so failed Admin UI sign-in attempts are limited per source address and username " - "only. Set it to the address ranges of the proxies in front of LiteLLM to also limit each " - "source address across usernames.", + "only. Set it to the address ranges of the proxies in front of LiteLLM, or to an empty list when " + "clients connect directly, to also limit each source address across usernames.", TRUSTED_PROXY_RANGES_KEY, ) +def declared_proxy_ranges(settings: Mapping[str, object]) -> tuple[str, ...] | None: + """What the operator says fronts LiteLLM: the proxy ranges, an empty tuple for none, None when unsaid. + + Only a declared topology makes the source address trustworthy enough to limit across usernames. + An unset key, or a value that is not a list of ranges, leaves it unknown and the source scope off. + """ + raw_ranges: Final = settings.get(TRUSTED_PROXY_RANGES_KEY) + if isinstance(raw_ranges, (list, tuple, set)) and not raw_ranges: + return () + cidrs: Final = tuple(normalize_cidr_ranges(raw_ranges, setting_name=TRUSTED_PROXY_RANGES_KEY)) + return cidrs or None + + def _positive_int(raw: object, key: str, default: int) -> int: if raw is None: return default @@ -221,8 +226,11 @@ class Block: @dataclass(frozen=True, slots=True) class LoginThrottle: - """Failed-login limits for one request's source address; ``source_limit`` is None when the - source scope is off because ``trusted_proxy_ranges`` is unset and the peer address is the ingress.""" + """Failed-login limits for one request's source address. + + ``source_limit`` is None when the source scope is off: ``trusted_proxy_ranges`` is unset, so the peer + address may be a shared ingress. An empty list means clients connect directly and the peer is the source. + """ client_ip: str source_limit: int | None @@ -242,15 +250,13 @@ class LoginThrottle: redis_cache: RedisCache | None, ) -> LoginThrottle: settings: Final = general_settings if general_settings is not None else _NO_SETTINGS - cidrs: Final = normalize_cidr_ranges( - settings.get(TRUSTED_PROXY_RANGES_KEY), setting_name=TRUSTED_PROXY_RANGES_KEY - ) + proxies: Final = declared_proxy_ranges(settings) resolved, _ = resolve_client_ip( - request, TrustedProxyConfig(use_forwarded_for=bool(cidrs), trusted_proxy_cidrs=cidrs) + request, TrustedProxyConfig(use_forwarded_for=bool(proxies), trusted_proxy_cidrs=proxies or ()) ) return cls( client_ip=resolved or _UNKNOWN_SOURCE, - source_limit=_source_limit(settings, resolved) if cidrs and resolved is not None else None, + source_limit=_source_limit(settings, resolved) if proxies is not None and resolved is not None else None, user_limit=_int_setting(settings, USER_LIMIT_KEY, DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_USER), window_seconds=_int_setting(settings, WINDOW_KEY, DEFAULT_FAILED_LOGIN_WINDOW_SECONDS), block_seconds=_int_setting(settings, BLOCK_KEY, DEFAULT_FAILED_LOGIN_BLOCK_SECONDS), @@ -270,18 +276,21 @@ class LoginThrottle: source_block=f"{_CACHE_KEY_PREFIX}:{{{group}}}:block:source", ) - @asynccontextmanager - async def attempt(self, username: str, *, exempt: bool = False) -> AsyncGenerator[LoginAttempt]: - if not self.enabled or exempt: - yield LoginAttempt(throttle=self, username=username, block=None) - return - keys: Final = self._keys(username) - block: Final = await self._active_block(keys) + async def attempt(self, username: str) -> LoginAttempt: + """Refuses a blocked key before any credential is looked at; otherwise hands back the attempt to settle.""" + if not self.enabled: + return LoginAttempt(throttle=self, username=username) + block: Final = await self._active_block(self._keys(username)) if block is None: - yield LoginAttempt(throttle=self, username=username, block=None) - return - slot: Final = keys.pair_block if block.scope == "user" else keys.source_block - yield LoginAttempt(throttle=self, username=username, block=block, slot=slot) + return LoginAttempt(throttle=self, username=username) + verbose_proxy_logger.warning( + "Admin UI sign-in refused: the %s is blocked for %s more seconds; username=%r source=%s", + block.scope, + block.retry_after, + username, + self.client_ip, + ) + self.refuse(block.retry_after) async def _active_block(self, keys: _Keys) -> Block | None: local: Final = self._local_block_ttls(keys) @@ -372,8 +381,6 @@ class LoginThrottle: class LoginAttempt: throttle: LoginThrottle username: str - block: Block | None - slot: str | None = None async def succeeded(self) -> None: if not self.throttle.enabled: @@ -383,8 +390,6 @@ class LoginAttempt: async def failed(self) -> None: if not self.throttle.enabled: return - if self.block is not None and self.slot is not None: - await self._hold_then_refuse(self.block, self.slot) user_block, source_block = await self.throttle.record_failure(self.username) if user_block == 0 and source_block == 0: return @@ -395,26 +400,3 @@ class LoginAttempt: self.username, self.throttle.client_ip, ) - - async def _hold_then_refuse(self, block: Block, slot: str) -> NoReturn: - held: Final = _HELD_ATTEMPTS.get(slot, 0) - if held >= MAX_HELD_ATTEMPTS_PER_KEY: - verbose_proxy_logger.warning( - "Admin UI sign-in refused at once: %s wrong attempts already held for a blocked %s; " - "username=%r source=%s", - held, - block.scope, - self.username, - self.throttle.client_ip, - ) - self.throttle.refuse(block.retry_after) - _HELD_ATTEMPTS[slot] = held + 1 - try: - await _sleep(BLOCKED_ATTEMPT_HOLD_SECONDS) - finally: - remaining: Final = _HELD_ATTEMPTS.get(slot, 1) - 1 - if remaining > 0: - _HELD_ATTEMPTS[slot] = remaining - else: - _HELD_ATTEMPTS.pop(slot, None) - self.throttle.refuse(max(block.retry_after - BLOCKED_ATTEMPT_HOLD_SECONDS, 1)) diff --git a/litellm/proxy/auth/login_utils.py b/litellm/proxy/auth/login_utils.py index 6f52babb255..e0d599b0017 100644 --- a/litellm/proxy/auth/login_utils.py +++ b/litellm/proxy/auth/login_utils.py @@ -186,9 +186,11 @@ async def authenticate_user( or if username/password login is disabled while SSO is configured Recovery: an admin locked out of the UI by - `disable_password_login_when_sso_enabled` can still administer the proxy over - the API with the master key (Authorization: Bearer ), which never - goes through this function. To restore UI username/password login, unset the + `disable_password_login_when_sso_enabled`, or by the failed sign-in block in + `throttle`, can still administer the proxy over the API with the master key + (Authorization: Bearer ), which never goes through this function. + No credential, the env admin credentials and the master key included, is + exempt from the block. To restore UI username/password login, unset the setting in config.yaml (or the DB-persisted general_settings) and restart the proxy; this is a deliberate, auditable config change rather than a hidden bypass. @@ -217,12 +219,8 @@ async def authenticate_user( code=500, ) - admin_credentials_match: Final = _admin_credentials_match(username, password, master_key, general_settings) - - async with throttle.attempt(username, exempt=admin_credentials_match) as attempt: - return await _sign_in( - username, password, master_key, prisma_client, attempt, general_settings, admin_credentials_match - ) + attempt: Final = await throttle.attempt(username) + return await _sign_in(username, password, master_key, prisma_client, attempt, general_settings) async def _sign_in( @@ -232,8 +230,8 @@ async def _sign_in( prisma_client: PrismaClient | None, attempt: LoginAttempt, general_settings: Mapping[str, object], - admin_credentials_match: bool, ) -> LoginResult: + admin_credentials_match: Final = _admin_credentials_match(username, password, master_key, general_settings) # Check if we can find the `username` in the db. On the UI, users can enter username=their email _user_row: LiteLLM_UserTable | None = None user_role: ( diff --git a/litellm/proxy/auth/network.py b/litellm/proxy/auth/network.py index 8a20e207113..4e8ab7512a7 100644 --- a/litellm/proxy/auth/network.py +++ b/litellm/proxy/auth/network.py @@ -1,6 +1,7 @@ from __future__ import annotations import ipaddress +from collections.abc import Sequence from typing import Any, Final from fastapi import Request @@ -19,7 +20,7 @@ class NetworkContext(BaseModel): class TrustedProxyConfig(BaseModel): use_forwarded_for: bool = False - trusted_proxy_cidrs: list[str] = Field(default_factory=list) + trusted_proxy_cidrs: Sequence[str] = Field(default_factory=tuple) def normalize_cidr_ranges(configured_ranges: Any, *, setting_name: str = "trusted_proxy_cidrs") -> list[str]: diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index d00c695d968..4075ac4aff7 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -328,8 +328,8 @@ from litellm.proxy.auth.fallback_model_access import router_fallback_access_chec from litellm.proxy.auth.handle_jwt import JWTHandler from litellm.proxy.auth.litellm_license import AUTO_ROUTER_LICENSE_REMEDY, LicenseCheck from litellm.proxy.auth.login_throttle import ( - TRUSTED_PROXY_RANGES_KEY, LoginThrottle, + declared_proxy_ranges, warn_login_counters_are_per_worker, warn_source_login_limit_is_off, ) @@ -5819,7 +5819,7 @@ class ProxyConfig: if os.getenv("NUM_WORKERS", "1") != "1" and redis_usage_cache is None: warn_login_counters_are_per_worker(os.getenv("NUM_WORKERS", "1")) - if not general_settings.get(TRUSTED_PROXY_RANGES_KEY): + if declared_proxy_ranges(general_settings) is None: warn_source_login_limit_is_off() _bg_hc_model_groups: Final = parse_background_health_check_model_groups(general_settings) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index d1fdb2d6a70..8670d7fef41 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -13,28 +13,6 @@ from unittest.mock import AsyncMock, MagicMock, patch import pytest -class _RecordedSleeps: - """A sleep that records what it was asked to wait for instead of waiting.""" - - def __init__(self): - self.seconds: list[float] = [] - - async def __call__(self, seconds: float) -> None: - self.seconds.append(seconds) - - -@pytest.fixture(autouse=True) -def login_delays(monkeypatch): - """Replace the hold on a blocked wrong password, so the suite pays no wall clock and can read it back.""" - from litellm.proxy.auth import login_throttle - - recorded = _RecordedSleeps() - monkeypatch.setattr(login_throttle, "_sleep", recorded) - login_throttle._HELD_ATTEMPTS.clear() - yield recorded - login_throttle._HELD_ATTEMPTS.clear() - - def _unlimited_throttle(): """A throttle wired to real in-memory stores with limits no test can reach.""" from litellm.caching.in_memory_cache import InMemoryCache @@ -749,8 +727,8 @@ def _local_count(throttle, key: str) -> int: @pytest.mark.asyncio async def test_too_many_failures_for_one_username_block_that_pair_and_carry_retry_after(monkeypatch): - """One failure past the pair limit blocks the source for that username; the next wrong guess is held - and answered 429 with the block's remaining time, and the counter is not touched by blocked guesses.""" + """One failure past the pair limit blocks the source for that username; the next guess is answered 429 + with the block's remaining time, and the counter is not touched by blocked guesses.""" from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") @@ -764,30 +742,17 @@ async def test_too_many_failures_for_one_username_block_that_pair_and_carry_retr with pytest.raises(ProxyException) as blocked: await _guess(throttle) assert blocked.value.code == "429" - assert blocked.value.headers.get("Retry-After") == "47", "the 30s hold is taken off the remaining block" + assert blocked.value.headers.get("Retry-After") == "77" assert _local_count(throttle, keys.pair_counter) == 3, "a blocked guess is not counted again" @pytest.mark.asyncio -async def test_a_wrong_password_from_a_blocked_key_is_held_before_it_is_refused(monkeypatch, login_delays): - """The hold is the rate cap: a blocked key gets one verified guess per held slot per 30 seconds.""" - from litellm.proxy.auth.login_throttle import BLOCKED_ATTEMPT_HOLD_SECONDS +async def test_a_blocked_key_is_refused_before_the_password_is_looked_at(monkeypatch): + """The block is the rate cap: once a key is blocked, nothing from it reaches the user lookup or the + password check, so a guessing script gets no verification work out of the proxy.""" + from litellm.proxy._types import ProxyException + from litellm.proxy.auth.login_utils import authenticate_user - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - throttle = _throttle(user_limit=1) - - assert [await _fail(throttle) for _ in range(2)] == ["401", "401"] - assert login_delays.seconds == [], "an unblocked wrong password is answered at once" - - assert await _fail(throttle) == "429" - assert login_delays.seconds == [BLOCKED_ATTEMPT_HOLD_SECONDS] - - -@pytest.mark.asyncio -async def test_a_correct_password_signs_in_while_its_pair_is_blocked(monkeypatch): - """The block is soft: the real user is still verified and gets in, so nobody can be locked out by - guessing at their account.""" monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") monkeypatch.setenv("DATABASE_URL", "postgresql://stub") @@ -795,13 +760,50 @@ async def test_a_correct_password_signs_in_while_its_pair_is_blocked(monkeypatch assert [await _fail(throttle, username="user@corp.com") for _ in range(3)] == ["401", "401", "429"] - result = await _db_login(throttle, "user@corp.com", "right", correct=True) - assert result.key == "sk-ui" + lookup = _known_user("user@corp.com") + verify = MagicMock(return_value=True) + with ( + patch( # test-quality-ok: the user lookup is the database boundary; a blocked attempt must not reach it + "litellm.proxy.auth.login_utils.UserRepository", lookup + ), + patch( # test-quality-ok: the password check is the expensive step; a blocked attempt must not reach it + "litellm.proxy.auth.login_utils.verify_password", verify + ), + pytest.raises(ProxyException) as refused, + ): + await authenticate_user( + username="user@corp.com", + password="right", + master_key="sk-master", + prisma_client=MagicMock(), + throttle=throttle, + ) + + assert refused.value.code == "429" + assert lookup.return_value.table.find_first.await_count == 0 + assert verify.call_count == 0 @pytest.mark.asyncio -async def test_a_correct_password_signs_in_while_its_source_is_blocked(monkeypatch): - """Same for the source-wide block: it slows guessing from that address, it does not refuse a user.""" +async def test_a_correct_password_is_refused_while_its_pair_is_blocked(monkeypatch): + """Letting the right password through would give a guesser unlimited tries, so the block is hard: the + real user waits it out, or uses the master key over the API, which never passes through here.""" + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + throttle = _throttle(user_limit=1, block_seconds=90) + + assert [await _fail(throttle, username="user@corp.com") for _ in range(3)] == ["401", "401", "429"] + + with pytest.raises(ProxyException) as refused: + await _db_login(throttle, "user@corp.com", "right", correct=True) + assert refused.value.code == "429" + assert refused.value.headers.get("Retry-After") == "90" + + +@pytest.mark.asyncio +async def test_a_correct_password_is_refused_while_its_source_is_blocked(monkeypatch): + """Same for the source-wide block: every username from that address is refused until it lapses.""" monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") monkeypatch.setenv("DATABASE_URL", "postgresql://stub") @@ -811,8 +813,9 @@ async def test_a_correct_password_signs_in_while_its_source_is_blocked(monkeypat assert await _fail(throttle, username=f"other-{i}@corp.com") == "401" assert await _fail(throttle, username="other-9@corp.com") == "429", "the source is blocked for everyone" - result = await _db_login(throttle, "user@corp.com", "right", correct=True) - assert result.key == "sk-ui" + with pytest.raises(ProxyException) as refused: + await _db_login(throttle, "user@corp.com", "right", correct=True) + assert refused.value.code == "429" @pytest.mark.asyncio @@ -903,6 +906,63 @@ async def test_without_trusted_proxy_ranges_the_source_scope_is_off(monkeypatch) assert [await _fail(throttle, username=f"user-{i}@corp.com") for i in range(6)] == ["401"] * 6 +@pytest.mark.asyncio +async def test_an_empty_trusted_proxy_ranges_means_the_peer_is_the_client_and_the_source_scope_is_on(monkeypatch): + """An explicit empty list says there are no proxies: the peer address is the client, the forwarded header + is ignored, and the source-wide limit applies. Only an unset key means the topology is unknown.""" + from litellm.proxy.auth.login_throttle import LoginThrottle + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.setenv("UI_PASSWORD", "right") + request = MagicMock() + request.headers = {"x-forwarded-for": "203.0.113.9"} + request.client = MagicMock() + request.client.host = "198.51.100.7" + throttle = LoginThrottle.from_request( + request, + general_settings={"trusted_proxy_ranges": [], "max_failed_login_attempts_per_source": 3}, + redis_cache=None, + ) + + assert throttle.client_ip == "198.51.100.7" + assert throttle.source_limit == 3 + assert [await _fail(throttle, username=f"user-{i}@corp.com") for i in range(4)] == ["401"] * 4 + assert await _fail(throttle, username="user-99@corp.com") == "429", "the spray is stopped by the source limit" + + +@pytest.mark.parametrize("configured", [None, 5, {"10.0.0.0/8": True}, ["", " "]]) +def test_a_trusted_proxy_ranges_value_that_names_no_ranges_leaves_the_topology_unknown(configured): + """Only a real list of ranges or an explicit empty list counts as a declaration; anything else is the same + as unset, so a typo cannot switch the source-wide block on behind a shared ingress.""" + from litellm.proxy.auth.login_throttle import LoginThrottle, declared_proxy_ranges + + settings = {"trusted_proxy_ranges": configured} if configured is not None else {} + assert declared_proxy_ranges(settings) is None + + request = MagicMock() + request.headers = {} + request.client = MagicMock() + request.client.host = "198.51.100.7" + throttle = LoginThrottle.from_request(request, general_settings=settings, redis_cache=None) + assert throttle.source_limit is None + assert throttle.client_ip == "198.51.100.7" + + +def test_declared_proxy_ranges_distinguishes_none_from_empty_from_configured(): + from litellm.proxy.auth.login_throttle import declared_proxy_ranges + + assert declared_proxy_ranges({}) is None + assert declared_proxy_ranges({"trusted_proxy_ranges": []}) == () + assert declared_proxy_ranges({"trusted_proxy_ranges": ["10.0.0.0/8", " 192.168.1.1 "]}) == ( + "10.0.0.0/8", + "192.168.1.1", + ) + assert declared_proxy_ranges({"trusted_proxy_ranges": "10.0.0.0/8,172.16.0.0/12"}) == ( + "10.0.0.0/8", + "172.16.0.0/12", + ) + + @pytest.mark.asyncio async def test_with_trusted_proxy_ranges_the_source_is_the_forwarded_client(monkeypatch): """The header is walked right to left past the trusted hops, so a forged left-most entry cannot pick the bucket.""" @@ -1056,9 +1116,10 @@ async def test_the_block_time_is_fixed_and_not_refreshed_by_blocked_guesses(monk @pytest.mark.asyncio -async def test_the_configured_admin_credentials_bypass_the_throttle(monkeypatch): - """The only account that can fix a misconfiguration is exempt: no hold, no slot, even while blocked.""" - from litellm.proxy.auth import login_throttle as lt +async def test_the_configured_admin_credentials_are_not_exempt_from_the_block(monkeypatch): + """Exempting the env credentials would make them the one password worth guessing without limit, so the + right UI_PASSWORD is refused while its pair is blocked, and signs in normally once the block lapses.""" + from litellm.proxy._types import ProxyException monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") @@ -1075,9 +1136,31 @@ async def test_the_configured_admin_credentials_bypass_the_throttle(monkeypatch) "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ), ): + with pytest.raises(ProxyException) as refused: + await _guess(throttle, password="right") + assert refused.value.code == "429" + + throttle.blocks.delete_cache(throttle._keys("admin").pair_block) result = await _guess(throttle, password="right") assert result.key == "sk-ui" - assert lt._HELD_ATTEMPTS == {} + + +@pytest.mark.asyncio +async def test_the_master_key_used_as_the_ui_password_is_not_exempt_from_the_block(monkeypatch): + """Without UI_PASSWORD the master key doubles as the admin password; it gets no special treatment here + either. Lockout recovery is the master key as a bearer token over the API, which never enters this path.""" + from litellm.proxy._types import ProxyException + + monkeypatch.setenv("UI_USERNAME", "admin") + monkeypatch.delenv("UI_PASSWORD", raising=False) + monkeypatch.setenv("DATABASE_URL", "postgresql://stub") + throttle = _throttle(user_limit=1) + + assert [await _fail(throttle) for _ in range(3)] == ["401", "401", "429"] + + with pytest.raises(ProxyException) as refused: + await _guess(throttle, password="sk-master") + assert refused.value.code == "429" @pytest.mark.asyncio @@ -1180,138 +1263,29 @@ async def test_a_wrong_password_for_a_known_user_also_counts(monkeypatch): @pytest.mark.asyncio -async def test_held_attempts_from_one_blocked_key_are_capped(monkeypatch): - """Holding a wrong guess open must not let one blocked key park unlimited sockets in password checks.""" - import asyncio - +async def test_a_source_block_outranks_a_pair_block_in_the_retry_after(monkeypatch): + """When both scopes are blocked, the answer carries the source block's time, which is the one that + still applies to every other username from that address.""" from litellm.proxy._types import ProxyException - from litellm.proxy.auth import login_throttle as lt - from litellm.proxy.auth.login_throttle import MAX_HELD_ATTEMPTS_PER_KEY monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - release = asyncio.Event() - - async def _park(_seconds: float) -> None: - await release.wait() - - monkeypatch.setattr(lt, "_sleep", _park) - throttle = _throttle(user_limit=1, client_ip="203.0.113.44") - slot = throttle._keys("admin").pair_block - assert [await _fail(throttle) for _ in range(2)] == ["401", "401"] - - held = [asyncio.create_task(_guess(throttle)) for _ in range(MAX_HELD_ATTEMPTS_PER_KEY)] - for _ in range(1000): - if lt._HELD_ATTEMPTS.get(slot) == MAX_HELD_ATTEMPTS_PER_KEY: - break - await asyncio.sleep(0) - assert lt._HELD_ATTEMPTS.get(slot) == MAX_HELD_ATTEMPTS_PER_KEY - - try: - with pytest.raises(ProxyException) as over_cap: - await _guess(throttle) - assert over_cap.value.code == "429" - assert over_cap.value.headers.get("Retry-After") == "300", "refused at once, for the whole block" - assert await _fail(throttle, username="someone-else@corp.com") == "401", "other keys are not affected" - finally: - release.set() - for task in held: - with pytest.raises(ProxyException): - await task - - assert lt._HELD_ATTEMPTS.get(slot) is None, "the slots are released once the held attempts answer" - - -@pytest.mark.asyncio -async def test_a_full_hold_pool_still_lets_the_right_password_in(monkeypatch): - """Five parked wrong guesses from the office must not turn the soft block into a lockout for the real user.""" - import asyncio - - from litellm.proxy.auth import login_throttle as lt - from litellm.proxy.auth.login_throttle import MAX_HELD_ATTEMPTS_PER_KEY - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - monkeypatch.setenv("DATABASE_URL", "postgresql://stub") - release = asyncio.Event() - - async def _park(_seconds: float) -> None: - await release.wait() - - monkeypatch.setattr(lt, "_sleep", _park) - throttle = _throttle(user_limit=1, source_limit=3, client_ip="203.0.113.46") - assert [await _fail(throttle, username=f"spray-{i}@corp.com") for i in range(4)] == ["401"] * 4 - source_slot = throttle._keys("known@example.com").source_block - assert throttle._local_block_ttl(source_slot) > 0, "the source is blocked" - - held = [ - asyncio.create_task(_guess(throttle, username="known@example.com")) for _ in range(MAX_HELD_ATTEMPTS_PER_KEY) - ] - for _ in range(1000): - if lt._HELD_ATTEMPTS.get(source_slot) == MAX_HELD_ATTEMPTS_PER_KEY: - break - await asyncio.sleep(0) - assert lt._HELD_ATTEMPTS == {source_slot: MAX_HELD_ATTEMPTS_PER_KEY} - - try: - signed_in = await _db_login(throttle, "known@example.com", "right", correct=True) - assert signed_in.user_id == "u-1" - finally: - release.set() - for task in held: - with pytest.raises(ProxyException): - await task - - -@pytest.mark.asyncio -async def test_a_blocked_source_shares_one_held_slot_pool_across_its_blocked_usernames(monkeypatch): - """Once the source is blocked, a pair block for a username must not hand that username its own five slots.""" - import asyncio - - from litellm.proxy._types import ProxyException - from litellm.proxy.auth import login_throttle as lt - from litellm.proxy.auth.login_throttle import MAX_HELD_ATTEMPTS_PER_KEY - - monkeypatch.setenv("UI_USERNAME", "admin") - monkeypatch.setenv("UI_PASSWORD", "right") - release = asyncio.Event() - - async def _park(_seconds: float) -> None: - await release.wait() - - monkeypatch.setattr(lt, "_sleep", _park) - throttle = _throttle(user_limit=1, source_limit=3, client_ip="203.0.113.45") + throttle = _throttle(user_limit=1, source_limit=3, block_seconds=120, client_ip="203.0.113.45") assert [await _fail(throttle) for _ in range(2)] == ["401", "401"], "the admin pair is now blocked" + throttle.blocks.set_cache(throttle._keys("admin").pair_block, 1, ttl=30) assert [await _fail(throttle, username=f"spray-{i}@corp.com") for i in range(3)] == ["401"] * 3 - source_slot = throttle._keys("admin").source_block - assert throttle._local_block_ttl(source_slot) > 0, "the source is now blocked as well" + assert throttle._local_block_ttl(throttle._keys("admin").source_block) == 120, "the source is now blocked too" - usernames = ["admin", *(f"fresh-{i}@corp.com" for i in range(MAX_HELD_ATTEMPTS_PER_KEY - 1))] - held = [asyncio.create_task(_guess(throttle, username=name)) for name in usernames] - for _ in range(1000): - if lt._HELD_ATTEMPTS.get(source_slot) == MAX_HELD_ATTEMPTS_PER_KEY: - break - await asyncio.sleep(0) - assert lt._HELD_ATTEMPTS == {source_slot: MAX_HELD_ATTEMPTS_PER_KEY} - - try: - for name in ("admin", "fresh-0@corp.com", "never-seen@corp.com"): - with pytest.raises(ProxyException) as over_cap: - await _guess(throttle, username=name) - assert over_cap.value.code == "429" - assert over_cap.value.headers.get("Retry-After") == "300" - finally: - release.set() - for task in held: - with pytest.raises(ProxyException): - await task - - assert lt._HELD_ATTEMPTS == {} + for name in ("admin", "spray-0@corp.com", "never-seen@corp.com"): + with pytest.raises(ProxyException) as refused: + await _guess(throttle, username=name) + assert refused.value.code == "429" + assert refused.value.headers.get("Retry-After") == "120", name @pytest.mark.asyncio -async def test_disabling_the_control_removes_the_hold_as_well(monkeypatch, login_delays): - """The escape hatch has to turn off the whole control, not only the refusal.""" +async def test_disabling_the_control_lets_every_attempt_through(monkeypatch): + """The escape hatch has to turn off the whole control: no counting and no refusal.""" import dataclasses monkeypatch.setenv("UI_USERNAME", "admin") @@ -1319,7 +1293,7 @@ async def test_disabling_the_control_removes_the_hold_as_well(monkeypatch, login throttle = dataclasses.replace(_throttle(user_limit=1), enabled=False) assert [await _fail(throttle) for _ in range(6)] == ["401"] * 6 - assert login_delays.seconds == [] + assert _local_count(throttle, throttle._keys("admin").pair_counter) == 0 class _FakeRedis: @@ -1404,12 +1378,15 @@ async def test_redis_is_the_only_counter_while_it_answers(monkeypatch): assert await _fail(second_worker, username="user@corp.com") == "429", "the second worker sees the block" + block_keys = [k for k in redis.values if ":block:user:" in k] + assert block_keys, "the block lives in Redis, where every worker reads it" + for key in block_keys: + await redis.async_delete_cache(key) await _db_login(second_worker, "user@corp.com", "right", correct=True) assert not [k for k in redis.values if ":user:" in k and ":block:" not in k], ( "success clears the shared pair counter" ) - assert [k for k in redis.values if ":block:user:" in k], "an active block is not lifted by one success" @pytest.mark.asyncio @@ -1436,7 +1413,7 @@ async def test_a_redis_outage_falls_back_to_this_workers_own_counter(monkeypatch verbose_proxy_logger.removeHandler(handler) assert blocked.value.code == "429" - assert blocked.value.headers.get("Retry-After") == "270" + assert blocked.value.headers.get("Retry-After") == "300" assert any("Redis failed while counting Admin UI sign-in attempts" in r.getMessage() for r in records) diff --git a/tests/test_litellm/proxy/proxy_server/conftest.py b/tests/test_litellm/proxy/proxy_server/conftest.py index a349d378985..d1adf2a5c02 100644 --- a/tests/test_litellm/proxy/proxy_server/conftest.py +++ b/tests/test_litellm/proxy/proxy_server/conftest.py @@ -522,16 +522,9 @@ def reset_login_throttle(monkeypatch): Only the throttle's own keys are removed, so other cache entries remain untouched. """ from litellm.proxy import proxy_server as ps - from litellm.proxy.auth import login_throttle from litellm.proxy.auth.login_throttle import _BLOCKS, _CACHE_KEY_PREFIX, _COUNTERS - async def _no_delay(_seconds: float) -> None: - """The hold on a rejected sign-in from a blocked key, replaced so the route tests stay fast.""" - - monkeypatch.setattr(login_throttle, "_sleep", _no_delay) - def _drop_throttle_keys() -> None: - login_throttle._HELD_ATTEMPTS.clear() for store in (_COUNTERS, _BLOCKS): for key in tuple(store.cache_dict) + tuple(store.ttl_dict): if key.startswith(_CACHE_KEY_PREFIX): diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py index 82e86086518..a82f078fcb9 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py @@ -553,14 +553,14 @@ def test_budget_is_shared_across_username_casing(client, monkeypatch, reset_logi def test_a_refused_attempt_carries_retry_after(client, monkeypatch, reset_login_throttle): - """The 429 tells the caller how long the block has left, after the 30 seconds it was already held.""" + """The 429 tells the caller how long the block has left.""" _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1, failed_login_block_seconds=77) assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] refused = client.post("/v2/login", json={"username": "admin", "password": "wrong"}) assert refused.status_code == 429 - assert refused.headers.get("retry-after") == "47" + assert refused.headers.get("retry-after") == "77" def test_the_form_returns_a_human_readable_lockout_page(client, monkeypatch, reset_login_throttle): @@ -572,8 +572,8 @@ def test_the_form_returns_a_human_readable_lockout_page(client, monkeypatch, res refused = client.post("/login", data={"username": "admin", "password": "wrong"}) assert refused.status_code == 429 assert refused.headers.get("content-type", "").startswith("text/html") - assert "Try again in about 47 seconds" in refused.text - assert refused.headers.get("retry-after") == "47" + assert "Try again in about 77 seconds" in refused.text + assert refused.headers.get("retry-after") == "77" def test_a_second_username_from_the_same_source_still_gets_through(client, monkeypatch, reset_login_throttle): @@ -608,8 +608,29 @@ def test_a_spray_across_usernames_is_not_blocked_without_trusted_proxy_ranges( assert sprayed == [401] * 8 -def test_the_configured_admin_password_still_signs_in_while_blocked(client, monkeypatch, reset_login_throttle): - """The operator must never be locked out of the console by traffic aimed at it.""" +def test_a_spray_across_usernames_is_blocked_on_the_source_with_an_empty_trusted_proxy_ranges( + client, monkeypatch, reset_login_throttle +): + """An explicit empty list says nothing fronts the proxy, so the peer address is the client and the + source scope is on. A forwarded header from an untrusted peer is ignored rather than trusted.""" + _install_real_auth(monkeypatch, trusted_proxy_ranges=[], max_failed_login_attempts_per_source=4) + + sprayed = [ + client.post( + "/v2/login", + json={"username": f"sprayed-{i}@corp.com", "password": "wrong"}, + headers={"x-forwarded-for": f"203.0.113.{i}"}, + ).status_code + for i in range(5) + ] + assert sprayed == [401] * 5 + + assert _json_login(client, "/v2/login", username="sprayed-6@corp.com") == 429 + + +def test_the_configured_admin_password_is_refused_while_blocked(client, monkeypatch, reset_login_throttle): + """The env credentials get no bypass: a bypass would make them the one password worth guessing without + limit. An operator who is blocked administers the proxy with the master key over the API meanwhile.""" from unittest.mock import AsyncMock, patch _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) @@ -625,18 +646,38 @@ def test_the_configured_admin_password_still_signs_in_while_blocked(client, monk "litellm.proxy.auth.login_utils.generate_key_helper_fn", new=AsyncMock(return_value={"token": "sk-ui"}) ), ): + assert _json_login(client, "/v2/login", password="right-password") == 429 + reset_login_throttle() assert _json_login(client, "/v2/login", password="right-password") == 200 -def test_a_database_users_correct_password_signs_in_while_blocked(client, monkeypatch, reset_login_throttle): - """The block is soft: guessing at an account slows the guesser down, it does not lock the owner out.""" +def test_the_master_key_as_a_bearer_token_still_works_while_the_ui_password_is_blocked( + client, monkeypatch, reset_login_throttle +): + """Lockout recovery: the API path with the master key never enters the sign-in throttle.""" _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) + + assert [_json_login(client, "/v2/login") for _ in range(3)] == [401, 401, 429] + + assert client.get("/models", headers={"Authorization": "Bearer sk-not-the-master"}).status_code >= 400 + assert client.get("/models", headers={"Authorization": "Bearer sk-test-master"}).status_code == 200 + assert _json_login(client, "/v2/login", password="right-password") == 429, "the UI block is unaffected" + + +def test_a_database_users_correct_password_is_refused_while_blocked(client, monkeypatch, reset_login_throttle): + """The block is hard: while it lasts, nothing from that source signs in as that user, right password or not, + and the block is not extended by the refused attempts.""" + _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1, failed_login_block_seconds=64) _db_user(monkeypatch, "user@corp.com") assert [_json_login(client, "/v2/login", username="user@corp.com") for _ in range(3)] == [401, 401, 429] + refused = client.post("/v2/login", json={"username": "user@corp.com", "password": "right-db-password"}) + assert refused.status_code == 429 + assert refused.headers.get("retry-after") == "64" + + reset_login_throttle() assert _json_login(client, "/v2/login", username="user@corp.com", password="right-db-password") == 200 - assert _json_login(client, "/v2/login", username="user@corp.com") == 429, "the block itself is still in force" def test_sign_in_succeeds_again_once_the_block_is_cleared(client, monkeypatch, reset_login_throttle): diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 80b1d3b1841..4c8cacf120f 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -3450,7 +3450,8 @@ async def test_load_config_warns_that_the_source_login_limit_is_off_without_trus tmp_path, monkeypatch, caplog ): """The per-source failed-login limit is skipped when the source cannot be attributed, and the - operator must be told so at startup; a configured range silences it.""" + operator must be told so at startup. Both a configured range and an explicit empty list (no + proxies, the peer is the source) silence it, since both keep the limit on.""" import logging from litellm.proxy.auth.login_throttle import warn_source_login_limit_is_off @@ -3465,12 +3466,13 @@ async def test_load_config_warns_that_the_source_login_limit_is_off_without_trus await ProxyConfig().load_config(router=MagicMock(), config_file_path=str(config_file)) assert "trusted_proxy_ranges is not set" in caplog.text - caplog.clear() - warn_source_login_limit_is_off.cache_clear() - config_file.write_text("model_list: []\ngeneral_settings:\n trusted_proxy_ranges: ['10.0.0.0/8']\n") - with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): - await ProxyConfig().load_config(router=MagicMock(), config_file_path=str(config_file)) - assert "trusted_proxy_ranges is not set" not in caplog.text + for configured in ("['10.0.0.0/8']", "[]"): + caplog.clear() + warn_source_login_limit_is_off.cache_clear() + config_file.write_text(f"model_list: []\ngeneral_settings:\n trusted_proxy_ranges: {configured}\n") + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + await ProxyConfig().load_config(router=MagicMock(), config_file_path=str(config_file)) + assert "trusted_proxy_ranges is not set" not in caplog.text, configured @pytest.mark.asyncio diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index ee6824ca02c..871cfea2d29 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -26512,7 +26512,7 @@ export interface components { enforce_fallback_model_access?: boolean | null; /** * Failed Login Block Seconds - * @description How long a blocked source address, or source address and username, stays blocked. The block is soft: a correct password still signs in, but each attempt from a blocked key takes one of 5 held slots per worker and a wrong password is held for 30 seconds before it is refused with 429. Set under `general_settings` in config.yaml. Defaults to 300 + * @description How long a blocked source address, or source address and username, stays blocked. Every attempt from a blocked key, right or wrong, is refused with 429 before the password is checked; the block is not extended by refused attempts. Set under `general_settings` in config.yaml. Defaults to 300 */ failed_login_block_seconds?: number | null; /** @@ -26566,7 +26566,7 @@ export interface components { max_batch_file_size_mb?: number | null; /** * Max Failed Login Attempts Per Source - * @description Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set, since otherwise every client behind an ingress shares one address. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10 + * @description Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and this limit is off. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10 */ max_failed_login_attempts_per_source?: number | null; /** @@ -26756,7 +26756,7 @@ export interface components { supported_db_objects?: components["schemas"]["SupportedDBObjectType"][] | null; /** * Trusted Proxy Ranges - * @description CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler. + * @description CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler, and whose X-Forwarded-For is used to attribute Admin UI sign-in attempts to a source address. Set it to an empty list when clients connect directly, so the peer address is the source. Left unset, the per-source sign-in limit is off. */ trusted_proxy_ranges?: string[] | null; /** From b227a8c4c99f0e24999323c6f4db5aeff2001b8e Mon Sep 17 00:00:00 2001 From: yucheng Date: Thu, 17 Sep 2026 16:42:04 +0000 Subject: [PATCH 063/141] refactor(proxy): move login throttle sentinels into constants Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 5 +++ litellm/proxy/auth/login_throttle.py | 37 ++++++++++--------- .../proxy/auth/test_login_utils.py | 20 ++++------ .../proxy/proxy_server/conftest.py | 5 ++- 4 files changed, 36 insertions(+), 31 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index 338fe0f6b85..a64932a0e4b 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1644,6 +1644,11 @@ LITELLM_EXPIRED_UI_SESSION_KEY_CLEANUP_INTERVAL_SECONDS: Final = int( LITELLM_EXPIRED_UI_SESSION_KEY_CLEANUP_BATCH_SIZE: Final = int( os.getenv("LITELLM_EXPIRED_UI_SESSION_KEY_CLEANUP_BATCH_SIZE", 1000) ) +LOGIN_THROTTLE_CACHE_KEY_PREFIX: Final = "login_fail" +LOGIN_THROTTLE_UNKNOWN_SOURCE: Final = "unknown" +LOGIN_THROTTLE_MAX_TRACKED_COUNTERS: Final = 20_000 +LOGIN_THROTTLE_MAX_TRACKED_BLOCKS: Final = 10_000 +LOGIN_THROTTLE_NOT_BLOCKED: Final = (0, 0) LITELLM_PROXY_ADMIN_NAME: Final = "default_user_id" LITELLM_PROXY_BUDGET_NAME: Final = "litellm-proxy-budget" GLOBAL_PROXY_SPEND_CACHE_KEY: Final = f"{LITELLM_PROXY_ADMIN_NAME}:spend" diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 0106d58e426..d16cd43e36d 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -17,7 +17,6 @@ import time from collections.abc import Mapping from dataclasses import dataclass from functools import cache -from types import MappingProxyType from typing import Final, Literal, NamedTuple, NoReturn, Protocol, TypeAlias from fastapi import Request, status @@ -27,6 +26,14 @@ from redis.exceptions import RedisError from litellm._logging import verbose_proxy_logger from litellm.caching.in_memory_cache import InMemoryCache from litellm.caching.redis_cache import RedisCache, RedisCircuitBreakerOpenError +from litellm.constants import ( + EMPTY_MAPPING, + LOGIN_THROTTLE_CACHE_KEY_PREFIX, + LOGIN_THROTTLE_MAX_TRACKED_BLOCKS, + LOGIN_THROTTLE_MAX_TRACKED_COUNTERS, + LOGIN_THROTTLE_NOT_BLOCKED, + LOGIN_THROTTLE_UNKNOWN_SOURCE, +) from litellm.proxy._types import ProxyErrorTypes, ProxyException from litellm.proxy.auth.network import TrustedProxyConfig, normalize_cidr_ranges, resolve_client_ip from litellm.secret_managers.main import get_secret_bool @@ -45,12 +52,6 @@ WINDOW_KEY: Final = "failed_login_window_seconds" BLOCK_KEY: Final = "failed_login_block_seconds" TRUSTED_PROXY_RANGES_KEY: Final = "trusted_proxy_ranges" -_CACHE_KEY_PREFIX: Final = "login_fail" -_UNKNOWN_SOURCE: Final = "unknown" -_MAX_TRACKED_COUNTERS: Final = 20_000 -_MAX_TRACKED_BLOCKS: Final = 10_000 -_NO_SETTINGS: Final[Mapping[str, object]] = MappingProxyType({}) -_NOT_BLOCKED: Final = (0, 0) _REDIS_FAILURES: Final = (RedisError, RedisCircuitBreakerOpenError, OSError, asyncio.TimeoutError) _LOCAL_BLOCK_EXPIRY: Final = TypeAdapter[float | None](float | None) _SOURCE_LIMIT_OVERRIDES: Final = TypeAdapter[Mapping[str, object]](Mapping[str, object]) @@ -94,9 +95,11 @@ _RECORD_FAILURE_LUA: Final = ( ) _COUNTERS: Final = InMemoryCache( - max_size_in_memory=_MAX_TRACKED_COUNTERS, default_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS + max_size_in_memory=LOGIN_THROTTLE_MAX_TRACKED_COUNTERS, default_ttl=DEFAULT_FAILED_LOGIN_WINDOW_SECONDS +) +_BLOCKS: Final = InMemoryCache( + max_size_in_memory=LOGIN_THROTTLE_MAX_TRACKED_BLOCKS, default_ttl=DEFAULT_FAILED_LOGIN_BLOCK_SECONDS ) -_BLOCKS: Final = InMemoryCache(max_size_in_memory=_MAX_TRACKED_BLOCKS, default_ttl=DEFAULT_FAILED_LOGIN_BLOCK_SECONDS) @cache @@ -249,13 +252,13 @@ class LoginThrottle: general_settings: Mapping[str, object] | None, redis_cache: RedisCache | None, ) -> LoginThrottle: - settings: Final = general_settings if general_settings is not None else _NO_SETTINGS + settings: Final = general_settings if general_settings is not None else EMPTY_MAPPING proxies: Final = declared_proxy_ranges(settings) resolved, _ = resolve_client_ip( request, TrustedProxyConfig(use_forwarded_for=bool(proxies), trusted_proxy_cidrs=proxies or ()) ) return cls( - client_ip=resolved or _UNKNOWN_SOURCE, + client_ip=resolved or LOGIN_THROTTLE_UNKNOWN_SOURCE, source_limit=_source_limit(settings, resolved) if proxies is not None and resolved is not None else None, user_limit=_int_setting(settings, USER_LIMIT_KEY, DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_USER), window_seconds=_int_setting(settings, WINDOW_KEY, DEFAULT_FAILED_LOGIN_WINDOW_SECONDS), @@ -270,10 +273,10 @@ class LoginThrottle: group: Final = source_group(self.client_ip) user: Final = hashlib.sha256(username.casefold().encode("utf-8")).hexdigest() return _Keys( - pair_counter=f"{_CACHE_KEY_PREFIX}:{{{group}}}:user:{user}", - pair_block=f"{_CACHE_KEY_PREFIX}:{{{group}}}:block:user:{user}", - source_counter=f"{_CACHE_KEY_PREFIX}:{{{group}}}:source", - source_block=f"{_CACHE_KEY_PREFIX}:{{{group}}}:block:source", + pair_counter=f"{LOGIN_THROTTLE_CACHE_KEY_PREFIX}:{{{group}}}:user:{user}", + pair_block=f"{LOGIN_THROTTLE_CACHE_KEY_PREFIX}:{{{group}}}:block:user:{user}", + source_counter=f"{LOGIN_THROTTLE_CACHE_KEY_PREFIX}:{{{group}}}:source", + source_block=f"{LOGIN_THROTTLE_CACHE_KEY_PREFIX}:{{{group}}}:block:source", ) async def attempt(self, username: str) -> LoginAttempt: @@ -305,14 +308,14 @@ class LoginThrottle: async def _shared_block_ttls(self, keys: _Keys) -> _BlockTtls: if self.redis_cache is None: - return _NOT_BLOCKED + return LOGIN_THROTTLE_NOT_BLOCKED try: return _LUA_BLOCK_TTLS.validate_python( await self.redis_cache.async_register_script(_BLOCK_TTLS_LUA)(keys, ()) ) except _REDIS_FAILURES as err: self._warn_redis(err) - return _NOT_BLOCKED + return LOGIN_THROTTLE_NOT_BLOCKED def _local_block_ttls(self, keys: _Keys) -> _BlockTtls: return self._local_block_ttl(keys.pair_block), self._local_block_ttl(keys.source_block) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 8670d7fef41..986760c3cc5 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1361,7 +1361,7 @@ class _DownRedis(_FakeRedis): @pytest.mark.asyncio async def test_redis_is_the_only_counter_while_it_answers(monkeypatch): """Every worker must spend the same budget, see the same block, and a success must clear the pair for all.""" - from litellm.proxy.auth.login_throttle import _CACHE_KEY_PREFIX + from litellm.constants import LOGIN_THROTTLE_CACHE_KEY_PREFIX monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") @@ -1371,7 +1371,7 @@ async def test_redis_is_the_only_counter_while_it_answers(monkeypatch): second_worker = _throttle(user_limit=2, stores=_stores(), redis_cache=redis) assert [await _fail(first_worker, username="user@corp.com") for _ in range(3)] == ["401"] * 3 - assert not [k for k in first_worker.counters.cache_dict if str(k).startswith(_CACHE_KEY_PREFIX)], ( + assert not [k for k in first_worker.counters.cache_dict if str(k).startswith(LOGIN_THROTTLE_CACHE_KEY_PREFIX)], ( "with Redis answering, no worker may keep a counter of its own" ) assert not first_worker.blocks.cache_dict @@ -1437,7 +1437,8 @@ async def test_a_failed_redis_delete_still_clears_this_workers_counter(monkeypat async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): """Regression: throttle entries must not evict cached credentials from user_api_key_cache.""" from litellm.proxy import proxy_server as ps - from litellm.proxy.auth.login_throttle import _CACHE_KEY_PREFIX, LoginThrottle + from litellm.constants import LOGIN_THROTTLE_CACHE_KEY_PREFIX + from litellm.proxy.auth.login_throttle import LoginThrottle monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") @@ -1453,7 +1454,7 @@ async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): assert await _fail(throttle, username=f"made-up-{i}@example.com") == "401" added = set(ps.user_api_key_cache.in_memory_cache.cache_dict) - auth_cache_keys_before - assert not [k for k in added if str(k).startswith(_CACHE_KEY_PREFIX)] + assert not [k for k in added if str(k).startswith(LOGIN_THROTTLE_CACHE_KEY_PREFIX)] def test_settings_that_arrive_as_environment_strings_are_honored(): @@ -1557,17 +1558,12 @@ async def test_a_username_spray_cannot_evict_an_active_block(monkeypatch): """Counters and blocks live in separate bounded stores, so a flood of made-up pairs fills the counter store while the blocks it already earned stay in force.""" from litellm.caching.in_memory_cache import InMemoryCache - from litellm.proxy.auth.login_throttle import ( - _BLOCKS, - _COUNTERS, - _MAX_TRACKED_BLOCKS, - _MAX_TRACKED_COUNTERS, - LoginThrottle, - ) + from litellm.constants import LOGIN_THROTTLE_MAX_TRACKED_BLOCKS, LOGIN_THROTTLE_MAX_TRACKED_COUNTERS + from litellm.proxy.auth.login_throttle import _BLOCKS, _COUNTERS, LoginThrottle monkeypatch.setenv("UI_USERNAME", "admin") monkeypatch.setenv("UI_PASSWORD", "right") - assert _MAX_TRACKED_COUNTERS >= 10_000 and _MAX_TRACKED_BLOCKS >= 10_000 + assert LOGIN_THROTTLE_MAX_TRACKED_COUNTERS >= 10_000 and LOGIN_THROTTLE_MAX_TRACKED_BLOCKS >= 10_000 assert _COUNTERS is not _BLOCKS counters, blocks = InMemoryCache(max_size_in_memory=50), InMemoryCache(max_size_in_memory=50) throttle = LoginThrottle( diff --git a/tests/test_litellm/proxy/proxy_server/conftest.py b/tests/test_litellm/proxy/proxy_server/conftest.py index d1adf2a5c02..ae1b42363ef 100644 --- a/tests/test_litellm/proxy/proxy_server/conftest.py +++ b/tests/test_litellm/proxy/proxy_server/conftest.py @@ -521,13 +521,14 @@ def reset_login_throttle(monkeypatch): window, so without this a failed sign-in test could block unrelated tests later. Only the throttle's own keys are removed, so other cache entries remain untouched. """ + from litellm.constants import LOGIN_THROTTLE_CACHE_KEY_PREFIX from litellm.proxy import proxy_server as ps - from litellm.proxy.auth.login_throttle import _BLOCKS, _CACHE_KEY_PREFIX, _COUNTERS + from litellm.proxy.auth.login_throttle import _BLOCKS, _COUNTERS def _drop_throttle_keys() -> None: for store in (_COUNTERS, _BLOCKS): for key in tuple(store.cache_dict) + tuple(store.ttl_dict): - if key.startswith(_CACHE_KEY_PREFIX): + if key.startswith(LOGIN_THROTTLE_CACHE_KEY_PREFIX): store.delete_cache(key) monkeypatch.setattr(ps, "redis_usage_cache", None) From 5f64dfd8dd4deffdee76c673325590176fc48001 Mon Sep 17 00:00:00 2001 From: yassin Date: Thu, 17 Sep 2026 18:14:04 +0000 Subject: [PATCH 064/141] fix(proxy): price Azure Speech fast transcription and limit unpriced batch writes to admins Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 4 + .../llm_passthrough_endpoints.py | 22 +++ ...zure_speech_passthrough_logging_handler.py | 30 +++- ...zure_speech_passthrough_logging_handler.py | 39 ++++- .../test_llm_pass_through_endpoints.py | 146 ++++++++++++++++-- 5 files changed, 220 insertions(+), 21 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index 15a1d054e26..d9da0cc0f64 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1574,13 +1574,17 @@ AZURE_SPEECH_CUSTOM_LLM_PROVIDER: Final = "azure_speech" AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX: Final = "/azure_speech" AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX: Final = "/speech/" AZURE_SPEECH_BATCH_PATH_PREFIX: Final = "/speechtotext/" +AZURE_SPEECH_FAST_TRANSCRIPTION_PATH: Final = "/speechtotext/transcriptions:transcribe" +AZURE_SPEECH_UNPRICED_WRITE_METHODS: Final = frozenset({"POST", "PUT"}) AZURE_SPEECH_STT_DOMAIN: Final = "stt.speech.microsoft.com" AZURE_SPEECH_COGNITIVE_SERVICES_DOMAIN: Final = "api.cognitive.microsoft.com" AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER: Final = "Ocp-Apim-Subscription-Key" AZURE_SPEECH_SHORT_AUDIO_MODEL: Final = "short-audio" AZURE_SPEECH_BATCH_MODEL: Final = "batch-transcription" +AZURE_SPEECH_FAST_TRANSCRIPTION_MODEL: Final = "fast-transcription" AZURE_SPEECH_PRICING_MODEL: Final = "azure/speech/azure-stt" AZURE_SPEECH_TICKS_PER_SECOND: Final = 10_000_000 +AZURE_SPEECH_MILLISECONDS_PER_SECOND: Final = 1_000 BASE_MCP_ROUTE: Final = "/mcp" diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index e64eac87a7f..b8966508f81 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -33,10 +33,12 @@ from litellm.constants import ( AZURE_SPEECH_BATCH_PATH_PREFIX, AZURE_SPEECH_COGNITIVE_SERVICES_DOMAIN, AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + AZURE_SPEECH_FAST_TRANSCRIPTION_PATH, AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX, AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, AZURE_SPEECH_STT_DOMAIN, AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER, + AZURE_SPEECH_UNPRICED_WRITE_METHODS, BEDROCK_AGENT_RUNTIME_PASS_THROUGH_ROUTES, ) from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix @@ -65,6 +67,7 @@ from litellm.proxy.common_utils.http_parsing_utils import ( get_request_body, is_json_content_type, ) +from litellm.proxy.common_utils.resource_ownership import is_proxy_admin from litellm.proxy.common_utils.sse_keepalive import ( wrap_passthrough_sse_bytes_with_keepalive_pings, ) @@ -1357,6 +1360,14 @@ def resolve_azure_speech_base_url(endpoint_path: str, api_base: str | None, regi return httpx.URL(f"https://{region}.{domain}") +def azure_speech_write_is_unpriced(method: str, endpoint_path: str) -> bool: + return ( + endpoint_path.startswith(AZURE_SPEECH_BATCH_PATH_PREFIX) + and endpoint_path != AZURE_SPEECH_FAST_TRANSCRIPTION_PATH + and method.upper() in AZURE_SPEECH_UNPRICED_WRITE_METHODS + ) + + @router.api_route( f"{AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX}/{{endpoint:path}}", methods=["GET", "POST", "PUT", "DELETE", "PATCH"], # mutable-ok: fastapi route methods must be a list @@ -1395,6 +1406,17 @@ async def azure_speech_proxy_route( "AZURE_SPEECH_REGION or AZURE_SPEECH_API_BASE in the proxy environment." ), ) + if azure_speech_write_is_unpriced( + method=request.method, endpoint_path=normalized_endpoint_path + ) and not is_proxy_admin(user_api_key_dict): + raise HTTPException( + status_code=403, + detail=( + f"{request.method} {normalized_endpoint_path} creates Azure Speech work whose cost is unknown at " + "request time, so it is limited to proxy admin keys. Use " + f"{AZURE_SPEECH_FAST_TRANSCRIPTION_PATH} for transcription that is priced per request." + ), + ) azure_speech_api_key: Final = passthrough_endpoint_router.get_credentials( custom_llm_provider=AZURE_SPEECH_CUSTOM_LLM_PROVIDER, region_name=None, diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py index 74587acd453..588b7cc8e56 100644 --- a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py @@ -9,6 +9,9 @@ from litellm._logging import verbose_proxy_logger from litellm.constants import ( AZURE_SPEECH_BATCH_MODEL, AZURE_SPEECH_CUSTOM_LLM_PROVIDER, + AZURE_SPEECH_FAST_TRANSCRIPTION_MODEL, + AZURE_SPEECH_FAST_TRANSCRIPTION_PATH, + AZURE_SPEECH_MILLISECONDS_PER_SECOND, AZURE_SPEECH_PRICING_MODEL, AZURE_SPEECH_SHORT_AUDIO_MODEL, AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, @@ -28,10 +31,16 @@ class AzureSpeechPassthroughLoggingHandler: def _is_short_audio_route(url_route: str) -> bool: return urlparse(url_route).path.startswith(AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX) + @staticmethod + def _is_fast_transcription_route(url_route: str) -> bool: + return urlparse(url_route).path.endswith(AZURE_SPEECH_FAST_TRANSCRIPTION_PATH) + @staticmethod def _model_from_url_route(url_route: str) -> str: if AzureSpeechPassthroughLoggingHandler._is_short_audio_route(url_route): return f"{AZURE_SPEECH_CUSTOM_LLM_PROVIDER}/{AZURE_SPEECH_SHORT_AUDIO_MODEL}" + if AzureSpeechPassthroughLoggingHandler._is_fast_transcription_route(url_route): + return f"{AZURE_SPEECH_CUSTOM_LLM_PROVIDER}/{AZURE_SPEECH_FAST_TRANSCRIPTION_MODEL}" return f"{AZURE_SPEECH_CUSTOM_LLM_PROVIDER}/{AZURE_SPEECH_BATCH_MODEL}" @staticmethod @@ -45,10 +54,25 @@ class AzureSpeechPassthroughLoggingHandler: return (offset + duration) / AZURE_SPEECH_TICKS_PER_SECOND @staticmethod - def _response_cost(url_route: str, response_body: Mapping[str, object] | Sequence[object] | None) -> float: - if not AzureSpeechPassthroughLoggingHandler._is_short_audio_route(url_route): + def _fast_transcription_audio_seconds(response_body: Mapping[str, object] | Sequence[object] | None) -> float: + if not isinstance(response_body, Mapping): return 0.0 - audio_seconds: Final = AzureSpeechPassthroughLoggingHandler._recognized_audio_seconds(response_body) + duration_milliseconds: Final = response_body.get("durationMilliseconds") + if not isinstance(duration_milliseconds, int): + return 0.0 + return duration_milliseconds / AZURE_SPEECH_MILLISECONDS_PER_SECOND + + @staticmethod + def _billed_audio_seconds(url_route: str, response_body: Mapping[str, object] | Sequence[object] | None) -> float: + if AzureSpeechPassthroughLoggingHandler._is_short_audio_route(url_route): + return AzureSpeechPassthroughLoggingHandler._recognized_audio_seconds(response_body) + if AzureSpeechPassthroughLoggingHandler._is_fast_transcription_route(url_route): + return AzureSpeechPassthroughLoggingHandler._fast_transcription_audio_seconds(response_body) + return 0.0 + + @staticmethod + def _response_cost(url_route: str, response_body: Mapping[str, object] | Sequence[object] | None) -> float: + audio_seconds: Final = AzureSpeechPassthroughLoggingHandler._billed_audio_seconds(url_route, response_body) if audio_seconds <= 0.0: return 0.0 try: diff --git a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py index 5ffb1f7785d..50d3de64f72 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py @@ -13,9 +13,19 @@ from litellm.proxy.pass_through_endpoints.success_handler import ( PassThroughEndpointLogging, ) -SHORT_AUDIO_URL = "https://eastus.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US" +SHORT_AUDIO_URL = ( + "https://eastus.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US" +) BATCH_URL = "https://eastus.api.cognitive.microsoft.com/speechtotext/v3.2/transcriptions" -TRANSCRIPT_BODY = {"RecognitionStatus": "Success", "Offset": 5000000, "Duration": 25000000, "DisplayText": "Hello world."} +FAST_URL = "https://eastus.api.cognitive.microsoft.com/speechtotext/transcriptions:transcribe?api-version=2024-11-15" +FAST_BODY = {"durationMilliseconds": 5061, "combinedPhrases": [{"text": "Hello world."}]} +FAST_AUDIO_SECONDS = 5.061 +TRANSCRIPT_BODY = { + "RecognitionStatus": "Success", + "Offset": 5000000, + "Duration": 25000000, + "DisplayText": "Hello world.", +} TRANSCRIPT = json.dumps(TRANSCRIPT_BODY) TRANSCRIPT_AUDIO_SECONDS = 3.0 PRICE_PER_SECOND = 0.5 @@ -52,6 +62,7 @@ class TestAzureSpeechPassthroughHandler: "url_route,expected_model,expected_cost", [ (SHORT_AUDIO_URL, "azure_speech/short-audio", TRANSCRIPT_AUDIO_SECONDS * PRICE_PER_SECOND), + (FAST_URL, "azure_speech/fast-transcription", FAST_AUDIO_SECONDS * PRICE_PER_SECOND), (BATCH_URL, "azure_speech/batch-transcription", 0.0), (f"{BATCH_URL}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab/files", "azure_speech/batch-transcription", 0.0), ], @@ -61,7 +72,7 @@ class TestAzureSpeechPassthroughHandler: handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( httpx_response=_make_response(url_route), - response_body=TRANSCRIPT_BODY, + response_body={**TRANSCRIPT_BODY, **FAST_BODY}, logging_obj=logging_obj, url_route=url_route, result=TRANSCRIPT, @@ -110,6 +121,28 @@ class TestAzureSpeechPassthroughHandler: assert handler_result["kwargs"]["model"] == "azure_speech/short-audio" assert handler_result["kwargs"]["response_cost"] == 0.0 + @pytest.mark.parametrize( + "response_body", + [{"durationMilliseconds": 0}, {"durationMilliseconds": "5061"}, {"duration": 5061}, {}, [], None], + ) + def test_fast_transcription_without_duration_milliseconds_logs_zero_cost( + self, response_body: dict[str, object] | list[dict[str, object]] | None + ): + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(FAST_URL), + response_body=response_body, + logging_obj=_make_logging_obj(), + url_route=FAST_URL, + result="", + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert handler_result["kwargs"]["model"] == "azure_speech/fast-transcription" + assert handler_result["kwargs"]["response_cost"] == 0.0 + def test_missing_price_entry_still_logs_the_row_at_zero_cost(self, monkeypatch: pytest.MonkeyPatch): monkeypatch.delitem(litellm.model_cost, "azure/speech/azure-stt") diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index 43fc28c34c4..eb0c607fbef 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -6141,6 +6141,7 @@ class TestAzureRelayDeploymentSegment: AZURE_SPEECH_SHORT_AUDIO_ENDPOINT: Final = "/speech/recognition/conversation/cognitiveservices/v1" AZURE_SPEECH_BATCH_ENDPOINT: Final = "/speechtotext/v3.2/transcriptions" +AZURE_SPEECH_FAST_ENDPOINT: Final = "/speechtotext/transcriptions:transcribe" AZURE_SPEECH_PCM16_HEADER: Final = ( b"RIFF\x24\x0c\x00\x00WAVEfmt \x10\x00\x00\x00\x01\x00\x01\x00\x80\x3e\x00\x00\x00\x7d\x00\x00\x02\x00\x10\x00data\x00\x0c\x00\x00" ) @@ -6149,8 +6150,7 @@ AZURE_SPEECH_NON_UTF8_WAV_BYTES: Final = AZURE_SPEECH_PCM16_HEADER + bytes(range AZURE_SPEECH_TRANSCRIPT: Final = {"RecognitionStatus": "Success", "DisplayText": "The eagle has landed."} -@pytest.fixture -def azure_speech_client(monkeypatch: pytest.MonkeyPatch) -> Iterator[TestClient]: +def _azure_speech_test_client(monkeypatch: pytest.MonkeyPatch, caller: UserAPIKeyAuth) -> TestClient: from litellm.proxy.proxy_server import app monkeypatch.setenv("AZURE_SPEECH_API_KEY", "server-subscription-key") @@ -6159,8 +6159,20 @@ def azure_speech_client(monkeypatch: pytest.MonkeyPatch) -> Iterator[TestClient] monkeypatch.delenv("SERVER_ROOT_PATH", raising=False) monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) litellm.in_memory_llm_clients_cache.flush_cache() - monkeypatch.setitem(app.dependency_overrides, user_api_key_auth, lambda: UserAPIKeyAuth(api_key="sk-virtual")) - yield TestClient(app) + monkeypatch.setitem(app.dependency_overrides, user_api_key_auth, lambda: caller) + return TestClient(app) + + +@pytest.fixture +def azure_speech_client(monkeypatch: pytest.MonkeyPatch) -> Iterator[TestClient]: + yield _azure_speech_test_client(monkeypatch, UserAPIKeyAuth(api_key="sk-virtual")) + + +@pytest.fixture +def azure_speech_admin_client(monkeypatch: pytest.MonkeyPatch) -> Iterator[TestClient]: + yield _azure_speech_test_client( + monkeypatch, UserAPIKeyAuth(api_key="sk-admin", user_role=LitellmUserRoles.PROXY_ADMIN) + ) class TestAzureSpeechProxyRoute: @@ -6193,17 +6205,19 @@ class TestAzureSpeechProxyRoute: assert "authorization" not in sent.headers assert "caller-supplied-key" not in repr(sent.headers) - def test_batch_json_goes_to_the_cognitive_services_host(self, azure_speech_client: TestClient) -> None: + def test_admin_batch_job_creation_goes_to_the_cognitive_services_host( + self, azure_speech_admin_client: TestClient + ) -> None: body: Final = {"contentUrls": ["https://example.com/a.wav"], "locale": "en-US", "displayName": "job"} with respx.mock(assert_all_called=True) as upstream: route = upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( return_value=httpx.Response(201, json={"self": "https://eastus.api.cognitive.microsoft.com/x"}) ) - response = azure_speech_client.post( + response = azure_speech_admin_client.post( f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", json=body, - headers={"Authorization": "Bearer sk-virtual"}, + headers={"Authorization": "Bearer sk-admin"}, ) assert response.status_code == 201 @@ -6212,22 +6226,80 @@ class TestAzureSpeechProxyRoute: assert sent.headers["ocp-apim-subscription-key"] == "server-subscription-key" assert "authorization" not in sent.headers - def test_batch_multipart_upload_is_forwarded_byte_for_byte(self, azure_speech_client: TestClient) -> None: + @pytest.mark.parametrize( + "method,endpoint", + [ + ("POST", AZURE_SPEECH_BATCH_ENDPOINT), + ("POST", "/speechtotext/v3.2/models"), + ("PUT", "/speechtotext/v3.2/endpoints/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab"), + ], + ) + def test_non_admin_key_cannot_create_unpriced_batch_work( + self, azure_speech_client: TestClient, method: str, endpoint: str + ) -> None: + with respx.mock(assert_all_called=False) as upstream: + catch_all = upstream.route().mock(return_value=httpx.Response(201, json={"status": "NotStarted"})) + + response = azure_speech_client.request( + method, + f"/azure_speech{endpoint}", + json={"contentUrls": ["https://example.com/a.wav"], "locale": "en-US"}, + headers={"Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 403, response.text + assert AZURE_SPEECH_FAST_ENDPOINT in response.text + assert not catch_all.called + + def test_non_admin_key_can_still_read_delete_and_fast_transcribe_in_the_batch_family( + self, azure_speech_client: TestClient + ) -> None: + job_path: Final = f"{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab" with respx.mock(assert_all_called=True) as upstream: - route = upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( - return_value=httpx.Response(201, json={"status": "NotStarted"}) + upstream.get(f"https://eastus.api.cognitive.microsoft.com{job_path}").mock( + return_value=httpx.Response(200, json={"status": "Succeeded"}) + ) + upstream.delete(f"https://eastus.api.cognitive.microsoft.com{job_path}").mock( + return_value=httpx.Response(204) + ) + upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_FAST_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"durationMilliseconds": 640, "combinedPhrases": []}) + ) + + statuses = [ + azure_speech_client.get(f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-virtual"}), + azure_speech_client.delete(f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-virtual"}), + azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_FAST_ENDPOINT}", + params={"api-version": "2024-11-15"}, + files={"audio": ("eagle.wav", AZURE_SPEECH_WAV_BYTES, "audio/wav")}, + data={"definition": json.dumps({"locales": ["en-US"]})}, + headers={"Authorization": "Bearer sk-virtual"}, + ), + ] + + assert [r.status_code for r in statuses] == [200, 204, 200] + + def test_fast_transcription_multipart_upload_is_forwarded_byte_for_byte( + self, azure_speech_client: TestClient + ) -> None: + with respx.mock(assert_all_called=True) as upstream: + route = upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_FAST_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"durationMilliseconds": 640, "combinedPhrases": []}) ) response = azure_speech_client.post( - f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", + f"/azure_speech{AZURE_SPEECH_FAST_ENDPOINT}", + params={"api-version": "2024-11-15"}, files={"audio": ("eagle.wav", AZURE_SPEECH_NON_UTF8_WAV_BYTES, "audio/wav")}, data={"definition": json.dumps({"locales": ["en-US"]})}, headers={"Authorization": "Bearer sk-virtual"}, ) - assert response.status_code == 201 + assert response.status_code == 200 sent = route.calls.last.request assert sent.headers["content-type"].startswith("multipart/form-data; boundary=") + assert dict(sent.url.params) == {"api-version": "2024-11-15"} assert AZURE_SPEECH_NON_UTF8_WAV_BYTES in sent.content assert b'name="definition"' in sent.content assert sent.headers["ocp-apim-subscription-key"] == "server-subscription-key" @@ -6247,7 +6319,7 @@ class TestAzureSpeechProxyRoute: @pytest.mark.parametrize("method", ["GET", "POST"]) def test_batch_requests_are_logged_as_azure_speech_not_assemblyai( - self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch, method: str + self, azure_speech_admin_client: TestClient, monkeypatch: pytest.MonkeyPatch, method: str ) -> None: from litellm.integrations.custom_logger import CustomLogger @@ -6266,11 +6338,11 @@ class TestAzureSpeechProxyRoute: return_value=httpx.Response(200, json={"values": []}) ) - response = azure_speech_client.request( + response = azure_speech_admin_client.request( method, f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", json={"locale": "en-US"} if method == "POST" else None, - headers={"Authorization": "Bearer sk-virtual"}, + headers={"Authorization": "Bearer sk-admin"}, ) assert response.status_code == 200 @@ -6278,6 +6350,50 @@ class TestAzureSpeechProxyRoute: ("azure_speech/batch-transcription", "azure_speech", 0.0) ] + def test_fast_transcription_spend_is_priced_from_duration_milliseconds( + self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch + ) -> None: + from litellm.integrations.custom_logger import CustomLogger + + class _Recorder(CustomLogger): + def __init__(self) -> None: + super().__init__() + self.payloads: list[dict[str, object]] = [] # mutable-ok: test recorder accumulates callback payloads + + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time) -> None: + self.payloads.append(kwargs["standard_logging_object"]) + + recorder: Final = _Recorder() + monkeypatch.setattr(litellm, "_async_success_callback", [*litellm._async_success_callback, recorder]) + monkeypatch.setitem( + litellm.model_cost, + "azure/speech/azure-stt", + { + "litellm_provider": "azure", + "mode": "audio_transcription", + "input_cost_per_second": 0.25, + "output_cost_per_second": 0.0, + }, + ) + with respx.mock(assert_all_called=True) as upstream: + upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_FAST_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"durationMilliseconds": 5061, "combinedPhrases": []}) + ) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_FAST_ENDPOINT}", + params={"api-version": "2024-11-15"}, + files={"audio": ("eagle.wav", AZURE_SPEECH_WAV_BYTES, "audio/wav")}, + data={"definition": json.dumps({"locales": ["en-US"]})}, + headers={"Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 200 + assert [(p["model"], p["custom_llm_provider"]) for p in recorder.payloads] == [ + ("azure_speech/fast-transcription", "azure_speech") + ] + assert recorder.payloads[0]["response_cost"] == pytest.approx(5.061 * 0.25) + def test_short_audio_spend_is_priced_from_the_recognized_duration( self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch ) -> None: From 25846d13417af204bb52154c125654dcc4da6411 Mon Sep 17 00:00:00 2001 From: yassin Date: Thu, 17 Sep 2026 20:29:51 +0000 Subject: [PATCH 065/141] fix(proxy): limit the whole Azure Speech batch API to proxy admin keys Ordinary keys could read, patch and delete batch transcription jobs that other keys created with the proxy's shared Azure subscription, so every /speechtotext/v3.2 method is now admin only while fast transcription stays open. Also clears SERVER_ROOT_PATH in the real-auth test helper because test_custom_proxy leaves it set at import time and the shared app then 404s pass-through routes Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 1 - .../llm_passthrough_endpoints.py | 15 ++--- .../test_llm_pass_through_endpoints.py | 65 ++++++++++++------- 3 files changed, 48 insertions(+), 33 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index d9da0cc0f64..62523c5d2c3 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1575,7 +1575,6 @@ AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX: Final = "/azure_speech" AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX: Final = "/speech/" AZURE_SPEECH_BATCH_PATH_PREFIX: Final = "/speechtotext/" AZURE_SPEECH_FAST_TRANSCRIPTION_PATH: Final = "/speechtotext/transcriptions:transcribe" -AZURE_SPEECH_UNPRICED_WRITE_METHODS: Final = frozenset({"POST", "PUT"}) AZURE_SPEECH_STT_DOMAIN: Final = "stt.speech.microsoft.com" AZURE_SPEECH_COGNITIVE_SERVICES_DOMAIN: Final = "api.cognitive.microsoft.com" AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER: Final = "Ocp-Apim-Subscription-Key" diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index b8966508f81..8d18de42f45 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -38,7 +38,6 @@ from litellm.constants import ( AZURE_SPEECH_SHORT_AUDIO_PATH_PREFIX, AZURE_SPEECH_STT_DOMAIN, AZURE_SPEECH_SUBSCRIPTION_KEY_HEADER, - AZURE_SPEECH_UNPRICED_WRITE_METHODS, BEDROCK_AGENT_RUNTIME_PASS_THROUGH_ROUTES, ) from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix @@ -1360,11 +1359,10 @@ def resolve_azure_speech_base_url(endpoint_path: str, api_base: str | None, regi return httpx.URL(f"https://{region}.{domain}") -def azure_speech_write_is_unpriced(method: str, endpoint_path: str) -> bool: +def azure_speech_path_manages_shared_resources(endpoint_path: str) -> bool: return ( endpoint_path.startswith(AZURE_SPEECH_BATCH_PATH_PREFIX) and endpoint_path != AZURE_SPEECH_FAST_TRANSCRIPTION_PATH - and method.upper() in AZURE_SPEECH_UNPRICED_WRITE_METHODS ) @@ -1406,15 +1404,14 @@ async def azure_speech_proxy_route( "AZURE_SPEECH_REGION or AZURE_SPEECH_API_BASE in the proxy environment." ), ) - if azure_speech_write_is_unpriced( - method=request.method, endpoint_path=normalized_endpoint_path - ) and not is_proxy_admin(user_api_key_dict): + if azure_speech_path_manages_shared_resources(normalized_endpoint_path) and not is_proxy_admin(user_api_key_dict): raise HTTPException( status_code=403, detail=( - f"{request.method} {normalized_endpoint_path} creates Azure Speech work whose cost is unknown at " - "request time, so it is limited to proxy admin keys. Use " - f"{AZURE_SPEECH_FAST_TRANSCRIPTION_PATH} for transcription that is priced per request." + f"{request.method} {normalized_endpoint_path} manages batch transcription resources that belong to " + "the proxy's Azure Speech subscription and whose cost is unknown at request time, so it is limited " + f"to proxy admin keys. Use {AZURE_SPEECH_FAST_TRANSCRIPTION_PATH} for transcription that is priced " + "per request." ), ) azure_speech_api_key: Final = passthrough_endpoint_router.get_credentials( diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index eb0c607fbef..0e8a2b9e0e9 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -6232,13 +6232,17 @@ class TestAzureSpeechProxyRoute: ("POST", AZURE_SPEECH_BATCH_ENDPOINT), ("POST", "/speechtotext/v3.2/models"), ("PUT", "/speechtotext/v3.2/endpoints/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab"), + ("GET", AZURE_SPEECH_BATCH_ENDPOINT), + ("GET", f"{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab/files"), + ("PATCH", f"{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab"), + ("DELETE", f"{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab"), ], ) - def test_non_admin_key_cannot_create_unpriced_batch_work( + def test_non_admin_key_cannot_manage_shared_batch_resources( self, azure_speech_client: TestClient, method: str, endpoint: str ) -> None: with respx.mock(assert_all_called=False) as upstream: - catch_all = upstream.route().mock(return_value=httpx.Response(201, json={"status": "NotStarted"})) + catch_all = upstream.route().mock(return_value=httpx.Response(200, json={"status": "Succeeded"})) response = azure_speech_client.request( method, @@ -6251,9 +6255,23 @@ class TestAzureSpeechProxyRoute: assert AZURE_SPEECH_FAST_ENDPOINT in response.text assert not catch_all.called - def test_non_admin_key_can_still_read_delete_and_fast_transcribe_in_the_batch_family( - self, azure_speech_client: TestClient - ) -> None: + def test_non_admin_key_can_still_fast_transcribe_in_the_batch_family(self, azure_speech_client: TestClient) -> None: + with respx.mock(assert_all_called=True) as upstream: + upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_FAST_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"durationMilliseconds": 640, "combinedPhrases": []}) + ) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_FAST_ENDPOINT}", + params={"api-version": "2024-11-15"}, + files={"audio": ("eagle.wav", AZURE_SPEECH_WAV_BYTES, "audio/wav")}, + data={"definition": json.dumps({"locales": ["en-US"]})}, + headers={"Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 200 + + def test_admin_key_reads_and_deletes_batch_jobs(self, azure_speech_admin_client: TestClient) -> None: job_path: Final = f"{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab" with respx.mock(assert_all_called=True) as upstream: upstream.get(f"https://eastus.api.cognitive.microsoft.com{job_path}").mock( @@ -6262,23 +6280,15 @@ class TestAzureSpeechProxyRoute: upstream.delete(f"https://eastus.api.cognitive.microsoft.com{job_path}").mock( return_value=httpx.Response(204) ) - upstream.post(f"https://eastus.api.cognitive.microsoft.com{AZURE_SPEECH_FAST_ENDPOINT}").mock( - return_value=httpx.Response(200, json={"durationMilliseconds": 640, "combinedPhrases": []}) - ) statuses = [ - azure_speech_client.get(f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-virtual"}), - azure_speech_client.delete(f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-virtual"}), - azure_speech_client.post( - f"/azure_speech{AZURE_SPEECH_FAST_ENDPOINT}", - params={"api-version": "2024-11-15"}, - files={"audio": ("eagle.wav", AZURE_SPEECH_WAV_BYTES, "audio/wav")}, - data={"definition": json.dumps({"locales": ["en-US"]})}, - headers={"Authorization": "Bearer sk-virtual"}, + azure_speech_admin_client.get(f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-admin"}), + azure_speech_admin_client.delete( + f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-admin"} ), ] - assert [r.status_code for r in statuses] == [200, 204, 200] + assert [r.status_code for r in statuses] == [200, 204] def test_fast_transcription_multipart_upload_is_forwarded_byte_for_byte( self, azure_speech_client: TestClient @@ -6305,14 +6315,16 @@ class TestAzureSpeechProxyRoute: assert sent.headers["ocp-apim-subscription-key"] == "server-subscription-key" assert "authorization" not in sent.headers - def test_batch_get_is_forwarded_with_the_job_id_path(self, azure_speech_client: TestClient) -> None: + def test_batch_get_is_forwarded_with_the_job_id_path(self, azure_speech_admin_client: TestClient) -> None: job_path: Final = f"{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab/files" with respx.mock(assert_all_called=True) as upstream: route = upstream.get(f"https://eastus.api.cognitive.microsoft.com{job_path}").mock( return_value=httpx.Response(200, json={"values": []}) ) - response = azure_speech_client.get(f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-virtual"}) + response = azure_speech_admin_client.get( + f"/azure_speech{job_path}", headers={"Authorization": "Bearer sk-admin"} + ) assert (response.status_code, response.json()) == (200, {"values": []}) assert route.calls.last.request.headers["ocp-apim-subscription-key"] == "server-subscription-key" @@ -6445,8 +6457,8 @@ class TestAzureSpeechProxyRoute: short_audio = upstream.post( f"https://my-speech.cognitiveservices.azure.com{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}" ).mock(return_value=httpx.Response(200, json=AZURE_SPEECH_TRANSCRIPT)) - batch = upstream.get(f"https://my-speech.cognitiveservices.azure.com{AZURE_SPEECH_BATCH_ENDPOINT}").mock( - return_value=httpx.Response(200, json={"values": []}) + fast = upstream.post(f"https://my-speech.cognitiveservices.azure.com{AZURE_SPEECH_FAST_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"durationMilliseconds": 640, "combinedPhrases": []}) ) azure_speech_client.post( @@ -6454,9 +6466,15 @@ class TestAzureSpeechProxyRoute: content=AZURE_SPEECH_WAV_BYTES, headers={"Content-Type": "audio/wav", "Authorization": "Bearer sk-virtual"}, ) - azure_speech_client.get(f"/azure_speech{AZURE_SPEECH_BATCH_ENDPOINT}", headers={"Authorization": "Bearer x"}) + azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_FAST_ENDPOINT}", + params={"api-version": "2024-11-15"}, + files={"audio": ("eagle.wav", AZURE_SPEECH_WAV_BYTES, "audio/wav")}, + data={"definition": json.dumps({"locales": ["en-US"]})}, + headers={"Authorization": "Bearer sk-virtual"}, + ) - assert short_audio.called and batch.called + assert short_audio.called and fast.called @pytest.mark.parametrize("endpoint", ["openai/deployments/whisper/audio/transcriptions", "speech", "speechtotext"]) def test_unknown_path_family_is_rejected_before_any_upstream_call( @@ -6543,6 +6561,7 @@ class TestAzureSpeechRawBodyThroughRealAuth: monkeypatch.setenv("AZURE_SPEECH_API_KEY", "server-subscription-key") monkeypatch.setenv("AZURE_SPEECH_REGION", "eastus") monkeypatch.delenv("AZURE_SPEECH_API_BASE", raising=False) + monkeypatch.delenv("SERVER_ROOT_PATH", raising=False) monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) litellm.in_memory_llm_clients_cache.flush_cache() with patch.multiple( # test-quality-ok: the real user_api_key_auth reads proxy_server module globals (master_key, caches) that have no injection seam From ea37596b884056b2c818f8945d71e394b5126c7d Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 01:01:26 +0000 Subject: [PATCH 066/141] feat(mcp): let proxy admins force-close live MCP sessions and revoke stored user credentials Adds an admin-only DELETE /v1/mcp/sessions that terminates stateful MCP gateway sessions on the current worker by session id prefix and/or by the LiteLLM user that opened them, tombstones the terminated ids so a client reusing one gets 404 instead of a silently recreated stateless session, and lets PROXY_ADMIN name a user_id on the BYOK and OAuth credential delete routes. Full and view-only admins can list every user's stored credential metadata for a server (never the secret). The dashboard gains Disconnect controls on the Live Connections tab and a User Credentials tab with Revoke controls, both hidden from read-only admins. Resolves LIT-8001 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 2 + litellm/proxy/_experimental/mcp_server/db.py | 32 ++ .../proxy/_experimental/mcp_server/server.py | 72 +++- litellm/proxy/_lazy_openapi_snapshot.json | 237 ++++++++++- litellm/proxy/_types.py | 10 + .../mcp_management_endpoints.py | 129 +++++- litellm/types/mcp.py | 8 + .../mcp_server/test_db_credentials.py | 47 +++ .../mcp_server/test_mcp_server.py | 177 ++++++++ .../test_mcp_management_endpoints.py | 398 ++++++++++++++++++ ...MCPGatewaySessionsTab.integration.test.tsx | 90 +++- .../_components/MCPGatewaySessionsTab.tsx | 146 ++++++- ...rUserCredentialsPanel.integration.test.tsx | 102 +++++ .../MCPServerUserCredentialsPanel.tsx | 212 ++++++++++ .../_components/mcp_server_view.tsx | 21 + .../mcp-servers/_components/mcp_servers.tsx | 4 +- .../src/components/mcp_tools/types.tsx | 20 + .../src/components/networking.tsx | 36 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 132 +++++- 19 files changed, 1826 insertions(+), 49 deletions(-) create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.integration.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.tsx diff --git a/litellm/constants.py b/litellm/constants.py index d4827bb7483..2ebc9beb632 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -183,6 +183,8 @@ MCP_TOOL_LISTING_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_TOOL_LISTING_TIME MCP_METADATA_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_METADATA_TIMEOUT", "10.0")) MCP_HEALTH_CHECK_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_HEALTH_CHECK_TIMEOUT", "10.0")) MCP_TOOL_LISTING_MAX_PAGES: Final = 1000 +MCP_GATEWAY_SESSION_ID_PREFIX_LENGTH: Final = 8 +MCP_ADMIN_TERMINATED_SESSION_IDS_MAX: Final = 1024 # Allowlist of commands permitted for MCP stdio transport. # Prevents arbitrary command execution via /mcp-rest/test/* endpoints or server creation. diff --git a/litellm/proxy/_experimental/mcp_server/db.py b/litellm/proxy/_experimental/mcp_server/db.py index 789b2ffaef4..a04e2f5c9b8 100644 --- a/litellm/proxy/_experimental/mcp_server/db.py +++ b/litellm/proxy/_experimental/mcp_server/db.py @@ -24,6 +24,7 @@ from litellm.proxy._types import ( MCPApprovalStatus, MCPEnvVar, MCPEnvVarScope, + MCPServerUserCredentialListItem, MCPSubmissionsSummary, NewMCPServerRequest, SpecialMCPServerName, @@ -1504,6 +1505,37 @@ async def get_user_oauth_credential( return _parse_oauth_payload(decoded) +def _server_user_credential_item( + row: "prisma_db_models.LiteLLM_MCPUserCredentials", +) -> MCPServerUserCredentialListItem: + oauth_payload: Final = _decode_oauth_payload(row.credential_b64) + if oauth_payload is None: + return MCPServerUserCredentialListItem( + user_id=row.user_id, + credential_type="byok", + updated_at=row.updated_at.isoformat(), + ) + return MCPServerUserCredentialListItem( + user_id=row.user_id, + credential_type="oauth2", + expires_at=oauth_payload.get("expires_at"), + connected_at=oauth_payload.get("connected_at"), + updated_at=row.updated_at.isoformat(), + ) + + +async def list_server_user_credentials( + prisma_client: PrismaClient, + server_id: str, +) -> tuple[MCPServerUserCredentialListItem, ...]: + """Every user's stored credential for one server, typed but without the secret, for admins.""" + rows: Final = await _db_find_user_credential_rows( + prisma_client, + {"server_id": server_id}, # mutable-ok: prisma where-inputs must be plain dicts + ) + return tuple(_server_user_credential_item(row) for row in rows) + + async def list_user_oauth_credentials( prisma_client: PrismaClient, user_id: str, diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py index 524bac747ad..80d9274859d 100644 --- a/litellm/proxy/_experimental/mcp_server/server.py +++ b/litellm/proxy/_experimental/mcp_server/server.py @@ -14,7 +14,7 @@ import time import traceback import types import uuid -from collections import Counter +from collections import Counter, deque from collections.abc import AsyncIterator, Callable, Mapping, Sequence from datetime import datetime from typing import TYPE_CHECKING, Any, Final, NoReturn, Protocol @@ -28,7 +28,11 @@ from starlette.types import Message, Receive, Scope, Send from typing_extensions import ReadOnly, TypedDict from litellm._logging import verbose_logger -from litellm.constants import MAXIMUM_TRACEBACK_LINES_TO_LOG +from litellm.constants import ( + MAXIMUM_TRACEBACK_LINES_TO_LOG, + MCP_ADMIN_TERMINATED_SESSION_IDS_MAX, + MCP_GATEWAY_SESSION_ID_PREFIX_LENGTH, +) from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.llms.custom_httpx.http_handler import ( get_async_httpx_client, @@ -91,6 +95,7 @@ from litellm.types.mcp import ( MCPGatewaySession, MCPGatewaySessionGroupCount, MCPGatewaySessionsResponse, + MCPGatewaySessionsTerminateResponse, MCPSpecVersion, ) from litellm.types.mcp_server.mcp_server_manager import MCPInfo, MCPServer @@ -618,6 +623,9 @@ if MCP_AVAILABLE: _stateful_session_locks: Final[dict[str, asyncio.Lock]] = {} _stateful_session_active_request_counts: Final[dict[str, int]] = {} _stateful_session_client_info: Final[dict[str, Implementation]] = {} # mutable-ok: cleared on session teardown + _admin_terminated_session_ids: Final[deque[str]] = deque( # mutable-ok: bounded ring, appended on admin termination + maxlen=MCP_ADMIN_TERMINATED_SESSION_IDS_MAX + ) class _TerminableTransport(Protocol): async def terminate(self) -> None: ... @@ -3850,7 +3858,7 @@ if MCP_AVAILABLE: client_info: Final = _stateful_session_client_info.get(session_id) key_auth: Final = auth_user.user_api_key_auth return MCPGatewaySession( - session_id_prefix=session_id[:8], + session_id_prefix=session_id[:MCP_GATEWAY_SESSION_ID_PREFIX_LENGTH], client_name=client_info.name if client_info is not None else None, client_version=client_info.version if client_info is not None else None, user_id=key_auth.user_id if key_auth is not None else None, @@ -3885,6 +3893,53 @@ if MCP_AVAILABLE: sessions=sessions, ) + def _session_matches_admin_selector( + session_id: str, + auth_user: MCPAuthenticatedUser, + session_id_prefix: str | None, + user_id: str | None, + ) -> bool: + if session_id_prefix is not None and not session_id.startswith(session_id_prefix): + return False + if user_id is None: + return True + key_auth: Final = auth_user.user_api_key_auth + return key_auth is not None and key_auth.user_id == user_id + + async def terminate_mcp_gateway_sessions( + *, + session_id_prefix: str | None = None, + user_id: str | None = None, + ) -> MCPGatewaySessionsTerminateResponse: + """Force-close every live stateful session on this worker matching the selector. + + The transport is terminated (open streams close), all per-session + tracking is dropped, and the id is remembered so a client that keeps + sending it receives 404 and has to ``initialize`` again, which re-runs + admission. Only sessions held by this worker process are affected. + """ + now: Final = time.monotonic() + server_instances: Final = _stateful_server_instances() + targets: Final = tuple( + (session_id, auth_user) + for session_id, auth_user in tuple(_stateful_session_auth_contexts.items()) + if session_id in server_instances + and _session_matches_admin_selector(session_id, auth_user, session_id_prefix, user_id) + ) + terminated: Final = tuple(_gateway_session_for(session_id, auth_user, now) for session_id, auth_user in targets) + for session_id, _ in targets: + _admin_terminated_session_ids.append(session_id) + transport = server_instances.pop(session_id, None) + _remove_stateful_session_tracking(session_id) + if transport is not None: + await transport.terminate() + verbose_logger.warning("MCP session '%s' terminated by an administrator.", session_id) + return MCPGatewaySessionsTerminateResponse( + worker_pid=os.getpid(), + terminated_sessions=len(terminated), + sessions=terminated, + ) + async def _read_request_body_for_routing( receive: Receive, ) -> tuple[list[Message], bytes]: @@ -4009,6 +4064,17 @@ if MCP_AVAILABLE: await success_response(scope, receive, send) return True + if _session_id in _admin_terminated_session_ids: + terminated_response: Final = JSONResponse( + status_code=404, + content={ # mutable-ok: JSONResponse content must be a plain dict + "error": "Not Found", + "details": "mcp-session-id was terminated by an administrator. Send initialize to start a new session.", + }, + ) + await terminated_response(scope, receive, send) + return True + # Non-DELETE: strip stale session ID to allow new session creation verbose_logger.warning( "MCP session ID '%s' not found in this worker's memory. " diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index 82b709feb92..1e9b1f5b743 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -27648,6 +27648,32 @@ "title": "MCPGatewaySessionsResponse", "type": "object" }, + "MCPGatewaySessionsTerminateResponse": { + "description": "Stateful sessions an administrator force-closed on this proxy worker.", + "properties": { + "sessions": { + "items": { + "$ref": "#/components/schemas/MCPGatewaySession" + }, + "title": "Sessions", + "type": "array" + }, + "terminated_sessions": { + "title": "Terminated Sessions", + "type": "integer" + }, + "worker_pid": { + "title": "Worker Pid", + "type": "integer" + } + }, + "required": [ + "worker_pid", + "terminated_sessions" + ], + "title": "MCPGatewaySessionsTerminateResponse", + "type": "object" + }, "MCPOAuthUserCredentialRequest": { "description": "Stores a user's OAuth2 token for an OpenAPI MCP server.", "properties": { @@ -27744,6 +27770,56 @@ "title": "MCPOAuthUserCredentialStatus", "type": "object" }, + "MCPServerUserCredentialListItem": { + "description": "One user's stored credential for an MCP server, as an admin sees it. Never carries the secret.", + "properties": { + "connected_at": { + "anyOf": [ + { + "type": "string" + }, + { + "type": "null" + } + ], + "title": "Connected At" + }, + "credential_type": { + "enum": [ + "oauth2", + "byok" + ], + "title": "Credential Type", + "type": "string" + }, + "expires_at": { + "anyOf": [ + { + "type": "string" + }, + { + "type": "null" + } + ], + "title": "Expires At" + }, + "updated_at": { + "title": "Updated At", + "type": "string" + }, + "user_id": { + "title": "User Id", + "type": "string" + } + }, + "required": [ + "user_id", + "credential_type", + "updated_at" + ], + "title": "MCPServerUserCredentialListItem", + "type": "object" + }, "MCPSubmissionsSummary": { "properties": { "active": { @@ -29920,7 +29996,7 @@ }, "/v1/mcp/server/{server_id}/oauth-user-credential": { "delete": { - "description": "Revoke the calling user's stored OAuth2 token for an MCP server", + "description": "Revoke the calling user's stored OAuth2 token for an MCP server. A proxy admin may pass user_id to revoke another user's stored token.", "operationId": "delete_mcp_oauth_user_credential_v1_mcp_server__server_id__oauth_user_credential_delete", "parameters": [ { @@ -29931,6 +30007,23 @@ "title": "Server Id", "type": "string" } + }, + { + "in": "query", + "name": "user_id", + "required": false, + "schema": { + "anyOf": [ + { + "minLength": 1, + "type": "string" + }, + { + "type": "null" + } + ], + "title": "User Id" + } } ], "responses": { @@ -30130,7 +30223,7 @@ }, "/v1/mcp/server/{server_id}/user-credential": { "delete": { - "description": "Delete the calling user's stored API key for a BYOK MCP server", + "description": "Delete the calling user's stored API key for a BYOK MCP server. A proxy admin may pass user_id to revoke another user's stored key.", "operationId": "delete_mcp_user_credential_v1_mcp_server__server_id__user_credential_delete", "parameters": [ { @@ -30141,6 +30234,23 @@ "title": "Server Id", "type": "string" } + }, + { + "in": "query", + "name": "user_id", + "required": false, + "schema": { + "anyOf": [ + { + "minLength": 1, + "type": "string" + }, + { + "type": "null" + } + ], + "title": "User Id" + } } ], "responses": { @@ -30232,6 +30342,58 @@ ] } }, + "/v1/mcp/server/{server_id}/user-credentials": { + "get": { + "description": "List every user's stored BYOK or OAuth2 credential for an MCP server (admin only, no secrets)", + "operationId": "list_mcp_server_user_credentials_v1_mcp_server__server_id__user_credentials_get", + "parameters": [ + { + "in": "path", + "name": "server_id", + "required": true, + "schema": { + "title": "Server Id", + "type": "string" + } + } + ], + "responses": { + "200": { + "content": { + "application/json": { + "schema": { + "items": { + "$ref": "#/components/schemas/MCPServerUserCredentialListItem" + }, + "title": "Response List Mcp Server User Credentials V1 Mcp Server Server Id User Credentials Get", + "type": "array" + } + } + }, + "description": "Successful Response" + }, + "422": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/HTTPValidationError" + } + } + }, + "description": "Validation Error" + } + }, + "security": [ + { + "APIKeyHeader": [] + } + ], + "summary": "List Mcp Server User Credentials", + "tags": [ + "mcp_management" + ] + } + }, "/v1/mcp/server/{server_id}/user-env-vars": { "delete": { "description": "Clear the calling user's per-user MCP env var values for this server.", @@ -30383,6 +30545,77 @@ } }, "/v1/mcp/sessions": { + "delete": { + "description": "Force-close live stateful MCP gateway sessions on this proxy worker, selected by session id prefix and/or by the LiteLLM user that opened them (proxy admin only).", + "operationId": "delete_mcp_gateway_sessions_v1_mcp_sessions_delete", + "parameters": [ + { + "in": "query", + "name": "session_id_prefix", + "required": false, + "schema": { + "anyOf": [ + { + "minLength": 8, + "type": "string" + }, + { + "type": "null" + } + ], + "title": "Session Id Prefix" + } + }, + { + "in": "query", + "name": "user_id", + "required": false, + "schema": { + "anyOf": [ + { + "minLength": 1, + "type": "string" + }, + { + "type": "null" + } + ], + "title": "User Id" + } + } + ], + "responses": { + "200": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/MCPGatewaySessionsTerminateResponse" + } + } + }, + "description": "Successful Response" + }, + "422": { + "content": { + "application/json": { + "schema": { + "$ref": "#/components/schemas/HTTPValidationError" + } + } + }, + "description": "Validation Error" + } + }, + "security": [ + { + "APIKeyHeader": [] + } + ], + "summary": "Delete Mcp Gateway Sessions", + "tags": [ + "mcp_management" + ] + }, "get": { "description": "Live stateful MCP gateway sessions on this proxy worker, grouped by AI client and by user.", "operationId": "get_mcp_gateway_sessions_v1_mcp_sessions_get", diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index ff32d4784df..ee92e64b046 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -1716,6 +1716,16 @@ class MCPUserCredentialListItem(LiteLLMPydanticObjectBase): connected_at: str | None = None # ISO-8601 +class MCPServerUserCredentialListItem(LiteLLMPydanticObjectBase): + """One user's stored credential for an MCP server, as an admin sees it. Never carries the secret.""" + + user_id: str + credential_type: Literal["oauth2", "byok"] + expires_at: str | None = None + connected_at: str | None = None + updated_at: str + + class MCPUserEnvVarsRequest(LiteLLMPydanticObjectBase): """Payload for storing the calling user's per-user env var values.""" diff --git a/litellm/proxy/management_endpoints/mcp_management_endpoints.py b/litellm/proxy/management_endpoints/mcp_management_endpoints.py index 82a1cdcdd00..728ba9568e8 100644 --- a/litellm/proxy/management_endpoints/mcp_management_endpoints.py +++ b/litellm/proxy/management_endpoints/mcp_management_endpoints.py @@ -47,7 +47,7 @@ except ImportError: import litellm from litellm._logging import verbose_logger, verbose_proxy_logger from litellm._uuid import uuid -from litellm.constants import LITELLM_PROXY_ADMIN_NAME +from litellm.constants import LITELLM_PROXY_ADMIN_NAME, MCP_GATEWAY_SESSION_ID_PREFIX_LENGTH from litellm.proxy._experimental.mcp_server.utils import ( LITELLM_MCP_SERVER_DESCRIPTION, LITELLM_MCP_SERVER_NAME, @@ -145,6 +145,7 @@ if MCP_AVAILABLE: get_user_env_vars, get_user_env_vars_bulk, get_user_oauth_credential, + list_server_user_credentials, list_user_oauth_credentials, mcp_oauth_token_identity, merge_user_env_vars, @@ -180,6 +181,7 @@ if MCP_AVAILABLE: MCPApprovalStatus, MCPOAuthUserCredentialRequest, MCPOAuthUserCredentialStatus, + MCPServerUserCredentialListItem, MCPSubmissionsSummary, MCPTransport, MCPUserCredentialListItem, @@ -221,6 +223,7 @@ if MCP_AVAILABLE: MCPAuth, MCPCredentials, MCPGatewaySessionsResponse, + MCPGatewaySessionsTerminateResponse, normalize_upstream_header_name, ) from litellm.types.mcp_server.mcp_server_manager import MCPServer @@ -662,6 +665,31 @@ if MCP_AVAILABLE: """ return user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN + def _resolve_credential_target_user_id(user_api_key_dict: UserAPIKeyAuth, requested_user_id: str | None) -> str: + """The user whose stored MCP credential a request acts on. + + Defaults to the caller. Naming another user is a revocation and needs + ``PROXY_ADMIN``; a read-only admin or a regular user gets 403. + """ + caller_user_id: Final = user_api_key_dict.user_id or "" + if requested_user_id is not None and requested_user_id != caller_user_id: + if not _user_is_full_admin(user_api_key_dict): + raise HTTPException( + status_code=status.HTTP_403_FORBIDDEN, + detail={ # mutable-ok: FastAPI HTTPException detail requires a plain dict + "error": "Proxy admin access required to revoke another user's MCP credential.", + }, + ) + return requested_user_id + if not caller_user_id: + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail={ + "error": "User ID not found in token" + }, # mutable-ok: FastAPI HTTPException detail requires a plain dict + ) + return caller_user_id + def _is_restricted_virtual_key_request(user_api_key_dict: UserAPIKeyAuth) -> bool: """Best-effort detection for route-restricted virtual keys. @@ -1373,6 +1401,41 @@ if MCP_AVAILABLE: return get_mcp_gateway_sessions_report() + @router.delete( + "/sessions", + description=( + "Force-close live stateful MCP gateway sessions on this proxy worker, selected by session id prefix " + "and/or by the LiteLLM user that opened them (proxy admin only)." + ), + dependencies=(Depends(user_api_key_auth),), + response_model=MCPGatewaySessionsTerminateResponse, + ) + @management_endpoint_wrapper + async def delete_mcp_gateway_sessions( + user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)], + session_id_prefix: Annotated[str | None, Query(min_length=MCP_GATEWAY_SESSION_ID_PREFIX_LENGTH)] = None, + user_id: Annotated[str | None, Query(min_length=1)] = None, + ) -> MCPGatewaySessionsTerminateResponse: + if not _user_is_full_admin(user_api_key_dict): + raise HTTPException( + status_code=status.HTTP_403_FORBIDDEN, + detail={ # mutable-ok: FastAPI HTTPException detail requires a plain dict + "error": "Proxy admin access required to terminate MCP gateway sessions.", + }, + ) + if session_id_prefix is None and user_id is None: + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail={ # mutable-ok: FastAPI HTTPException detail requires a plain dict + "error": "Provide session_id_prefix and/or user_id to select the sessions to terminate.", + }, + ) + from litellm.proxy._experimental.mcp_server.server import ( + terminate_mcp_gateway_sessions, + ) + + return await terminate_mcp_gateway_sessions(session_id_prefix=session_id_prefix, user_id=user_id) + @router.get( "/server/submissions", description="Returns all MCP servers submitted by non-admin users (admin review queue). Mirrors GET /guardrails/submissions.", @@ -2261,7 +2324,10 @@ if MCP_AVAILABLE: @router.delete( "/server/{server_id}/user-credential", - description="Delete the calling user's stored API key for a BYOK MCP server", + description=( + "Delete the calling user's stored API key for a BYOK MCP server. " + "A proxy admin may pass user_id to revoke another user's stored key." + ), dependencies=[Depends(user_api_key_auth)], response_model=MCPUserCredentialResponse, ) @@ -2269,24 +2335,20 @@ if MCP_AVAILABLE: async def delete_mcp_user_credential( server_id: str, user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth), + user_id: Annotated[str | None, Query(min_length=1)] = None, ): - """Remove the calling user's BYOK credential.""" + """Remove the target user's BYOK credential (the caller unless an admin names another user).""" prisma_client: Final = get_prisma_client_or_throw("Database not connected. Connect a database to your proxy") - user_id: Final = user_api_key_dict.user_id or "" - if not user_id: - raise HTTPException( - status_code=status.HTTP_400_BAD_REQUEST, - detail={"error": "User ID not found in token"}, - ) + target_user_id: Final = _resolve_credential_target_user_id(user_api_key_dict, user_id) try: - await delete_user_credential(prisma_client, user_id, server_id) + await delete_user_credential(prisma_client, target_user_id, server_id) except RecordNotFoundError: pass # Already deleted or didn't exist from litellm.proxy._experimental.mcp_server.server import ( _invalidate_byok_cred_cache, ) - _invalidate_byok_cred_cache(user_id, server_id) + _invalidate_byok_cred_cache(target_user_id, server_id) return MCPUserCredentialResponse(server_id=server_id, has_credential=False) # ── OAuth2 user-credential endpoints ────────────────────────────────────── @@ -2362,7 +2424,10 @@ if MCP_AVAILABLE: @router.delete( "/server/{server_id}/oauth-user-credential", - description="Revoke the calling user's stored OAuth2 token for an MCP server", + description=( + "Revoke the calling user's stored OAuth2 token for an MCP server. " + "A proxy admin may pass user_id to revoke another user's stored token." + ), dependencies=[Depends(user_api_key_auth)], response_model=MCPOAuthUserCredentialStatus, ) @@ -2370,29 +2435,25 @@ if MCP_AVAILABLE: async def delete_mcp_oauth_user_credential( server_id: str, user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth), + user_id: Annotated[str | None, Query(min_length=1)] = None, ): - """Revoke/delete the user's OAuth2 credential.""" + """Revoke the target user's OAuth2 credential (the caller unless an admin names another user).""" prisma_client: Final = get_prisma_client_or_throw("Database not connected. Connect a database to your proxy") - user_id: Final = user_api_key_dict.user_id or "" - if not user_id: - raise HTTPException( - status_code=status.HTTP_400_BAD_REQUEST, - detail={"error": "User ID not found in token"}, - ) + target_user_id: Final = _resolve_credential_target_user_id(user_api_key_dict, user_id) # Only delete if the stored credential is actually an OAuth2 token. # This prevents accidentally deleting a BYOK credential if one exists # for the same (user_id, server_id) pair. - cred_to_delete: Final = await get_user_oauth_credential(prisma_client, user_id, server_id) + cred_to_delete: Final = await get_user_oauth_credential(prisma_client, target_user_id, server_id) if cred_to_delete is not None: try: - await delete_user_credential(prisma_client, user_id, server_id) + await delete_user_credential(prisma_client, target_user_id, server_id) except RecordNotFoundError: pass # Already gone — treat as a successful delete from litellm.proxy._experimental.mcp_server.mcp_server_manager import ( # noqa: PLC0415 global_mcp_server_manager, ) - await global_mcp_server_manager.invalidate_user_oauth_token_cache(user_id, server_id) + await global_mcp_server_manager.invalidate_user_oauth_token_cache(target_user_id, server_id) return MCPOAuthUserCredentialStatus( server_id=server_id, has_credential=False, @@ -2481,6 +2542,30 @@ if MCP_AVAILABLE: ) return items + @router.get( + "/server/{server_id}/user-credentials", + description="List every user's stored BYOK or OAuth2 credential for an MCP server (admin only, no secrets)", + dependencies=(Depends(user_api_key_auth),), + response_model=list[MCPServerUserCredentialListItem], + ) + @management_endpoint_wrapper + async def list_mcp_server_user_credentials( + server_id: str, + user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)], + ) -> tuple[MCPServerUserCredentialListItem, ...]: + if user_api_key_dict.user_role not in ( + LitellmUserRoles.PROXY_ADMIN, + LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY, + ): + raise HTTPException( + status_code=status.HTTP_403_FORBIDDEN, + detail={ # mutable-ok: FastAPI HTTPException detail requires a plain dict + "error": "Admin access required to view MCP server user credentials.", + }, + ) + prisma_client: Final = get_prisma_client_or_throw("Database not connected. Connect a database to your proxy") + return await list_server_user_credentials(prisma_client, server_id) + # ── Per-user MCP env var endpoints ──────────────────────────────────────── async def _authorize_and_fetch_mcp_server( diff --git a/litellm/types/mcp.py b/litellm/types/mcp.py index 2d06bb9a009..c5a26c997b7 100644 --- a/litellm/types/mcp.py +++ b/litellm/types/mcp.py @@ -464,3 +464,11 @@ class MCPGatewaySessionsResponse(BaseModel): by_client: list[MCPGatewaySessionGroupCount] = Field(default_factory=list) by_user: list[MCPGatewaySessionGroupCount] = Field(default_factory=list) sessions: list[MCPGatewaySession] = Field(default_factory=list) + + +class MCPGatewaySessionsTerminateResponse(BaseModel): + """Stateful sessions an administrator force-closed on this proxy worker.""" + + worker_pid: int + terminated_sessions: int + sessions: list[MCPGatewaySession] = Field(default_factory=list) diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py index 60a5e1a22bb..cfcff73b857 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py @@ -212,6 +212,53 @@ async def test_purge_user_oauth_credentials_for_server_invalidates_each_user(): assert set(invalidations) == {("alice", "srv-1"), ("bob", "srv-1")} +@pytest.mark.asyncio +async def test_list_server_user_credentials_types_each_row_without_leaking_the_secret(): + """The admin view of one server's stored credentials names the user and the kind of + credential (OAuth2 vs BYOK) and echoes OAuth expiry, but never the token or key itself.""" + from litellm.proxy._experimental.mcp_server.db import list_server_user_credentials + + oauth_row = _legacy_row( + json.dumps( + { + "type": "oauth2", + "access_token": "tok-alice", + "expires_at": "2026-12-31T00:00:00+00:00", + "connected_at": "2026-01-01T00:00:00+00:00", + } + ) + ) + oauth_row.user_id = "alice" + oauth_row.updated_at = datetime(2026, 1, 1, tzinfo=timezone.utc) + byok_row = _byok_row("carol") + byok_row.updated_at = datetime(2026, 2, 1, tzinfo=timezone.utc) + prisma = MagicMock() + prisma.db.litellm_mcpusercredentials.find_many = AsyncMock(return_value=[oauth_row, byok_row]) + + items = await list_server_user_credentials(prisma, "srv-1") + + prisma.db.litellm_mcpusercredentials.find_many.assert_awaited_once_with(where={"server_id": "srv-1"}) + assert [item.model_dump() for item in items] == [ + { + "user_id": "alice", + "credential_type": "oauth2", + "expires_at": "2026-12-31T00:00:00+00:00", + "connected_at": "2026-01-01T00:00:00+00:00", + "updated_at": "2026-01-01T00:00:00+00:00", + }, + { + "user_id": "carol", + "credential_type": "byok", + "expires_at": None, + "connected_at": None, + "updated_at": "2026-02-01T00:00:00+00:00", + }, + ] + serialized = "".join(item.model_dump_json() for item in items) + assert "tok-alice" not in serialized + assert "sk-byok-carol" not in serialized + + @pytest.mark.asyncio async def test_purge_user_oauth_credentials_for_server_spares_byok_rows(): """Regression: the purge used to delete_many on server_id alone, wiping BYOK API keys that share diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py index ef424255f04..4311a69d465 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py @@ -2871,6 +2871,183 @@ def test_remove_stateful_session_tracking_drops_client_info(): assert session_id not in mcp_server._stateful_session_client_info +def _admin_terminate_fixture(mcp_server): + def auth_user(user_id: str): + return mcp_server.MCPAuthenticatedUser( + user_api_key_auth=UserAPIKeyAuth(api_key=f"key-{user_id}", user_id=user_id), + ) + + contexts = { + "alice-session-1": auth_user("alice"), + "alice-session-2": auth_user("alice"), + "bob-session-1": auth_user("bob"), + "anon-session-1": mcp_server.MCPAuthenticatedUser(user_api_key_auth=None), + "gone-session-1": auth_user("alice"), + } + transports = { + session_id: MagicMock(terminate=AsyncMock()) + for session_id in ("alice-session-1", "alice-session-2", "bob-session-1", "anon-session-1") + } + return contexts, transports + + +@pytest.mark.asyncio +async def test_terminate_mcp_gateway_sessions_by_user_closes_every_live_session_of_that_user(): + try: + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy._experimental.mcp_server.server import session_manager_stateful + except ImportError: + pytest.skip("MCP server not available") + + contexts, transports = _admin_terminate_fixture(mcp_server) + live_transports = dict(transports) + last_seen = {session_id: 100.0 for session_id in contexts} + locks = {session_id: asyncio.Lock() for session_id in contexts} + + with ( + patch.object( # test-quality-ok: the transport registry is a module-level singleton; the suite's only seam + session_manager_stateful, "_server_instances", live_transports + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_contexts, contexts, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_context_last_seen, last_seen, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_locks, locks, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_owners, {session_id: "owner" for session_id in contexts}, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_active_request_counts, {}, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_client_info, {}, clear=True + ), + ): + result = await mcp_server.terminate_mcp_gateway_sessions(user_id="alice") + + assert set(live_transports) == {"bob-session-1", "anon-session-1"} + assert set(mcp_server._stateful_session_auth_contexts) == {"bob-session-1", "anon-session-1", "gone-session-1"} + assert set(mcp_server._stateful_session_locks) == {"bob-session-1", "anon-session-1", "gone-session-1"} + assert set(mcp_server._stateful_session_owners) == {"bob-session-1", "anon-session-1", "gone-session-1"} + assert set(mcp_server._stateful_session_auth_context_last_seen) == { + "bob-session-1", + "anon-session-1", + "gone-session-1", + } + + transports["alice-session-1"].terminate.assert_awaited_once() + transports["alice-session-2"].terminate.assert_awaited_once() + transports["bob-session-1"].terminate.assert_not_awaited() + transports["anon-session-1"].terminate.assert_not_awaited() + assert result.terminated_sessions == 2 + assert sorted(session.session_id_prefix for session in result.sessions) == ["alice-se", "alice-se"] + assert {session.user_id for session in result.sessions} == {"alice"} + assert "key-alice" not in result.model_dump_json() + assert "alice-session-1" not in result.model_dump_json() + + +@pytest.mark.asyncio +async def test_terminate_mcp_gateway_sessions_prefix_and_user_must_both_match(): + try: + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy._experimental.mcp_server.server import session_manager_stateful + except ImportError: + pytest.skip("MCP server not available") + + contexts, transports = _admin_terminate_fixture(mcp_server) + live_transports = dict(transports) + + with ( + patch.object( # test-quality-ok: the transport registry is a module-level singleton; the suite's only seam + session_manager_stateful, "_server_instances", live_transports + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_contexts, contexts, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_client_info, {}, clear=True + ), + ): + mismatch = await mcp_server.terminate_mcp_gateway_sessions(session_id_prefix="alice-session-1", user_id="bob") + assert mismatch.terminated_sessions == 0 + assert set(live_transports) == set(transports) + + stale = await mcp_server.terminate_mcp_gateway_sessions(session_id_prefix="gone-session-1") + assert stale.terminated_sessions == 0 + + exact = await mcp_server.terminate_mcp_gateway_sessions(session_id_prefix="alice-session-1", user_id="alice") + assert exact.terminated_sessions == 1 + assert set(live_transports) == {"alice-session-2", "bob-session-1", "anon-session-1"} + + +@pytest.mark.asyncio +async def test_admin_terminated_session_id_gets_404_instead_of_a_fresh_stateless_session(): + """Once an admin closes a session, a client replaying its id must not be silently upgraded to a + new stateless session by the stale-header path; it gets 404 and has to initialize again.""" + try: + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy._experimental.mcp_server.server import session_manager_stateful + from starlette.types import Scope + except ImportError: + pytest.skip("MCP server not available") + + session_id = "admin-closed-session-1" + live_transports = {session_id: MagicMock(terminate=AsyncMock())} + contexts = { + session_id: mcp_server.MCPAuthenticatedUser( + user_api_key_auth=UserAPIKeyAuth(api_key="key-alice", user_id="alice"), + ) + } + + def scope_with_session_header() -> Scope: + return { + "type": "http", + "method": "POST", + "headers": [(b"content-type", b"application/json"), (b"mcp-session-id", session_id.encode())], + } + + try: + with ( + patch.object( # test-quality-ok: the transport registry is a module-level singleton; the suite's only seam + session_manager_stateful, "_server_instances", live_transports + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_contexts, contexts, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_client_info, {}, clear=True + ), + ): + await mcp_server.terminate_mcp_gateway_sessions(session_id_prefix=session_id) + + terminated_scope = scope_with_session_header() + send = AsyncMock() + handled = await mcp_server._handle_stale_mcp_session( + terminated_scope, AsyncMock(), send, session_manager_stateful + ) + + assert handled is True + statuses = [m["status"] for (m,), _ in send.await_args_list if m["type"] == "http.response.start"] + assert statuses == [404] + assert [k for k, _ in terminated_scope["headers"]] == [b"content-type", b"mcp-session-id"] + + unknown_scope = scope_with_session_header() + unknown_scope["headers"][1] = (b"mcp-session-id", b"never-seen-session") + assert ( + await mcp_server._handle_stale_mcp_session( + unknown_scope, AsyncMock(), AsyncMock(), session_manager_stateful + ) + is False + ) + assert [k for k, _ in unknown_scope["headers"]] == [b"content-type"] + finally: + mcp_server._admin_terminated_session_ids.clear() + + @pytest.mark.asyncio async def test_initialize_request_with_existing_session_tracks_new_session(): try: diff --git a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py index 7c874aff3df..2f33018599f 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py @@ -5136,6 +5136,261 @@ async def test_delete_mcp_oauth_user_credential_invalidates_when_record_already_ assert result.has_credential is False +def _make_admin_auth(role: LitellmUserRoles = LitellmUserRoles.PROXY_ADMIN) -> "UserAPIKeyAuth": + return UserAPIKeyAuth(api_key="sk-admin", user_id="admin-user", user_role=role) + + +@pytest.mark.asyncio +async def test_admin_revokes_another_users_byok_credential(): + """A proxy admin naming user_id deletes and cache-invalidates that user's stored key, not their own.""" + if not mgmt_endpoints.MCP_AVAILABLE: + pytest.skip("MCP module not installed") + + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_user_credential, + ) + + delete_mock = AsyncMock(return_value=None) + invalidate_mock = MagicMock() + with ( + patch( # test-quality-ok: endpoint test stubs the Prisma client lookup + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", + return_value=_make_prisma_client(), + ), + patch( # test-quality-ok: endpoint test stubs the credential row delete + "litellm.proxy.management_endpoints.mcp_management_endpoints.delete_user_credential", + new=delete_mock, + ), + patch.object( # test-quality-ok: the cache invalidator is module scoped; the suite's only seam + mcp_server, "_invalidate_byok_cred_cache", new=invalidate_mock + ), + ): + result = await delete_mcp_user_credential( + server_id="srv-byok-admin", + user_api_key_dict=_make_admin_auth(), + user_id="mallory", + ) + + delete_mock.assert_awaited_once() + assert delete_mock.await_args.args[1:] == ("mallory", "srv-byok-admin") + invalidate_mock.assert_called_once_with("mallory", "srv-byok-admin") + assert result.has_credential is False + + +@pytest.mark.asyncio +@pytest.mark.parametrize("role", [LitellmUserRoles.INTERNAL_USER, LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY]) +async def test_non_full_admin_cannot_revoke_another_users_byok_credential(role): + if not mgmt_endpoints.MCP_AVAILABLE: + pytest.skip("MCP module not installed") + + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_user_credential, + ) + + delete_mock = AsyncMock(return_value=None) + with ( + patch( # test-quality-ok: endpoint test stubs the Prisma client lookup + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", + return_value=_make_prisma_client(), + ), + patch( # test-quality-ok: endpoint test stubs the credential row delete + "litellm.proxy.management_endpoints.mcp_management_endpoints.delete_user_credential", + new=delete_mock, + ), + ): + with pytest.raises(HTTPException) as exc_info: + await delete_mcp_user_credential( + server_id="srv-byok-forbidden", + user_api_key_dict=_make_admin_auth(role), + user_id="mallory", + ) + + assert exc_info.value.status_code == 403 + delete_mock.assert_not_awaited() + + +@pytest.mark.asyncio +async def test_user_naming_themselves_still_deletes_own_byok_credential(): + if not mgmt_endpoints.MCP_AVAILABLE: + pytest.skip("MCP module not installed") + + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_user_credential, + ) + + delete_mock = AsyncMock(return_value=None) + with ( + patch( # test-quality-ok: endpoint test stubs the Prisma client lookup + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", + return_value=_make_prisma_client(), + ), + patch( # test-quality-ok: endpoint test stubs the credential row delete + "litellm.proxy.management_endpoints.mcp_management_endpoints.delete_user_credential", + new=delete_mock, + ), + patch.object( # test-quality-ok: the cache invalidator is module scoped; the suite's only seam + mcp_server, "_invalidate_byok_cred_cache", new=MagicMock() + ), + ): + await delete_mcp_user_credential( + server_id="srv-byok-self", + user_api_key_dict=_make_user_auth("user-self"), + user_id="user-self", + ) + + assert delete_mock.await_args.args[1:] == ("user-self", "srv-byok-self") + + +@pytest.mark.asyncio +async def test_admin_revokes_another_users_oauth_credential(): + """A proxy admin naming user_id reads, deletes, and cache-invalidates that user's OAuth token.""" + if not mgmt_endpoints.MCP_AVAILABLE: + pytest.skip("MCP module not installed") + + from litellm.proxy._experimental.mcp_server import mcp_server_manager as manager_module + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_oauth_user_credential, + ) + + get_mock = AsyncMock(return_value={"type": "oauth2", "access_token": "mallory-tok"}) + delete_mock = AsyncMock(return_value=None) + invalidate_mock = AsyncMock(return_value=None) + with ( + patch( # test-quality-ok: endpoint test stubs the Prisma client lookup + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", + return_value=_make_prisma_client(), + ), + patch( # test-quality-ok: endpoint test stubs the stored OAuth token read + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_user_oauth_credential", + new=get_mock, + ), + patch( # test-quality-ok: endpoint test stubs the credential row delete + "litellm.proxy.management_endpoints.mcp_management_endpoints.delete_user_credential", + new=delete_mock, + ), + patch.object( # test-quality-ok: the OAuth cache lives on the global manager; the suite's only seam + manager_module.global_mcp_server_manager, + "invalidate_user_oauth_token_cache", + new=invalidate_mock, + ), + ): + result = await delete_mcp_oauth_user_credential( + server_id="srv-oauth-admin", + user_api_key_dict=_make_admin_auth(), + user_id="mallory", + ) + + assert get_mock.await_args.args[1:] == ("mallory", "srv-oauth-admin") + assert delete_mock.await_args.args[1:] == ("mallory", "srv-oauth-admin") + invalidate_mock.assert_awaited_once_with("mallory", "srv-oauth-admin") + assert result.has_credential is False + + +@pytest.mark.asyncio +@pytest.mark.parametrize("role", [LitellmUserRoles.INTERNAL_USER, LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY]) +async def test_non_full_admin_cannot_revoke_another_users_oauth_credential(role): + if not mgmt_endpoints.MCP_AVAILABLE: + pytest.skip("MCP module not installed") + + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_oauth_user_credential, + ) + + get_mock = AsyncMock(return_value={"type": "oauth2", "access_token": "mallory-tok"}) + delete_mock = AsyncMock(return_value=None) + with ( + patch( # test-quality-ok: endpoint test stubs the Prisma client lookup + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", + return_value=_make_prisma_client(), + ), + patch( # test-quality-ok: endpoint test stubs the stored OAuth token read + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_user_oauth_credential", + new=get_mock, + ), + patch( # test-quality-ok: endpoint test stubs the credential row delete + "litellm.proxy.management_endpoints.mcp_management_endpoints.delete_user_credential", + new=delete_mock, + ), + ): + with pytest.raises(HTTPException) as exc_info: + await delete_mcp_oauth_user_credential( + server_id="srv-oauth-forbidden", + user_api_key_dict=_make_admin_auth(role), + user_id="mallory", + ) + + assert exc_info.value.status_code == 403 + get_mock.assert_not_awaited() + delete_mock.assert_not_awaited() + + +@pytest.mark.asyncio +@pytest.mark.parametrize("role", [LitellmUserRoles.PROXY_ADMIN, LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY]) +async def test_admin_lists_every_users_credential_for_a_server(role): + if not mgmt_endpoints.MCP_AVAILABLE: + pytest.skip("MCP module not installed") + + from litellm.proxy._types import MCPServerUserCredentialListItem + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + list_mcp_server_user_credentials, + ) + + items = ( + MCPServerUserCredentialListItem(user_id="alice", credential_type="byok", updated_at="2026-01-01T00:00:00"), + MCPServerUserCredentialListItem(user_id="bob", credential_type="oauth2", updated_at="2026-01-02T00:00:00"), + ) + list_mock = AsyncMock(return_value=items) + with ( + patch( # test-quality-ok: endpoint test stubs the Prisma client lookup + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", + return_value=_make_prisma_client(), + ), + patch( # test-quality-ok: endpoint test stubs the credential row listing + "litellm.proxy.management_endpoints.mcp_management_endpoints.list_server_user_credentials", + new=list_mock, + ), + ): + result = await list_mcp_server_user_credentials( + server_id="srv-list-admin", + user_api_key_dict=_make_admin_auth(role), + ) + + assert list_mock.await_args.args[1:] == ("srv-list-admin",) + assert [(item.user_id, item.credential_type) for item in result] == [("alice", "byok"), ("bob", "oauth2")] + + +@pytest.mark.asyncio +async def test_non_admin_cannot_list_a_servers_user_credentials(): + if not mgmt_endpoints.MCP_AVAILABLE: + pytest.skip("MCP module not installed") + + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + list_mcp_server_user_credentials, + ) + + list_mock = AsyncMock(return_value=()) + with ( + patch( # test-quality-ok: endpoint test stubs the Prisma client lookup + "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", + return_value=_make_prisma_client(), + ), + patch( # test-quality-ok: endpoint test stubs the credential row listing + "litellm.proxy.management_endpoints.mcp_management_endpoints.list_server_user_credentials", + new=list_mock, + ), + ): + with pytest.raises(HTTPException) as exc_info: + await list_mcp_server_user_credentials( + server_id="srv-list-forbidden", + user_api_key_dict=_make_user_auth("user-plain"), + ) + + assert exc_info.value.status_code == 403 + list_mock.assert_not_awaited() + + @pytest.mark.asyncio async def test_list_mcp_user_credentials_batch_server_fetch(): """list_mcp_user_credentials uses a single batch DB call, not N+1 queries.""" @@ -7321,3 +7576,146 @@ class TestGetMCPGatewaySessions: assert [(group.label, group.count) for group in result.by_client] == [("cursor", 1)] assert [(group.label, group.count) for group in result.by_user] == [("alice", 1)] assert "sk-live-secret" not in result.model_dump_json() + + +class TestDeleteMCPGatewaySessions: + @pytest.fixture(autouse=True) + def _forget_admin_terminated_ids(self): + from litellm.proxy._experimental.mcp_server import server as mcp_server + + yield + mcp_server._admin_terminated_session_ids.clear() + + @pytest.mark.asyncio + @pytest.mark.parametrize("role", [LitellmUserRoles.INTERNAL_USER, LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY]) + async def test_non_full_admin_forbidden_before_any_session_is_touched(self, role): + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_gateway_sessions, + ) + + session_id = "gateway-terminate-forbidden-1" + transport = MagicMock(terminate=AsyncMock()) + auth_user = mcp_server.MCPAuthenticatedUser( + user_api_key_auth=UserAPIKeyAuth(api_key="sk-live", user_id="alice"), + ) + with ( + patch.object( # test-quality-ok: the transport registry is a module-level singleton; the suite's only seam + mcp_server.session_manager_stateful, "_server_instances", {session_id: transport} + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_contexts, {session_id: auth_user}, clear=True + ), + ): + with pytest.raises(HTTPException) as exc_info: + await delete_mcp_gateway_sessions( + user_api_key_dict=generate_mock_user_api_key_auth(user_role=role), + session_id_prefix=session_id, + user_id=None, + ) + assert exc_info.value.status_code == 403 + transport.terminate.assert_not_awaited() + assert session_id in mcp_server._stateful_session_auth_contexts + + @pytest.mark.asyncio + async def test_requires_a_selector(self): + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_gateway_sessions, + ) + + with pytest.raises(HTTPException) as exc_info: + await delete_mcp_gateway_sessions( + user_api_key_dict=generate_mock_user_api_key_auth(user_role=LitellmUserRoles.PROXY_ADMIN), + session_id_prefix=None, + user_id=None, + ) + assert exc_info.value.status_code == 400 + + @pytest.mark.asyncio + async def test_admin_terminates_only_the_selected_session(self): + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_gateway_sessions, + ) + from litellm.types.mcp import MCPGatewaySessionsTerminateResponse + + target_id = "11111111-target-session" + other_id = "22222222-other-session" + target_transport = MagicMock(terminate=AsyncMock()) + other_transport = MagicMock(terminate=AsyncMock()) + transports = {target_id: target_transport, other_id: other_transport} + contexts = { + target_id: mcp_server.MCPAuthenticatedUser( + user_api_key_auth=UserAPIKeyAuth(api_key="sk-live-target", user_id="alice"), + ), + other_id: mcp_server.MCPAuthenticatedUser( + user_api_key_auth=UserAPIKeyAuth(api_key="sk-live-other", user_id="bob"), + ), + } + with ( + patch.object( # test-quality-ok: the transport registry is a module-level singleton; the suite's only seam + mcp_server.session_manager_stateful, "_server_instances", transports + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_contexts, contexts, clear=True + ), + ): + result = await delete_mcp_gateway_sessions( + user_api_key_dict=generate_mock_user_api_key_auth(user_role=LitellmUserRoles.PROXY_ADMIN), + session_id_prefix=target_id[:8], + user_id=None, + ) + assert target_id not in transports + assert other_id in transports + assert target_id not in mcp_server._stateful_session_auth_contexts + assert other_id in mcp_server._stateful_session_auth_contexts + + target_transport.terminate.assert_awaited_once() + other_transport.terminate.assert_not_awaited() + assert isinstance(result, MCPGatewaySessionsTerminateResponse) + assert result.terminated_sessions == 1 + assert [(s.session_id_prefix, s.user_id) for s in result.sessions] == [(target_id[:8], "alice")] + assert target_id not in result.model_dump_json() + assert "sk-live-target" not in result.model_dump_json() + + @pytest.mark.asyncio + async def test_admin_terminates_every_session_of_the_selected_user(self): + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy.management_endpoints.mcp_management_endpoints import ( + delete_mcp_gateway_sessions, + ) + + def auth_user(user_id: str): + return mcp_server.MCPAuthenticatedUser( + user_api_key_auth=UserAPIKeyAuth(api_key=f"sk-live-{user_id}", user_id=user_id), + ) + + transports = { + "bob-session-1": MagicMock(terminate=AsyncMock()), + "bob-session-2": MagicMock(terminate=AsyncMock()), + "alice-session-1": MagicMock(terminate=AsyncMock()), + } + contexts = { + "bob-session-1": auth_user("bob"), + "bob-session-2": auth_user("bob"), + "alice-session-1": auth_user("alice"), + } + with ( + patch.object( # test-quality-ok: the transport registry is a module-level singleton; the suite's only seam + mcp_server.session_manager_stateful, "_server_instances", transports + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_contexts, contexts, clear=True + ), + ): + result = await delete_mcp_gateway_sessions( + user_api_key_dict=generate_mock_user_api_key_auth(user_role=LitellmUserRoles.PROXY_ADMIN), + session_id_prefix=None, + user_id="bob", + ) + assert set(transports) == {"alice-session-1"} + assert set(mcp_server._stateful_session_auth_contexts) == {"alice-session-1"} + + assert result.terminated_sessions == 2 + assert {s.user_id for s in result.sessions} == {"bob"} + assert "sk-live-bob" not in result.model_dump_json() diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.integration.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.integration.test.tsx index 11328ff1a3d..f1ddf709038 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.integration.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.integration.test.tsx @@ -1,13 +1,15 @@ import React from "react"; import { render, screen, within } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; import { describe, it, expect, vi, beforeEach } from "vitest"; import { QueryClient, QueryClientProvider } from "@tanstack/react-query"; -import { MCPGatewaySessionsTab, formatIdleSeconds } from "./MCPGatewaySessionsTab"; +import { MCPGatewaySessionsTab, describeTerminateResult, formatIdleSeconds } from "./MCPGatewaySessionsTab"; import * as networking from "@/components/networking"; -import type { MCPGatewaySessionsResponse } from "@/components/mcp_tools/types"; +import type { MCPGatewaySessionsResponse, MCPGatewaySessionsTerminateResponse } from "@/components/mcp_tools/types"; vi.mock("@/components/networking", () => ({ fetchMCPGatewaySessions: vi.fn(), + terminateMCPGatewaySessions: vi.fn(), })); const REPORT: MCPGatewaySessionsResponse = { @@ -64,11 +66,11 @@ const REPORT: MCPGatewaySessionsResponse = { ], }; -const renderTab = () => { +const renderTab = ({ canTerminate = false }: { canTerminate?: boolean } = {}) => { const queryClient = new QueryClient({ defaultOptions: { queries: { retry: false, gcTime: 0 } } }); return render( - + , ); }; @@ -83,6 +85,17 @@ describe("formatIdleSeconds", () => { }); }); +describe("describeTerminateResult", () => { + it("pluralizes the session count and names the worker", () => { + expect(describeTerminateResult({ worker_pid: 9, terminated_sessions: 1, sessions: [] })).toBe( + "Disconnected 1 session on worker pid 9.", + ); + expect(describeTerminateResult({ worker_pid: 9, terminated_sessions: 0, sessions: [] })).toBe( + "Disconnected 0 sessions on worker pid 9.", + ); + }); +}); + describe("MCPGatewaySessionsTab", () => { beforeEach(() => { vi.clearAllMocks(); @@ -135,4 +148,73 @@ describe("MCPGatewaySessionsTab", () => { expect(alert).toHaveTextContent("Could not load live connections"); expect(alert).toHaveTextContent("Admin access required"); }); + + it("hides every disconnect control from a read-only admin", async () => { + vi.mocked(networking.fetchMCPGatewaySessions).mockResolvedValue(REPORT); + renderTab({ canTerminate: false }); + + await screen.findByRole("region", { name: "Live sessions" }); + expect(screen.queryByRole("button", { name: /^Disconnect/ })).not.toBeInTheDocument(); + }); + + it("disconnects one session by its displayed prefix after confirmation and refetches", async () => { + const user = userEvent.setup(); + const terminated: MCPGatewaySessionsTerminateResponse = { + worker_pid: 4242, + terminated_sessions: 1, + sessions: [REPORT.sessions[1]], + }; + vi.mocked(networking.fetchMCPGatewaySessions).mockResolvedValue(REPORT); + vi.mocked(networking.terminateMCPGatewaySessions).mockResolvedValue(terminated); + renderTab({ canTerminate: true }); + + await user.click(await screen.findByRole("button", { name: "Disconnect session bbbb2222" })); + expect(networking.terminateMCPGatewaySessions).not.toHaveBeenCalled(); + const dialog = await screen.findByRole("alertdialog"); + expect(dialog).toHaveTextContent("session bbbb2222"); + await user.click(within(dialog).getByRole("button", { name: "Disconnect" })); + + const status = await screen.findByText("Disconnected 1 session on worker pid 4242.", { exact: false }); + expect(status).toBeInTheDocument(); + expect(networking.terminateMCPGatewaySessions).toHaveBeenCalledWith("token", { session_id_prefix: "bbbb2222" }); + expect(networking.fetchMCPGatewaySessions).toHaveBeenCalledTimes(2); + }); + + it("disconnects every session of a user from the by-user table", async () => { + const user = userEvent.setup(); + vi.mocked(networking.fetchMCPGatewaySessions).mockResolvedValue(REPORT); + vi.mocked(networking.terminateMCPGatewaySessions).mockResolvedValue({ + worker_pid: 4242, + terminated_sessions: 2, + sessions: [REPORT.sessions[0], REPORT.sessions[1]], + }); + renderTab({ canTerminate: true }); + + const byUser = await screen.findByRole("region", { name: "Sessions by user" }); + expect(within(byUser).queryByRole("button", { name: /\(unknown\)/ })).not.toBeInTheDocument(); + await user.click(within(byUser).getByRole("button", { name: "Disconnect all sessions for user alice" })); + const dialog = await screen.findByRole("alertdialog"); + expect(dialog).toHaveTextContent("every live session opened by user alice"); + await user.click(within(dialog).getByRole("button", { name: "Disconnect" })); + + expect(await screen.findByText(/Disconnected 2 sessions on worker pid 4242\./)).toBeInTheDocument(); + expect(networking.terminateMCPGatewaySessions).toHaveBeenCalledWith("token", { user_id: "alice" }); + }); + + it("shows the API error when a disconnect is refused", async () => { + const user = userEvent.setup(); + vi.mocked(networking.fetchMCPGatewaySessions).mockResolvedValue(REPORT); + vi.mocked(networking.terminateMCPGatewaySessions).mockRejectedValue( + new Error("Proxy admin access required to terminate MCP gateway sessions."), + ); + renderTab({ canTerminate: true }); + + await user.click(await screen.findByRole("button", { name: "Disconnect session aaaa1111" })); + await user.click(within(await screen.findByRole("alertdialog")).getByRole("button", { name: "Disconnect" })); + + const alert = await screen.findByRole("alert"); + expect(alert).toHaveTextContent("Could not disconnect"); + expect(alert).toHaveTextContent("Proxy admin access required to terminate MCP gateway sessions."); + expect(screen.getByRole("region", { name: "Live sessions" })).toBeInTheDocument(); + }); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.tsx index 18f44d5d090..04a095f8792 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPGatewaySessionsTab.tsx @@ -1,14 +1,27 @@ "use client"; -import React from "react"; -import { useQuery } from "@tanstack/react-query"; -import { RefreshCw } from "lucide-react"; +import React, { useState } from "react"; +import { useMutation, useQuery, useQueryClient } from "@tanstack/react-query"; +import { RefreshCw, Unplug } from "lucide-react"; import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert"; +import { + AlertDialog, + AlertDialogContent, + AlertDialogDescription, + AlertDialogFooter, + AlertDialogHeader, + AlertDialogTitle, +} from "@/components/ui/alert-dialog"; import { Button } from "@/components/ui/button"; import { Table, TableBody, TableCell, TableHead, TableHeader, TableRow } from "@/components/ui/table"; import { UiLoadingSpinner } from "@/components/ui/ui-loading-spinner"; -import { fetchMCPGatewaySessions } from "@/components/networking"; -import type { MCPGatewaySessionGroupCount, MCPGatewaySessionsResponse } from "@/components/mcp_tools/types"; +import { fetchMCPGatewaySessions, terminateMCPGatewaySessions } from "@/components/networking"; +import type { + MCPGatewaySessionGroupCount, + MCPGatewaySessionSelector, + MCPGatewaySessionsResponse, + MCPGatewaySessionsTerminateResponse, +} from "@/components/mcp_tools/types"; import { createQueryKeys } from "@/app/(dashboard)/hooks/common/queryKeysFactory"; const mcpGatewaySessionKeys = createQueryKeys("mcpGatewaySessions"); @@ -28,6 +41,16 @@ function groupLabel(label: string | null): string { return label === "" ? '""' : label; } +export function describeSelector(selector: MCPGatewaySessionSelector): string { + if (selector.user_id !== undefined) return `every live session opened by user ${groupLabel(selector.user_id)}`; + return `session ${selector.session_id_prefix}`; +} + +export function describeTerminateResult(result: MCPGatewaySessionsTerminateResponse): string { + const noun = result.terminated_sessions === 1 ? "session" : "sessions"; + return `Disconnected ${result.terminated_sessions} ${noun} on worker pid ${result.worker_pid}.`; +} + function StatCard({ label, value }: { label: string; value: number }) { return (

@@ -37,14 +60,37 @@ function StatCard({ label, value }: { label: string; value: number }) { ); } +function DisconnectUserButton({ + userId, + onDisconnectUser, +}: { + userId: string | null; + onDisconnectUser: (userId: string) => void; +}) { + if (userId === null || userId === "") return null; + return ( + + ); +} + function GroupCountTable({ title, groups, labelHeader, + onDisconnectUser, }: { title: string; groups: MCPGatewaySessionGroupCount[]; labelHeader: string; + onDisconnectUser?: (userId: string) => void; }) { return (
@@ -54,6 +100,7 @@ function GroupCountTable({ {labelHeader} Sessions + {onDisconnectUser ? Actions : null} @@ -61,6 +108,11 @@ function GroupCountTable({ {groupLabel(group.label)} {group.count} + {onDisconnectUser ? ( + + + + ) : null} ))} @@ -73,10 +125,12 @@ function SessionsBody({ data, error, isLoading, + onDisconnect, }: { data: MCPGatewaySessionsResponse | undefined; error: Error | null; isLoading: boolean; + onDisconnect: ((selector: MCPGatewaySessionSelector) => void) | null; }) { if (isLoading) { return ( @@ -117,7 +171,12 @@ function SessionsBody({
- + onDisconnect({ user_id: userId }) : undefined} + />

@@ -134,11 +193,12 @@ function SessionsBody({ Client IP Idle In flight + {onDisconnect ? Actions : null} - {data.sessions.map((session) => ( - + {data.sessions.map((session, index) => ( + {session.session_id_prefix} {session.client_name === null ? ( @@ -169,6 +229,19 @@ function SessionsBody({ {session.client_ip || "-"} {formatIdleSeconds(session.idle_seconds)} {session.in_flight_requests} + {onDisconnect ? ( + + + + ) : null} ))} @@ -180,9 +253,12 @@ function SessionsBody({ interface MCPGatewaySessionsTabProps { accessToken: string | null; + canTerminate: boolean; } -export function MCPGatewaySessionsTab({ accessToken }: MCPGatewaySessionsTabProps) { +export function MCPGatewaySessionsTab({ accessToken, canTerminate }: MCPGatewaySessionsTabProps) { + const queryClient = useQueryClient(); + const [pendingSelector, setPendingSelector] = useState(null); const queryOptions = { queryKey: mcpGatewaySessionKeys.lists(), queryFn: () => fetchMCPGatewaySessions(accessToken!), @@ -190,6 +266,15 @@ export function MCPGatewaySessionsTab({ accessToken }: MCPGatewaySessionsTabProp refetchInterval: REFETCH_INTERVAL_MS, }; const { data, error, isLoading, isFetching, refetch } = useQuery(queryOptions); + const terminate = useMutation({ + mutationFn: (selector) => terminateMCPGatewaySessions(accessToken!, selector), + onSettled: () => queryClient.invalidateQueries({ queryKey: mcpGatewaySessionKeys.lists() }), + }); + const confirmDisconnect = () => { + if (pendingSelector === null) return; + terminate.mutate(pendingSelector); + setPendingSelector(null); + }; return (
@@ -214,7 +299,48 @@ export function MCPGatewaySessionsTab({ accessToken }: MCPGatewaySessionsTabProp
- + {terminate.isError ? ( + + Could not disconnect + {terminate.error.message} + + ) : null} + {terminate.isSuccess ? ( + + Disconnected + + {describeTerminateResult(terminate.data)} Clients holding those sessions must send a new initialize request, + which re-runs authentication. Sessions on other proxy workers are not affected. + + + ) : null} + + + + !open && setPendingSelector(null)}> + + + Disconnect MCP session + + {pendingSelector ? `This force-closes ${describeSelector(pendingSelector)} on this proxy worker. ` : ""} + In-flight requests fail and the client must initialize again before it can call tools. + + + + + + + +

); } diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.integration.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.integration.test.tsx new file mode 100644 index 00000000000..a20dd032d33 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.integration.test.tsx @@ -0,0 +1,102 @@ +import React from "react"; +import { render, screen, within } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; +import { describe, it, expect, vi, beforeEach } from "vitest"; +import { QueryClient, QueryClientProvider } from "@tanstack/react-query"; +import { MCPServerUserCredentialsPanel } from "./MCPServerUserCredentialsPanel"; +import * as networking from "@/components/networking"; +import type { MCPServerUserCredentialListItem } from "@/components/mcp_tools/types"; + +vi.mock("@/components/networking", () => ({ + fetchMCPServerUserCredentials: vi.fn(), + revokeMCPServerUserCredential: vi.fn(), +})); + +const ITEMS: MCPServerUserCredentialListItem[] = [ + { + user_id: "alice", + credential_type: "oauth2", + expires_at: "2026-12-31T00:00:00+00:00", + connected_at: "2026-01-01T00:00:00+00:00", + updated_at: "2026-01-01T00:00:00+00:00", + }, + { + user_id: "carol", + credential_type: "byok", + expires_at: null, + connected_at: null, + updated_at: "2026-02-01T00:00:00+00:00", + }, +]; + +const renderPanel = ({ canRevoke = false }: { canRevoke?: boolean } = {}) => { + const queryClient = new QueryClient({ defaultOptions: { queries: { retry: false, gcTime: 0 } } }); + return render( + + + , + ); +}; + +describe("MCPServerUserCredentialsPanel", () => { + beforeEach(() => { + vi.clearAllMocks(); + }); + + it("lists each user's credential type without a revoke control for a read-only admin", async () => { + vi.mocked(networking.fetchMCPServerUserCredentials).mockResolvedValue(ITEMS); + renderPanel({ canRevoke: false }); + + const table = await screen.findByRole("region", { name: "Stored user credentials" }); + expect(within(table).getByRole("row", { name: /alice/ })).toHaveTextContent("OAuth2"); + expect(within(table).getByRole("row", { name: /carol/ })).toHaveTextContent("BYOK API key"); + expect(screen.queryByRole("button", { name: /^Revoke credential/ })).not.toBeInTheDocument(); + expect(networking.fetchMCPServerUserCredentials).toHaveBeenCalledWith("token", "srv-1"); + }); + + it("revokes the selected user's credential through the route for its type and refetches", async () => { + const user = userEvent.setup(); + vi.mocked(networking.fetchMCPServerUserCredentials).mockResolvedValueOnce(ITEMS).mockResolvedValueOnce([ITEMS[1]]); + vi.mocked(networking.revokeMCPServerUserCredential).mockResolvedValue(undefined); + renderPanel({ canRevoke: true }); + + await user.click(await screen.findByRole("button", { name: "Revoke credential for user alice" })); + expect(networking.revokeMCPServerUserCredential).not.toHaveBeenCalled(); + const dialog = await screen.findByRole("alertdialog"); + expect(dialog).toHaveTextContent("OAuth2 credential stored for user alice"); + await user.click(within(dialog).getByRole("button", { name: "Revoke" })); + + expect(await screen.findByText(/OAuth2 credential for user alice was deleted/)).toBeInTheDocument(); + expect(networking.revokeMCPServerUserCredential).toHaveBeenCalledWith("token", "srv-1", "alice", "oauth2"); + const table = await screen.findByRole("region", { name: "Stored user credentials" }); + expect(within(table).queryByRole("row", { name: /alice/ })).not.toBeInTheDocument(); + expect(within(table).getByRole("row", { name: /carol/ })).toBeInTheDocument(); + }); + + it("shows the API error when a revoke is refused and keeps the list", async () => { + const user = userEvent.setup(); + vi.mocked(networking.fetchMCPServerUserCredentials).mockResolvedValue(ITEMS); + vi.mocked(networking.revokeMCPServerUserCredential).mockRejectedValue( + new Error("Proxy admin access required to revoke another user's MCP credential."), + ); + renderPanel({ canRevoke: true }); + + await user.click(await screen.findByRole("button", { name: "Revoke credential for user carol" })); + await user.click(within(await screen.findByRole("alertdialog")).getByRole("button", { name: "Revoke" })); + + const alert = await screen.findByRole("alert"); + expect(alert).toHaveTextContent("Could not revoke credential"); + expect(alert).toHaveTextContent("Proxy admin access required to revoke another user's MCP credential."); + expect(networking.revokeMCPServerUserCredential).toHaveBeenCalledWith("token", "srv-1", "carol", "byok"); + expect(screen.getByRole("region", { name: "Stored user credentials" })).toBeInTheDocument(); + }); + + it("shows the API error when the list cannot be loaded", async () => { + vi.mocked(networking.fetchMCPServerUserCredentials).mockRejectedValue(new Error("Admin access required")); + renderPanel(); + + const alert = await screen.findByRole("alert"); + expect(alert).toHaveTextContent("Could not load user credentials"); + expect(alert).toHaveTextContent("Admin access required"); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.tsx new file mode 100644 index 00000000000..20b679450c3 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/MCPServerUserCredentialsPanel.tsx @@ -0,0 +1,212 @@ +"use client"; + +import React, { useState } from "react"; +import { useMutation, useQuery, useQueryClient } from "@tanstack/react-query"; +import { RefreshCw, ShieldOff } from "lucide-react"; +import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert"; +import { + AlertDialog, + AlertDialogContent, + AlertDialogDescription, + AlertDialogFooter, + AlertDialogHeader, + AlertDialogTitle, +} from "@/components/ui/alert-dialog"; +import { Badge } from "@/components/ui/badge"; +import { Button } from "@/components/ui/button"; +import { Table, TableBody, TableCell, TableHead, TableHeader, TableRow } from "@/components/ui/table"; +import { UiLoadingSpinner } from "@/components/ui/ui-loading-spinner"; +import { fetchMCPServerUserCredentials, revokeMCPServerUserCredential } from "@/components/networking"; +import type { MCPServerUserCredentialListItem } from "@/components/mcp_tools/types"; +import { createQueryKeys } from "@/app/(dashboard)/hooks/common/queryKeysFactory"; + +const mcpServerUserCredentialKeys = createQueryKeys("mcpServerUserCredentials"); + +export function credentialTypeLabel(credentialType: MCPServerUserCredentialListItem["credential_type"]): string { + return credentialType === "oauth2" ? "OAuth2" : "BYOK API key"; +} + +export function formatTimestamp(value: string | null): string { + if (value === null) return "-"; + const parsed = new Date(value); + return Number.isNaN(parsed.getTime()) ? value : parsed.toLocaleString(); +} + +function CredentialsBody({ + items, + error, + isLoading, + onRevoke, +}: { + items: MCPServerUserCredentialListItem[] | undefined; + error: Error | null; + isLoading: boolean; + onRevoke: ((item: MCPServerUserCredentialListItem) => void) | null; +}) { + if (isLoading) { + return ( +
+ +

Loading user credentials...

+
+ ); + } + if (error) { + return ( + + Could not load user credentials + {error.message} + + ); + } + if (!items) return null; + if (items.length === 0) { + return ( +
+

No user has a stored credential for this server.

+
+ ); + } + return ( +
+ + + + User + Type + Connected + Expires + Updated + {onRevoke ? Actions : null} + + + + {items.map((item) => ( + + {item.user_id} + + {credentialTypeLabel(item.credential_type)} + + {formatTimestamp(item.connected_at)} + {formatTimestamp(item.expires_at)} + {formatTimestamp(item.updated_at)} + {onRevoke ? ( + + + + ) : null} + + ))} + +
+
+ ); +} + +interface MCPServerUserCredentialsPanelProps { + serverId: string; + accessToken: string | null; + canRevoke: boolean; +} + +export function MCPServerUserCredentialsPanel({ + serverId, + accessToken, + canRevoke, +}: MCPServerUserCredentialsPanelProps) { + const queryClient = useQueryClient(); + const [pendingItem, setPendingItem] = useState(null); + const queryKey = mcpServerUserCredentialKeys.detail(serverId); + const { data, error, isLoading, isFetching, refetch } = useQuery({ + queryKey, + queryFn: () => fetchMCPServerUserCredentials(accessToken!, serverId), + enabled: !!accessToken, + }); + const revoke = useMutation({ + mutationFn: (item) => revokeMCPServerUserCredential(accessToken!, serverId, item.user_id, item.credential_type), + onSettled: () => queryClient.invalidateQueries({ queryKey }), + }); + const confirmRevoke = () => { + if (pendingItem === null) return; + revoke.mutate(pendingItem); + setPendingItem(null); + }; + + return ( +
+
+
+

User Credentials

+

+ Per-user OAuth2 tokens and BYOK API keys stored for this server. Revoking one deletes it from the database + and clears the cached copy, so the user must connect again before the gateway will call this server for + them. +

+
+ +
+ + {revoke.isError ? ( + + Could not revoke credential + {revoke.error.message} + + ) : null} + {revoke.isSuccess ? ( + + Credential revoked + + The stored {credentialTypeLabel(revoke.variables.credential_type)} credential for user{" "} + {revoke.variables.user_id} was deleted. + + + ) : null} + + + + !open && setPendingItem(null)}> + + + Revoke stored credential + + {pendingItem + ? `This deletes the ${credentialTypeLabel(pendingItem.credential_type)} credential stored for user ${pendingItem.user_id}. ` + : ""} + Their next MCP request to this server fails until they connect again. + + + + + + + + +
+ ); +} + +export default MCPServerUserCredentialsPanel; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx index 475392620d4..c23b48ef672 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx @@ -9,7 +9,9 @@ import { MCPServer, handleTransport, handleAuth } from "@/components/mcp_tools/t // TODO: Move Tools viewer from index file import { MCPToolsViewer } from "."; import MCPServerEdit, { EDIT_OAUTH_UI_STATE_KEY } from "./mcp_server_edit"; +import { MCPServerUserCredentialsPanel } from "./MCPServerUserCredentialsPanel"; import { getSecureItem } from "@/utils/secureStorage"; +import { isProxyAdminRole, isProxyAdminTierRole } from "@/utils/roles"; import MCPServerCostDisplay from "./mcp_server_cost_display"; import { getMaskedAndFullUrl } from "./utils"; import { copyToClipboard as utilCopyToClipboard } from "@/utils/dataUtils"; @@ -63,6 +65,8 @@ export const MCPServerView: React.FC = ({ const [showFullUrl, setShowFullUrl] = useState(false); const [copiedStates, setCopiedStates] = useState>({}); const [selectedTabIndex, setSelectedTabIndex] = useState(returningFromEditOAuth ? 2 : initialTabIndex); + const canViewUserCredentials = userRole !== null && isProxyAdminTierRole(userRole); + const canRevokeUserCredentials = userRole !== null && isProxyAdminRole(userRole); const handleSuccess = (updated: MCPServer) => { setEditing(false); @@ -142,6 +146,11 @@ export const MCPServerView: React.FC = ({ Settings )} + {canViewUserCredentials && ( + + User Credentials + + )} {/* Overview Panel */} @@ -387,6 +396,18 @@ export const MCPServerView: React.FC = ({ )} + + {canViewUserCredentials && ( + + + + + + )}
); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx index 00d79022103..3a197786774 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx @@ -1,4 +1,4 @@ -import { isAdminRole, isProxyAdminTierRole } from "@/utils/roles"; +import { isAdminRole, isProxyAdminRole, isProxyAdminTierRole } from "@/utils/roles"; import { CircleHelp, Search } from "lucide-react"; import { Badge } from "@/components/ui/badge"; import { Button } from "@/components/ui/button"; @@ -755,7 +755,7 @@ const MCPServers: React.FC = ({ accessToken, userRole, userID }) )} {isProxyAdminTierRole(userRole) && ( - + )} diff --git a/ui/litellm-dashboard/src/components/mcp_tools/types.tsx b/ui/litellm-dashboard/src/components/mcp_tools/types.tsx index fe04eb4969b..2f6a3f1d0d9 100644 --- a/ui/litellm-dashboard/src/components/mcp_tools/types.tsx +++ b/ui/litellm-dashboard/src/components/mcp_tools/types.tsx @@ -587,3 +587,23 @@ export interface MCPGatewaySessionsResponse { by_user: MCPGatewaySessionGroupCount[]; sessions: MCPGatewaySession[]; } + +export interface MCPGatewaySessionsTerminateResponse { + worker_pid: number; + terminated_sessions: number; + sessions: MCPGatewaySession[]; +} + +export type MCPGatewaySessionSelector = + | { session_id_prefix: string; user_id?: undefined } + | { user_id: string; session_id_prefix?: undefined }; + +export type MCPServerUserCredentialType = "oauth2" | "byok"; + +export interface MCPServerUserCredentialListItem { + user_id: string; + credential_type: MCPServerUserCredentialType; + expires_at: string | null; + connected_at: string | null; + updated_at: string; +} diff --git a/ui/litellm-dashboard/src/components/networking.tsx b/ui/litellm-dashboard/src/components/networking.tsx index a8ea1b66488..9381e2c07c9 100644 --- a/ui/litellm-dashboard/src/components/networking.tsx +++ b/ui/litellm-dashboard/src/components/networking.tsx @@ -97,7 +97,14 @@ import type { ModelBudgetUsage, ModelMaxBudget } from "./key_team_helpers/ModelM import type { ObjectPermission } from "./object_permission_types"; import type { components } from "@/lib/http/schema"; import { jsonFields } from "./common_components/check_openapi_schema"; -import type { MCPGatewaySessionsResponse, MCPUserEnvVarsStatus } from "./mcp_tools/types"; +import type { + MCPGatewaySessionSelector, + MCPGatewaySessionsResponse, + MCPGatewaySessionsTerminateResponse, + MCPServerUserCredentialListItem, + MCPServerUserCredentialType, + MCPUserEnvVarsStatus, +} from "./mcp_tools/types"; import type { CoordinationRedisSettings, CoordinationRedisSettingsResponse, @@ -5112,6 +5119,33 @@ export const fetchMCPSubmissions = async (accessToken: string) => { export const fetchMCPGatewaySessions = async (accessToken: string): Promise => apiClient.get(`/v1/mcp/sessions`, { accessToken }); +export const terminateMCPGatewaySessions = async ( + accessToken: string, + selector: MCPGatewaySessionSelector, +): Promise => + apiClient.delete(`/v1/mcp/sessions`, { accessToken, query: { ...selector } }); + +export const fetchMCPServerUserCredentials = async ( + accessToken: string, + serverId: string, +): Promise => + apiClient.get(`/v1/mcp/server/${encodeURIComponent(serverId)}/user-credentials`, { + accessToken, + }); + +export const revokeMCPServerUserCredential = async ( + accessToken: string, + serverId: string, + userId: string, + credentialType: MCPServerUserCredentialType, +): Promise => { + const route = credentialType === "oauth2" ? "oauth-user-credential" : "user-credential"; + await apiClient.delete(`/v1/mcp/server/${encodeURIComponent(serverId)}/${route}`, { + accessToken, + query: { user_id: userId }, + }); +}; + export const approveMCPServer = async (accessToken: string, serverId: string) => { try { const url = (proxyBaseUrl ? `${proxyBaseUrl}` : "") + `/v1/mcp/server/${encodeURIComponent(serverId)}/approve`; diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index f1779c7cb5a..43e1e16125c 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -18902,7 +18902,7 @@ export interface paths { post: operations["store_mcp_oauth_user_credential_v1_mcp_server__server_id__oauth_user_credential_post"]; /** * Delete Mcp Oauth User Credential - * @description Revoke the calling user's stored OAuth2 token for an MCP server + * @description Revoke the calling user's stored OAuth2 token for an MCP server. A proxy admin may pass user_id to revoke another user's stored token. */ delete: operations["delete_mcp_oauth_user_credential_v1_mcp_server__server_id__oauth_user_credential_delete"]; options?: never; @@ -18966,7 +18966,7 @@ export interface paths { post: operations["store_mcp_user_credential_v1_mcp_server__server_id__user_credential_post"]; /** * Delete Mcp User Credential - * @description Delete the calling user's stored API key for a BYOK MCP server + * @description Delete the calling user's stored API key for a BYOK MCP server. A proxy admin may pass user_id to revoke another user's stored key. */ delete: operations["delete_mcp_user_credential_v1_mcp_server__server_id__user_credential_delete"]; options?: never; @@ -18974,6 +18974,26 @@ export interface paths { patch?: never; trace?: never; }; + "/v1/mcp/server/{server_id}/user-credentials": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + /** + * List Mcp Server User Credentials + * @description List every user's stored BYOK or OAuth2 credential for an MCP server (admin only, no secrets) + */ + get: operations["list_mcp_server_user_credentials_v1_mcp_server__server_id__user_credentials_get"]; + put?: never; + post?: never; + delete?: never; + options?: never; + head?: never; + patch?: never; + trace?: never; + }; "/v1/mcp/server/{server_id}/user-env-vars": { parameters: { query?: never; @@ -19016,7 +19036,11 @@ export interface paths { get: operations["get_mcp_gateway_sessions_v1_mcp_sessions_get"]; put?: never; post?: never; - delete?: never; + /** + * Delete Mcp Gateway Sessions + * @description Force-close live stateful MCP gateway sessions on this proxy worker, selected by session id prefix and/or by the LiteLLM user that opened them (proxy admin only). + */ + delete: operations["delete_mcp_gateway_sessions_v1_mcp_sessions_delete"]; options?: never; head?: never; patch?: never; @@ -32383,6 +32407,18 @@ export interface components { /** Worker Pid */ worker_pid: number; }; + /** + * MCPGatewaySessionsTerminateResponse + * @description Stateful sessions an administrator force-closed on this proxy worker. + */ + MCPGatewaySessionsTerminateResponse: { + /** Sessions */ + sessions?: components["schemas"]["MCPGatewaySession"][]; + /** Terminated Sessions */ + terminated_sessions: number; + /** Worker Pid */ + worker_pid: number; + }; /** * MCPOAuthUserCredentialRequest * @description Stores a user's OAuth2 token for an OpenAPI MCP server. @@ -32487,6 +32523,25 @@ export interface components { [key: string]: unknown; }; }; + /** + * MCPServerUserCredentialListItem + * @description One user's stored credential for an MCP server, as an admin sees it. Never carries the secret. + */ + MCPServerUserCredentialListItem: { + /** Connected At */ + connected_at?: string | null; + /** + * Credential Type + * @enum {string} + */ + credential_type: "oauth2" | "byok"; + /** Expires At */ + expires_at?: string | null; + /** Updated At */ + updated_at: string; + /** User Id */ + user_id: string; + }; /** MCPSubmissionsSummary */ MCPSubmissionsSummary: { /** Active */ @@ -65109,7 +65164,9 @@ export interface operations { }; delete_mcp_oauth_user_credential_v1_mcp_server__server_id__oauth_user_credential_delete: { parameters: { - query?: never; + query?: { + user_id?: string | null; + }; header?: never; path: { server_id: string; @@ -65241,7 +65298,9 @@ export interface operations { }; delete_mcp_user_credential_v1_mcp_server__server_id__user_credential_delete: { parameters: { - query?: never; + query?: { + user_id?: string | null; + }; header?: never; path: { server_id: string; @@ -65270,6 +65329,37 @@ export interface operations { }; }; }; + list_mcp_server_user_credentials_v1_mcp_server__server_id__user_credentials_get: { + parameters: { + query?: never; + header?: never; + path: { + server_id: string; + }; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["MCPServerUserCredentialListItem"][]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; get_mcp_user_env_vars_v1_mcp_server__server_id__user_env_vars_get: { parameters: { query?: never; @@ -65387,6 +65477,38 @@ export interface operations { }; }; }; + delete_mcp_gateway_sessions_v1_mcp_sessions_delete: { + parameters: { + query?: { + session_id_prefix?: string | null; + user_id?: string | null; + }; + header?: never; + path?: never; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["MCPGatewaySessionsTerminateResponse"]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; get_mcp_tools_v1_mcp_tools_get: { parameters: { query?: never; From aea33b4b507d139bc8b5a1d93aa055205602a67f Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 01:37:12 +0000 Subject: [PATCH 067/141] fix(ui): hide MCP disconnect and revoke controls from view-only admin sessions The auth hook normalizes proxy_admin_viewer to Admin for page access, so the role check alone let a view-only admin see Disconnect and Revoke buttons that the backend refuses with 403. Thread isViewOnly from useAuthorized into the MCP servers page and gate both mutation controls on it Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../_components/mcp_server_view.test.tsx | 62 +++++++++++++++---- .../_components/mcp_server_view.tsx | 4 +- .../_components/mcp_servers.test.tsx | 48 ++++++++++++++ .../mcp-servers/_components/mcp_servers.tsx | 8 ++- .../src/app/(dashboard)/mcp-servers/page.tsx | 4 +- .../src/components/mcp_tools/types.tsx | 1 + 6 files changed, 110 insertions(+), 17 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.test.tsx index 02d168bf7f4..da564f23de5 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.test.tsx @@ -1,7 +1,9 @@ -import { render, screen } from "@testing-library/react"; +import { render, screen, within } from "@testing-library/react"; import userEvent from "@testing-library/user-event"; import { describe, it, expect, vi, beforeEach } from "vitest"; +import { QueryClient, QueryClientProvider } from "@tanstack/react-query"; import { MCPServerView } from "./mcp_server_view"; +import * as networking from "@/components/networking"; import type { MCPServer } from "@/components/mcp_tools/types"; vi.mock(".", () => ({ @@ -13,6 +15,12 @@ vi.mock("./mcp_server_edit", () => ({ EDIT_OAUTH_UI_STATE_KEY: "litellm-mcp-oauth-edit-state", })); +vi.mock("@/components/networking", async (importOriginal) => ({ + ...(await importOriginal()), + fetchMCPServerUserCredentials: vi.fn(), + revokeMCPServerUserCredential: vi.fn(), +})); + const baseServer = { server_id: "srv-1", server_name: "demo server", @@ -25,19 +33,38 @@ const baseServer = { const renderView = (overrides: Partial = {}, props: Record = {}) => render( - , + + + , ); +const openUserCredentials = async (props: Record) => { + vi.mocked(networking.fetchMCPServerUserCredentials).mockResolvedValue([ + { + user_id: "alice", + credential_type: "byok", + expires_at: null, + connected_at: null, + updated_at: "2026-01-01T00:00:00+00:00", + }, + ]); + renderView({}, props); + await userEvent.click(screen.getByRole("tab", { name: "User Credentials" })); + return within(await screen.findByRole("region", { name: "Stored user credentials" })).getByRole("row", { + name: /alice/, + }); +}; + describe("MCPServerView", () => { beforeEach(() => { vi.clearAllMocks(); @@ -149,4 +176,15 @@ describe("MCPServerView", () => { expect(await screen.findByText("All tools enabled")).toBeInTheDocument(); }); + + it("lets a full admin revoke a stored user credential", async () => { + const row = await openUserCredentials({}); + expect(within(row).getByRole("button", { name: "Revoke credential for user alice" })).toBeInTheDocument(); + }); + + it("shows stored credentials to a view-only admin session without a revoke control", async () => { + const row = await openUserCredentials({ isViewOnly: true }); + expect(row).toHaveTextContent("BYOK API key"); + expect(within(row).queryByRole("button", { name: /^Revoke credential/ })).not.toBeInTheDocument(); + }); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx index c23b48ef672..a346c7d986b 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_server_view.tsx @@ -25,6 +25,7 @@ interface MCPServerViewProps { accessToken: string | null; userRole: string | null; userID: string | null; + isViewOnly?: boolean; availableAccessGroups: string[]; initialTabIndex?: number; } @@ -55,6 +56,7 @@ export const MCPServerView: React.FC = ({ accessToken, userRole, userID, + isViewOnly = false, availableAccessGroups, initialTabIndex = 0, }) => { @@ -66,7 +68,7 @@ export const MCPServerView: React.FC = ({ const [copiedStates, setCopiedStates] = useState>({}); const [selectedTabIndex, setSelectedTabIndex] = useState(returningFromEditOAuth ? 2 : initialTabIndex); const canViewUserCredentials = userRole !== null && isProxyAdminTierRole(userRole); - const canRevokeUserCredentials = userRole !== null && isProxyAdminRole(userRole); + const canRevokeUserCredentials = userRole !== null && isProxyAdminRole(userRole) && !isViewOnly; const handleSuccess = (updated: MCPServer) => { setEditing(false); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.test.tsx index 8abb8855e3d..1394a923174 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.test.tsx @@ -17,6 +17,8 @@ vi.mock("@/components/networking", () => ({ updateConfigFieldSetting: vi.fn().mockResolvedValue(undefined), deleteConfigFieldSetting: vi.fn().mockResolvedValue(undefined), listMCPUserEnvVarStatus: vi.fn().mockResolvedValue([]), + fetchMCPGatewaySessions: vi.fn(), + terminateMCPGatewaySessions: vi.fn(), })); const createQueryClient = () => @@ -400,4 +402,50 @@ describe("MCPServers", () => { // The server list refresh must NOT trigger a second health check expect(networking.fetchMCPServerHealth).toHaveBeenCalledTimes(1); }); + + const liveSessionsReport = { + worker_pid: 4242, + total_sessions: 1, + by_client: [{ label: "claude-code", count: 1 }], + by_user: [{ label: "alice", count: 1 }], + sessions: [ + { + session_id_prefix: "aaaa1111", + client_name: "claude-code", + client_version: "1.0.0", + user_id: "alice", + user_email: "alice@example.com", + key_alias: "alice-key", + team_id: null, + team_alias: null, + client_ip: "10.0.0.1", + idle_seconds: 5, + in_flight_requests: 0, + }, + ], + }; + + const openLiveConnections = async (props: { isViewOnly?: boolean }) => { + vi.mocked(networking.fetchMCPServers).mockResolvedValue([]); + vi.mocked(networking.fetchMCPGatewaySessions).mockResolvedValue(liveSessionsReport); + render( + + + , + ); + await userEvent.click(await screen.findByRole("tab", { name: "Live Connections" })); + return within(await screen.findByRole("region", { name: "Live sessions" })).getByRole("row", { name: /aaaa1111/ }); + }; + + it("lets a full admin disconnect a live session", async () => { + const row = await openLiveConnections({ isViewOnly: false }); + expect(within(row).getByRole("button", { name: "Disconnect session aaaa1111" })).toBeInTheDocument(); + }); + + it("shows live sessions to a view-only admin session without any disconnect control", async () => { + const row = await openLiveConnections({ isViewOnly: true }); + expect(row).toHaveTextContent("alice@example.com"); + expect(within(row).queryByRole("button", { name: /^Disconnect/ })).not.toBeInTheDocument(); + expect(screen.queryByRole("button", { name: /^Disconnect all/ })).not.toBeInTheDocument(); + }); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx index 3a197786774..aa0a031c55c 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_servers.tsx @@ -109,7 +109,7 @@ const readToolsOAuthServerId = (): string | null => { } }; -const MCPServers: React.FC = ({ accessToken, userRole, userID }) => { +const MCPServers: React.FC = ({ accessToken, userRole, userID, isViewOnly = false }) => { const { data: mcpServers, isLoading: isLoadingServers, refetch } = useMCPServers(); // Fetch health status for all servers @@ -578,6 +578,7 @@ const MCPServers: React.FC = ({ accessToken, userRole, userID }) accessToken={accessToken} userID={userID} userRole={userRole} + isViewOnly={isViewOnly} availableAccessGroups={uniqueMcpAccessGroups} initialTabIndex={selectedServerId === toolsTabServerId ? 1 : 0} /> @@ -755,7 +756,10 @@ const MCPServers: React.FC = ({ accessToken, userRole, userID }) )} {isProxyAdminTierRole(userRole) && ( - + )} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/page.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/page.tsx index 462c48360cd..c297dcb8d33 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/page.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/page.tsx @@ -4,6 +4,6 @@ import { MCPServers } from "./_components"; import useAuthorized from "@/app/(dashboard)/hooks/useAuthorized"; export default function McpServers() { - const { accessToken, userRole, userId } = useAuthorized(); - return ; + const { accessToken, userRole, userId, isViewOnly } = useAuthorized(); + return ; } diff --git a/ui/litellm-dashboard/src/components/mcp_tools/types.tsx b/ui/litellm-dashboard/src/components/mcp_tools/types.tsx index 2f6a3f1d0d9..f009429693b 100644 --- a/ui/litellm-dashboard/src/components/mcp_tools/types.tsx +++ b/ui/litellm-dashboard/src/components/mcp_tools/types.tsx @@ -517,6 +517,7 @@ export interface MCPServerProps { accessToken: string | null; userRole: string | null; userID: string | null; + isViewOnly?: boolean; } export interface MCPToolsetTool { From fd834f6f8b18c5b2a5d642627631f1d4bf566ab4 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 02:10:50 +0000 Subject: [PATCH 068/141] fix(mcp): broadcast BYOK and OAuth credential eviction to peer workers and expire admin session tombstones Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 3 +- .../mcp_server/byok_credential_cache.py | 38 ++++++ .../mcp_server/byok_oauth_endpoints.py | 2 +- .../mcp_server/oauth2_token_cache.py | 7 +- .../proxy/_experimental/mcp_server/server.py | 121 +++++++++--------- .../mcp_management_endpoints.py | 4 +- litellm/proxy/proxy_server.py | 3 +- .../mcp_server/test_byok_credential_cache.py | 57 +++++++++ .../mcp_server/test_byok_oauth_endpoints.py | 32 ++++- .../mcp_server/test_mcp_server.py | 70 +++++++++- .../mcp_server/test_oauth2_token_cache.py | 26 ++++ .../test_mcp_management_endpoints.py | 6 +- tests/test_litellm/proxy/test_proxy_server.py | 71 ++++++++++ 13 files changed, 366 insertions(+), 74 deletions(-) create mode 100644 litellm/proxy/_experimental/mcp_server/byok_credential_cache.py create mode 100644 tests/test_litellm/proxy/_experimental/mcp_server/test_byok_credential_cache.py diff --git a/litellm/constants.py b/litellm/constants.py index 2ebc9beb632..72fc6ed5160 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -184,7 +184,8 @@ MCP_METADATA_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_METADATA_TIMEOUT", "1 MCP_HEALTH_CHECK_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_HEALTH_CHECK_TIMEOUT", "10.0")) MCP_TOOL_LISTING_MAX_PAGES: Final = 1000 MCP_GATEWAY_SESSION_ID_PREFIX_LENGTH: Final = 8 -MCP_ADMIN_TERMINATED_SESSION_IDS_MAX: Final = 1024 +MCP_BYOK_CREDENTIAL_CACHE_TTL_SECONDS: Final = 60 +MCP_BYOK_CREDENTIAL_CACHE_MAX_SIZE: Final = 4096 # Allowlist of commands permitted for MCP stdio transport. # Prevents arbitrary command execution via /mcp-rest/test/* endpoints or server creation. diff --git a/litellm/proxy/_experimental/mcp_server/byok_credential_cache.py b/litellm/proxy/_experimental/mcp_server/byok_credential_cache.py new file mode 100644 index 00000000000..3892015c405 --- /dev/null +++ b/litellm/proxy/_experimental/mcp_server/byok_credential_cache.py @@ -0,0 +1,38 @@ +"""Per-worker cache of stored BYOK credentials, keyed so peer workers can evict it over the auth cache pub/sub.""" + +from dataclasses import dataclass +from typing import Final + +from litellm.caching.in_memory_cache import InMemoryCache +from litellm.constants import MCP_BYOK_CREDENTIAL_CACHE_MAX_SIZE, MCP_BYOK_CREDENTIAL_CACHE_TTL_SECONDS + +_CACHE_KEY_PREFIX: Final = "mcp_byok_credential" + + +@dataclass(frozen=True, slots=True) +class CachedByokCredential: + credential: str | None + + +byok_credential_cache: Final = InMemoryCache( + max_size_in_memory=MCP_BYOK_CREDENTIAL_CACHE_MAX_SIZE, + default_ttl=MCP_BYOK_CREDENTIAL_CACHE_TTL_SECONDS, +) + + +def byok_credential_cache_key(user_id: str, server_id: str) -> str: + return f"{_CACHE_KEY_PREFIX}:{user_id}:{server_id}" + + +def get_cached_byok_credential(user_id: str, server_id: str) -> CachedByokCredential | None: + cached: Final = byok_credential_cache.get_cache( # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # InMemoryCache is untyped + byok_credential_cache_key(user_id, server_id) + ) + return cached if isinstance(cached, CachedByokCredential) else None + + +def cache_byok_credential(user_id: str, server_id: str, credential: str | None) -> None: + byok_credential_cache.set_cache( # pyright: ignore[reportUnknownMemberType] # InMemoryCache is untyped + byok_credential_cache_key(user_id, server_id), + CachedByokCredential(credential=credential), + ) diff --git a/litellm/proxy/_experimental/mcp_server/byok_oauth_endpoints.py b/litellm/proxy/_experimental/mcp_server/byok_oauth_endpoints.py index 0ab76588b1f..2c63e0a96d8 100644 --- a/litellm/proxy/_experimental/mcp_server/byok_oauth_endpoints.py +++ b/litellm/proxy/_experimental/mcp_server/byok_oauth_endpoints.py @@ -865,7 +865,7 @@ async def byok_token( _invalidate_byok_cred_cache, ) - _invalidate_byok_cred_cache(user_id, server_id) + await _invalidate_byok_cred_cache(user_id, server_id) except Exception as exc: verbose_proxy_logger.error( "byok_token: failed to store user credential for user=%s server=%s: %s", diff --git a/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py b/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py index 42edc2999ab..3742d7b4ccc 100644 --- a/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py +++ b/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py @@ -295,12 +295,15 @@ class MCPPerUserTokenCache: ) async def delete(self, user_id: str, server_id: str) -> None: - """Invalidate the cached token (removes from both in-memory and Redis layers).""" + """Invalidate the cached token in Redis, here, and in every peer worker's in-memory layer.""" try: + from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import ( # noqa: PLC0415 # proxy import cycle + evict_and_broadcast, + ) from litellm.proxy.proxy_server import user_api_key_cache # noqa: PLC0415 key: Final = self._cache_key(user_id, server_id) - await user_api_key_cache.async_delete_cache(key) + await evict_and_broadcast((key,), user_api_key_cache) except Exception as exc: verbose_logger.debug( "MCPPerUserTokenCache.delete failed for user=%s server=%s: %s", diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py index 80d9274859d..a7ca2775fb1 100644 --- a/litellm/proxy/_experimental/mcp_server/server.py +++ b/litellm/proxy/_experimental/mcp_server/server.py @@ -14,7 +14,7 @@ import time import traceback import types import uuid -from collections import Counter, deque +from collections import Counter from collections.abc import AsyncIterator, Callable, Mapping, Sequence from datetime import datetime from typing import TYPE_CHECKING, Any, Final, NoReturn, Protocol @@ -30,7 +30,6 @@ from typing_extensions import ReadOnly, TypedDict from litellm._logging import verbose_logger from litellm.constants import ( MAXIMUM_TRACEBACK_LINES_TO_LOG, - MCP_ADMIN_TERMINATED_SESSION_IDS_MAX, MCP_GATEWAY_SESSION_ID_PREFIX_LENGTH, ) from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj @@ -42,6 +41,12 @@ from litellm.proxy._experimental.mcp_server.auth.user_api_key_auth_mcp import ( MCPRequestHandler, _is_mcp_admitted_user_subject, ) +from litellm.proxy._experimental.mcp_server.byok_credential_cache import ( + byok_credential_cache, + byok_credential_cache_key, + cache_byok_credential, + get_cached_byok_credential, +) from litellm.proxy._experimental.mcp_server.discoverable_endpoints import ( get_request_base_url, ) @@ -86,6 +91,9 @@ from litellm.proxy._types import ( UserAPIKeyAuth, ) from litellm.proxy.auth.ip_address_utils import IPAddressUtils +from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import ( + publish_auth_cache_invalidation, +) from litellm.proxy.litellm_pre_call_utils import ( LiteLLMProxyRequestSetup, get_chain_id_from_headers, @@ -107,13 +115,6 @@ if TYPE_CHECKING: from litellm.proxy._experimental.mcp_server.db import OAuthCredentialPayload -# Short-lived in-memory cache for BYOK credentials. -# Keyed by (user_id, server_id); value is (credential_or_None, monotonic_timestamp). -# Storing the credential value (not just a bool) means _get_byok_credential and -# _check_byok_credential share a single DB round-trip per TTL window. -_byok_cred_cache: Final[dict[tuple[str, str], tuple[str | None, float]]] = {} -_BYOK_CRED_CACHE_TTL: Final = 60 # seconds -_BYOK_CRED_CACHE_MAX_SIZE: Final = 4096 # cap to prevent unbounded growth _STATEFUL_SESSION_IDLE_TIMEOUT_SECONDS: Final = 30 * 60 # Upper bound on concurrent stateful sessions a single caller may hold. Each # `initialize` creates a session that survives until the idle timeout, so @@ -132,20 +133,11 @@ _MCP_TRANSPORT_SPAN_SCOPE_KEY: Final = "litellm_otel_transport_span" _MCP_DESTINATIONS_SCOPE_KEY: Final = "litellm_otel_request_destinations" -def _invalidate_byok_cred_cache(user_id: str, server_id: str) -> None: - """Remove a (user_id, server_id) entry from the BYOK credential cache. - - Call this after storing or deleting a credential so subsequent calls - see the fresh value rather than a stale cached result. - """ - _byok_cred_cache.pop((user_id, server_id), None) - - -def _write_byok_cred_cache(user_id: str, server_id: str, credential: str | None) -> None: - """Write a credential value to the cache, evicting all entries if at capacity.""" - if len(_byok_cred_cache) >= _BYOK_CRED_CACHE_MAX_SIZE: - _byok_cred_cache.clear() - _byok_cred_cache[(user_id, server_id)] = (credential, time.monotonic()) +async def _invalidate_byok_cred_cache(user_id: str, server_id: str) -> None: + """Drop a stored-or-deleted BYOK credential from this worker's cache and from every peer worker's.""" + cache_key: Final = byok_credential_cache_key(user_id, server_id) + byok_credential_cache.delete_cache(cache_key) + await publish_auth_cache_invalidation(cache_key=cache_key) # Check if MCP is available @@ -623,9 +615,7 @@ if MCP_AVAILABLE: _stateful_session_locks: Final[dict[str, asyncio.Lock]] = {} _stateful_session_active_request_counts: Final[dict[str, int]] = {} _stateful_session_client_info: Final[dict[str, Implementation]] = {} # mutable-ok: cleared on session teardown - _admin_terminated_session_ids: Final[deque[str]] = deque( # mutable-ok: bounded ring, appended on admin termination - maxlen=MCP_ADMIN_TERMINATED_SESSION_IDS_MAX - ) + _admin_terminated_session_ids: Final[dict[str, float]] = {} # mutable-ok: admin-closed id -> last replay class _TerminableTransport(Protocol): async def terminate(self) -> None: ... @@ -697,6 +687,7 @@ if MCP_AVAILABLE: for session_id in list(_stateful_session_auth_context_last_seen): if session_id not in _stateful_session_auth_contexts: _remove_stateful_session_tracking(session_id) + _forget_expired_admin_terminated_session_ids(now) async def _enforce_stateful_session_cap_for_owner(owner: str) -> bool: """ @@ -2819,35 +2810,28 @@ if MCP_AVAILABLE: mcp_server: MCPServer, user_api_key_auth: UserAPIKeyAuth | None, ) -> str | None: - """Retrieve the stored BYOK credential for a user+server pair. - - Uses the shared _byok_cred_cache to avoid a DB round-trip on every - tool call within the TTL window. - """ + """Retrieve the stored BYOK credential for a user+server pair, served from the worker cache within its TTL.""" if not mcp_server.is_byok: return None user_id: Final = (user_api_key_auth.user_id if user_api_key_auth else None) or "" if not user_id: return None - cache_key: Final = (user_id, mcp_server.server_id) - cached: Final = _byok_cred_cache.get(cache_key) + cached: Final = get_cached_byok_credential(user_id, mcp_server.server_id) if cached is not None: - credential, ts = cached - if time.monotonic() - ts < _BYOK_CRED_CACHE_TTL: - return credential + return cached.credential from litellm.proxy._experimental.mcp_server.db import get_user_credential from litellm.proxy.proxy_server import prisma_client if prisma_client is None: return None - credential = await get_user_credential( + credential: Final = await get_user_credential( prisma_client=prisma_client, user_id=user_id, server_id=mcp_server.server_id, ) - _write_byok_cred_cache(user_id, mcp_server.server_id, credential) + cache_byok_credential(user_id, mcp_server.server_id, credential) return credential async def _check_byok_credential( @@ -2876,27 +2860,23 @@ if MCP_AVAILABLE: headers={"WWW-Authenticate": get_byok_www_authenticate()}, ) - # Check shared credential cache before hitting the DB. - cache_key: Final = (user_id, mcp_server.server_id) - cached: Final = _byok_cred_cache.get(cache_key) + cached: Final = get_cached_byok_credential(user_id, mcp_server.server_id) if cached is not None: - cached_cred, ts = cached - if time.monotonic() - ts < _BYOK_CRED_CACHE_TTL: - if cached_cred is None: - raise HTTPException( - status_code=401, - detail={ - "error": "byok_auth_required", - "server_id": mcp_server.server_id, - "server_name": mcp_server.server_name or mcp_server.name, - "message": ( - "No stored credential found for this BYOK server. " - "Complete the OAuth authorization flow to provide your API key." - ), - }, - headers={"WWW-Authenticate": get_byok_www_authenticate()}, - ) - return + if cached.credential is None: + raise HTTPException( + status_code=401, + detail={ + "error": "byok_auth_required", + "server_id": mcp_server.server_id, + "server_name": mcp_server.server_name or mcp_server.name, + "message": ( + "No stored credential found for this BYOK server. " + "Complete the OAuth authorization flow to provide your API key." + ), + }, + headers={"WWW-Authenticate": get_byok_www_authenticate()}, + ) + return from litellm.proxy._experimental.mcp_server.db import get_user_credential from litellm.proxy.proxy_server import prisma_client @@ -2920,7 +2900,7 @@ if MCP_AVAILABLE: user_id=user_id, server_id=mcp_server.server_id, ) - _write_byok_cred_cache(user_id, mcp_server.server_id, credential) + cache_byok_credential(user_id, mcp_server.server_id, credential) if credential is None: raise HTTPException( status_code=401, @@ -3906,6 +3886,24 @@ if MCP_AVAILABLE: key_auth: Final = auth_user.user_api_key_auth return key_auth is not None and key_auth.user_id == user_id + def _forget_expired_admin_terminated_session_ids(now: float) -> None: + for session_id in [ + session_id + for session_id, last_replayed in _admin_terminated_session_ids.items() + if now - last_replayed >= _STATEFUL_SESSION_IDLE_TIMEOUT_SECONDS + ]: + del _admin_terminated_session_ids[session_id] + + def _is_admin_terminated_session_id(session_id: str, now: float) -> bool: + last_replayed: Final = _admin_terminated_session_ids.get(session_id) + if last_replayed is None: + return False + if now - last_replayed >= _STATEFUL_SESSION_IDLE_TIMEOUT_SECONDS: + del _admin_terminated_session_ids[session_id] + return False + _admin_terminated_session_ids[session_id] = now + return True + async def terminate_mcp_gateway_sessions( *, session_id_prefix: str | None = None, @@ -3919,6 +3917,7 @@ if MCP_AVAILABLE: admission. Only sessions held by this worker process are affected. """ now: Final = time.monotonic() + _forget_expired_admin_terminated_session_ids(now) server_instances: Final = _stateful_server_instances() targets: Final = tuple( (session_id, auth_user) @@ -3928,7 +3927,7 @@ if MCP_AVAILABLE: ) terminated: Final = tuple(_gateway_session_for(session_id, auth_user, now) for session_id, auth_user in targets) for session_id, _ in targets: - _admin_terminated_session_ids.append(session_id) + _admin_terminated_session_ids[session_id] = now transport = server_instances.pop(session_id, None) _remove_stateful_session_tracking(session_id) if transport is not None: @@ -4064,7 +4063,7 @@ if MCP_AVAILABLE: await success_response(scope, receive, send) return True - if _session_id in _admin_terminated_session_ids: + if _is_admin_terminated_session_id(_session_id, time.monotonic()): terminated_response: Final = JSONResponse( status_code=404, content={ # mutable-ok: JSONResponse content must be a plain dict diff --git a/litellm/proxy/management_endpoints/mcp_management_endpoints.py b/litellm/proxy/management_endpoints/mcp_management_endpoints.py index 728ba9568e8..5326cf3415f 100644 --- a/litellm/proxy/management_endpoints/mcp_management_endpoints.py +++ b/litellm/proxy/management_endpoints/mcp_management_endpoints.py @@ -2317,7 +2317,7 @@ if MCP_AVAILABLE: _invalidate_byok_cred_cache, ) - _invalidate_byok_cred_cache(user_id, server_id) + await _invalidate_byok_cred_cache(user_id, server_id) return MCPUserCredentialResponse(server_id=server_id, has_credential=True) # save=False: credential not persisted return MCPUserCredentialResponse(server_id=server_id, has_credential=False) @@ -2348,7 +2348,7 @@ if MCP_AVAILABLE: _invalidate_byok_cred_cache, ) - _invalidate_byok_cred_cache(target_user_id, server_id) + await _invalidate_byok_cred_cache(target_user_id, server_id) return MCPUserCredentialResponse(server_id=server_id, has_credential=False) # ── OAuth2 user-credential endpoints ────────────────────────────────────── diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index d7d8413d2ce..48122549aea 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -307,6 +307,7 @@ from litellm.litellm_core_utils.sensitive_data_masker import ( from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler from litellm.llms.openai_like.model_info import MODEL_INFO_REFRESH_SECONDS from litellm.llms.vertex_ai.vertex_llm_base import VertexBase +from litellm.proxy._experimental.mcp_server.byok_credential_cache import byok_credential_cache from litellm.proxy._lazy_features import attach_lazy_features, reserve_lazy_slot from litellm.proxy._types import * from litellm.proxy.analytics_endpoints.analytics_endpoints import ( @@ -7535,7 +7536,7 @@ class ProxyConfig: subscriber: Final = AuthCacheInvalidationSubscriber( redis_cache=redis_cache, user_api_key_cache=user_api_key_cache, - additional_in_memory_caches=(spend_counter_cache.in_memory_cache,), + additional_in_memory_caches=(spend_counter_cache.in_memory_cache, byok_credential_cache), ) self.auth_cache_invalidation_subscriber = subscriber subscriber.start() diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_credential_cache.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_credential_cache.py new file mode 100644 index 00000000000..8ec5b8642bc --- /dev/null +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_credential_cache.py @@ -0,0 +1,57 @@ +import json + +import pytest + +from litellm.proxy._experimental.mcp_server.byok_credential_cache import ( + CachedByokCredential, + byok_credential_cache, + byok_credential_cache_key, + cache_byok_credential, + get_cached_byok_credential, +) +from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import AuthCacheInvalidationSubscriber +from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache + + +class _FakeRedisCache: + namespace = None + + def init_async_client(self) -> object: + return object() + + +@pytest.fixture(autouse=True) +def _empty_cache(): + byok_credential_cache.flush_cache() + yield + byok_credential_cache.flush_cache() + + +def test_a_cached_negative_lookup_is_distinguishable_from_a_miss(): + assert get_cached_byok_credential("u-1", "srv-1") is None + cache_byok_credential("u-1", "srv-1", None) + assert get_cached_byok_credential("u-1", "srv-1") == CachedByokCredential(credential=None) + cache_byok_credential("u-1", "srv-1", "sk-stored") + assert get_cached_byok_credential("u-1", "srv-1") == CachedByokCredential(credential="sk-stored") + assert get_cached_byok_credential("u-1", "srv-2") is None + + +def test_peer_worker_invalidation_message_evicts_the_cached_credential(): + """The key a mutating worker broadcasts must be the key every other worker caches under.""" + cache_byok_credential("mallory", "srv-byok", "sk-revoked") + cache_byok_credential("alice", "srv-byok", "sk-kept") + subscriber = AuthCacheInvalidationSubscriber( + redis_cache=_FakeRedisCache(), # pyright: ignore[reportArgumentType] # subscriber is never started; only its message handler runs + user_api_key_cache=UserApiKeyCache(), + additional_in_memory_caches=(byok_credential_cache,), + ) + + subscriber._apply_message( # pyright: ignore[reportPrivateUsage] # exercising the real cross-worker message handler + { + "type": "message", + "data": json.dumps({"cache_key": byok_credential_cache_key("mallory", "srv-byok")}).encode(), + } + ) + + assert get_cached_byok_credential("mallory", "srv-byok") is None + assert get_cached_byok_credential("alice", "srv-byok") == CachedByokCredential(credential="sk-kept") diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py index 55accfb169d..df0d7916baf 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py @@ -592,7 +592,7 @@ async def test_check_byok_credential_missing_credential(monkeypatch): monkeypatch.delenv("PROXY_BASE_URL", raising=False) monkeypatch.delenv("SERVER_ROOT_PATH", raising=False) - monkeypatch.setattr(server_module, "_byok_cred_cache", {}) + server_module.byok_credential_cache.flush_cache() mock_prisma = MagicMock() with ( @@ -628,7 +628,7 @@ async def test_execute_byok_tool_missing_credential_advertises_api_key_flow(monk from litellm.types.mcp_server.mcp_server_manager import MCPServer monkeypatch.setenv("PROXY_BASE_URL", "https://gateway.example.com/proxy") - monkeypatch.setattr(mcp_module, "_byok_cred_cache", {}) + mcp_module.byok_credential_cache.flush_cache() server = MCPServer(server_id="byok-discovery", name="byok-discovery", transport=MCPTransport.http, is_byok=True) prisma = MagicMock() prisma.db.litellm_mcpusercredentials.find_unique = AsyncMock(return_value=None) @@ -677,6 +677,34 @@ async def test_check_byok_credential_has_credential(): await _check_byok_credential(server, user_auth) +@pytest.mark.asyncio +async def test_invalidate_byok_cred_cache_evicts_locally_and_broadcasts_the_same_key(): + """A revoked credential must stop being served here and on every peer worker within the TTL.""" + from litellm.proxy._experimental.mcp_server import server as server_module + from litellm.proxy._experimental.mcp_server.byok_credential_cache import byok_credential_cache_key + from litellm.proxy._types import UserAPIKeyAuth + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer(server_id="byok-revoke", name="byok-server", transport=MCPTransport.http, is_byok=True) + user_auth = UserAPIKeyAuth(user_id="mallory", api_key="sk-test") + server_module.byok_credential_cache.flush_cache() + db_lookup = AsyncMock(side_effect=["sk-before-revoke", None]) + publish = AsyncMock() + + with ( + patch("litellm.proxy._experimental.mcp_server.db.get_user_credential", new=db_lookup), + patch("litellm.proxy.proxy_server.prisma_client", MagicMock()), + patch.object(server_module, "publish_auth_cache_invalidation", new=publish), + ): + assert await server_module._get_byok_credential(server, user_auth) == "sk-before-revoke" + assert await server_module._get_byok_credential(server, user_auth) == "sk-before-revoke" + await server_module._invalidate_byok_cred_cache("mallory", "byok-revoke") + assert await server_module._get_byok_credential(server, user_auth) is None + + assert db_lookup.await_count == 2 + publish.assert_awaited_once_with(cache_key=byok_credential_cache_key("mallory", "byok-revoke")) + + @pytest.mark.asyncio async def test_check_byok_credential_db_unavailable_fails_closed(): """BYOK server with no prisma_client → 503, not silent pass. diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py index 4311a69d465..f496f22d5d2 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py @@ -2989,9 +2989,10 @@ async def test_admin_terminated_session_id_gets_404_instead_of_a_fresh_stateless """Once an admin closes a session, a client replaying its id must not be silently upgraded to a new stateless session by the stale-header path; it gets 404 and has to initialize again.""" try: + from starlette.types import Scope + from litellm.proxy._experimental.mcp_server import server as mcp_server from litellm.proxy._experimental.mcp_server.server import session_manager_stateful - from starlette.types import Scope except ImportError: pytest.skip("MCP server not available") @@ -3048,6 +3049,73 @@ async def test_admin_terminated_session_id_gets_404_instead_of_a_fresh_stateless mcp_server._admin_terminated_session_ids.clear() +@pytest.mark.asyncio +async def test_admin_terminated_session_id_stays_refused_while_replayed_and_is_forgotten_like_an_idle_session(): + """The refusal window slides on every replay, so a client that keeps retrying is never silently + upgraded to a stateless session no matter how many other sessions an admin closes later; an id + nobody has replayed for a full idle timeout is dropped from the table by the idle sweep.""" + try: + from starlette.types import Scope + + from litellm.proxy._experimental.mcp_server import server as mcp_server + from litellm.proxy._experimental.mcp_server.server import session_manager_stateful + except ImportError: + pytest.skip("MCP server not available") + + idle_timeout = mcp_server._STATEFUL_SESSION_IDLE_TIMEOUT_SECONDS + retrying_id, silent_id = "admin-closed-retrying", "admin-closed-silent" + contexts = { + session_id: mcp_server.MCPAuthenticatedUser( + user_api_key_auth=UserAPIKeyAuth(api_key="key-alice", user_id="alice"), + ) + for session_id in (retrying_id, silent_id) + } + live_transports = {session_id: MagicMock(terminate=AsyncMock()) for session_id in contexts} + + async def replay(session_id: str, now: float) -> tuple[bool, list[bytes]]: + scope: Scope = { + "type": "http", + "method": "POST", + "headers": [(b"content-type", b"application/json"), (b"mcp-session-id", session_id.encode())], + } + with patch.object(mcp_server.time, "monotonic", return_value=now): + handled = await mcp_server._handle_stale_mcp_session( + scope, AsyncMock(), AsyncMock(), session_manager_stateful + ) + return handled, [k for k, _ in scope["headers"]] + + try: + with ( + patch.object( # test-quality-ok: the transport registry is a module-level singleton; the suite's only seam + session_manager_stateful, "_server_instances", live_transports + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_contexts, contexts, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_client_info, {}, clear=True + ), + patch.dict( # test-quality-ok: the session tables are module-level singletons; the suite's only seam + mcp_server._stateful_session_auth_context_last_seen, {}, clear=True + ), + ): + with patch.object(mcp_server.time, "monotonic", return_value=1000.0): + closed = await mcp_server.terminate_mcp_gateway_sessions(user_id="alice") + assert closed.terminated_sessions == 2 + + for elapsed in (idle_timeout - 1, 2 * idle_timeout - 2, 3 * idle_timeout - 3): + assert await replay(retrying_id, 1000.0 + elapsed) == (True, [b"content-type", b"mcp-session-id"]) + + await mcp_server._purge_expired_stateful_session_auth_contexts(now=1000.0 + idle_timeout) + assert set(mcp_server._admin_terminated_session_ids) == {retrying_id} + + assert await replay(silent_id, 1000.0 + idle_timeout) == (False, [b"content-type"]) + assert await replay(retrying_id, 1000.0 + 4 * idle_timeout) == (False, [b"content-type"]) + assert mcp_server._admin_terminated_session_ids == {} + finally: + mcp_server._admin_terminated_session_ids.clear() + + @pytest.mark.asyncio async def test_initialize_request_with_existing_session_tracks_new_session(): try: diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py index f7567efcabc..301b3887922 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py @@ -395,6 +395,32 @@ async def test_invalidate_clears_every_identity_for_a_server(): assert mock_client.post.call_count == 3 +@pytest.mark.asyncio +async def test_per_user_token_delete_evicts_locally_and_broadcasts_to_peer_workers(): + """Revoking a user's OAuth token must not leave peer workers serving it from their in-memory layer.""" + from litellm.proxy import proxy_server + from litellm.proxy._experimental.mcp_server.oauth2_token_cache import MCPPerUserTokenCache + from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache + + local_cache = UserApiKeyCache() + publish = AsyncMock() + token_cache = MCPPerUserTokenCache() + key = token_cache._cache_key("mallory", "srv-oauth") # pyright: ignore[reportPrivateUsage] # asserting the broadcast names the stored key + local_cache.in_memory_cache.set_cache(key, "encrypted-token") + + with ( + patch.object(proxy_server, "user_api_key_cache", local_cache), + patch( + "litellm.proxy.common_utils.auth_cache_invalidation_pubsub.publish_auth_cache_invalidation", + new=publish, + ), + ): + await token_cache.delete("mallory", "srv-oauth") + + assert local_cache.in_memory_cache.get_cache(key) is None + publish.assert_awaited_once_with(cache_key=key) + + @pytest.mark.asyncio async def test_m2m_mint_uses_admin_entered_token_url_when_issuer_yield_empties_resolved(): """A pinned issuer empties the resolved token_url while configured_token_url keeps the diff --git a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py index 2f33018599f..a69cfe0ac7d 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py @@ -5152,7 +5152,7 @@ async def test_admin_revokes_another_users_byok_credential(): ) delete_mock = AsyncMock(return_value=None) - invalidate_mock = MagicMock() + invalidate_mock = AsyncMock() with ( patch( # test-quality-ok: endpoint test stubs the Prisma client lookup "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", @@ -5174,7 +5174,7 @@ async def test_admin_revokes_another_users_byok_credential(): delete_mock.assert_awaited_once() assert delete_mock.await_args.args[1:] == ("mallory", "srv-byok-admin") - invalidate_mock.assert_called_once_with("mallory", "srv-byok-admin") + invalidate_mock.assert_awaited_once_with("mallory", "srv-byok-admin") assert result.has_credential is False @@ -5231,7 +5231,7 @@ async def test_user_naming_themselves_still_deletes_own_byok_credential(): new=delete_mock, ), patch.object( # test-quality-ok: the cache invalidator is module scoped; the suite's only seam - mcp_server, "_invalidate_byok_cred_cache", new=MagicMock() + mcp_server, "_invalidate_byok_cred_cache", new=AsyncMock() ), ): await delete_mcp_user_credential( diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 41c4956dba6..acd13fef9a6 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -13919,3 +13919,74 @@ async def test_token_counter_loads_a_custom_tokenizer_once_per_identifier_revisi ] finally: litellm.utils._select_custom_tokenizer_helper.cache_clear() + + +@pytest.mark.asyncio +async def test_auth_cache_invalidation_subscriber_evicts_byok_credentials_cached_by_this_worker(): + """A peer worker's BYOK revocation broadcast must reach this worker's BYOK credential cache.""" + from redis.asyncio import Redis + + from litellm.proxy._experimental.mcp_server.byok_credential_cache import ( + byok_credential_cache, + byok_credential_cache_key, + cache_byok_credential, + get_cached_byok_credential, + ) + from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache + + class _QueuePubSub: + def __init__(self, messages: list[object]) -> None: + self.queue: asyncio.Queue[object] = asyncio.Queue() + for message in messages: + self.queue.put_nowait(message) + + async def subscribe(self, *channels: str) -> None: + return None + + async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> object | None: + try: + return await asyncio.wait_for(self.queue.get(), timeout) + except asyncio.TimeoutError: + return None + + async def aclose(self) -> None: + return None + + class _PubSubRedisClient(Redis): + def __init__(self, pubsub: _QueuePubSub) -> None: + self._scripted_pubsub = pubsub + + def pubsub(self) -> _QueuePubSub: + return self._scripted_pubsub + + class _FakeRedisCache: + namespace = None + + def __init__(self, client: object) -> None: + self._client = client + + def init_async_client(self) -> object: + return self._client + + byok_credential_cache.flush_cache() + cache_byok_credential("mallory", "srv-byok", "sk-revoked-elsewhere") + message: Final = { + "type": "message", + "data": json.dumps({"cache_key": byok_credential_cache_key("mallory", "srv-byok")}).encode(), + } + proxy_config: Final = proxy_server_module.ProxyConfig() + proxy_config.start_auth_cache_invalidation_subscriber( + redis_cache=_FakeRedisCache(_PubSubRedisClient(_QueuePubSub([message]))), # pyright: ignore[reportArgumentType] # fake pub/sub capable redis; no live redis in this unit test + user_api_key_cache=UserApiKeyCache(), + ) + try: + for _ in range(200): + if get_cached_byok_credential("mallory", "srv-byok") is None: + break + await asyncio.sleep(0.01) + evicted: Final = get_cached_byok_credential("mallory", "srv-byok") is None + finally: + await proxy_config.stop_auth_cache_invalidation_subscriber() + byok_credential_cache.flush_cache() + + assert evicted, "the subscriber does not evict the BYOK credential cache on a peer worker's broadcast" From ce48a3fbcce70cc17f877de30b9c2a64ed4bc278 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 02:49:41 +0000 Subject: [PATCH 069/141] test(mcp): give the new cache and tombstone patches TQ008 reasons and match the keyword eviction call Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/mcp_tests/test_per_user_oauth_cache.py | 4 +--- .../mcp_server/test_byok_oauth_endpoints.py | 12 +++++++++--- .../_experimental/mcp_server/test_mcp_server.py | 8 ++++++-- .../mcp_server/test_oauth2_token_cache.py | 6 ++++-- 4 files changed, 20 insertions(+), 10 deletions(-) diff --git a/tests/mcp_tests/test_per_user_oauth_cache.py b/tests/mcp_tests/test_per_user_oauth_cache.py index 141b906fce9..ac453df8fa5 100644 --- a/tests/mcp_tests/test_per_user_oauth_cache.py +++ b/tests/mcp_tests/test_per_user_oauth_cache.py @@ -331,9 +331,7 @@ class TestMCPPerUserTokenCache: with patch("litellm.proxy.proxy_server.user_api_key_cache", mock_dual_cache): await cache.delete("alice", "slack-test") - mock_dual_cache.async_delete_cache.assert_called_once_with( - "mcp:per_user_token:alice:slack-test" - ) + mock_dual_cache.async_delete_cache.assert_called_once_with(key="mcp:per_user_token:alice:slack-test") mock_dual_cache.async_set_cache.assert_not_called() @pytest.mark.asyncio diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py index df0d7916baf..87e23893616 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py @@ -692,9 +692,15 @@ async def test_invalidate_byok_cred_cache_evicts_locally_and_broadcasts_the_same publish = AsyncMock() with ( - patch("litellm.proxy._experimental.mcp_server.db.get_user_credential", new=db_lookup), - patch("litellm.proxy.proxy_server.prisma_client", MagicMock()), - patch.object(server_module, "publish_auth_cache_invalidation", new=publish), + patch( # test-quality-ok: the DB row lookup is the only seam below the credential resolver; no Prisma fake exists + "litellm.proxy._experimental.mcp_server.db.get_user_credential", new=db_lookup + ), + patch( # test-quality-ok: the resolver reads the module-level prisma_client singleton; the suite's only seam + "litellm.proxy.proxy_server.prisma_client", MagicMock() + ), + patch.object( # test-quality-ok: the redis publisher is module-level; asserting the broadcast without a redis + server_module, "publish_auth_cache_invalidation", new=publish + ), ): assert await server_module._get_byok_credential(server, user_auth) == "sk-before-revoke" assert await server_module._get_byok_credential(server, user_auth) == "sk-before-revoke" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py index f496f22d5d2..0e2ceb98987 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py @@ -3078,7 +3078,9 @@ async def test_admin_terminated_session_id_stays_refused_while_replayed_and_is_f "method": "POST", "headers": [(b"content-type", b"application/json"), (b"mcp-session-id", session_id.encode())], } - with patch.object(mcp_server.time, "monotonic", return_value=now): + with patch.object( # test-quality-ok: the stale-session handler reads the clock directly; no injectable now + mcp_server.time, "monotonic", return_value=now + ): handled = await mcp_server._handle_stale_mcp_session( scope, AsyncMock(), AsyncMock(), session_manager_stateful ) @@ -3099,7 +3101,9 @@ async def test_admin_terminated_session_id_stays_refused_while_replayed_and_is_f mcp_server._stateful_session_auth_context_last_seen, {}, clear=True ), ): - with patch.object(mcp_server.time, "monotonic", return_value=1000.0): + with patch.object( # test-quality-ok: termination stamps the tombstone from the clock directly; no injectable now + mcp_server.time, "monotonic", return_value=1000.0 + ): closed = await mcp_server.terminate_mcp_gateway_sessions(user_id="alice") assert closed.terminated_sessions == 2 diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py index 301b3887922..30d0f17a099 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py @@ -409,8 +409,10 @@ async def test_per_user_token_delete_evicts_locally_and_broadcasts_to_peer_worke local_cache.in_memory_cache.set_cache(key, "encrypted-token") with ( - patch.object(proxy_server, "user_api_key_cache", local_cache), - patch( + patch.object( # test-quality-ok: the token cache reads the module-level user_api_key_cache singleton; the suite's only seam + proxy_server, "user_api_key_cache", local_cache + ), + patch( # test-quality-ok: the redis publisher is module-level; asserting the broadcast without a redis "litellm.proxy.common_utils.auth_cache_invalidation_pubsub.publish_auth_cache_invalidation", new=publish, ), From 85d338da8bc9bc1f9ea8fa7cc4ccb0f1940e4407 Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 04:25:44 +0000 Subject: [PATCH 070/141] feat(dashscope): add qwen3.8-omni-flash and qwen3.8-flash to the model cost map Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 74 +++++++++++++++++++ model_prices_and_context_window.json | 74 +++++++++++++++++++ 2 files changed, 148 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 355e2b90e96..3409d9b4e46 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -16690,6 +16690,43 @@ "supports_tool_choice": true, "supports_vision": true }, + "dashscope/qwen3.8-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "dashscope", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_web_search": true + }, + "dashscope/qwen3.8-omni-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "dashscope", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, + "supports_web_search": true + }, "dashscope/qwq-plus": { "input_cost_per_token": 8e-07, "litellm_provider": "dashscope", @@ -18594,6 +18631,43 @@ "supports_tool_choice": true, "supports_vision": true }, + "qwen_ai_platform/qwen3.8-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "qwen_ai_platform", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_web_search": true + }, + "qwen_ai_platform/qwen3.8-omni-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "qwen_ai_platform", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, + "supports_web_search": true + }, "qwen_ai_platform/qwq-plus": { "input_cost_per_token": 8e-07, "litellm_provider": "qwen_ai_platform", diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 355e2b90e96..3409d9b4e46 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -16690,6 +16690,43 @@ "supports_tool_choice": true, "supports_vision": true }, + "dashscope/qwen3.8-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "dashscope", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_web_search": true + }, + "dashscope/qwen3.8-omni-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "dashscope", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, + "supports_web_search": true + }, "dashscope/qwq-plus": { "input_cost_per_token": 8e-07, "litellm_provider": "dashscope", @@ -18594,6 +18631,43 @@ "supports_tool_choice": true, "supports_vision": true }, + "qwen_ai_platform/qwen3.8-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "qwen_ai_platform", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_web_search": true + }, + "qwen_ai_platform/qwen3.8-omni-flash": { + "cache_read_input_token_cost": 1.6e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "qwen_ai_platform", + "max_input_tokens": 991808, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.7e-07, + "source": "https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-pricing", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, + "supports_web_search": true + }, "qwen_ai_platform/qwq-plus": { "input_cost_per_token": 8e-07, "litellm_provider": "qwen_ai_platform", From 4863e1775ba4630df7d6596b164b509d836357a2 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 04:33:16 +0000 Subject: [PATCH 071/141] feat(guardrails): add TypeSafe Jev relevance-based compaction guardrail Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_lazy_openapi_snapshot.json | 2 +- .../guardrails/auto_router_compression.py | 2 +- .../guardrail_hooks/typesafe/__init__.py | 80 ++++ .../guardrail_hooks/typesafe/typesafe.py | 390 ++++++++++++++++++ litellm/types/guardrails.py | 7 +- .../guardrails/guardrail_hooks/typesafe.py | 58 +++ .../guardrail_hooks/test_typesafe.py | 274 ++++++++++++ .../add_model/buildAutoRouterCompression.ts | 2 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 +- 9 files changed, 812 insertions(+), 5 deletions(-) create mode 100644 litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py create mode 100644 litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py create mode 100644 litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py create mode 100644 tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index b244678e201..a215cc573d7 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -10006,7 +10006,7 @@ }, "unreachable_fallback": { "default": "fail_closed", - "description": "Behavior when a guardrail endpoint is unreachable due to network errors. Implemented by guardrail='generic_guardrail_api', 'agent_365', 'akto', 'vigil_guard', 'repelloai', 'headroom', and 'compresr'. 'fail_closed' raises an error (default). 'fail_open' logs a critical error and allows the request to proceed.", + "description": "Behavior when a guardrail endpoint is unreachable due to network errors. Implemented by guardrail='generic_guardrail_api', 'agent_365', 'akto', 'vigil_guard', 'repelloai', 'headroom', 'compresr', and 'typesafe'. 'fail_closed' raises an error (default). 'fail_open' logs a critical error and allows the request to proceed.", "enum": [ "fail_closed", "fail_open" diff --git a/litellm/proxy/guardrails/auto_router_compression.py b/litellm/proxy/guardrails/auto_router_compression.py index c37b9fff1f0..335419c6372 100644 --- a/litellm/proxy/guardrails/auto_router_compression.py +++ b/litellm/proxy/guardrails/auto_router_compression.py @@ -21,7 +21,7 @@ if TYPE_CHECKING: from litellm.integrations.custom_guardrail import CustomGuardrail from litellm.router import Router -COMPRESSION_GUARDRAIL_PROVIDERS: Final = frozenset({"headroom", "compresr"}) +COMPRESSION_GUARDRAIL_PROVIDERS: Final = frozenset({"headroom", "compresr", "typesafe"}) _NO_COMPRESSION: Final = "none" # A ContextVar, not metadata: metadata reaches spend logs the caller can read, and a diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py new file mode 100644 index 00000000000..c347c863a1b --- /dev/null +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py @@ -0,0 +1,80 @@ +from __future__ import annotations + +from typing import TYPE_CHECKING, Final, cast + +from litellm.types.guardrails import ( + GuardrailEventHooks, + Mode, + SupportedGuardrailIntegrations, +) + +from .typesafe import TypeSafeGuardrail + +if TYPE_CHECKING: + from litellm.types.guardrails import Guardrail, LitellmParams + + +def _coerce_event_hook( + mode: str | list[str] | Mode, +) -> GuardrailEventHooks | list[GuardrailEventHooks] | Mode: + if isinstance(mode, Mode): + return mode + if isinstance(mode, list): + return [GuardrailEventHooks(item) for item in mode] + return GuardrailEventHooks(mode) + + +def _get_optional_value(litellm_params: LitellmParams, optional_params: object | None, attribute_name: str) -> object: + if optional_params is not None: + value: Final = getattr(optional_params, attribute_name, None) + if value is not None: + return cast(object, value) + return cast(object, getattr(litellm_params, attribute_name, None)) + + +def _optional_float(litellm_params: LitellmParams, optional_params: object | None, attribute_name: str) -> float | None: + value: Final = _get_optional_value(litellm_params, optional_params, attribute_name) + if isinstance(value, bool) or not isinstance(value, (int, float)): + return None + return float(value) + + +def _optional_int(litellm_params: LitellmParams, optional_params: object | None, attribute_name: str) -> int | None: + value: Final = _get_optional_value(litellm_params, optional_params, attribute_name) + if isinstance(value, bool) or not isinstance(value, int): + return None + return value + + +def initialize_guardrail(litellm_params: LitellmParams, guardrail: Guardrail) -> TypeSafeGuardrail: + import litellm + + optional_params: Final = getattr(litellm_params, "optional_params", None) + + _callback: Final = TypeSafeGuardrail( + api_base=litellm_params.api_base, + api_key=litellm_params.api_key, + model=litellm_params.model, + relevance_threshold=_optional_float(litellm_params, optional_params, "relevance_threshold"), + min_chars_to_evaluate=_optional_int(litellm_params, optional_params, "min_chars_to_evaluate"), + max_result_chars_in_state=_optional_int(litellm_params, optional_params, "max_result_chars_in_state"), + guardrail_name=guardrail["guardrail_name"], + event_hook=_coerce_event_hook(litellm_params.mode), + default_on=litellm_params.default_on or False, + unreachable_fallback=( + litellm_params.unreachable_fallback if "unreachable_fallback" in litellm_params.model_fields_set else None + ), + ) + litellm.logging_callback_manager.add_litellm_callback( # pyright: ignore[reportUnknownMemberType] # callback manager is untyped + _callback + ) + return _callback + + +guardrail_initializer_registry: Final = { + SupportedGuardrailIntegrations.TYPESAFE.value: initialize_guardrail, +} + +guardrail_class_registry: Final = { + SupportedGuardrailIntegrations.TYPESAFE.value: TypeSafeGuardrail, +} diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py new file mode 100644 index 00000000000..ef192030e95 --- /dev/null +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py @@ -0,0 +1,390 @@ +"""TypeSafe (Jev) relevance-based compaction guardrail. + +Instead of summarizing tool output, the guardrail asks TypeSafe's Jev model +one yes/no question per completed tool exchange ("is this result still needed +for the current task?") over ``POST {api_base}/v1/systemone`` and blanks the +tool results Jev judges no longer relevant. The assistant tool-call rows stay +intact, so the conversation remains well-formed while the dead context stops +consuming input tokens. + +Exchanges follow litellm's own compression protection policy: system rows, the +last user row, and the last assistant row (which, expanded over its tool +exchange, covers the most recent exchange) are never evaluated or rewritten. +""" + +from __future__ import annotations + +import asyncio +import time +from collections.abc import Mapping, Sequence +from typing import TYPE_CHECKING, Annotated, Final, Literal, TypeGuard, cast + +import httpx +from fastapi import HTTPException +from httpx import Response as HttpxResponse +from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError + +import litellm +from litellm._logging import verbose_proxy_logger +from litellm.compression.compress import get_protected_indices +from litellm.integrations.custom_guardrail import ( + CustomGuardrail, + log_guardrail_information, # pyright: ignore[reportUnknownVariableType] # decorator is untyped in custom_guardrail +) +from litellm.litellm_core_utils.prompt_templates.factory import group_tool_exchanges +from litellm.llms.custom_httpx.http_handler import ( + AsyncHTTPHandler, + get_async_httpx_client, # pyright: ignore[reportUnknownVariableType] # helper is untyped in http_handler + httpxSpecialProvider, +) +from litellm.proxy.guardrails.guardrail_hooks.content_text import content_to_text +from litellm.secret_managers.main import get_secret_str +from litellm.types.guardrails import GuardrailEventHooks, Mode +from litellm.types.utils import GenericGuardrailAPIInputs + +if TYPE_CHECKING: + from litellm.litellm_core_utils.litellm_logging import ( + Logging as LiteLLMLoggingObj, + ) + from litellm.types.proxy.guardrails.guardrail_hooks.typesafe import ( + TypeSafeGuardrailConfigModel, + ) + +DEFAULT_API_BASE: Final = "https://api.typesafe.ai" +DEFAULT_MODEL: Final = "jev-latest" +DEFAULT_RELEVANCE_THRESHOLD: Final = 0.2 +DEFAULT_MIN_CHARS_TO_EVALUATE: Final = 200 +DEFAULT_MAX_RESULT_CHARS_IN_STATE: Final = 4000 +_MAX_EXCHANGES_EVALUATED: Final = 200 +# The shared GuardrailCallback client carries no per-call bound; an on-request +# guardrail must not hold the caller's request for the client's pooled timeout. +_JEV_TIMEOUT_SECONDS: Final = 30.0 +DROPPED_RESULT_TEXT: Final = ( + "[Tool result removed by TypeSafe compaction: judged no longer relevant to the current task]" +) + + +def _is_str_object_dict(value: object) -> TypeGuard[dict[str, object]]: # guard-ok: isinstance narrows correctly; predicate is trivially correct # fmt: skip + return isinstance(value, dict) + + +def _is_object_list(value: object) -> TypeGuard[list[object]]: # guard-ok: isinstance narrows correctly; predicate is trivially correct # fmt: skip + return isinstance(value, list) + + +def _safe_response_text(response: object, limit: int = 500) -> str: + try: + text: Final = getattr(response, "text", "") + except httpx.DecodingError: + return "" + return (text or "")[:limit] + + +class _JevNoulAnswer(BaseModel): + model_config = ConfigDict(frozen=True, allow_inf_nan=False) + + type: Literal["noul"] + noul: Annotated[float, Field(ge=0.0, le=1.0)] + + +class _JevSystemOneResponse(BaseModel): + model_config = ConfigDict(frozen=True) + + answers: Mapping[str, _JevNoulAnswer] + + +_JEV_RESPONSE_ADAPTER: Final = TypeAdapter(_JevSystemOneResponse) + + +def _question_instructions(question_id: str) -> str: + return ( + f"Is tool exchange `{question_id}` in `tool_exchanges` still needed by the assistant to " + "complete `task`? Answer yes if its result contains information the assistant has not yet " + "fully used or will need again; answer no if it is off-topic, superseded, or already " + "incorporated into later messages." + ) + + +def _tool_call_entries(assistant_message: Mapping[str, object]) -> list[dict[str, object]]: + tool_calls: Final = assistant_message.get("tool_calls") + if not _is_object_list(tool_calls): + return [] + entries: Final[list[dict[str, object]]] = [] + for tool_call in tool_calls: + if not _is_str_object_dict(tool_call): + continue + function = tool_call.get("function") + fn = function if _is_str_object_dict(function) else tool_call + entries.append({"name": fn.get("name"), "arguments": fn.get("arguments")}) + return entries + + +def _protected_indices(messages: Sequence[Mapping[str, object]]) -> frozenset[int]: + """Rows typesafe must not rewrite, expanded over whole tool exchanges. + + ``get_protected_indices`` covers system rows, the last user row, the last + assistant row, and cache_control prefixes. Expanding over exchanges keeps an + exchange atomic: the last assistant row protects its own tool results too, + so the most recent exchange is never evaluated. + """ + protected: Final = frozenset(get_protected_indices(messages)) + return protected | frozenset( + index + for group in group_tool_exchanges(messages) + if any(member in protected for member in group) + for index in group + ) + + +class TypeSafeGuardrail(CustomGuardrail): + def __init__( + self, + api_base: str | None = None, + api_key: str | None = None, + model: str | None = None, + relevance_threshold: float | None = None, + min_chars_to_evaluate: int | None = None, + max_result_chars_in_state: int | None = None, + unreachable_fallback: str | None = None, + guardrail_name: str | None = None, + event_hook: GuardrailEventHooks | list[GuardrailEventHooks] | Mode | None = None, + default_on: bool = False, + async_handler: AsyncHTTPHandler | None = None, + ): + raw_api_base: Final = (api_base or get_secret_str("TYPESAFE_API_BASE") or DEFAULT_API_BASE).rstrip("/") + self.typesafe_api_base = raw_api_base + self.typesafe_api_key = api_key or get_secret_str("TYPESAFE_API_KEY") + if not self.typesafe_api_key: + raise ValueError( + "TypeSafe guardrail requires an API key. Set `api_key` in the " + "guardrail config or the TYPESAFE_API_KEY env var." + ) + self.jev_model = model or DEFAULT_MODEL + self.relevance_threshold = DEFAULT_RELEVANCE_THRESHOLD if relevance_threshold is None else relevance_threshold + self.min_chars_to_evaluate = ( + DEFAULT_MIN_CHARS_TO_EVALUATE if min_chars_to_evaluate is None else min_chars_to_evaluate + ) + self.max_result_chars_in_state = ( + DEFAULT_MAX_RESULT_CHARS_IN_STATE if max_result_chars_in_state is None else max_result_chars_in_state + ) + self.unreachable_fallback: Literal["fail_closed", "fail_open"] = ( + "fail_closed" if unreachable_fallback == "fail_closed" else "fail_open" + ) + self.async_handler: AsyncHTTPHandler = async_handler or get_async_httpx_client( + llm_provider=httpxSpecialProvider.GuardrailCallback, + ) + super().__init__( # pyright: ignore[reportUnknownMemberType] # CustomGuardrail.__init__ is untyped + guardrail_name=guardrail_name, + event_hook=event_hook, + default_on=default_on, + ) + + def _handle_failure(self, error: str, log_detail: dict[str, object]) -> None: + """fail_open logs and the caller forwards uncompacted; fail_closed raises. + Upstream bodies go to server logs only; the raised HTTPException is generic.""" + if self.unreachable_fallback == "fail_open": + verbose_proxy_logger.warning( + "TypeSafe: %s; fail_open configured, forwarding request uncompacted. detail=%s", + error, + log_detail, + ) + return + verbose_proxy_logger.error("TypeSafe: %s. detail=%s", error, log_detail) + raise HTTPException(status_code=500, detail={"error": error}) + + def _candidate_exchanges(self, messages: list[dict[str, object]]) -> list[tuple[int, ...]]: + """Message-index groups eligible for relevance evaluation, oldest first. + + A candidate is a completed tool exchange: an assistant row that made + tool calls plus at least one ``tool``/``function`` row answering it, + with no member protected, and enough combined tool-result text to be + worth an evaluation call. + """ + protected: Final = _protected_indices(messages) + candidates: Final[list[tuple[int, ...]]] = [] + for group in group_tool_exchanges(messages): + if len(group) < 2: + continue + if messages[group[0]].get("role") != "assistant": + continue + if any(member in protected for member in group): + continue + tool_text = "".join( + content_to_text(messages[index].get("content")) + for index in group[1:] + if messages[index].get("role") in ("tool", "function") + ) + if not tool_text or len(tool_text) < self.min_chars_to_evaluate: + continue + candidates.append(group) + return candidates[-_MAX_EXCHANGES_EVALUATED:] + + def _build_state(self, messages: list[dict[str, object]], candidates: list[tuple[int, ...]]) -> dict[str, object]: + task: Final = next( + ( + content_to_text(messages[index].get("content")) + for index in range(len(messages) - 1, -1, -1) + if messages[index].get("role") == "user" + ), + "", + ) + system: Final = "\n\n".join( + content_to_text(message.get("content")) for message in messages if message.get("role") == "system" + ) + tool_exchanges: Final[dict[str, object]] = {} + for ordinal, group in enumerate(candidates): + result_text = "".join( + content_to_text(messages[index].get("content")) + for index in group[1:] + if messages[index].get("role") in ("tool", "function") + ) + tool_exchanges[f"e{ordinal}"] = { + "tool_calls": _tool_call_entries(messages[group[0]]), + "result": result_text[: self.max_result_chars_in_state], + } + return {"task": task, "system": system, "tool_exchanges": tool_exchanges} + + async def _call_systemone(self, state: dict[str, object], question_ids: list[str]) -> _JevSystemOneResponse | None: + """Evaluate each exchange. Returns the response, or None when the service + failed and fail_open applies.""" + payload: Final[dict[str, object]] = { + "model": self.jev_model, + "state": state, + "questions": { + question_id: {"type": "noul", "instructions": _question_instructions(question_id)} + for question_id in question_ids + }, + } + try: + raw_response: HttpxResponse = await self.async_handler.post( # pyright: ignore[reportUnknownMemberType] # AsyncHTTPHandler.post is untyped + url=f"{self.typesafe_api_base}/v1/systemone", + json=payload, + headers={ + "Authorization": f"Bearer {self.typesafe_api_key}", + "Content-Type": "application/json", + }, + timeout=_JEV_TIMEOUT_SECONDS, + ) + except asyncio.CancelledError: + raise + except httpx.HTTPStatusError as e: + resp: Final = getattr(e, "response", None) + self._handle_failure( + "TypeSafe evaluation service returned an error", + {"status_code": getattr(resp, "status_code", None), "body": _safe_response_text(resp)}, + ) + return None + except (httpx.RequestError, litellm.Timeout) as e: + self._handle_failure("TypeSafe evaluation service request failed", {"detail": str(e)}) + return None + except Exception as e: + self._handle_failure("TypeSafe evaluation service request failed", {"detail": str(e)}) + return None + if not 200 <= raw_response.status_code < 300: + self._handle_failure( + "TypeSafe evaluation service returned an error", + {"status_code": raw_response.status_code, "body": _safe_response_text(raw_response)}, + ) + return None + try: + body: Final = cast(object, raw_response.json()) + except (ValueError, httpx.DecodingError, RecursionError): + self._handle_failure( + "TypeSafe evaluation service returned an unreadable response", + {"body": _safe_response_text(raw_response)}, + ) + return None + try: + return _JEV_RESPONSE_ADAPTER.validate_python(body) + except ValidationError: + self._handle_failure( + "TypeSafe evaluation service returned unexpected response shape", + {"body": _safe_response_text(raw_response)}, + ) + return None + + @log_guardrail_information + async def apply_guardrail( + self, + inputs: GenericGuardrailAPIInputs, + request_data: dict[str, object], + input_type: Literal["request", "response"], + logging_obj: LiteLLMLoggingObj | None = None, + ) -> GenericGuardrailAPIInputs: + if input_type != "request": + return inputs + + structured_messages: Final = inputs.get("structured_messages") + if not _is_object_list(structured_messages) or not structured_messages: + return inputs + messages: Final = [m for m in structured_messages if _is_str_object_dict(m)] + if len(messages) != len(structured_messages): + return inputs + + candidates: Final = self._candidate_exchanges(messages) + if not candidates: + verbose_proxy_logger.debug("TypeSafe: no completed tool exchanges eligible for evaluation") + return inputs + + question_ids: Final = [f"e{ordinal}" for ordinal in range(len(candidates))] + state: Final = self._build_state(messages, candidates) + + start_time: Final = time.monotonic() + response: Final = await self._call_systemone(state, question_ids) + end_time: Final = time.monotonic() + if response is None: + return inputs + + dropped_ordinals: Final = frozenset( + ordinal + for ordinal in range(len(candidates)) + if (answer := response.answers.get(f"e{ordinal}")) is not None and answer.noul < self.relevance_threshold + ) + dropped_tool_indices: Final[frozenset[int]] = frozenset( + index + for ordinal in dropped_ordinals + for index in candidates[ordinal][1:] + if messages[index].get("role") in ("tool", "function") + ) + if not dropped_tool_indices: + verbose_proxy_logger.debug("TypeSafe: all evaluated exchanges still relevant; request unchanged") + return inputs + + compacted_messages: Final = [ + {**message, "content": DROPPED_RESULT_TEXT} if index in dropped_tool_indices else message + for index, message in enumerate(messages) + ] + chars_removed: Final = sum( + len(content_to_text(messages[index].get("content"))) - len(DROPPED_RESULT_TEXT) + for index in dropped_tool_indices + ) + exchanges_dropped: Final = len(dropped_ordinals) + verbose_proxy_logger.info( + "TypeSafe: evaluated %s tool exchange(s), dropped %s, ~%s chars removed", + len(candidates), + exchanges_dropped, + chars_removed, + ) + self.add_standard_logging_guardrail_information_to_request_data( # pyright: ignore[reportUnknownMemberType] # untyped base helper + guardrail_json_response={ + "exchanges_evaluated": len(candidates), + "exchanges_dropped": exchanges_dropped, + "chars_removed": chars_removed, + "model": self.jev_model, + }, + request_data=request_data, + guardrail_status="success", + guardrail_provider="typesafe", + start_time=start_time, + end_time=end_time, + duration=end_time - start_time, + ) + return {**inputs, "structured_messages": compacted_messages} # pyright: ignore[reportReturnType] # plain dicts satisfy AllMessageValues at runtime + + @staticmethod + def get_config_model() -> type[TypeSafeGuardrailConfigModel] | None: + from litellm.types.proxy.guardrails.guardrail_hooks.typesafe import ( + TypeSafeGuardrailConfigModel, + ) + + return TypeSafeGuardrailConfigModel diff --git a/litellm/types/guardrails.py b/litellm/types/guardrails.py index 6346e13f3ba..400cadd69e7 100644 --- a/litellm/types/guardrails.py +++ b/litellm/types/guardrails.py @@ -62,6 +62,9 @@ from litellm.types.proxy.guardrails.guardrail_hooks.singulr import ( from litellm.types.proxy.guardrails.guardrail_hooks.tool_permission import ( ToolPermissionGuardrailConfigModel, ) +from litellm.types.proxy.guardrails.guardrail_hooks.typesafe import ( + TypeSafeGuardrailConfigModel, +) from litellm.types.proxy.guardrails.guardrail_hooks.vigil_guard import ( VigilGuardGuardrailConfigModel, ) @@ -138,6 +141,7 @@ class SupportedGuardrailIntegrations(Enum): SINGULR = "singulr" HEADROOM = "headroom" COMPRESR = "compresr" + TYPESAFE = "typesafe" STRAIKER = "straiker" ALICE = "alice" AGENT_365 = "agent_365" @@ -1055,7 +1059,7 @@ class BaseLitellmParams(ContentFilterConfigModel): # works for new and patch up default="fail_closed", description=( "Behavior when a guardrail endpoint is unreachable due to network errors. " - "Implemented by guardrail='generic_guardrail_api', 'agent_365', 'akto', 'vigil_guard', 'repelloai', 'headroom', and 'compresr'. " + "Implemented by guardrail='generic_guardrail_api', 'agent_365', 'akto', 'vigil_guard', 'repelloai', 'headroom', 'compresr', and 'typesafe'. " "'fail_closed' raises an error (default). 'fail_open' logs a critical error and allows the request to proceed." ), ) @@ -1171,6 +1175,7 @@ class LitellmParams( # pyright: ignore[reportIncompatibleVariableOverride] # o LakeraV2GuardrailConfigModel, HeadroomGuardrailConfigModel, CompresrGuardrailConfigModel, + TypeSafeGuardrailConfigModel, RepelloAIGuardrailConfigModel, LassoGuardrailConfigModel, DeepKeepGuardrailConfigModel, diff --git a/litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py b/litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py new file mode 100644 index 00000000000..4e742bfc7be --- /dev/null +++ b/litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py @@ -0,0 +1,58 @@ +from typing import Literal + +from pydantic import BaseModel, Field + +from .base import GuardrailConfigModel + + +class TypeSafeGuardrailOptionalParams(BaseModel): + """Optional tuning knobs for the TypeSafe (Jev) compaction guardrail.""" + + relevance_threshold: float | None = Field( + default=None, + description=( + "Relevance cutoff in [0, 1]. A completed tool exchange is dropped when Jev " + "scores the probability that it is still needed below this value. Defaults to 0.2." + ), + ) + min_chars_to_evaluate: int | None = Field( + default=None, + description=( + "Skip tool exchanges whose combined tool-result text is shorter than this many characters. Defaults to 200." + ), + ) + max_result_chars_in_state: int | None = Field( + default=None, + description=( + "Tool result text is truncated to this many characters when sent to the Jev evaluator. Defaults to 4000." + ), + ) + + +class TypeSafeGuardrailConfigModel(GuardrailConfigModel[TypeSafeGuardrailOptionalParams]): + api_key: str | None = Field( + default=None, + description="TypeSafe API key, sent as a Bearer token. Falls back to the TYPESAFE_API_KEY env var.", + ) + api_base: str | None = Field( + default=None, + description=( + "Base URL of the TypeSafe API. Falls back to the TYPESAFE_API_BASE env var, then https://api.typesafe.ai." + ), + ) + model: str | None = Field( + default=None, + description="TypeSafe evaluation model (not the LLM). Defaults to 'jev-latest'.", + ) + unreachable_fallback: Literal["fail_closed", "fail_open"] = Field( + default="fail_open", + description=( + "Behavior when the TypeSafe evaluation service is unreachable or errors. " + "'fail_open' (default) forwards the request uncompacted. 'fail_closed' " + "raises an error instead." + ), + ) + + @staticmethod + def ui_friendly_name() -> str: + return "TypeSafe (Jev) Compaction" diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py new file mode 100644 index 00000000000..293e4f8c344 --- /dev/null +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py @@ -0,0 +1,274 @@ +""" +Unit tests for the TypeSafe (Jev) compaction guardrail. + +Tests cover: +- exchanges scored below relevance_threshold have their tool rows blanked while + assistant tool-call rows and kept exchanges pass through verbatim, without + mutating the caller's message list +- protected rows (system, last user, and the last tool exchange via the + last-assistant rule) are never sent to Jev even when long +- exchanges under min_chars_to_evaluate are skipped +- request shape: POST {api_base}/v1/systemone with Bearer auth, one noul + question per candidate keyed e, task = last user text, results truncated + to max_result_chars_in_state +- identity return when there are no candidates or nothing is dropped +- fail_open forwards uncompacted on service failure; fail_closed raises +- response input_type passthrough and initialize_guardrail wiring +""" + +from unittest.mock import AsyncMock, MagicMock + +import pytest +from fastapi import HTTPException + +from litellm.proxy.guardrails.guardrail_hooks.typesafe import ( + TypeSafeGuardrail, + guardrail_class_registry, + guardrail_initializer_registry, + initialize_guardrail, +) +from litellm.proxy.guardrails.guardrail_hooks.typesafe.typesafe import DROPPED_RESULT_TEXT +from litellm.types.guardrails import SupportedGuardrailIntegrations +from litellm.types.utils import GenericGuardrailAPIInputs + +FAKE_API_BASE = "https://typesafe.example.com" +FAKE_API_KEY = "ts_test-key" + +SYSTEM_TEXT = "You are a research assistant." +USER_TEXT = "Which 2026 EV has the longest range?" +TOOL_OUTPUT_LONG = "Result: EV range comparison. " * 40 # > 200 chars +TOOL_OUTPUT_SHORT = "short" + + +def _exchange(call_id: str, tool_text: str, name: str = "web_search") -> list[dict]: + return [ + { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": call_id, + "type": "function", + "function": {"name": name, "arguments": '{"query": "ev"}'}, + } + ], + }, + {"role": "tool", "tool_call_id": call_id, "name": name, "content": tool_text}, + ] + + +def _messages(*, tail: list | None = None) -> list[dict]: + base = [ + {"role": "system", "content": SYSTEM_TEXT}, + {"role": "user", "content": USER_TEXT}, + ] + return base + (tail or []) + + +def _make_guardrail(handler: MagicMock | None = None, **kwargs) -> TypeSafeGuardrail: + defaults = dict( + api_base=FAKE_API_BASE, + api_key=FAKE_API_KEY, + guardrail_name="typesafe", + default_on=True, + async_handler=handler or _make_handler({"e0": 0.9}), + ) + defaults.update(kwargs) + return TypeSafeGuardrail(**defaults) + + +def _make_handler(answers: dict[str, float], status: int = 200) -> MagicMock: + response = MagicMock() + response.status_code = status + response.json.return_value = { + "model": "jev-1.13.0", + "answers": {qid: {"type": "noul", "noul": score} for qid, score in answers.items()}, + "usage": {"input_tokens": 10, "output_tokens": 1}, + } + response.text = "" + handler = MagicMock() + handler.post = AsyncMock(return_value=response) + return handler + + +def _inputs(messages: list) -> GenericGuardrailAPIInputs: + return GenericGuardrailAPIInputs(structured_messages=messages) + + +async def _apply(guardrail: TypeSafeGuardrail, messages: list, input_type: str = "request"): + return await guardrail.apply_guardrail( + inputs=_inputs(messages), + request_data={}, + input_type=input_type, # pyright: ignore[reportArgumentType] # test uses the same literal domain + logging_obj=None, + ) + + +@pytest.mark.asyncio +async def test_low_noul_exchange_blanked_high_kept_and_input_not_mutated(): + handler = _make_handler({"e0": 0.1, "e1": 0.95}) + guardrail = _make_guardrail(handler) + messages = _messages( + tail=[ + *_exchange("call_1", TOOL_OUTPUT_LONG), + *_exchange("call_2", TOOL_OUTPUT_LONG), + {"role": "assistant", "content": "still thinking"}, + ] + ) + snapshot = [dict(m) for m in messages] + + result = await _apply(guardrail, messages) + out = result["structured_messages"] + + assert out[3]["content"] == DROPPED_RESULT_TEXT + assert out[3]["tool_call_id"] == "call_1" + assert out[3]["role"] == "tool" + assert out[5]["content"] == TOOL_OUTPUT_LONG + assert out[2] == messages[2] + assert out[4] == messages[4] + assert out[6]["content"] == "still thinking" + assert messages == snapshot + + +@pytest.mark.asyncio +async def test_last_exchange_and_protected_rows_never_evaluated(): + handler = _make_handler({"e0": 0.05}) + guardrail = _make_guardrail(handler) + # Ends on a tool result: the last assistant row is protected, so the whole + # last exchange is out of scope even though its text is long. + messages = _messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), *_exchange("call_2", TOOL_OUTPUT_LONG)]) + + result = await _apply(guardrail, messages) + + payload = handler.post.call_args.kwargs["json"] + assert list(payload["questions"]) == ["e0"] + assert list(payload["state"]["tool_exchanges"]) == ["e0"] + assert payload["state"]["task"] == USER_TEXT + assert payload["state"]["system"] == SYSTEM_TEXT + out = result["structured_messages"] + assert out[3]["content"] == DROPPED_RESULT_TEXT + assert out[5]["content"] == TOOL_OUTPUT_LONG + + +@pytest.mark.asyncio +async def test_short_exchange_not_sent(): + handler = _make_handler({"e0": 0.9}) + guardrail = _make_guardrail(handler) + messages = _messages( + tail=[ + *_exchange("call_1", TOOL_OUTPUT_SHORT), + *_exchange("call_2", TOOL_OUTPUT_LONG), + {"role": "assistant", "content": "done"}, + ] + ) + result = await _apply(guardrail, messages) + payload = handler.post.call_args.kwargs["json"] + assert list(payload["questions"]) == ["e0"] + exchange = payload["state"]["tool_exchanges"]["e0"] + assert exchange["result"] == TOOL_OUTPUT_LONG + assert result is not None + + +@pytest.mark.asyncio +async def test_request_body_shape_and_truncation(): + handler = _make_handler({"e0": 0.9}) + guardrail = _make_guardrail(handler, max_result_chars_in_state=50) + messages = _messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "done"}]) + await _apply(guardrail, messages) + + kwargs = handler.post.call_args.kwargs + assert kwargs["url"].endswith("/v1/systemone") + assert kwargs["url"].startswith(FAKE_API_BASE) + assert kwargs["headers"]["Authorization"] == f"Bearer {FAKE_API_KEY}" + assert kwargs["headers"]["Content-Type"] == "application/json" + payload = kwargs["json"] + assert payload["model"] == "jev-latest" + assert list(payload["questions"]) == ["e0"] + assert payload["questions"]["e0"]["type"] == "noul" + assert "e0" in payload["questions"]["e0"]["instructions"] + assert payload["state"]["task"] == USER_TEXT + exchange = payload["state"]["tool_exchanges"]["e0"] + assert exchange["result"] == TOOL_OUTPUT_LONG[:50] + assert exchange["tool_calls"] == [{"name": "web_search", "arguments": '{"query": "ev"}'}] + + +@pytest.mark.asyncio +async def test_no_candidates_returns_identity_and_skips_http(): + handler = _make_handler({}) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[{"role": "assistant", "content": "plain answer"}])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert result is inputs + handler.post.assert_not_called() + + +@pytest.mark.asyncio +async def test_all_above_threshold_returns_identity(): + handler = _make_handler({"e0": 0.9}) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert result is inputs + + +@pytest.mark.asyncio +async def test_fail_open_returns_inputs_on_exception(): + handler = MagicMock() + handler.post = AsyncMock(side_effect=Exception("connection refused")) + guardrail = _make_guardrail(handler, unreachable_fallback="fail_open") + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert result is inputs + + +@pytest.mark.asyncio +async def test_fail_closed_raises_http_exception(): + handler = MagicMock() + handler.post = AsyncMock(side_effect=Exception("connection refused")) + guardrail = _make_guardrail(handler, unreachable_fallback="fail_closed") + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + with pytest.raises(HTTPException): + await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + + +@pytest.mark.asyncio +async def test_fail_open_on_non_2xx(): + handler = _make_handler({"e0": 0.9}, status=500) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert result is inputs + + +@pytest.mark.asyncio +async def test_response_input_type_passthrough(): + handler = _make_handler({"e0": 0.05}) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG)])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="response", logging_obj=None) + assert result is inputs + handler.post.assert_not_called() + + +def test_initialize_guardrail_applies_optional_params_and_registry_keys(): + from litellm.types.guardrails import LitellmParams + + litellm_params = LitellmParams( + guardrail="typesafe", + mode="pre_call", + api_key=FAKE_API_KEY, + api_base=FAKE_API_BASE, + optional_params={ + "relevance_threshold": 0.5, + "min_chars_to_evaluate": 10, + "max_result_chars_in_state": 100, + }, + ) + callback = initialize_guardrail(litellm_params, {"guardrail_name": "jev-compaction"}) + assert isinstance(callback, TypeSafeGuardrail) + assert callback.relevance_threshold == 0.5 + assert callback.min_chars_to_evaluate == 10 + assert callback.max_result_chars_in_state == 100 + assert callback.unreachable_fallback == "fail_open" + assert guardrail_initializer_registry[SupportedGuardrailIntegrations.TYPESAFE.value] is initialize_guardrail + assert guardrail_class_registry[SupportedGuardrailIntegrations.TYPESAFE.value] is TypeSafeGuardrail diff --git a/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.ts b/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.ts index c3503f78afb..292b497f6df 100644 --- a/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.ts +++ b/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.ts @@ -14,7 +14,7 @@ export const NO_COMPRESSION = "none"; /** Guardrail providers that compress prompts, mirroring COMPRESSION_GUARDRAIL_PROVIDERS in * litellm/proxy/guardrails/auto_router_compression.py. Both are selectable per hop. */ -export const COMPRESSION_GUARDRAIL_PROVIDERS: readonly string[] = ["headroom", "compresr"]; +export const COMPRESSION_GUARDRAIL_PROVIDERS: readonly string[] = ["headroom", "compresr", "typesafe"]; export const isCompressionGuardrailProvider = (provider: unknown): boolean => typeof provider === "string" && COMPRESSION_GUARDRAIL_PROVIDERS.includes(provider.toLowerCase()); diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index fd882937e79..16d61bd9b0f 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -24272,7 +24272,7 @@ export interface components { timeout?: number | null; /** * Unreachable Fallback - * @description Behavior when a guardrail endpoint is unreachable due to network errors. Implemented by guardrail='generic_guardrail_api', 'agent_365', 'akto', 'vigil_guard', 'repelloai', 'headroom', and 'compresr'. 'fail_closed' raises an error (default). 'fail_open' logs a critical error and allows the request to proceed. + * @description Behavior when a guardrail endpoint is unreachable due to network errors. Implemented by guardrail='generic_guardrail_api', 'agent_365', 'akto', 'vigil_guard', 'repelloai', 'headroom', 'compresr', and 'typesafe'. 'fail_closed' raises an error (default). 'fail_open' logs a critical error and allows the request to proceed. * @default fail_closed * @enum {string} */ From dffb6a38d95f5e2fa2c259f06f50243c46ca2ab2 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 04:35:58 +0000 Subject: [PATCH 072/141] refactor(guardrails): tighten typesafe guardrail typing and error handling Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/typesafe/__init__.py | 42 +++++-------- .../guardrail_hooks/typesafe/typesafe.py | 63 +++++++------------ .../guardrail_hooks/test_typesafe.py | 3 +- 3 files changed, 40 insertions(+), 68 deletions(-) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py index c347c863a1b..40f040eb676 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py @@ -1,12 +1,17 @@ from __future__ import annotations -from typing import TYPE_CHECKING, Final, cast +from typing import TYPE_CHECKING, Final + +from pydantic import BaseModel from litellm.types.guardrails import ( GuardrailEventHooks, Mode, SupportedGuardrailIntegrations, ) +from litellm.types.proxy.guardrails.guardrail_hooks.typesafe import ( + TypeSafeGuardrailOptionalParams, +) from .typesafe import TypeSafeGuardrail @@ -24,40 +29,27 @@ def _coerce_event_hook( return GuardrailEventHooks(mode) -def _get_optional_value(litellm_params: LitellmParams, optional_params: object | None, attribute_name: str) -> object: - if optional_params is not None: - value: Final = getattr(optional_params, attribute_name, None) - if value is not None: - return cast(object, value) - return cast(object, getattr(litellm_params, attribute_name, None)) - - -def _optional_float(litellm_params: LitellmParams, optional_params: object | None, attribute_name: str) -> float | None: - value: Final = _get_optional_value(litellm_params, optional_params, attribute_name) - if isinstance(value, bool) or not isinstance(value, (int, float)): - return None - return float(value) - - -def _optional_int(litellm_params: LitellmParams, optional_params: object | None, attribute_name: str) -> int | None: - value: Final = _get_optional_value(litellm_params, optional_params, attribute_name) - if isinstance(value, bool) or not isinstance(value, int): - return None - return value +def _optional_params(litellm_params: LitellmParams) -> TypeSafeGuardrailOptionalParams: + value: Final = litellm_params.optional_params + if isinstance(value, TypeSafeGuardrailOptionalParams): + return value + if isinstance(value, BaseModel): + return TypeSafeGuardrailOptionalParams.model_validate(value.model_dump()) + return TypeSafeGuardrailOptionalParams() def initialize_guardrail(litellm_params: LitellmParams, guardrail: Guardrail) -> TypeSafeGuardrail: import litellm - optional_params: Final = getattr(litellm_params, "optional_params", None) + optional_params: Final = _optional_params(litellm_params) _callback: Final = TypeSafeGuardrail( api_base=litellm_params.api_base, api_key=litellm_params.api_key, model=litellm_params.model, - relevance_threshold=_optional_float(litellm_params, optional_params, "relevance_threshold"), - min_chars_to_evaluate=_optional_int(litellm_params, optional_params, "min_chars_to_evaluate"), - max_result_chars_in_state=_optional_int(litellm_params, optional_params, "max_result_chars_in_state"), + relevance_threshold=optional_params.relevance_threshold, + min_chars_to_evaluate=optional_params.min_chars_to_evaluate, + max_result_chars_in_state=optional_params.max_result_chars_in_state, guardrail_name=guardrail["guardrail_name"], event_hook=_coerce_event_hook(litellm_params.mode), default_on=litellm_params.default_on or False, diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py index ef192030e95..360219e0ad5 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py @@ -3,13 +3,7 @@ Instead of summarizing tool output, the guardrail asks TypeSafe's Jev model one yes/no question per completed tool exchange ("is this result still needed for the current task?") over ``POST {api_base}/v1/systemone`` and blanks the -tool results Jev judges no longer relevant. The assistant tool-call rows stay -intact, so the conversation remains well-formed while the dead context stops -consuming input tokens. - -Exchanges follow litellm's own compression protection policy: system rows, the -last user row, and the last assistant row (which, expanded over its tool -exchange, covers the most recent exchange) are never evaluated or rewritten. +tool results Jev judges no longer relevant. """ from __future__ import annotations @@ -24,7 +18,6 @@ from fastapi import HTTPException from httpx import Response as HttpxResponse from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError -import litellm from litellm._logging import verbose_proxy_logger from litellm.compression.compress import get_protected_indices from litellm.integrations.custom_guardrail import ( @@ -56,8 +49,6 @@ DEFAULT_RELEVANCE_THRESHOLD: Final = 0.2 DEFAULT_MIN_CHARS_TO_EVALUATE: Final = 200 DEFAULT_MAX_RESULT_CHARS_IN_STATE: Final = 4000 _MAX_EXCHANGES_EVALUATED: Final = 200 -# The shared GuardrailCallback client carries no per-call bound; an on-request -# guardrail must not hold the caller's request for the client's pooled timeout. _JEV_TIMEOUT_SECONDS: Final = 30.0 DROPPED_RESULT_TEXT: Final = ( "[Tool result removed by TypeSafe compaction: judged no longer relevant to the current task]" @@ -72,9 +63,11 @@ def _is_object_list(value: object) -> TypeGuard[list[object]]: # guard-ok: isin return isinstance(value, list) -def _safe_response_text(response: object, limit: int = 500) -> str: +def _safe_response_text(response: HttpxResponse | None, limit: int = 500) -> str: + if response is None: + return "" try: - text: Final = getattr(response, "text", "") + text: Final = response.text except httpx.DecodingError: return "" return (text or "")[:limit] @@ -120,13 +113,7 @@ def _tool_call_entries(assistant_message: Mapping[str, object]) -> list[dict[str def _protected_indices(messages: Sequence[Mapping[str, object]]) -> frozenset[int]: - """Rows typesafe must not rewrite, expanded over whole tool exchanges. - - ``get_protected_indices`` covers system rows, the last user row, the last - assistant row, and cache_control prefixes. Expanding over exchanges keeps an - exchange atomic: the last assistant row protects its own tool results too, - so the most recent exchange is never evaluated. - """ + """``get_protected_indices`` expanded over whole tool exchanges, so the most recent exchange is never evaluated.""" protected: Final = frozenset(get_protected_indices(messages)) return protected | frozenset( index @@ -180,8 +167,7 @@ class TypeSafeGuardrail(CustomGuardrail): ) def _handle_failure(self, error: str, log_detail: dict[str, object]) -> None: - """fail_open logs and the caller forwards uncompacted; fail_closed raises. - Upstream bodies go to server logs only; the raised HTTPException is generic.""" + """fail_open logs and returns; fail_closed raises a generic 502 (upstream bodies stay in server logs).""" if self.unreachable_fallback == "fail_open": verbose_proxy_logger.warning( "TypeSafe: %s; fail_open configured, forwarding request uncompacted. detail=%s", @@ -190,16 +176,10 @@ class TypeSafeGuardrail(CustomGuardrail): ) return verbose_proxy_logger.error("TypeSafe: %s. detail=%s", error, log_detail) - raise HTTPException(status_code=500, detail={"error": error}) + raise HTTPException(status_code=502, detail={"error": error}) def _candidate_exchanges(self, messages: list[dict[str, object]]) -> list[tuple[int, ...]]: - """Message-index groups eligible for relevance evaluation, oldest first. - - A candidate is a completed tool exchange: an assistant row that made - tool calls plus at least one ``tool``/``function`` row answering it, - with no member protected, and enough combined tool-result text to be - worth an evaluation call. - """ + """Completed tool exchanges eligible for evaluation: unprotected, and long enough to be worth a call.""" protected: Final = _protected_indices(messages) candidates: Final[list[tuple[int, ...]]] = [] for group in group_tool_exchanges(messages): @@ -245,8 +225,7 @@ class TypeSafeGuardrail(CustomGuardrail): return {"task": task, "system": system, "tool_exchanges": tool_exchanges} async def _call_systemone(self, state: dict[str, object], question_ids: list[str]) -> _JevSystemOneResponse | None: - """Evaluate each exchange. Returns the response, or None when the service - failed and fail_open applies.""" + """Returns the response, or None when the service failed and fail_open applies.""" payload: Final[dict[str, object]] = { "model": self.jev_model, "state": state, @@ -267,18 +246,18 @@ class TypeSafeGuardrail(CustomGuardrail): ) except asyncio.CancelledError: raise - except httpx.HTTPStatusError as e: - resp: Final = getattr(e, "response", None) - self._handle_failure( - "TypeSafe evaluation service returned an error", - {"status_code": getattr(resp, "status_code", None), "body": _safe_response_text(resp)}, - ) - return None - except (httpx.RequestError, litellm.Timeout) as e: - self._handle_failure("TypeSafe evaluation service request failed", {"detail": str(e)}) - return None except Exception as e: - self._handle_failure("TypeSafe evaluation service request failed", {"detail": str(e)}) + detail: Final[dict[str, object]] = ( + { + "error_type": type(e).__name__, + "detail": str(e), + "status_code": e.response.status_code, + "body": _safe_response_text(e.response), + } + if isinstance(e, httpx.HTTPStatusError) + else {"error_type": type(e).__name__, "detail": str(e)} + ) + self._handle_failure("TypeSafe evaluation service request failed", detail) return None if not 200 <= raw_response.status_code < 300: self._handle_failure( diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py index 293e4f8c344..8b732091d7c 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py @@ -227,8 +227,9 @@ async def test_fail_closed_raises_http_exception(): handler.post = AsyncMock(side_effect=Exception("connection refused")) guardrail = _make_guardrail(handler, unreachable_fallback="fail_closed") inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) - with pytest.raises(HTTPException): + with pytest.raises(HTTPException) as exc_info: await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert exc_info.value.status_code == 502 @pytest.mark.asyncio From 27b566590123e494107d1b39bab5a411a8d045b0 Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 04:45:28 +0000 Subject: [PATCH 073/141] fix(dashscope): mark qwen3.8-flash as vision and video capable with explicit cache creation cost Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/model_prices_and_context_window_backup.json | 6 ++++++ model_prices_and_context_window.json | 6 ++++++ 2 files changed, 12 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 3409d9b4e46..e903b239881 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -16691,6 +16691,7 @@ "supports_vision": true }, "dashscope/qwen3.8-flash": { + "cache_creation_input_token_cost": 2e-07, "cache_read_input_token_cost": 1.6e-08, "input_cost_per_token": 1.5e-07, "litellm_provider": "dashscope", @@ -16705,6 +16706,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, "supports_web_search": true }, "dashscope/qwen3.8-omni-flash": { @@ -18632,6 +18635,7 @@ "supports_vision": true }, "qwen_ai_platform/qwen3.8-flash": { + "cache_creation_input_token_cost": 2e-07, "cache_read_input_token_cost": 1.6e-08, "input_cost_per_token": 1.5e-07, "litellm_provider": "qwen_ai_platform", @@ -18646,6 +18650,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, "supports_web_search": true }, "qwen_ai_platform/qwen3.8-omni-flash": { diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 3409d9b4e46..e903b239881 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -16691,6 +16691,7 @@ "supports_vision": true }, "dashscope/qwen3.8-flash": { + "cache_creation_input_token_cost": 2e-07, "cache_read_input_token_cost": 1.6e-08, "input_cost_per_token": 1.5e-07, "litellm_provider": "dashscope", @@ -16705,6 +16706,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, "supports_web_search": true }, "dashscope/qwen3.8-omni-flash": { @@ -18632,6 +18635,7 @@ "supports_vision": true }, "qwen_ai_platform/qwen3.8-flash": { + "cache_creation_input_token_cost": 2e-07, "cache_read_input_token_cost": 1.6e-08, "input_cost_per_token": 1.5e-07, "litellm_provider": "qwen_ai_platform", @@ -18646,6 +18650,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, + "supports_video_input": true, + "supports_vision": true, "supports_web_search": true }, "qwen_ai_platform/qwen3.8-omni-flash": { From ca8c9062d80f1ee5f2e34765d0ba3dbc4fd7af72 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 04:48:11 +0000 Subject: [PATCH 074/141] fix(guardrails): satisfy strict lint gates in typesafe guardrail Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/typesafe/typesafe.py | 44 ++++++++++++------- 1 file changed, 28 insertions(+), 16 deletions(-) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py index 360219e0ad5..9fca6eaed20 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py @@ -11,7 +11,7 @@ from __future__ import annotations import asyncio import time from collections.abc import Mapping, Sequence -from typing import TYPE_CHECKING, Annotated, Final, Literal, TypeGuard, cast +from typing import TYPE_CHECKING, Annotated, Final, Literal import httpx from fastapi import HTTPException @@ -55,12 +55,22 @@ DROPPED_RESULT_TEXT: Final = ( ) -def _is_str_object_dict(value: object) -> TypeGuard[dict[str, object]]: # guard-ok: isinstance narrows correctly; predicate is trivially correct # fmt: skip - return isinstance(value, dict) +_STR_OBJECT_DICT_ADAPTER: Final = TypeAdapter(dict[str, object]) +_OBJECT_LIST_ADAPTER: Final = TypeAdapter(list[object]) -def _is_object_list(value: object) -> TypeGuard[list[object]]: # guard-ok: isinstance narrows correctly; predicate is trivially correct # fmt: skip - return isinstance(value, list) +def _as_str_object_dict(value: object) -> dict[str, object] | None: + try: + return _STR_OBJECT_DICT_ADAPTER.validate_python(value) + except ValidationError: + return None + + +def _as_object_list(value: object) -> list[object] | None: + try: + return _OBJECT_LIST_ADAPTER.validate_python(value) + except ValidationError: + return None def _safe_response_text(response: HttpxResponse | None, limit: int = 500) -> str: @@ -99,15 +109,16 @@ def _question_instructions(question_id: str) -> str: def _tool_call_entries(assistant_message: Mapping[str, object]) -> list[dict[str, object]]: - tool_calls: Final = assistant_message.get("tool_calls") - if not _is_object_list(tool_calls): + tool_calls: Final = _as_object_list(assistant_message.get("tool_calls")) + if tool_calls is None: return [] entries: Final[list[dict[str, object]]] = [] for tool_call in tool_calls: - if not _is_str_object_dict(tool_call): + parsed_call = _as_str_object_dict(tool_call) + if parsed_call is None: continue - function = tool_call.get("function") - fn = function if _is_str_object_dict(function) else tool_call + function = _as_str_object_dict(parsed_call.get("function")) + fn = function if function is not None else parsed_call entries.append({"name": fn.get("name"), "arguments": fn.get("arguments")}) return entries @@ -137,7 +148,7 @@ class TypeSafeGuardrail(CustomGuardrail): event_hook: GuardrailEventHooks | list[GuardrailEventHooks] | Mode | None = None, default_on: bool = False, async_handler: AsyncHTTPHandler | None = None, - ): + ) -> None: raw_api_base: Final = (api_base or get_secret_str("TYPESAFE_API_BASE") or DEFAULT_API_BASE).rstrip("/") self.typesafe_api_base = raw_api_base self.typesafe_api_key = api_key or get_secret_str("TYPESAFE_API_KEY") @@ -266,7 +277,7 @@ class TypeSafeGuardrail(CustomGuardrail): ) return None try: - body: Final = cast(object, raw_response.json()) + body: Final[object] = raw_response.json() # pyright: ignore[reportAny] # httpx Response.json() is untyped except (ValueError, httpx.DecodingError, RecursionError): self._handle_failure( "TypeSafe evaluation service returned an unreadable response", @@ -293,12 +304,13 @@ class TypeSafeGuardrail(CustomGuardrail): if input_type != "request": return inputs - structured_messages: Final = inputs.get("structured_messages") - if not _is_object_list(structured_messages) or not structured_messages: + structured_messages: Final = _as_object_list(inputs.get("structured_messages")) + if not structured_messages: return inputs - messages: Final = [m for m in structured_messages if _is_str_object_dict(m)] - if len(messages) != len(structured_messages): + parsed_messages: Final = [_as_str_object_dict(m) for m in structured_messages] + if any(m is None for m in parsed_messages): return inputs + messages: Final = [m for m in parsed_messages if m is not None] candidates: Final = self._candidate_exchanges(messages) if not candidates: From fd4476b1308feba1da8e86cd162401ce1bb2da4c Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 04:49:50 +0000 Subject: [PATCH 075/141] fix(guardrails): bound typesafe tuning params, preserve result tail, log fail-open status Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/typesafe/typesafe.py | 26 +++++++++++++++++- .../guardrails/guardrail_hooks/typesafe.py | 7 ++++- .../guardrail_hooks/test_typesafe.py | 27 ++++++++++++------- 3 files changed, 49 insertions(+), 11 deletions(-) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py index 9fca6eaed20..b3cf5bcf800 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py @@ -53,6 +53,7 @@ _JEV_TIMEOUT_SECONDS: Final = 30.0 DROPPED_RESULT_TEXT: Final = ( "[Tool result removed by TypeSafe compaction: judged no longer relevant to the current task]" ) +_ELISION_MARKER: Final = "\n... [middle truncated] ...\n" _STR_OBJECT_DICT_ADAPTER: Final = TypeAdapter(dict[str, object]) @@ -99,6 +100,17 @@ class _JevSystemOneResponse(BaseModel): _JEV_RESPONSE_ADAPTER: Final = TypeAdapter(_JevSystemOneResponse) +def _truncate_for_state(text: str, max_chars: int) -> str: + """Keeps the head and tail within ``max_chars`` so Jev sees both ends of a long result.""" + if len(text) <= max_chars: + return text + if max_chars <= len(_ELISION_MARKER): + return text[:max_chars] + budget: Final = max_chars - len(_ELISION_MARKER) + head: Final = budget // 2 + return text[:head] + _ELISION_MARKER + text[len(text) - (budget - head) :] + + def _question_instructions(question_id: str) -> str: return ( f"Is tool exchange `{question_id}` in `tool_exchanges` still needed by the assistant to " @@ -231,7 +243,7 @@ class TypeSafeGuardrail(CustomGuardrail): ) tool_exchanges[f"e{ordinal}"] = { "tool_calls": _tool_call_entries(messages[group[0]]), - "result": result_text[: self.max_result_chars_in_state], + "result": _truncate_for_state(result_text, self.max_result_chars_in_state), } return {"task": task, "system": system, "tool_exchanges": tool_exchanges} @@ -324,6 +336,18 @@ class TypeSafeGuardrail(CustomGuardrail): response: Final = await self._call_systemone(state, question_ids) end_time: Final = time.monotonic() if response is None: + self.add_standard_logging_guardrail_information_to_request_data( # pyright: ignore[reportUnknownMemberType] # untyped base helper + guardrail_json_response={ + "error": "TypeSafe evaluation unavailable; request forwarded uncompacted", + "model": self.jev_model, + }, + request_data=request_data, + guardrail_status="guardrail_failed_to_respond", + guardrail_provider="typesafe", + start_time=start_time, + end_time=end_time, + duration=end_time - start_time, + ) return inputs dropped_ordinals: Final = frozenset( diff --git a/litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py b/litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py index 4e742bfc7be..59482d2e190 100644 --- a/litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py +++ b/litellm/types/proxy/guardrails/guardrail_hooks/typesafe.py @@ -10,6 +10,8 @@ class TypeSafeGuardrailOptionalParams(BaseModel): relevance_threshold: float | None = Field( default=None, + ge=0.0, + le=1.0, description=( "Relevance cutoff in [0, 1]. A completed tool exchange is dropped when Jev " "scores the probability that it is still needed below this value. Defaults to 0.2." @@ -17,14 +19,17 @@ class TypeSafeGuardrailOptionalParams(BaseModel): ) min_chars_to_evaluate: int | None = Field( default=None, + ge=0, description=( "Skip tool exchanges whose combined tool-result text is shorter than this many characters. Defaults to 200." ), ) max_result_chars_in_state: int | None = Field( default=None, + ge=1, description=( - "Tool result text is truncated to this many characters when sent to the Jev evaluator. Defaults to 4000." + "Tool result text is truncated to this many characters when sent to the Jev evaluator, " + "keeping the head and tail. Defaults to 4000." ), ) diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py index 8b732091d7c..e2e5fc2fefc 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py @@ -40,7 +40,7 @@ TOOL_OUTPUT_LONG = "Result: EV range comparison. " * 40 # > 200 chars TOOL_OUTPUT_SHORT = "short" -def _exchange(call_id: str, tool_text: str, name: str = "web_search") -> list[dict]: +def _exchange(call_id: str, tool_text: str, name: str = "web_search") -> list[dict[str, object]]: return [ { "role": "assistant", @@ -57,7 +57,7 @@ def _exchange(call_id: str, tool_text: str, name: str = "web_search") -> list[di ] -def _messages(*, tail: list | None = None) -> list[dict]: +def _messages(*, tail: list[dict[str, object]] | None = None) -> list[dict[str, object]]: base = [ {"role": "system", "content": SYSTEM_TEXT}, {"role": "user", "content": USER_TEXT}, @@ -65,16 +65,21 @@ def _messages(*, tail: list | None = None) -> list[dict]: return base + (tail or []) -def _make_guardrail(handler: MagicMock | None = None, **kwargs) -> TypeSafeGuardrail: - defaults = dict( +def _make_guardrail( + handler: MagicMock | None = None, + *, + max_result_chars_in_state: int | None = None, + unreachable_fallback: str | None = None, +) -> TypeSafeGuardrail: + return TypeSafeGuardrail( api_base=FAKE_API_BASE, api_key=FAKE_API_KEY, guardrail_name="typesafe", default_on=True, async_handler=handler or _make_handler({"e0": 0.9}), + max_result_chars_in_state=max_result_chars_in_state, + unreachable_fallback=unreachable_fallback, ) - defaults.update(kwargs) - return TypeSafeGuardrail(**defaults) def _make_handler(answers: dict[str, float], status: int = 200) -> MagicMock: @@ -91,11 +96,13 @@ def _make_handler(answers: dict[str, float], status: int = 200) -> MagicMock: return handler -def _inputs(messages: list) -> GenericGuardrailAPIInputs: +def _inputs(messages: list[dict[str, object]]) -> GenericGuardrailAPIInputs: return GenericGuardrailAPIInputs(structured_messages=messages) -async def _apply(guardrail: TypeSafeGuardrail, messages: list, input_type: str = "request"): +async def _apply( + guardrail: TypeSafeGuardrail, messages: list[dict[str, object]], input_type: str = "request" +) -> GenericGuardrailAPIInputs: return await guardrail.apply_guardrail( inputs=_inputs(messages), request_data={}, @@ -188,7 +195,9 @@ async def test_request_body_shape_and_truncation(): assert "e0" in payload["questions"]["e0"]["instructions"] assert payload["state"]["task"] == USER_TEXT exchange = payload["state"]["tool_exchanges"]["e0"] - assert exchange["result"] == TOOL_OUTPUT_LONG[:50] + assert len(exchange["result"]) == 50 + assert exchange["result"].startswith(TOOL_OUTPUT_LONG[:10]) + assert exchange["result"].endswith(TOOL_OUTPUT_LONG[-11:]) assert exchange["tool_calls"] == [{"name": "web_search", "arguments": '{"query": "ev"}'}] From 91d4c579e43bd5429741254ea0ff94956edc96fc Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 05:06:48 +0000 Subject: [PATCH 076/141] refactor(guardrails): freeze or suppress mutable constructions in typesafe guardrail Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/typesafe/__init__.py | 15 +- .../guardrail_hooks/typesafe/typesafe.py | 148 ++++++++++-------- .../guardrail_hooks/test_typesafe.py | 2 +- 3 files changed, 91 insertions(+), 74 deletions(-) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py index 40f040eb676..c1b05597a6c 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py @@ -1,5 +1,6 @@ from __future__ import annotations +from types import MappingProxyType from typing import TYPE_CHECKING, Final from pydantic import BaseModel @@ -25,7 +26,9 @@ def _coerce_event_hook( if isinstance(mode, Mode): return mode if isinstance(mode, list): - return [GuardrailEventHooks(item) for item in mode] + return [ + GuardrailEventHooks(item) for item in mode + ] # mutable-ok: CustomGuardrail event_hook contract wants a list return GuardrailEventHooks(mode) @@ -63,10 +66,8 @@ def initialize_guardrail(litellm_params: LitellmParams, guardrail: Guardrail) -> return _callback -guardrail_initializer_registry: Final = { - SupportedGuardrailIntegrations.TYPESAFE.value: initialize_guardrail, -} +guardrail_initializer_registry: Final = MappingProxyType( + {SupportedGuardrailIntegrations.TYPESAFE.value: initialize_guardrail} +) -guardrail_class_registry: Final = { - SupportedGuardrailIntegrations.TYPESAFE.value: TypeSafeGuardrail, -} +guardrail_class_registry: Final = MappingProxyType({SupportedGuardrailIntegrations.TYPESAFE.value: TypeSafeGuardrail}) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py index b3cf5bcf800..384bfd2c54b 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py @@ -11,6 +11,7 @@ from __future__ import annotations import asyncio import time from collections.abc import Mapping, Sequence +from types import MappingProxyType from typing import TYPE_CHECKING, Annotated, Final, Literal import httpx @@ -120,19 +121,20 @@ def _question_instructions(question_id: str) -> str: ) -def _tool_call_entries(assistant_message: Mapping[str, object]) -> list[dict[str, object]]: +def _tool_call_entry(tool_call: object) -> dict[str, object] | None: + parsed_call = _as_str_object_dict(tool_call) + if parsed_call is None: + return None + function = _as_str_object_dict(parsed_call.get("function")) + fn = function if function is not None else parsed_call + return {"name": fn.get("name"), "arguments": fn.get("arguments")} # mutable-ok: serialized to JSON + + +def _tool_call_entries(assistant_message: Mapping[str, object]) -> tuple[dict[str, object], ...]: tool_calls: Final = _as_object_list(assistant_message.get("tool_calls")) if tool_calls is None: - return [] - entries: Final[list[dict[str, object]]] = [] - for tool_call in tool_calls: - parsed_call = _as_str_object_dict(tool_call) - if parsed_call is None: - continue - function = _as_str_object_dict(parsed_call.get("function")) - fn = function if function is not None else parsed_call - entries.append({"name": fn.get("name"), "arguments": fn.get("arguments")}) - return entries + return () + return tuple(entry for tool_call in tool_calls if (entry := _tool_call_entry(tool_call)) is not None) def _protected_indices(messages: Sequence[Mapping[str, object]]) -> frozenset[int]: @@ -199,30 +201,32 @@ class TypeSafeGuardrail(CustomGuardrail): ) return verbose_proxy_logger.error("TypeSafe: %s. detail=%s", error, log_detail) - raise HTTPException(status_code=502, detail={"error": error}) + raise HTTPException(status_code=502, detail={"error": error}) # mutable-ok: FastAPI wants a dict detail - def _candidate_exchanges(self, messages: list[dict[str, object]]) -> list[tuple[int, ...]]: + def _candidate_exchanges(self, messages: Sequence[dict[str, object]]) -> tuple[tuple[int, ...], ...]: """Completed tool exchanges eligible for evaluation: unprotected, and long enough to be worth a call.""" protected: Final = _protected_indices(messages) - candidates: Final[list[tuple[int, ...]]] = [] - for group in group_tool_exchanges(messages): - if len(group) < 2: - continue - if messages[group[0]].get("role") != "assistant": - continue - if any(member in protected for member in group): - continue - tool_text = "".join( - content_to_text(messages[index].get("content")) - for index in group[1:] - if messages[index].get("role") in ("tool", "function") - ) - if not tool_text or len(tool_text) < self.min_chars_to_evaluate: - continue - candidates.append(group) + candidates: Final = tuple( + group + for group in group_tool_exchanges(messages) + if len(group) >= 2 + and messages[group[0]].get("role") == "assistant" + and not any(member in protected for member in group) + and len(self._exchange_tool_text(messages, group)) >= self.min_chars_to_evaluate + ) return candidates[-_MAX_EXCHANGES_EVALUATED:] - def _build_state(self, messages: list[dict[str, object]], candidates: list[tuple[int, ...]]) -> dict[str, object]: + @staticmethod + def _exchange_tool_text(messages: Sequence[dict[str, object]], group: tuple[int, ...]) -> str: + return "".join( + content_to_text(messages[index].get("content")) + for index in group[1:] + if messages[index].get("role") in ("tool", "function") + ) + + def _build_state( + self, messages: Sequence[dict[str, object]], candidates: tuple[tuple[int, ...], ...] + ) -> dict[str, object]: task: Final = next( ( content_to_text(messages[index].get("content")) @@ -234,26 +238,29 @@ class TypeSafeGuardrail(CustomGuardrail): system: Final = "\n\n".join( content_to_text(message.get("content")) for message in messages if message.get("role") == "system" ) - tool_exchanges: Final[dict[str, object]] = {} - for ordinal, group in enumerate(candidates): - result_text = "".join( - content_to_text(messages[index].get("content")) - for index in group[1:] - if messages[index].get("role") in ("tool", "function") - ) - tool_exchanges[f"e{ordinal}"] = { + tool_exchanges: Final = { # mutable-ok: accumulated once, serialized to JSON + f"e{ordinal}": { # mutable-ok: serialized to JSON "tool_calls": _tool_call_entries(messages[group[0]]), - "result": _truncate_for_state(result_text, self.max_result_chars_in_state), + "result": _truncate_for_state( + self._exchange_tool_text(messages, group), self.max_result_chars_in_state + ), } - return {"task": task, "system": system, "tool_exchanges": tool_exchanges} + for ordinal, group in enumerate(candidates) + } + return {"task": task, "system": system, "tool_exchanges": tool_exchanges} # mutable-ok: serialized to JSON - async def _call_systemone(self, state: dict[str, object], question_ids: list[str]) -> _JevSystemOneResponse | None: + async def _call_systemone( + self, state: dict[str, object], question_ids: Sequence[str] + ) -> _JevSystemOneResponse | None: """Returns the response, or None when the service failed and fail_open applies.""" - payload: Final[dict[str, object]] = { + payload: Final[dict[str, object]] = { # mutable-ok: serialized to JSON by httpx "model": self.jev_model, "state": state, - "questions": { - question_id: {"type": "noul", "instructions": _question_instructions(question_id)} + "questions": { # mutable-ok: serialized to JSON + question_id: { + "type": "noul", + "instructions": _question_instructions(question_id), + } # mutable-ok: serialized to JSON for question_id in question_ids }, } @@ -261,7 +268,7 @@ class TypeSafeGuardrail(CustomGuardrail): raw_response: HttpxResponse = await self.async_handler.post( # pyright: ignore[reportUnknownMemberType] # AsyncHTTPHandler.post is untyped url=f"{self.typesafe_api_base}/v1/systemone", json=payload, - headers={ + headers={ # mutable-ok: httpx header contract is a dict "Authorization": f"Bearer {self.typesafe_api_key}", "Content-Type": "application/json", }, @@ -271,21 +278,24 @@ class TypeSafeGuardrail(CustomGuardrail): raise except Exception as e: detail: Final[dict[str, object]] = ( - { + { # mutable-ok: log detail record "error_type": type(e).__name__, "detail": str(e), "status_code": e.response.status_code, "body": _safe_response_text(e.response), } if isinstance(e, httpx.HTTPStatusError) - else {"error_type": type(e).__name__, "detail": str(e)} + else {"error_type": type(e).__name__, "detail": str(e)} # mutable-ok: log detail record ) self._handle_failure("TypeSafe evaluation service request failed", detail) return None if not 200 <= raw_response.status_code < 300: self._handle_failure( "TypeSafe evaluation service returned an error", - {"status_code": raw_response.status_code, "body": _safe_response_text(raw_response)}, + { + "status_code": raw_response.status_code, + "body": _safe_response_text(raw_response), + }, # mutable-ok: log detail record ) return None try: @@ -293,7 +303,7 @@ class TypeSafeGuardrail(CustomGuardrail): except (ValueError, httpx.DecodingError, RecursionError): self._handle_failure( "TypeSafe evaluation service returned an unreadable response", - {"body": _safe_response_text(raw_response)}, + {"body": _safe_response_text(raw_response)}, # mutable-ok: log detail record ) return None try: @@ -301,7 +311,7 @@ class TypeSafeGuardrail(CustomGuardrail): except ValidationError: self._handle_failure( "TypeSafe evaluation service returned unexpected response shape", - {"body": _safe_response_text(raw_response)}, + {"body": _safe_response_text(raw_response)}, # mutable-ok: log detail record ) return None @@ -319,17 +329,17 @@ class TypeSafeGuardrail(CustomGuardrail): structured_messages: Final = _as_object_list(inputs.get("structured_messages")) if not structured_messages: return inputs - parsed_messages: Final = [_as_str_object_dict(m) for m in structured_messages] + parsed_messages: Final = tuple(_as_str_object_dict(m) for m in structured_messages) if any(m is None for m in parsed_messages): return inputs - messages: Final = [m for m in parsed_messages if m is not None] + messages: Final = tuple(m for m in parsed_messages if m is not None) candidates: Final = self._candidate_exchanges(messages) if not candidates: verbose_proxy_logger.debug("TypeSafe: no completed tool exchanges eligible for evaluation") return inputs - question_ids: Final = [f"e{ordinal}" for ordinal in range(len(candidates))] + question_ids: Final = tuple(f"e{ordinal}" for ordinal in range(len(candidates))) state: Final = self._build_state(messages, candidates) start_time: Final = time.monotonic() @@ -337,10 +347,12 @@ class TypeSafeGuardrail(CustomGuardrail): end_time: Final = time.monotonic() if response is None: self.add_standard_logging_guardrail_information_to_request_data( # pyright: ignore[reportUnknownMemberType] # untyped base helper - guardrail_json_response={ - "error": "TypeSafe evaluation unavailable; request forwarded uncompacted", - "model": self.jev_model, - }, + guardrail_json_response=MappingProxyType( + { + "error": "TypeSafe evaluation unavailable; request forwarded uncompacted", + "model": self.jev_model, + } + ), request_data=request_data, guardrail_status="guardrail_failed_to_respond", guardrail_provider="typesafe", @@ -365,8 +377,10 @@ class TypeSafeGuardrail(CustomGuardrail): verbose_proxy_logger.debug("TypeSafe: all evaluated exchanges still relevant; request unchanged") return inputs - compacted_messages: Final = [ - {**message, "content": DROPPED_RESULT_TEXT} if index in dropped_tool_indices else message + compacted_messages: Final = [ # mutable-ok: structured_messages contract is a list of dicts + {**message, "content": DROPPED_RESULT_TEXT} + if index in dropped_tool_indices + else message # mutable-ok: JSON message row for index, message in enumerate(messages) ] chars_removed: Final = sum( @@ -381,12 +395,14 @@ class TypeSafeGuardrail(CustomGuardrail): chars_removed, ) self.add_standard_logging_guardrail_information_to_request_data( # pyright: ignore[reportUnknownMemberType] # untyped base helper - guardrail_json_response={ - "exchanges_evaluated": len(candidates), - "exchanges_dropped": exchanges_dropped, - "chars_removed": chars_removed, - "model": self.jev_model, - }, + guardrail_json_response=MappingProxyType( + { + "exchanges_evaluated": len(candidates), + "exchanges_dropped": exchanges_dropped, + "chars_removed": chars_removed, + "model": self.jev_model, + } + ), request_data=request_data, guardrail_status="success", guardrail_provider="typesafe", @@ -394,7 +410,7 @@ class TypeSafeGuardrail(CustomGuardrail): end_time=end_time, duration=end_time - start_time, ) - return {**inputs, "structured_messages": compacted_messages} # pyright: ignore[reportReturnType] # plain dicts satisfy AllMessageValues at runtime + return {**inputs, "structured_messages": compacted_messages} # pyright: ignore[reportReturnType] # mutable-ok: inputs protocol is a plain dict # plain dicts satisfy AllMessageValues at runtime @staticmethod def get_config_model() -> type[TypeSafeGuardrailConfigModel] | None: diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py index e2e5fc2fefc..92840c489c3 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py @@ -198,7 +198,7 @@ async def test_request_body_shape_and_truncation(): assert len(exchange["result"]) == 50 assert exchange["result"].startswith(TOOL_OUTPUT_LONG[:10]) assert exchange["result"].endswith(TOOL_OUTPUT_LONG[-11:]) - assert exchange["tool_calls"] == [{"name": "web_search", "arguments": '{"query": "ev"}'}] + assert list(exchange["tool_calls"]) == [{"name": "web_search", "arguments": '{"query": "ev"}'}] @pytest.mark.asyncio From 351afc85196a26b6180e82183c2794b5be3b8958 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 05:10:48 +0000 Subject: [PATCH 077/141] test(guardrails): cover typesafe failure paths and edge shapes Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/test_typesafe.py | 116 +++++++++++++++++- 1 file changed, 115 insertions(+), 1 deletion(-) diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py index 92840c489c3..a936d725bd3 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py @@ -16,7 +16,7 @@ Tests cover: - response input_type passthrough and initialize_guardrail wiring """ -from unittest.mock import AsyncMock, MagicMock +from unittest.mock import AsyncMock, MagicMock, PropertyMock import pytest from fastapi import HTTPException @@ -282,3 +282,117 @@ def test_initialize_guardrail_applies_optional_params_and_registry_keys(): assert callback.unreachable_fallback == "fail_open" assert guardrail_initializer_registry[SupportedGuardrailIntegrations.TYPESAFE.value] is initialize_guardrail assert guardrail_class_registry[SupportedGuardrailIntegrations.TYPESAFE.value] is TypeSafeGuardrail + + +def test_missing_api_key_raises(monkeypatch): + monkeypatch.delenv("TYPESAFE_API_KEY", raising=False) + with pytest.raises(ValueError, match="requires an API key"): + TypeSafeGuardrail(api_key=None) + + +def test_get_config_model_and_ui_name(): + from litellm.types.proxy.guardrails.guardrail_hooks.typesafe import ( + TypeSafeGuardrailConfigModel, + ) + + assert TypeSafeGuardrail.get_config_model() is TypeSafeGuardrailConfigModel + assert TypeSafeGuardrailConfigModel.ui_friendly_name() == "TypeSafe (Jev) Compaction" + + +@pytest.mark.asyncio +async def test_non_list_and_non_dict_messages_return_identity(): + guardrail = _make_guardrail() + not_a_list = GenericGuardrailAPIInputs(structured_messages={"role": "user"}) + assert await guardrail.apply_guardrail(inputs=not_a_list, request_data={}, input_type="request", logging_obj=None) is not_a_list + with_bad_row = _inputs(_messages(tail=[["not", "a", "dict"]])) + assert await guardrail.apply_guardrail(inputs=with_bad_row, request_data={}, input_type="request", logging_obj=None) is with_bad_row + + +def test_odd_tool_call_shapes_yield_no_entries(): + from litellm.proxy.guardrails.guardrail_hooks.typesafe.typesafe import _tool_call_entries + + assert _tool_call_entries({"tool_calls": "not-a-list"}) == () + assert _tool_call_entries({"tool_calls": None}) == () + assert list(_tool_call_entries({"tool_calls": [42]})) == [] + entries = _tool_call_entries({"tool_calls": [{"function": {"name": "web_search", "arguments": "{}"}}]}) + assert list(entries) == [{"name": "web_search", "arguments": "{}"}] + + +@pytest.mark.asyncio +async def test_short_max_chars_uses_prefix_slice(): + handler = _make_handler({"e0": 0.9}) + guardrail = _make_guardrail(handler, max_result_chars_in_state=5) + await _apply(guardrail, _messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + result = handler.post.call_args.kwargs["json"]["state"]["tool_exchanges"]["e0"]["result"] + assert result == TOOL_OUTPUT_LONG[:5] + + +@pytest.mark.asyncio +async def test_unreadable_json_body_fails_open(): + handler = MagicMock() + response = MagicMock() + response.status_code = 200 + response.text = "not json" + response.json.side_effect = ValueError("no json") + handler.post = AsyncMock(return_value=response) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert result is inputs + + +@pytest.mark.asyncio +async def test_malformed_answers_shape_fails_open(): + handler = MagicMock() + response = MagicMock() + response.status_code = 200 + response.text = '{"answers": "oops"}' + response.json.return_value = {"answers": "oops"} + handler.post = AsyncMock(return_value=response) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert result is inputs + + +@pytest.mark.asyncio +async def test_http_status_error_includes_status_and_undecodable_body(): + import httpx + + response = MagicMock() + response.status_code = 503 + type(response).text = PropertyMock(side_effect=httpx.DecodingError("bad codec")) + handler = MagicMock() + handler.post = AsyncMock( + side_effect=httpx.HTTPStatusError("unavailable", request=MagicMock(), response=response) + ) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + assert result is inputs + + +@pytest.mark.asyncio +async def test_cancelled_jev_call_propagates(): + import asyncio + + handler = MagicMock() + handler.post = AsyncMock(side_effect=asyncio.CancelledError()) + guardrail = _make_guardrail(handler) + inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + with pytest.raises(asyncio.CancelledError): + await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) + + +def test_optional_params_defaults_and_event_hook_coercion(): + from litellm.proxy.guardrails.guardrail_hooks.typesafe import _coerce_event_hook, _optional_params + from litellm.types.guardrails import GuardrailEventHooks, LitellmParams + + assert _coerce_event_hook("pre_call") is GuardrailEventHooks.pre_call + assert _coerce_event_hook(["pre_call", "post_call"]) == [ + GuardrailEventHooks.pre_call, + GuardrailEventHooks.post_call, + ] + litellm_params = LitellmParams(guardrail="typesafe", mode="pre_call", api_key=FAKE_API_KEY) + params = _optional_params(litellm_params) + assert params.relevance_threshold is None From 87e1c6b3ba19b184735601323eef6d3002f5a0c2 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 05:25:20 +0000 Subject: [PATCH 078/141] fix(guardrails): pin mutable-ok suppressions to constructed literals Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrails/guardrail_hooks/typesafe/__init__.py | 4 ++-- .../guardrails/guardrail_hooks/typesafe/typesafe.py | 12 ++++++------ 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py index c1b05597a6c..abfc60c669a 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py @@ -26,9 +26,9 @@ def _coerce_event_hook( if isinstance(mode, Mode): return mode if isinstance(mode, list): - return [ + return [ # mutable-ok: CustomGuardrail event_hook contract wants a list GuardrailEventHooks(item) for item in mode - ] # mutable-ok: CustomGuardrail event_hook contract wants a list + ] return GuardrailEventHooks(mode) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py index 384bfd2c54b..cb4f21b3261 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py @@ -257,10 +257,10 @@ class TypeSafeGuardrail(CustomGuardrail): "model": self.jev_model, "state": state, "questions": { # mutable-ok: serialized to JSON - question_id: { + question_id: { # mutable-ok: serialized to JSON "type": "noul", "instructions": _question_instructions(question_id), - } # mutable-ok: serialized to JSON + } for question_id in question_ids }, } @@ -292,10 +292,10 @@ class TypeSafeGuardrail(CustomGuardrail): if not 200 <= raw_response.status_code < 300: self._handle_failure( "TypeSafe evaluation service returned an error", - { + { # mutable-ok: log detail record "status_code": raw_response.status_code, "body": _safe_response_text(raw_response), - }, # mutable-ok: log detail record + }, ) return None try: @@ -378,9 +378,9 @@ class TypeSafeGuardrail(CustomGuardrail): return inputs compacted_messages: Final = [ # mutable-ok: structured_messages contract is a list of dicts - {**message, "content": DROPPED_RESULT_TEXT} + {**message, "content": DROPPED_RESULT_TEXT} # mutable-ok: JSON message row if index in dropped_tool_indices - else message # mutable-ok: JSON message row + else message for index, message in enumerate(messages) ] chars_removed: Final = sum( From 22e6947fef68c2e3af137cdca78b3d8811fbb8a4 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 05:44:25 +0000 Subject: [PATCH 079/141] fix(guardrails): keep typesafe guardrail log payloads JSON-serializable Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/typesafe/typesafe.py | 25 ++++++++----------- 1 file changed, 10 insertions(+), 15 deletions(-) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py index cb4f21b3261..9df5c204a77 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py @@ -11,7 +11,6 @@ from __future__ import annotations import asyncio import time from collections.abc import Mapping, Sequence -from types import MappingProxyType from typing import TYPE_CHECKING, Annotated, Final, Literal import httpx @@ -347,12 +346,10 @@ class TypeSafeGuardrail(CustomGuardrail): end_time: Final = time.monotonic() if response is None: self.add_standard_logging_guardrail_information_to_request_data( # pyright: ignore[reportUnknownMemberType] # untyped base helper - guardrail_json_response=MappingProxyType( - { - "error": "TypeSafe evaluation unavailable; request forwarded uncompacted", - "model": self.jev_model, - } - ), + guardrail_json_response={ # mutable-ok: must stay JSON-serializable for shared logging + "error": "TypeSafe evaluation unavailable; request forwarded uncompacted", + "model": self.jev_model, + }, request_data=request_data, guardrail_status="guardrail_failed_to_respond", guardrail_provider="typesafe", @@ -395,14 +392,12 @@ class TypeSafeGuardrail(CustomGuardrail): chars_removed, ) self.add_standard_logging_guardrail_information_to_request_data( # pyright: ignore[reportUnknownMemberType] # untyped base helper - guardrail_json_response=MappingProxyType( - { - "exchanges_evaluated": len(candidates), - "exchanges_dropped": exchanges_dropped, - "chars_removed": chars_removed, - "model": self.jev_model, - } - ), + guardrail_json_response={ # mutable-ok: must stay JSON-serializable for shared logging + "exchanges_evaluated": len(candidates), + "exchanges_dropped": exchanges_dropped, + "chars_removed": chars_removed, + "model": self.jev_model, + }, request_data=request_data, guardrail_status="success", guardrail_provider="typesafe", From 0765f6d571b3a27de6bc9a7456da977bf62705e4 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 06:05:43 +0000 Subject: [PATCH 080/141] fix(guardrails): keep typesafe registries as dicts so guardrail discovery finds them Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/typesafe/__init__.py | 11 +++---- .../guardrail_hooks/test_typesafe.py | 29 +++++++++++++------ 2 files changed, 26 insertions(+), 14 deletions(-) diff --git a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py index abfc60c669a..dcea75d3a98 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py +++ b/litellm/proxy/guardrails/guardrail_hooks/typesafe/__init__.py @@ -1,6 +1,5 @@ from __future__ import annotations -from types import MappingProxyType from typing import TYPE_CHECKING, Final from pydantic import BaseModel @@ -66,8 +65,10 @@ def initialize_guardrail(litellm_params: LitellmParams, guardrail: Guardrail) -> return _callback -guardrail_initializer_registry: Final = MappingProxyType( - {SupportedGuardrailIntegrations.TYPESAFE.value: initialize_guardrail} -) +guardrail_initializer_registry: Final = { # mutable-ok: guardrail_registry discovery checks isinstance(registry, dict) + SupportedGuardrailIntegrations.TYPESAFE.value: initialize_guardrail, +} -guardrail_class_registry: Final = MappingProxyType({SupportedGuardrailIntegrations.TYPESAFE.value: TypeSafeGuardrail}) +guardrail_class_registry: Final = { # mutable-ok: guardrail_registry discovery checks isinstance(registry, dict) + SupportedGuardrailIntegrations.TYPESAFE.value: TypeSafeGuardrail, +} diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py index a936d725bd3..2d1db07a1a0 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_typesafe.py @@ -36,7 +36,7 @@ FAKE_API_KEY = "ts_test-key" SYSTEM_TEXT = "You are a research assistant." USER_TEXT = "Which 2026 EV has the longest range?" -TOOL_OUTPUT_LONG = "Result: EV range comparison. " * 40 # > 200 chars +TOOL_OUTPUT_LONG = "Result: EV range comparison. " * 40 TOOL_OUTPUT_SHORT = "short" @@ -141,8 +141,6 @@ async def test_low_noul_exchange_blanked_high_kept_and_input_not_mutated(): async def test_last_exchange_and_protected_rows_never_evaluated(): handler = _make_handler({"e0": 0.05}) guardrail = _make_guardrail(handler) - # Ends on a tool result: the last assistant row is protected, so the whole - # last exchange is out of scope even though its text is long. messages = _messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), *_exchange("call_2", TOOL_OUTPUT_LONG)]) result = await _apply(guardrail, messages) @@ -303,9 +301,15 @@ def test_get_config_model_and_ui_name(): async def test_non_list_and_non_dict_messages_return_identity(): guardrail = _make_guardrail() not_a_list = GenericGuardrailAPIInputs(structured_messages={"role": "user"}) - assert await guardrail.apply_guardrail(inputs=not_a_list, request_data={}, input_type="request", logging_obj=None) is not_a_list + assert ( + await guardrail.apply_guardrail(inputs=not_a_list, request_data={}, input_type="request", logging_obj=None) + is not_a_list + ) with_bad_row = _inputs(_messages(tail=[["not", "a", "dict"]])) - assert await guardrail.apply_guardrail(inputs=with_bad_row, request_data={}, input_type="request", logging_obj=None) is with_bad_row + assert ( + await guardrail.apply_guardrail(inputs=with_bad_row, request_data={}, input_type="request", logging_obj=None) + is with_bad_row + ) def test_odd_tool_call_shapes_yield_no_entries(): @@ -322,7 +326,9 @@ def test_odd_tool_call_shapes_yield_no_entries(): async def test_short_max_chars_uses_prefix_slice(): handler = _make_handler({"e0": 0.9}) guardrail = _make_guardrail(handler, max_result_chars_in_state=5) - await _apply(guardrail, _messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) + await _apply( + guardrail, _messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}]) + ) result = handler.post.call_args.kwargs["json"]["state"]["tool_exchanges"]["e0"]["result"] assert result == TOOL_OUTPUT_LONG[:5] @@ -363,9 +369,7 @@ async def test_http_status_error_includes_status_and_undecodable_body(): response.status_code = 503 type(response).text = PropertyMock(side_effect=httpx.DecodingError("bad codec")) handler = MagicMock() - handler.post = AsyncMock( - side_effect=httpx.HTTPStatusError("unavailable", request=MagicMock(), response=response) - ) + handler.post = AsyncMock(side_effect=httpx.HTTPStatusError("unavailable", request=MagicMock(), response=response)) guardrail = _make_guardrail(handler) inputs = _inputs(_messages(tail=[*_exchange("call_1", TOOL_OUTPUT_LONG), {"role": "assistant", "content": "x"}])) result = await guardrail.apply_guardrail(inputs=inputs, request_data={}, input_type="request", logging_obj=None) @@ -396,3 +400,10 @@ def test_optional_params_defaults_and_event_hook_coercion(): litellm_params = LitellmParams(guardrail="typesafe", mode="pre_call", api_key=FAKE_API_KEY) params = _optional_params(litellm_params) assert params.relevance_threshold is None + + +def test_typesafe_initializer_discoverable_via_hook_registries(): + from litellm.proxy.guardrails.guardrail_registry import get_guardrail_initializer_from_hooks + + initializers = get_guardrail_initializer_from_hooks() + assert initializers["typesafe"] is initialize_guardrail From a052da69744e6d09ba312e07a181b451552edffc Mon Sep 17 00:00:00 2001 From: shivam Date: Fri, 11 Sep 2026 23:54:35 +0000 Subject: [PATCH 081/141] feat(keys): let team service account keys use key management endpoints for their own team Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 9 + litellm/proxy/auth/route_checks.py | 5 +- .../key_management_endpoints.py | 41 ++++- .../team_member_permission_checks.py | 36 ++-- .../proxy/auth/test_route_checks.py | 69 +++++++ .../test_key_management_endpoints.py | 130 ++++++++++++- .../test_team_member_permission_checks.py | 172 ++++++++++++++++-- 7 files changed, 415 insertions(+), 47 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index b0d31df92ce..cc44a86c691 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -3270,6 +3270,15 @@ class UserAPIKeyAuth(LiteLLM_VerificationTokenView): # the expected response ob user_role=LitellmUserRoles.PROXY_ADMIN, ) + @property + def is_team_service_account(self) -> bool: + return ( + self.user_id is None + and self.team_id is not None + and bool(self.metadata) + and self.metadata.get("service_account_id") is not None + ) + def user_api_key_has_admin_view(user_api_key_dict: UserAPIKeyAuth) -> bool: """Return True if the caller's role grants unscoped read access to all diff --git a/litellm/proxy/auth/route_checks.py b/litellm/proxy/auth/route_checks.py index 0a6b618805d..001ab9a30b0 100644 --- a/litellm/proxy/auth/route_checks.py +++ b/litellm/proxy/auth/route_checks.py @@ -326,7 +326,10 @@ class RouteChecks: pass elif route.startswith("/v1/mcp/") or route.startswith("/mcp-rest/"): pass # authN/authZ handled by api itself - elif RouteChecks.check_passthrough_route_access(route=route, user_api_key_dict=valid_token): + elif RouteChecks.check_passthrough_route_access(route=route, user_api_key_dict=valid_token) or ( + valid_token.is_team_service_account + and RouteChecks.check_route_access(route=route, allowed_routes=LiteLLMRoutes.key_management_routes.value) + ): pass elif valid_token.allowed_routes is not None: # check if route is in allowed_routes (exact match or prefix match) diff --git a/litellm/proxy/management_endpoints/key_management_endpoints.py b/litellm/proxy/management_endpoints/key_management_endpoints.py index 802a7c3e469..42fbe95487f 100644 --- a/litellm/proxy/management_endpoints/key_management_endpoints.py +++ b/litellm/proxy/management_endpoints/key_management_endpoints.py @@ -509,6 +509,16 @@ def _get_user_in_team(team_table: LiteLLM_TeamTableCachedObj, user_id: str | Non return None +def _get_caller_team_role( + team_table: LiteLLM_TeamTableCachedObj, + user_api_key_dict: UserAPIKeyAuth, +) -> Literal["admin", "user"] | None: + if user_api_key_dict.is_team_service_account and user_api_key_dict.team_id == team_table.team_id: + return "user" + member: Final = _get_user_in_team(team_table=team_table, user_id=user_api_key_dict.user_id) + return None if member is None else member.role + + def _calculate_key_rotation_time(rotation_interval: str) -> datetime: """ Helper function to calculate the next rotation time for a key based on the rotation interval. @@ -603,7 +613,7 @@ def _team_key_operation_team_member_check( detail=f"User={assigned_user_id} not assigned to team={team_table.team_id}", ) - team_member_object: Final = _get_user_in_team(team_table=team_table, user_id=user_api_key_dict.user_id) + caller_team_role: Final = _get_caller_team_role(team_table=team_table, user_api_key_dict=user_api_key_dict) is_admin: Final = ( user_api_key_dict.user_role is not None and user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN.value @@ -611,22 +621,22 @@ def _team_key_operation_team_member_check( if is_admin: return True - elif team_member_object is None: + elif caller_team_role is None: raise HTTPException( status_code=400, detail=f"User={user_api_key_dict.user_id} not assigned to team={team_table.team_id}", ) elif ( "allowed_team_member_roles" in team_key_generation - and team_member_object.role not in team_key_generation["allowed_team_member_roles"] + and caller_team_role not in team_key_generation["allowed_team_member_roles"] ): raise HTTPException( status_code=400, - detail=f"Team member role {team_member_object.role} not in allowed_team_member_roles={team_key_generation['allowed_team_member_roles']}", + detail=f"Team member role {caller_team_role} not in allowed_team_member_roles={team_key_generation['allowed_team_member_roles']}", ) TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( - team_member_object=team_member_object, + team_member_role=caller_team_role, team_table=team_table, route=route, ) @@ -747,6 +757,12 @@ def key_generation_check( Check if admin has restricted key creation to certain roles for teams or individuals """ + if user_api_key_dict.is_team_service_account and data.team_id != user_api_key_dict.team_id: + raise HTTPException( + status_code=403, + detail=f"Service account keys can only create keys for their own team. team_id={user_api_key_dict.team_id}", + ) + ## check if key is for team or individual is_team_key: Final = _is_team_key(data=data) _is_admin: Final = ( @@ -2223,6 +2239,10 @@ async def generate_service_account_key_fn( prisma_client=prisma_client, ) + if data.metadata is None or data.metadata.get("service_account_id") is None: + service_account_id: Final = (data.metadata or {}).get("service_account_id") or data.key_alias or str(uuid.uuid4()) + data.metadata = {**(data.metadata or {}), "service_account_id": service_account_id} # rebind-ok: stamping the generated service_account_id onto the request model so it persists on the key + verbose_proxy_logger.debug("entered /key/generate") custom_key_generate_hook: Final[Callable[..., Awaitable[Mapping[str, object]]] | None] = _custom_key_generate_hook( @@ -3837,8 +3857,10 @@ async def validate_key_team_change( detail=f"Key={key.token} has a rpm_limit={key.rpm_limit} which is greater than the team's rpm_limit={team.rpm_limit}.", ) + team_table: Final = cast(LiteLLM_TeamTableCachedObj, team) + # Check if the key's user_id is a member of the team - member_object: Final = _get_user_in_team(team_table=cast(LiteLLM_TeamTableCachedObj, team), user_id=key.user_id) + member_object: Final = _get_user_in_team(team_table=team_table, user_id=key.user_id) if key.user_id is not None: if not member_object: raise HTTPException( @@ -3854,8 +3876,11 @@ async def validate_key_team_change( team_obj=team, ) or TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( - team_member_object=member_object, - team_table=cast(LiteLLM_TeamTableCachedObj, team), + team_member_role=_get_caller_team_role( + team_table=team_table, + user_api_key_dict=change_initiated_by, + ), + team_table=team_table, route=KeyManagementRoutes.KEY_UPDATE.value, ) ): diff --git a/litellm/proxy/management_helpers/team_member_permission_checks.py b/litellm/proxy/management_helpers/team_member_permission_checks.py index 86c7a7bd947..a7e75caeb69 100644 --- a/litellm/proxy/management_helpers/team_member_permission_checks.py +++ b/litellm/proxy/management_helpers/team_member_permission_checks.py @@ -1,4 +1,4 @@ -from typing import Final +from typing import Final, Literal from litellm.proxy._types import ( KeyManagementRoutes, @@ -6,7 +6,6 @@ from litellm.proxy._types import ( LiteLLM_VerificationToken, LiteLLMRoutes, LitellmUserRoles, - Member, ProxyErrorTypes, ProxyException, UserAPIKeyAuth, @@ -27,7 +26,6 @@ DEFAULT_TEAM_MEMBER_PERMISSIONS: Final = BASELINE_TEAM_MEMBER_PERMISSIONS class TeamMemberPermissionChecks: @staticmethod def get_permissions_for_team_member( - team_member_object: Member, team_table: LiteLLM_TeamTableCachedObj, ) -> list[KeyManagementRoutes]: """ @@ -67,7 +65,7 @@ class TeamMemberPermissionChecks: Main handler for checking if a team member can update a key """ from litellm.proxy.management_endpoints.key_management_endpoints import ( - _get_user_in_team, + _get_caller_team_role, ) # 1. Don't execute these checks if the user role is proxy admin @@ -87,12 +85,12 @@ class TeamMemberPermissionChecks: check_db_only=True, ) - # 4. Extract `Member` object from `team_table` - key_assigned_user_in_team: Final = _get_user_in_team(team_table=team_table, user_id=user_api_key_dict.user_id) + # 4. Resolve the caller's role in the key's team (service accounts act as "user") + caller_team_role: Final = _get_caller_team_role(team_table=team_table, user_api_key_dict=user_api_key_dict) # 5. Check if the team member has permissions for the endpoint has_permission: Final = TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( - team_member_object=key_assigned_user_in_team, + team_member_role=caller_team_role, team_table=team_table, route=route, ) @@ -106,7 +104,7 @@ class TeamMemberPermissionChecks: @staticmethod def does_team_member_have_permissions_for_endpoint( - team_member_object: Member | None, + team_member_role: Literal["admin", "user"] | None, team_table: LiteLLM_TeamTableCachedObj, route: str, ) -> bool | None: @@ -116,13 +114,12 @@ class TeamMemberPermissionChecks: # permission checks only run for non-admin users # Non-Admin user trying to access information about a team's key - if team_member_object is None: + if team_member_role is None: return False - if team_member_object.role == "admin": + if team_member_role == "admin": return True _team_member_permissions: Final = TeamMemberPermissionChecks.get_permissions_for_team_member( - team_member_object=team_member_object, team_table=team_table, ) team_member_permissions = TeamMemberPermissionChecks._get_list_of_route_enum_as_str(_team_member_permissions) @@ -156,7 +153,7 @@ class TeamMemberPermissionChecks: from fastapi import HTTPException from litellm.proxy.management_endpoints.key_management_endpoints import ( - _get_user_in_team, + _get_caller_team_role, ) # No-op when the request does not assign any access groups. @@ -177,20 +174,19 @@ class TeamMemberPermissionChecks: ), ) - team_member_object: Final = _get_user_in_team(team_table=team_table, user_id=user_api_key_dict.user_id) + caller_team_role: Final = _get_caller_team_role(team_table=team_table, user_api_key_dict=user_api_key_dict) # Team admins always bypass (consistent with other member-permission checks). - if team_member_object is not None and team_member_object.role == "admin": + if caller_team_role == "admin": return permissions: Final = ( TeamMemberPermissionChecks._get_list_of_route_enum_as_str( TeamMemberPermissionChecks.get_permissions_for_team_member( - team_member_object=team_member_object, team_table=team_table, ) ) - if team_member_object is not None + if caller_team_role is not None else [] ) @@ -214,7 +210,7 @@ class TeamMemberPermissionChecks: Returns True if the user belongs to the team that the key is assigned to """ from litellm.proxy.management_endpoints.key_management_endpoints import ( - _get_user_in_team, + _get_caller_team_role, ) from litellm.proxy.proxy_server import prisma_client, user_api_key_cache @@ -228,9 +224,9 @@ class TeamMemberPermissionChecks: check_db_only=True, ) - # 4. Extract `Member` object from `team_table` - team_member_object: Final = _get_user_in_team(team_table=team_table, user_id=user_api_key_dict.user_id) - return team_member_object is not None + # 4. Resolve the caller's role in the key's team (service accounts act as "user") + caller_team_role: Final = _get_caller_team_role(team_table=team_table, user_api_key_dict=user_api_key_dict) + return caller_team_role is not None @staticmethod def get_all_available_team_member_permissions() -> list[str]: diff --git a/tests/test_litellm/proxy/auth/test_route_checks.py b/tests/test_litellm/proxy/auth/test_route_checks.py index 72c7011e6f5..b1405fc35cb 100644 --- a/tests/test_litellm/proxy/auth/test_route_checks.py +++ b/tests/test_litellm/proxy/auth/test_route_checks.py @@ -3967,3 +3967,72 @@ def test_auto_router_session_read_grant_rejects_other_methods_paths_and_scopes( RouteChecks.should_call_route(route, valid_token, request) assert error.value.status_code == 403 + + +@pytest.mark.parametrize("route", ["/key/generate", "/key/update"]) +def test_team_service_account_key_allowed_key_management_routes(route): + """A service account key (user_id=None, team_id set, metadata.service_account_id) + can reach key-management routes; team scoping is enforced in the handlers.""" + valid_token = UserAPIKeyAuth( + api_key="sk", + team_id="t1", + user_id=None, + metadata={"service_account_id": "ci"}, + ) + request = MagicMock(spec=Request) + request.query_params = {} + + result = RouteChecks.non_proxy_admin_allowed_routes_check( + user_obj=None, + _user_role=None, + route=route, + request=request, + valid_token=valid_token, + request_data={}, + ) + assert result is None + + +def test_team_service_account_key_rejected_for_non_key_management_route(): + """The service account carve-out does not extend past key-management routes.""" + valid_token = UserAPIKeyAuth( + api_key="sk", + team_id="t1", + user_id=None, + metadata={"service_account_id": "ci"}, + ) + request = MagicMock(spec=Request) + request.query_params = {} + + with pytest.raises(Exception, match="Only proxy admin can be used to generate, delete, update"): + RouteChecks.non_proxy_admin_allowed_routes_check( + user_obj=None, + _user_role=None, + route="/team/new", + request=request, + valid_token=valid_token, + request_data={}, + ) + + +def test_team_key_without_service_account_marker_still_rejected(): + """A team key without metadata.service_account_id is not a service account + and still cannot reach key-management routes.""" + valid_token = UserAPIKeyAuth( + api_key="sk", + team_id="t1", + user_id=None, + metadata={}, + ) + request = MagicMock(spec=Request) + request.query_params = {} + + with pytest.raises(Exception, match="Only proxy admin can be used to generate, delete, update"): + RouteChecks.non_proxy_admin_allowed_routes_check( + user_obj=None, + _user_role=None, + route="/key/generate", + request=request, + valid_token=valid_token, + request_data={}, + ) diff --git a/tests/test_litellm/proxy/management_endpoints/test_key_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_key_management_endpoints.py index cc0a7631b59..f865e959a78 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_key_management_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_key_management_endpoints.py @@ -18,6 +18,7 @@ import inspect from litellm.proxy._types import ( GenerateKeyRequest, + KeyManagementRoutes, NewUserRequest, LiteLLM_BudgetTable, LiteLLM_ObjectPermissionBase, @@ -3099,7 +3100,7 @@ async def test_validate_key_team_change_with_member_permissions(): # Verify the permission check was called with correct parameters mock_has_perms.assert_called_once_with( - team_member_object=mock_member_object, + team_member_role=mock_member_object.role, team_table=mock_team, route=KeyManagementRoutes.KEY_UPDATE.value, ) @@ -19733,3 +19734,130 @@ async def test_bulk_update_team_keys_runs_custom_key_policy_per_key(monkeypatch) assert [policy_request.operation for policy_request in received] == ["update", "update"] assert [policy_request.effective_key.max_budget for policy_request in received] == [50.0, 50.0] assert [policy_request.effective_key.team_id for policy_request in received] == ["team-abc", "team-abc"] + + +class TestServiceAccountKeyGenerationCheck: + """Service account keys (user_id=None, team_id set, metadata.service_account_id) + may only create keys for their own team.""" + + def _service_account_token(self, team_id: str) -> UserAPIKeyAuth: + return UserAPIKeyAuth( + api_key="sk-sa", + user_id=None, + team_id=team_id, + metadata={"service_account_id": "sa-1"}, + ) + + def test_other_team_denied(self): + data = GenerateKeyRequest(team_id="team-b") + with pytest.raises(HTTPException) as exc_info: + key_generation_check( + team_table=None, + user_api_key_dict=self._service_account_token(team_id="team-a"), + data=data, + route=KeyManagementRoutes.KEY_GENERATE, + ) + assert exc_info.value.status_code == 403 + + def test_personal_key_denied(self): + """team_id=None would mint a personal key; service accounts may only + create keys for their own team.""" + data = GenerateKeyRequest() + with pytest.raises(HTTPException) as exc_info: + key_generation_check( + team_table=None, + user_api_key_dict=self._service_account_token(team_id="team-a"), + data=data, + route=KeyManagementRoutes.KEY_GENERATE, + ) + assert exc_info.value.status_code == 403 + + def test_own_team_with_permission_allowed(self): + team_table = LiteLLM_TeamTableCachedObj( + team_id="team-a", + members_with_roles=[], + team_member_permissions=["/key/generate"], + ) + data = GenerateKeyRequest(team_id="team-a") + assert ( + key_generation_check( + team_table=team_table, + user_api_key_dict=self._service_account_token(team_id="team-a"), + data=data, + route=KeyManagementRoutes.KEY_GENERATE, + ) + is True + ) + + def test_own_team_without_permission_denied(self): + team_table = LiteLLM_TeamTableCachedObj( + team_id="team-a", + members_with_roles=[], + team_member_permissions=["/key/info"], + ) + data = GenerateKeyRequest(team_id="team-a") + with pytest.raises(ProxyException) as exc_info: + key_generation_check( + team_table=team_table, + user_api_key_dict=self._service_account_token(team_id="team-a"), + data=data, + route=KeyManagementRoutes.KEY_GENERATE, + ) + assert str(exc_info.value.code) == "401" + + +def _stub_service_account_generation(monkeypatch): + """Stub the DB lookups generate_service_account_key_fn needs so the test + exercises only the service_account_id stamping and user_id clearing.""" + from litellm.proxy import proxy_server + from litellm.proxy.management_endpoints import key_management_endpoints as kme + + mock_helper = AsyncMock(return_value=MagicMock()) + monkeypatch.setattr(proxy_server, "prisma_client", MagicMock()) + monkeypatch.setattr(kme, "validate_team_id_used_in_service_account_request", AsyncMock()) + monkeypatch.setattr(kme, "_common_key_generation_helper", mock_helper) + return mock_helper + + +@pytest.mark.asyncio +async def test_generate_service_account_key_stamps_service_account_id(monkeypatch): + """generate_service_account_key_fn must stamp metadata.service_account_id + (key_alias fallback) so the key is identifiable as a service account by + is_team_service_account and check_if_token_is_service_account.""" + from litellm.proxy.management_endpoints.key_management_endpoints import ( + generate_service_account_key_fn, + ) + + mock_helper = _stub_service_account_generation(monkeypatch) + data = GenerateKeyRequest(team_id="team-a", key_alias="sa-alias") + + await generate_service_account_key_fn( + data=data, + user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-1"), + litellm_changed_by=None, + ) + + assert data.metadata is not None + assert data.metadata["service_account_id"] == "sa-alias" + assert data.user_id is None + mock_helper.assert_awaited_once() + + +@pytest.mark.asyncio +async def test_generate_service_account_key_generates_uuid_when_no_alias(monkeypatch): + """Without key_alias, service_account_id falls back to a generated uuid.""" + from litellm.proxy.management_endpoints.key_management_endpoints import ( + generate_service_account_key_fn, + ) + + _stub_service_account_generation(monkeypatch) + data = GenerateKeyRequest(team_id="team-a") + + await generate_service_account_key_fn( + data=data, + user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-1"), + litellm_changed_by=None, + ) + + assert data.metadata is not None + assert data.metadata["service_account_id"] diff --git a/tests/test_litellm/proxy/management_helpers/test_team_member_permission_checks.py b/tests/test_litellm/proxy/management_helpers/test_team_member_permission_checks.py index 36c61eddbb2..55d29724e38 100644 --- a/tests/test_litellm/proxy/management_helpers/test_team_member_permission_checks.py +++ b/tests/test_litellm/proxy/management_helpers/test_team_member_permission_checks.py @@ -3,7 +3,12 @@ from unittest.mock import MagicMock import pytest -from litellm.proxy._types import KeyManagementRoutes, Member, ProxyException +from litellm.proxy._types import ( + KeyManagementRoutes, + Member, + ProxyException, + UserAPIKeyAuth, +) from litellm.proxy.management_helpers.team_member_permission_checks import ( BASELINE_TEAM_MEMBER_PERMISSIONS, TeamMemberPermissionChecks, @@ -21,22 +26,16 @@ class TestGetPermissionsForTeamMember: def test_none_permissions_returns_defaults(self): """When team_member_permissions is None, return DEFAULT_TEAM_MEMBER_PERMISSIONS.""" team = _make_team_table(None) - member = MagicMock(spec=Member) - result = TeamMemberPermissionChecks.get_permissions_for_team_member( - team_member_object=member, team_table=team - ) + result = TeamMemberPermissionChecks.get_permissions_for_team_member(team_table=team) assert set(result) == set(BASELINE_TEAM_MEMBER_PERMISSIONS) def test_empty_list_includes_baseline(self): """When team_member_permissions is [], baseline permissions are still included.""" team = _make_team_table([]) - member = MagicMock(spec=Member) - result = TeamMemberPermissionChecks.get_permissions_for_team_member( - team_member_object=member, team_table=team - ) + result = TeamMemberPermissionChecks.get_permissions_for_team_member(team_table=team) assert KeyManagementRoutes.KEY_INFO in result assert KeyManagementRoutes.KEY_HEALTH in result @@ -44,11 +43,8 @@ class TestGetPermissionsForTeamMember: def test_explicit_permissions_include_baseline(self): """When explicit permissions are set, baseline is always included.""" team = _make_team_table(["/key/generate", "/key/delete"]) - member = MagicMock(spec=Member) - result = TeamMemberPermissionChecks.get_permissions_for_team_member( - team_member_object=member, team_table=team - ) + result = TeamMemberPermissionChecks.get_permissions_for_team_member(team_table=team) assert KeyManagementRoutes.KEY_GENERATE in result assert KeyManagementRoutes.KEY_DELETE in result @@ -58,11 +54,8 @@ class TestGetPermissionsForTeamMember: def test_explicit_permissions_with_baseline_no_duplicates(self): """When explicit permissions already include baseline, no duplicates.""" team = _make_team_table(["/key/info", "/key/generate"]) - member = MagicMock(spec=Member) - result = TeamMemberPermissionChecks.get_permissions_for_team_member( - team_member_object=member, team_table=team - ) + result = TeamMemberPermissionChecks.get_permissions_for_team_member(team_table=team) # Using set ensures no duplicates from the implementation assert KeyManagementRoutes.KEY_INFO in result @@ -402,3 +395,148 @@ class TestEnforceMemberCanAssignAccessGroups: team_table=self._team(["/key/generate", self.AG_PERMISSION]), access_group_ids=["ag-1"], ) + + +class TestDoesTeamMemberHavePermissionsForEndpoint: + def _team(self, team_member_permissions, team_id="team-a"): + team = MagicMock() + team.team_id = team_id + team.team_member_permissions = team_member_permissions + return team + + def test_none_role_returns_false(self): + """A caller with no team membership is denied.""" + result = TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( + team_member_role=None, + team_table=self._team(["/key/update"]), + route=KeyManagementRoutes.KEY_UPDATE.value, + ) + assert result is False + + def test_admin_role_always_allowed(self): + """Team admins bypass the member permission list.""" + result = TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( + team_member_role="admin", + team_table=self._team([]), + route=KeyManagementRoutes.KEY_UPDATE.value, + ) + assert result is True + + def test_user_role_with_permission_allowed(self): + result = TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( + team_member_role="user", + team_table=self._team(["/key/update"]), + route=KeyManagementRoutes.KEY_UPDATE.value, + ) + assert result is True + + def test_user_role_without_permission_raises(self): + with pytest.raises(ProxyException) as exc: + TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( + team_member_role="user", + team_table=self._team(["/key/generate"]), + route=KeyManagementRoutes.KEY_UPDATE.value, + ) + assert str(exc.value.code) == "401" + assert exc.value.type == "team_member_permission_error" + + +class TestCanTeamMemberExecuteKeyManagementEndpointServiceAccount: + def _service_account_token(self, team_id: str) -> UserAPIKeyAuth: + return UserAPIKeyAuth( + api_key="sk-test", + user_id=None, + team_id=team_id, + metadata={"service_account_id": "sa-1"}, + ) + + @pytest.mark.asyncio + async def test_service_account_same_team_with_permission(self, monkeypatch): + """A service account key can manage keys in its own team when the + team grants the route via team_member_permissions.""" + from litellm.proxy.management_helpers import ( + team_member_permission_checks as module, + ) + + async def _mock_get_team_object(**kwargs): + team = MagicMock() + team.team_id = "team-a" + team.members_with_roles = [] + team.team_member_permissions = ["/key/update"] + return team + + monkeypatch.setattr(module, "get_team_object", _mock_get_team_object) + + existing_key_row = MagicMock() + existing_key_row.team_id = "team-a" + + result = await TeamMemberPermissionChecks.can_team_member_execute_key_management_endpoint( + user_api_key_dict=self._service_account_token(team_id="team-a"), + route=KeyManagementRoutes.KEY_UPDATE, + prisma_client=MagicMock(), + user_api_key_cache=MagicMock(), + existing_key_row=existing_key_row, + ) + assert result is None + + @pytest.mark.asyncio + async def test_service_account_same_team_without_permission(self, monkeypatch): + """A service account key is denied when the team's + team_member_permissions does not include the route.""" + from litellm.proxy.management_helpers import ( + team_member_permission_checks as module, + ) + + async def _mock_get_team_object(**kwargs): + team = MagicMock() + team.team_id = "team-a" + team.members_with_roles = [] + team.team_member_permissions = ["/key/generate"] + return team + + monkeypatch.setattr(module, "get_team_object", _mock_get_team_object) + + existing_key_row = MagicMock() + existing_key_row.team_id = "team-a" + + with pytest.raises(ProxyException) as exc: + await TeamMemberPermissionChecks.can_team_member_execute_key_management_endpoint( + user_api_key_dict=self._service_account_token(team_id="team-a"), + route=KeyManagementRoutes.KEY_UPDATE, + prisma_client=MagicMock(), + user_api_key_cache=MagicMock(), + existing_key_row=existing_key_row, + ) + assert str(exc.value.code) == "401" + assert exc.value.type == "team_member_permission_error" + + @pytest.mark.asyncio + async def test_service_account_different_team_denied(self, monkeypatch): + """A service account key cannot manage keys in another team, even if + that team grants the route to its members.""" + from litellm.proxy.management_helpers import ( + team_member_permission_checks as module, + ) + + async def _mock_get_team_object(**kwargs): + team = MagicMock() + team.team_id = "team-b" + team.members_with_roles = [] + team.team_member_permissions = ["/key/update"] + return team + + monkeypatch.setattr(module, "get_team_object", _mock_get_team_object) + + existing_key_row = MagicMock() + existing_key_row.team_id = "team-b" + + with pytest.raises(ProxyException) as exc: + await TeamMemberPermissionChecks.can_team_member_execute_key_management_endpoint( + user_api_key_dict=self._service_account_token(team_id="team-a"), + route=KeyManagementRoutes.KEY_UPDATE, + prisma_client=MagicMock(), + user_api_key_cache=MagicMock(), + existing_key_row=existing_key_row, + ) + assert str(exc.value.code) == "401" + assert exc.value.type == "team_member_permission_error" From 062e17a7ac7fb1b78373e5198c40962c6252c699 Mon Sep 17 00:00:00 2001 From: shivam Date: Fri, 11 Sep 2026 23:55:49 +0000 Subject: [PATCH 082/141] refactor(keys): keep validate_key_team_change permission check on the key's assigned member Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../proxy/management_endpoints/key_management_endpoints.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/litellm/proxy/management_endpoints/key_management_endpoints.py b/litellm/proxy/management_endpoints/key_management_endpoints.py index 42fbe95487f..f0a689d6c50 100644 --- a/litellm/proxy/management_endpoints/key_management_endpoints.py +++ b/litellm/proxy/management_endpoints/key_management_endpoints.py @@ -3876,10 +3876,7 @@ async def validate_key_team_change( team_obj=team, ) or TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( - team_member_role=_get_caller_team_role( - team_table=team_table, - user_api_key_dict=change_initiated_by, - ), + team_member_role=None if member_object is None else member_object.role, team_table=team_table, route=KeyManagementRoutes.KEY_UPDATE.value, ) From 2542b9cfe8f6cf84f27943f5926c218fd027fa27 Mon Sep 17 00:00:00 2001 From: shivam Date: Sat, 12 Sep 2026 00:04:53 +0000 Subject: [PATCH 083/141] style(keys): fix ruff format and drop comment that restates the call Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../management_endpoints/key_management_endpoints.py | 9 +++++++-- .../management_helpers/team_member_permission_checks.py | 4 +--- 2 files changed, 8 insertions(+), 5 deletions(-) diff --git a/litellm/proxy/management_endpoints/key_management_endpoints.py b/litellm/proxy/management_endpoints/key_management_endpoints.py index f0a689d6c50..705ea01f690 100644 --- a/litellm/proxy/management_endpoints/key_management_endpoints.py +++ b/litellm/proxy/management_endpoints/key_management_endpoints.py @@ -2240,8 +2240,13 @@ async def generate_service_account_key_fn( ) if data.metadata is None or data.metadata.get("service_account_id") is None: - service_account_id: Final = (data.metadata or {}).get("service_account_id") or data.key_alias or str(uuid.uuid4()) - data.metadata = {**(data.metadata or {}), "service_account_id": service_account_id} # rebind-ok: stamping the generated service_account_id onto the request model so it persists on the key + service_account_id: Final = ( + (data.metadata or {}).get("service_account_id") or data.key_alias or str(uuid.uuid4()) + ) + data.metadata = { # rebind-ok: stamping the generated service_account_id onto the request model so it persists on the key + **(data.metadata or {}), + "service_account_id": service_account_id, + } verbose_proxy_logger.debug("entered /key/generate") diff --git a/litellm/proxy/management_helpers/team_member_permission_checks.py b/litellm/proxy/management_helpers/team_member_permission_checks.py index a7e75caeb69..a076d8240c6 100644 --- a/litellm/proxy/management_helpers/team_member_permission_checks.py +++ b/litellm/proxy/management_helpers/team_member_permission_checks.py @@ -85,10 +85,9 @@ class TeamMemberPermissionChecks: check_db_only=True, ) - # 4. Resolve the caller's role in the key's team (service accounts act as "user") caller_team_role: Final = _get_caller_team_role(team_table=team_table, user_api_key_dict=user_api_key_dict) - # 5. Check if the team member has permissions for the endpoint + # 4. Check if the team member has permissions for the endpoint has_permission: Final = TeamMemberPermissionChecks.does_team_member_have_permissions_for_endpoint( team_member_role=caller_team_role, team_table=team_table, @@ -224,7 +223,6 @@ class TeamMemberPermissionChecks: check_db_only=True, ) - # 4. Resolve the caller's role in the key's team (service accounts act as "user") caller_team_role: Final = _get_caller_team_role(team_table=team_table, user_api_key_dict=user_api_key_dict) return caller_team_role is not None From b78793b153116c24ced32ef88e207160d35da606 Mon Sep 17 00:00:00 2001 From: ryan Date: Fri, 18 Sep 2026 09:16:57 +0000 Subject: [PATCH 084/141] fix(auth): limit team service account route carve-out to /key/generate and /key/update The key_management_routes group also contains /spend/logs, /team/daily/activity and other routes whose handlers scope non-admin callers by user_id. A userless service account key would have reached them unscoped, so the route check now uses a dedicated two-route allowlist Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 5 +++++ litellm/proxy/auth/route_checks.py | 4 +++- tests/test_litellm/proxy/auth/test_route_checks.py | 8 +++++--- 3 files changed, 13 insertions(+), 4 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index cc44a86c691..a31b6f834b3 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -660,6 +660,11 @@ class LiteLLMRoutes(enum.Enum): KeyManagementRoutes.AUTO_ROUTER_MANAGE.value, ] + team_service_account_key_routes = [ + KeyManagementRoutes.KEY_GENERATE.value, + KeyManagementRoutes.KEY_UPDATE.value, + ] + management_routes = ( [ # user diff --git a/litellm/proxy/auth/route_checks.py b/litellm/proxy/auth/route_checks.py index 001ab9a30b0..1b9fd7c42bf 100644 --- a/litellm/proxy/auth/route_checks.py +++ b/litellm/proxy/auth/route_checks.py @@ -328,7 +328,9 @@ class RouteChecks: pass # authN/authZ handled by api itself elif RouteChecks.check_passthrough_route_access(route=route, user_api_key_dict=valid_token) or ( valid_token.is_team_service_account - and RouteChecks.check_route_access(route=route, allowed_routes=LiteLLMRoutes.key_management_routes.value) + and RouteChecks.check_route_access( + route=route, allowed_routes=LiteLLMRoutes.team_service_account_key_routes.value + ) ): pass elif valid_token.allowed_routes is not None: diff --git a/tests/test_litellm/proxy/auth/test_route_checks.py b/tests/test_litellm/proxy/auth/test_route_checks.py index b1405fc35cb..72c59223549 100644 --- a/tests/test_litellm/proxy/auth/test_route_checks.py +++ b/tests/test_litellm/proxy/auth/test_route_checks.py @@ -3993,8 +3993,10 @@ def test_team_service_account_key_allowed_key_management_routes(route): assert result is None -def test_team_service_account_key_rejected_for_non_key_management_route(): - """The service account carve-out does not extend past key-management routes.""" +@pytest.mark.parametrize("route", ["/team/new", "/spend/logs", "/key/delete", "/key/regenerate"]) +def test_team_service_account_key_rejected_outside_generate_and_update(route): + """The service account carve-out covers only /key/generate and /key/update; other + key-management routes lack team scoping for a userless caller and stay denied.""" valid_token = UserAPIKeyAuth( api_key="sk", team_id="t1", @@ -4008,7 +4010,7 @@ def test_team_service_account_key_rejected_for_non_key_management_route(): RouteChecks.non_proxy_admin_allowed_routes_check( user_obj=None, _user_role=None, - route="/team/new", + route=route, request=request, valid_token=valid_token, request_data={}, From fade26b969880633ce4f87d0a81d01a2372f76ac Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 09:25:34 +0000 Subject: [PATCH 085/141] feat(proxy): derive the per-username sign-in allowance from the address limit The per-address-and-username allowance is now half the effective address allowance, rounded up, instead of a separate max_failed_login_attempts_per_user setting. A per-address override therefore raises or effectively removes both limits for that address, and no second override table is needed Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 9 +-- litellm/proxy/auth/login_throttle.py | 13 +++-- .../proxy/auth/test_login_utils.py | 56 ++++++++++++++++++- .../proxy_server/test_routes_login_sso.py | 18 +++--- tests/test_litellm/proxy/test_proxy_server.py | 2 - ui/litellm-dashboard/src/lib/http/schema.d.ts | 9 +-- 6 files changed, 76 insertions(+), 31 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 80147863304..801babbbb2d 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2755,16 +2755,11 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): max_failed_login_attempts_per_source: int | None = Field( None, ge=1, - description="Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and this limit is off. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10", + description="Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Half this value, rounded up, is the allowance for one username from that address; one more blocks that address for that username only, and its further failures stop counting toward the address limit, so a script stuck on one account does not block everyone behind a shared address. The per-address limit is only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and only the per-username half runs. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10", ) max_failed_login_attempts_per_source_overrides: dict[str, int] | None = Field( None, - description="Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins. Set under `general_settings` in config.yaml", - ) - max_failed_login_attempts_per_user: int | None = Field( - None, - ge=1, - description="Failed Admin UI sign-in attempts allowed from one source address for one username within `failed_login_window_seconds`. One more blocks that address for that username for `failed_login_block_seconds`, and its further failures stop counting against `max_failed_login_attempts_per_source`, so a script stuck on one account does not block everyone behind the same address. Set under `general_settings` in config.yaml. Defaults to 5", + description="Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A very large value opts the address out of both limits. Set under `general_settings` in config.yaml", ) failed_login_window_seconds: int | None = Field( None, diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index d16cd43e36d..012d557f334 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -39,7 +39,6 @@ from litellm.proxy.auth.network import TrustedProxyConfig, normalize_cidr_ranges from litellm.secret_managers.main import get_secret_bool DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE: Final = 10 -DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_USER: Final = 5 DEFAULT_FAILED_LOGIN_WINDOW_SECONDS: Final = 60 DEFAULT_FAILED_LOGIN_BLOCK_SECONDS: Final = 300 @@ -47,7 +46,6 @@ IPV6_SOURCE_PREFIX_LENGTH: Final = 64 SOURCE_LIMIT_KEY: Final = "max_failed_login_attempts_per_source" SOURCE_LIMIT_OVERRIDES_KEY: Final = "max_failed_login_attempts_per_source_overrides" -USER_LIMIT_KEY: Final = "max_failed_login_attempts_per_user" WINDOW_KEY: Final = "failed_login_window_seconds" BLOCK_KEY: Final = "failed_login_block_seconds" TRUSTED_PROXY_RANGES_KEY: Final = "trusted_proxy_ranges" @@ -204,6 +202,11 @@ def _source_limit(settings: Mapping[str, object], client_ip: str) -> int: return matches[-1][1] if matches else default +def user_limit_for(source_limit: int) -> int: + """Failures allowed for one username from one address: half the address allowance, rounded up.""" + return (source_limit + 1) // 2 + + def source_group(client_ip: str) -> str: """The bucket an address is counted in: IPv4 as is, IPv6 by its /64, so one prefix holder cannot rotate.""" address: Final = _parse_address(client_ip) @@ -233,6 +236,7 @@ class LoginThrottle: ``source_limit`` is None when the source scope is off: ``trusted_proxy_ranges`` is unset, so the peer address may be a shared ingress. An empty list means clients connect directly and the peer is the source. + ``user_limit`` is derived from the address allowance either way, see ``user_limit_for``. """ client_ip: str @@ -257,10 +261,11 @@ class LoginThrottle: resolved, _ = resolve_client_ip( request, TrustedProxyConfig(use_forwarded_for=bool(proxies), trusted_proxy_cidrs=proxies or ()) ) + source_limit: Final = _source_limit(settings, resolved or LOGIN_THROTTLE_UNKNOWN_SOURCE) return cls( client_ip=resolved or LOGIN_THROTTLE_UNKNOWN_SOURCE, - source_limit=_source_limit(settings, resolved) if proxies is not None and resolved is not None else None, - user_limit=_int_setting(settings, USER_LIMIT_KEY, DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_USER), + source_limit=source_limit if proxies is not None and resolved is not None else None, + user_limit=user_limit_for(source_limit), window_seconds=_int_setting(settings, WINDOW_KEY, DEFAULT_FAILED_LOGIN_WINDOW_SECONDS), block_seconds=_int_setting(settings, BLOCK_KEY, DEFAULT_FAILED_LOGIN_BLOCK_SECONDS), counters=_COUNTERS, diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 986760c3cc5..d6a094d56cb 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1471,7 +1471,6 @@ def test_settings_that_arrive_as_environment_strings_are_honored(): general_settings={ "trusted_proxy_ranges": "10.0.0.0/8", "max_failed_login_attempts_per_source": " 70 ", - "max_failed_login_attempts_per_user": "7", "failed_login_window_seconds": "not-a-number", "failed_login_block_seconds": "-5", }, @@ -1479,7 +1478,7 @@ def test_settings_that_arrive_as_environment_strings_are_honored(): ) assert throttle.source_limit == 70 - assert throttle.user_limit == 7 + assert throttle.user_limit == 35, "the per-username allowance is half the address allowance" assert throttle.window_seconds == 60, "garbage falls back to the default" assert throttle.block_seconds == 300, "a value below one would block nothing or forever" @@ -1503,6 +1502,59 @@ def test_the_defaults_are_the_agreed_ones(): ) +@pytest.mark.parametrize( + ("source_limit", "expected_user_limit"), + [(1, 1), (2, 1), (3, 2), (10, 5), (1_000_000, 500_000)], + ids=["one-stays-one", "two-halves-to-one", "odd-rounds-up", "default", "opt-out"], +) +def test_the_per_username_allowance_is_half_the_address_allowance_rounded_up(source_limit, expected_user_limit): + from litellm.proxy.auth.login_throttle import user_limit_for + + assert user_limit_for(source_limit) == expected_user_limit + + +def test_a_per_address_override_also_raises_that_address_per_username_allowance(): + """One override opts an address out of both limits, so operators need no second override table.""" + from litellm.proxy.auth.login_throttle import LoginThrottle + + settings = { + "trusted_proxy_ranges": ["10.0.0.0/8"], + "max_failed_login_attempts_per_source": 10, + "max_failed_login_attempts_per_source_overrides": {"203.0.113.0/24": 1_000_000}, + } + + def _from(client_ip: str) -> LoginThrottle: + request = MagicMock() + request.headers = {"x-forwarded-for": client_ip} + request.client = MagicMock() + request.client.host = "10.0.0.1" + return LoginThrottle.from_request(request, general_settings=settings, redis_cache=None) + + exempt = _from("203.0.113.9") + assert (exempt.source_limit, exempt.user_limit) == (1_000_000, 500_000) + + ordinary = _from("198.51.100.4") + assert (ordinary.source_limit, ordinary.user_limit) == (10, 5) + + +def test_the_per_username_allowance_follows_the_peer_override_when_the_source_scope_is_off(): + """Without trusted_proxy_ranges the address is not blocked, but its override still sizes the pair limit.""" + from litellm.proxy.auth.login_throttle import LoginThrottle + + request = MagicMock() + request.headers = {} + request.client = MagicMock() + request.client.host = "192.0.2.8" + throttle = LoginThrottle.from_request( + request, + general_settings={"max_failed_login_attempts_per_source_overrides": {"192.0.2.8": 40}}, + redis_cache=None, + ) + + assert throttle.source_limit is None + assert throttle.user_limit == 20 + + def test_the_disable_flag_is_read_once_not_per_login_attempt(monkeypatch): """Regression: the kill switch was read through the secret manager on every unauthenticated request.""" from litellm.proxy.auth import login_throttle diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py index a82f078fcb9..88f8be4e49a 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py @@ -531,7 +531,7 @@ def test_budget_is_shared_across_every_login_endpoint(client, monkeypatch, reset """ _install_real_auth( monkeypatch, - max_failed_login_attempts_per_user=10, + max_failed_login_attempts_per_source=20, control_plane_url="https://cp.example.com", ) @@ -544,7 +544,7 @@ def test_budget_is_shared_across_every_login_endpoint(client, monkeypatch, reset def test_budget_is_shared_across_username_casing(client, monkeypatch, reset_login_throttle): """The database lookup is case-insensitive, so casing must not partition the counter.""" - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=3) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=6) assert [_json_login(client, "/v2/login", username="admin@corp.com") for _ in range(2)] == [401] * 2 assert [_json_login(client, "/v2/login", username="ADMIN@corp.com") for _ in range(2)] == [401] * 2 @@ -554,7 +554,7 @@ def test_budget_is_shared_across_username_casing(client, monkeypatch, reset_logi def test_a_refused_attempt_carries_retry_after(client, monkeypatch, reset_login_throttle): """The 429 tells the caller how long the block has left.""" - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1, failed_login_block_seconds=77) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=2, failed_login_block_seconds=77) assert [_json_login(client, "/v2/login") for _ in range(2)] == [401, 401] @@ -565,7 +565,7 @@ def test_a_refused_attempt_carries_retry_after(client, monkeypatch, reset_login_ def test_the_form_returns_a_human_readable_lockout_page(client, monkeypatch, reset_login_throttle): """The no-JavaScript form must render a wait page when its POST is throttled.""" - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1, failed_login_block_seconds=77) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=2, failed_login_block_seconds=77) assert [_form_login(client) for _ in range(2)] == [401, 401] @@ -578,7 +578,7 @@ def test_the_form_returns_a_human_readable_lockout_page(client, monkeypatch, res def test_a_second_username_from_the_same_source_still_gets_through(client, monkeypatch, reset_login_throttle): """The pair block is per username, so one account's block cannot take the office down with it.""" - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=2) assert [_json_login(client, "/v2/login", username="admin") for _ in range(3)] == [401, 401, 429] @@ -633,7 +633,7 @@ def test_the_configured_admin_password_is_refused_while_blocked(client, monkeypa limit. An operator who is blocked administers the proxy with the master key over the API meanwhile.""" from unittest.mock import AsyncMock, patch - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=2) monkeypatch.setenv("DATABASE_URL", "postgresql://stub") assert [_json_login(client, "/v2/login") for _ in range(3)] == [401, 401, 429] @@ -655,7 +655,7 @@ def test_the_master_key_as_a_bearer_token_still_works_while_the_ui_password_is_b client, monkeypatch, reset_login_throttle ): """Lockout recovery: the API path with the master key never enters the sign-in throttle.""" - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=2) assert [_json_login(client, "/v2/login") for _ in range(3)] == [401, 401, 429] @@ -667,7 +667,7 @@ def test_the_master_key_as_a_bearer_token_still_works_while_the_ui_password_is_b def test_a_database_users_correct_password_is_refused_while_blocked(client, monkeypatch, reset_login_throttle): """The block is hard: while it lasts, nothing from that source signs in as that user, right password or not, and the block is not extended by the refused attempts.""" - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1, failed_login_block_seconds=64) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=2, failed_login_block_seconds=64) _db_user(monkeypatch, "user@corp.com") assert [_json_login(client, "/v2/login", username="user@corp.com") for _ in range(3)] == [401, 401, 429] @@ -682,7 +682,7 @@ def test_a_database_users_correct_password_is_refused_while_blocked(client, monk def test_sign_in_succeeds_again_once_the_block_is_cleared(client, monkeypatch, reset_login_throttle): """A cleared store lets the same username straight back to a plain credential check.""" - _install_real_auth(monkeypatch, max_failed_login_attempts_per_user=1) + _install_real_auth(monkeypatch, max_failed_login_attempts_per_source=2) assert [_json_login(client, "/v2/login") for _ in range(3)] == [401, 401, 429] diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 4c8cacf120f..9e2b85e765d 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -13519,13 +13519,11 @@ async def test_login_throttle_settings_are_not_hot_applied_from_the_database(): ps.general_settings.clear() await ProxyConfig()._update_general_settings( db_general_settings={ - "max_failed_login_attempts_per_user": 999, "max_failed_login_attempts_per_source": 999, "failed_login_window_seconds": 1, "failed_login_block_seconds": 1, } ) - assert "max_failed_login_attempts_per_user" not in ps.general_settings assert "max_failed_login_attempts_per_source" not in ps.general_settings assert "failed_login_window_seconds" not in ps.general_settings assert "failed_login_block_seconds" not in ps.general_settings diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 871cfea2d29..98bb3182916 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -26566,21 +26566,16 @@ export interface components { max_batch_file_size_mb?: number | null; /** * Max Failed Login Attempts Per Source - * @description Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and this limit is off. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10 + * @description Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Half this value, rounded up, is the allowance for one username from that address; one more blocks that address for that username only, and its further failures stop counting toward the address limit, so a script stuck on one account does not block everyone behind a shared address. The per-address limit is only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and only the per-username half runs. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10 */ max_failed_login_attempts_per_source?: number | null; /** * Max Failed Login Attempts Per Source Overrides - * @description Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins. Set under `general_settings` in config.yaml + * @description Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A very large value opts the address out of both limits. Set under `general_settings` in config.yaml */ max_failed_login_attempts_per_source_overrides?: { [key: string]: number; } | null; - /** - * Max Failed Login Attempts Per User - * @description Failed Admin UI sign-in attempts allowed from one source address for one username within `failed_login_window_seconds`. One more blocks that address for that username for `failed_login_block_seconds`, and its further failures stop counting against `max_failed_login_attempts_per_source`, so a script stuck on one account does not block everyone behind the same address. Set under `general_settings` in config.yaml. Defaults to 5 - */ - max_failed_login_attempts_per_user?: number | null; /** * Max File Size Mb * @description max file size in MB for /v1/files uploads, for any purpose, if a file is larger than this size it will be rejected before being forwarded to the provider From e68ea077f169d179e13610edbca937415563937c Mon Sep 17 00:00:00 2001 From: ryan Date: Fri, 18 Sep 2026 09:26:29 +0000 Subject: [PATCH 086/141] style(keys): keep rebind-ok reason within the line limit Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/management_endpoints/key_management_endpoints.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/litellm/proxy/management_endpoints/key_management_endpoints.py b/litellm/proxy/management_endpoints/key_management_endpoints.py index 705ea01f690..6d4b7f05532 100644 --- a/litellm/proxy/management_endpoints/key_management_endpoints.py +++ b/litellm/proxy/management_endpoints/key_management_endpoints.py @@ -2243,7 +2243,7 @@ async def generate_service_account_key_fn( service_account_id: Final = ( (data.metadata or {}).get("service_account_id") or data.key_alias or str(uuid.uuid4()) ) - data.metadata = { # rebind-ok: stamping the generated service_account_id onto the request model so it persists on the key + data.metadata = { # rebind-ok: stamp the service_account_id onto the request so it persists on the key **(data.metadata or {}), "service_account_id": service_account_id, } From b50265d75f719bfd2dd27b46c237a0981e6f1cbe Mon Sep 17 00:00:00 2001 From: ryan Date: Fri, 18 Sep 2026 09:34:44 +0000 Subject: [PATCH 087/141] refactor(keys): freeze the service account route tuple and stamp metadata without seeding mutable literals Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 4 ++-- .../management_endpoints/key_management_endpoints.py | 9 ++++----- 2 files changed, 6 insertions(+), 7 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index a31b6f834b3..2ca6abf7873 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -660,10 +660,10 @@ class LiteLLMRoutes(enum.Enum): KeyManagementRoutes.AUTO_ROUTER_MANAGE.value, ] - team_service_account_key_routes = [ + team_service_account_key_routes = ( KeyManagementRoutes.KEY_GENERATE.value, KeyManagementRoutes.KEY_UPDATE.value, - ] + ) management_routes = ( [ diff --git a/litellm/proxy/management_endpoints/key_management_endpoints.py b/litellm/proxy/management_endpoints/key_management_endpoints.py index 6d4b7f05532..f6f47a8ffc9 100644 --- a/litellm/proxy/management_endpoints/key_management_endpoints.py +++ b/litellm/proxy/management_endpoints/key_management_endpoints.py @@ -2240,13 +2240,12 @@ async def generate_service_account_key_fn( ) if data.metadata is None or data.metadata.get("service_account_id") is None: - service_account_id: Final = ( - (data.metadata or {}).get("service_account_id") or data.key_alias or str(uuid.uuid4()) - ) - data.metadata = { # rebind-ok: stamp the service_account_id onto the request so it persists on the key - **(data.metadata or {}), + service_account_id: Final = data.key_alias or str(uuid.uuid4()) + stamped_metadata: Final = { # mutable-ok: GenerateKeyRequest.metadata is a plain dict field + **(data.metadata or MappingProxyType({})), "service_account_id": service_account_id, } + data.metadata = stamped_metadata # rebind-ok: the request carries the stamp so it persists on the key verbose_proxy_logger.debug("entered /key/generate") From b6bb212248ff27676e63917ac3050a4809d827ea Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 09:40:05 +0000 Subject: [PATCH 088/141] refactor(proxy): raise the sign-in block explicitly and type the empty settings mapping Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 10 +++++----- tests/test_litellm/proxy/auth/test_login_utils.py | 2 +- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 012d557f334..2ad75463629 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -17,7 +17,7 @@ import time from collections.abc import Mapping from dataclasses import dataclass from functools import cache -from typing import Final, Literal, NamedTuple, NoReturn, Protocol, TypeAlias +from typing import Final, Literal, NamedTuple, Protocol, TypeAlias from fastapi import Request, status from pydantic import TypeAdapter, ValidationError @@ -256,7 +256,7 @@ class LoginThrottle: general_settings: Mapping[str, object] | None, redis_cache: RedisCache | None, ) -> LoginThrottle: - settings: Final = general_settings if general_settings is not None else EMPTY_MAPPING + settings: Final[Mapping[str, object]] = general_settings if general_settings is not None else EMPTY_MAPPING proxies: Final = declared_proxy_ranges(settings) resolved, _ = resolve_client_ip( request, TrustedProxyConfig(use_forwarded_for=bool(proxies), trusted_proxy_cidrs=proxies or ()) @@ -298,7 +298,7 @@ class LoginThrottle: username, self.client_ip, ) - self.refuse(block.retry_after) + raise self.refused(block.retry_after) async def _active_block(self, keys: _Keys) -> Block | None: local: Final = self._local_block_ttls(keys) @@ -375,8 +375,8 @@ class LoginThrottle: ) @staticmethod - def refuse(retry_after: int) -> NoReturn: - raise ProxyException( + def refused(retry_after: int) -> ProxyException: + return ProxyException( message="Too many failed sign-in attempts. Try again later.", type=ProxyErrorTypes.auth_error, param="username", diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index d6a094d56cb..4d4a10a9fa4 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1436,8 +1436,8 @@ async def test_a_failed_redis_delete_still_clears_this_workers_counter(monkeypat @pytest.mark.asyncio async def test_counters_do_not_share_the_key_authentication_cache(monkeypatch): """Regression: throttle entries must not evict cached credentials from user_api_key_cache.""" - from litellm.proxy import proxy_server as ps from litellm.constants import LOGIN_THROTTLE_CACHE_KEY_PREFIX + from litellm.proxy import proxy_server as ps from litellm.proxy.auth.login_throttle import LoginThrottle monkeypatch.setenv("UI_USERNAME", "admin") From 0da2f5b96cb88031fc359a65506b7cc137677564 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 09:54:54 +0000 Subject: [PATCH 089/141] feat(proxy): round the per-username sign-in allowance down and exempt an address with an override of 0 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 4 +- litellm/proxy/auth/login_throttle.py | 27 ++++++--- .../proxy/auth/test_login_utils.py | 58 ++++++++++++++----- ui/litellm-dashboard/src/lib/http/schema.d.ts | 4 +- 4 files changed, 66 insertions(+), 27 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 801babbbb2d..04ea7712545 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2755,11 +2755,11 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): max_failed_login_attempts_per_source: int | None = Field( None, ge=1, - description="Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Half this value, rounded up, is the allowance for one username from that address; one more blocks that address for that username only, and its further failures stop counting toward the address limit, so a script stuck on one account does not block everyone behind a shared address. The per-address limit is only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and only the per-username half runs. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10", + description="Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Half this value, rounded down but at least 1, is the allowance for one username from that address; one more blocks that address for that username only, and its further failures stop counting toward the address limit, so a script stuck on one account does not block everyone behind a shared address. The per-address limit is only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and only the per-username half runs. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10", ) max_failed_login_attempts_per_source_overrides: dict[str, int] | None = Field( None, - description="Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A very large value opts the address out of both limits. Set under `general_settings` in config.yaml", + description="Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A value of 0 exempts the address from both limits. Set under `general_settings` in config.yaml", ) failed_login_window_seconds: int | None = Field( None, diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 2ad75463629..273cf8b9f93 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -43,6 +43,7 @@ DEFAULT_FAILED_LOGIN_WINDOW_SECONDS: Final = 60 DEFAULT_FAILED_LOGIN_BLOCK_SECONDS: Final = 300 IPV6_SOURCE_PREFIX_LENGTH: Final = 64 +EXEMPT: Final = 0 SOURCE_LIMIT_KEY: Final = "max_failed_login_attempts_per_source" SOURCE_LIMIT_OVERRIDES_KEY: Final = "max_failed_login_attempts_per_source_overrides" @@ -157,6 +158,13 @@ def _int_setting(settings: Mapping[str, object], key: str, default: int) -> int: return _positive_int(settings.get(key), key, default) +def _override_limit(raw: object, default: int) -> int: + """A per-address override: a limit of 1 or more, or ``EXEMPT`` (0) to leave that address unlimited.""" + if str(raw).strip() == str(EXEMPT): + return EXEMPT + return _positive_int(raw, SOURCE_LIMIT_OVERRIDES_KEY, default) + + def _parse_address(client_ip: str) -> ipaddress.IPv4Address | ipaddress.IPv6Address | None: """The address as it is limited and counted: an IPv4-mapped IPv6 address is its IPv4 address.""" try: @@ -179,7 +187,10 @@ def _parse_network(raw_range: str) -> _Network | None: def _source_limit(settings: Mapping[str, object], client_ip: str) -> int: - """Failure allowance for this address: the most specific configured range containing it, else the default.""" + """Failure allowance for this address: the most specific configured range containing it, else the default. + + ``EXEMPT`` (0) means the operator opted this address out of both limits. + """ default: Final = _int_setting(settings, SOURCE_LIMIT_KEY, DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE) raw_overrides: Final = settings.get(SOURCE_LIMIT_OVERRIDES_KEY) if raw_overrides is None: @@ -195,7 +206,7 @@ def _source_limit(settings: Mapping[str, object], client_ip: str) -> int: if address is None: return default matches: Final = sorted( - (network.prefixlen, _positive_int(raw_limit, SOURCE_LIMIT_OVERRIDES_KEY, default)) + (network.prefixlen, _override_limit(raw_limit, default)) for raw_range, raw_limit in overrides.items() if (network := _parse_network(raw_range)) is not None and address in network ) @@ -203,8 +214,8 @@ def _source_limit(settings: Mapping[str, object], client_ip: str) -> int: def user_limit_for(source_limit: int) -> int: - """Failures allowed for one username from one address: half the address allowance, rounded up.""" - return (source_limit + 1) // 2 + """Failures allowed for one username from one address: half the address allowance, rounded down, at least 1.""" + return max(source_limit // 2, 1) def source_group(client_ip: str) -> str: @@ -236,7 +247,8 @@ class LoginThrottle: ``source_limit`` is None when the source scope is off: ``trusted_proxy_ranges`` is unset, so the peer address may be a shared ingress. An empty list means clients connect directly and the peer is the source. - ``user_limit`` is derived from the address allowance either way, see ``user_limit_for``. + ``user_limit`` is derived from the address allowance either way, see ``user_limit_for``. An address whose + override is ``EXEMPT`` gets a disabled throttle: nothing is counted or blocked for it. """ client_ip: str @@ -262,16 +274,17 @@ class LoginThrottle: request, TrustedProxyConfig(use_forwarded_for=bool(proxies), trusted_proxy_cidrs=proxies or ()) ) source_limit: Final = _source_limit(settings, resolved or LOGIN_THROTTLE_UNKNOWN_SOURCE) + exempt: Final = source_limit == EXEMPT return cls( client_ip=resolved or LOGIN_THROTTLE_UNKNOWN_SOURCE, - source_limit=source_limit if proxies is not None and resolved is not None else None, + source_limit=source_limit if proxies is not None and resolved is not None and not exempt else None, user_limit=user_limit_for(source_limit), window_seconds=_int_setting(settings, WINDOW_KEY, DEFAULT_FAILED_LOGIN_WINDOW_SECONDS), block_seconds=_int_setting(settings, BLOCK_KEY, DEFAULT_FAILED_LOGIN_BLOCK_SECONDS), counters=_COUNTERS, blocks=_BLOCKS, redis_cache=redis_cache, - enabled=not _rate_limit_disabled(), + enabled=not exempt and not _rate_limit_disabled(), ) def _keys(self, username: str) -> _Keys: diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 4d4a10a9fa4..d6e0a489939 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -6,6 +6,7 @@ to login_utils.py for better reusability. """ import os +from collections.abc import Mapping from contextlib import ExitStack from typing import Final from unittest.mock import AsyncMock, MagicMock, patch @@ -1504,39 +1505,64 @@ def test_the_defaults_are_the_agreed_ones(): @pytest.mark.parametrize( ("source_limit", "expected_user_limit"), - [(1, 1), (2, 1), (3, 2), (10, 5), (1_000_000, 500_000)], - ids=["one-stays-one", "two-halves-to-one", "odd-rounds-up", "default", "opt-out"], + [(1, 1), (2, 1), (3, 1), (10, 5), (11, 5), (70, 35)], + ids=["one-stays-one", "two-halves-to-one", "odd-rounds-down", "default", "eleven-rounds-down", "even"], ) -def test_the_per_username_allowance_is_half_the_address_allowance_rounded_up(source_limit, expected_user_limit): +def test_the_per_username_allowance_is_half_the_address_allowance_rounded_down_at_least_one( + source_limit, expected_user_limit +): from litellm.proxy.auth.login_throttle import user_limit_for assert user_limit_for(source_limit) == expected_user_limit -def test_a_per_address_override_also_raises_that_address_per_username_allowance(): - """One override opts an address out of both limits, so operators need no second override table.""" +def _throttle_behind_trusted_proxy(client_ip: str, settings: Mapping[str, object]) -> "LoginThrottle": from litellm.proxy.auth.login_throttle import LoginThrottle + request = MagicMock() + request.headers = {"x-forwarded-for": client_ip} + request.client = MagicMock() + request.client.host = "10.0.0.1" + return LoginThrottle.from_request(request, general_settings=settings, redis_cache=None) + + +def test_a_per_address_override_also_raises_that_address_per_username_allowance(): + """One override sizes both limits for an address, so operators need no second override table.""" settings = { "trusted_proxy_ranges": ["10.0.0.0/8"], "max_failed_login_attempts_per_source": 10, - "max_failed_login_attempts_per_source_overrides": {"203.0.113.0/24": 1_000_000}, + "max_failed_login_attempts_per_source_overrides": {"203.0.113.0/24": 50}, } - def _from(client_ip: str) -> LoginThrottle: - request = MagicMock() - request.headers = {"x-forwarded-for": client_ip} - request.client = MagicMock() - request.client.host = "10.0.0.1" - return LoginThrottle.from_request(request, general_settings=settings, redis_cache=None) + raised = _throttle_behind_trusted_proxy("203.0.113.9", settings) + assert (raised.source_limit, raised.user_limit) == (50, 25) - exempt = _from("203.0.113.9") - assert (exempt.source_limit, exempt.user_limit) == (1_000_000, 500_000) - - ordinary = _from("198.51.100.4") + ordinary = _throttle_behind_trusted_proxy("198.51.100.4", settings) assert (ordinary.source_limit, ordinary.user_limit) == (10, 5) +@pytest.mark.asyncio +async def test_an_override_of_zero_exempts_that_address_from_both_limits(): + """Regression: opting an address out used to mean guessing a large enough number.""" + settings = { + "trusted_proxy_ranges": ["10.0.0.0/8"], + "max_failed_login_attempts_per_source": 1, + "max_failed_login_attempts_per_source_overrides": {"203.0.113.7": 0, "203.0.113.0/24": 3}, + } + + exempt = _throttle_behind_trusted_proxy("203.0.113.7", settings) + assert exempt.enabled is False + assert exempt.source_limit is None + attempt = await exempt.attempt("scanner@example.com") + for _ in range(5): + await attempt.failed() + await exempt.attempt("scanner@example.com") + + sibling = _throttle_behind_trusted_proxy("203.0.113.8", settings) + assert sibling.enabled is True + assert (sibling.source_limit, sibling.user_limit) == (3, 1) + + def test_the_per_username_allowance_follows_the_peer_override_when_the_source_scope_is_off(): """Without trusted_proxy_ranges the address is not blocked, but its override still sizes the pair limit.""" from litellm.proxy.auth.login_throttle import LoginThrottle diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 98bb3182916..e6e1627c4ca 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -26566,12 +26566,12 @@ export interface components { max_batch_file_size_mb?: number | null; /** * Max Failed Login Attempts Per Source - * @description Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Half this value, rounded up, is the allowance for one username from that address; one more blocks that address for that username only, and its further failures stop counting toward the address limit, so a script stuck on one account does not block everyone behind a shared address. The per-address limit is only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and only the per-username half runs. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10 + * @description Failed Admin UI sign-in attempts allowed from one source address, across every username, within `failed_login_window_seconds`. One more blocks that address for `failed_login_block_seconds`. Half this value, rounded down but at least 1, is the allowance for one username from that address; one more blocks that address for that username only, and its further failures stop counting toward the address limit, so a script stuck on one account does not block everyone behind a shared address. The per-address limit is only enforced when `trusted_proxy_ranges` is set: to the proxies in front of LiteLLM, or to an empty list when clients connect directly. Left unset, the peer address may be a shared ingress and only the per-username half runs. IPv6 addresses are grouped by /64. Set under `general_settings` in config.yaml. Defaults to 10 */ max_failed_login_attempts_per_source?: number | null; /** * Max Failed Login Attempts Per Source Overrides - * @description Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A very large value opts the address out of both limits. Set under `general_settings` in config.yaml + * @description Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A value of 0 exempts the address from both limits. Set under `general_settings` in config.yaml */ max_failed_login_attempts_per_source_overrides?: { [key: string]: number; From 65765d6550fea8c7d1ac9fc700ea5d8171da8b54 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 09:57:35 +0000 Subject: [PATCH 090/141] test(proxy): import LoginThrottle under TYPE_CHECKING for the throttle helper annotation Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/proxy/auth/test_login_utils.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index d6e0a489939..52ae3ca092b 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -8,11 +8,14 @@ to login_utils.py for better reusability. import os from collections.abc import Mapping from contextlib import ExitStack -from typing import Final +from typing import TYPE_CHECKING, Final from unittest.mock import AsyncMock, MagicMock, patch import pytest +if TYPE_CHECKING: + from litellm.proxy.auth.login_throttle import LoginThrottle + def _unlimited_throttle(): """A throttle wired to real in-memory stores with limits no test can reach.""" From cfdf4fa5dc043f378d38c09a5b6ea87874076041 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 10:20:39 +0000 Subject: [PATCH 091/141] fix(proxy): break ties between equivalent login limit overrides deterministically Two spellings of one network share a prefix length, so the exemption wins the tie, then the higher limit, regardless of mapping order Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 2 +- litellm/proxy/auth/login_throttle.py | 12 +++++++++--- .../test_litellm/proxy/auth/test_login_utils.py | 17 +++++++++++++++++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 +- 4 files changed, 28 insertions(+), 5 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 04ea7712545..851b2e7b28e 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2759,7 +2759,7 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): ) max_failed_login_attempts_per_source_overrides: dict[str, int] | None = Field( None, - description="Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A value of 0 exempts the address from both limits. Set under `general_settings` in config.yaml", + description="Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins (between equivalent keys such as '1.2.3.4' and '1.2.3.4/32', an exemption wins, then the higher limit), and the per-username allowance for that address follows as half the override. A value of 0 exempts the address from both limits. Set under `general_settings` in config.yaml", ) failed_login_window_seconds: int | None = Field( None, diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 273cf8b9f93..e4d9ffd4c5b 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -186,10 +186,16 @@ def _parse_network(raw_range: str) -> _Network | None: return None +def _precedence(network: _Network, limit: int) -> tuple[int, bool, int]: + """Sort key for competing overrides: the longest prefix wins, then an exemption, then the higher limit.""" + return (network.prefixlen, limit == EXEMPT, limit) + + def _source_limit(settings: Mapping[str, object], client_ip: str) -> int: """Failure allowance for this address: the most specific configured range containing it, else the default. - ``EXEMPT`` (0) means the operator opted this address out of both limits. + ``EXEMPT`` (0) means the operator opted this address out of both limits. Between equivalent keys such as + ``1.2.3.4`` and ``1.2.3.4/32`` an exemption wins, then the higher limit. """ default: Final = _int_setting(settings, SOURCE_LIMIT_KEY, DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE) raw_overrides: Final = settings.get(SOURCE_LIMIT_OVERRIDES_KEY) @@ -206,11 +212,11 @@ def _source_limit(settings: Mapping[str, object], client_ip: str) -> int: if address is None: return default matches: Final = sorted( - (network.prefixlen, _override_limit(raw_limit, default)) + _precedence(network, _override_limit(raw_limit, default)) for raw_range, raw_limit in overrides.items() if (network := _parse_network(raw_range)) is not None and address in network ) - return matches[-1][1] if matches else default + return matches[-1][-1] if matches else default def user_limit_for(source_limit: int) -> int: diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 52ae3ca092b..0cac0364219 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -1023,6 +1023,23 @@ def test_source_overrides_pick_the_most_specific_matching_range(): assert _limit("::ffff:203.0.113.10") == 200 +@pytest.mark.parametrize( + ("overrides", "expected"), + [ + ({"203.0.113.7": 0, "203.0.113.7/32": 5}, None), + ({"203.0.113.7/32": 5, "203.0.113.7": 0}, None), + ({"203.0.113.0/24": 3, "203.0.113.9/24": 8}, 8), + ({"203.0.113.9/24": 8, "203.0.113.0/24": 3}, 8), + ], + ids=["exact-then-slash32", "slash32-then-exact", "low-then-high", "high-then-low"], +) +def test_equivalent_override_keys_resolve_to_the_exemption_then_the_higher_limit(overrides, expected): + """Two spellings of the same network are a config mistake, so precedence must not depend on dict order.""" + settings = {"trusted_proxy_ranges": ["10.0.0.0/8"], "max_failed_login_attempts_per_source_overrides": overrides} + + assert _throttle_behind_trusted_proxy("203.0.113.7", settings).source_limit == expected + + def test_ipv6_sources_are_grouped_by_their_64_bit_prefix(): """A /64 holder has 2^64 addresses; counting each one separately would hand them unlimited fresh buckets.""" from litellm.proxy.auth.login_throttle import source_group diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index e6e1627c4ca..5b1532c3629 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -26571,7 +26571,7 @@ export interface components { max_failed_login_attempts_per_source?: number | null; /** * Max Failed Login Attempts Per Source Overrides - * @description Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins, and the per-username allowance for that address follows as half the override. A value of 0 exempts the address from both limits. Set under `general_settings` in config.yaml + * @description Per-address overrides of `max_failed_login_attempts_per_source`, keyed by IP address or CIDR range, e.g. {'1.2.3.4': 200, '5.6.0.0/24': 500}. The most specific matching range wins (between equivalent keys such as '1.2.3.4' and '1.2.3.4/32', an exemption wins, then the higher limit), and the per-username allowance for that address follows as half the override. A value of 0 exempts the address from both limits. Set under `general_settings` in config.yaml */ max_failed_login_attempts_per_source_overrides?: { [key: string]: number; From cacd12b87dc778f335ad6ac41dad02ba00312000 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 11:28:35 +0000 Subject: [PATCH 092/141] fix(proxy): treat a malformed trusted_proxy_ranges entry as an undeclared topology A list with an entry that is not an address or CIDR range no longer switches the per-source Admin UI sign-in limit on against the direct peer address, so a typo cannot make a shared ingress address the bucket for every user behind it Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_types.py | 2 +- litellm/proxy/auth/login_throttle.py | 19 ++++++++++--------- .../proxy/auth/test_login_utils.py | 19 ++++++++++++++++--- ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 +- 4 files changed, 28 insertions(+), 14 deletions(-) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 851b2e7b28e..eee37016347 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -2880,7 +2880,7 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase): ) trusted_proxy_ranges: list[str] | None = Field( None, - description="CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler, and whose X-Forwarded-For is used to attribute Admin UI sign-in attempts to a source address. Set it to an empty list when clients connect directly, so the peer address is the source. Left unset, the per-source sign-in limit is off.", + description="CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler, and whose X-Forwarded-For is used to attribute Admin UI sign-in attempts to a source address. Set it to an empty list when clients connect directly, so the peer address is the source. Left unset, or containing an entry that is not an address or CIDR range, the per-source sign-in limit is off.", ) store_model_in_db: bool | None = Field( None, diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index e4d9ffd4c5b..49eb7a5f20d 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -120,9 +120,9 @@ def warn_login_counters_are_per_worker(num_workers: str) -> None: @cache def warn_source_login_limit_is_off() -> None: verbose_proxy_logger.warning( - "%s is not set, so failed Admin UI sign-in attempts are limited per source address and username " - "only. Set it to the address ranges of the proxies in front of LiteLLM, or to an empty list when " - "clients connect directly, to also limit each source address across usernames.", + "%s is not set or not a valid list of ranges, so failed Admin UI sign-in attempts are limited per " + "source address and username only. Set it to the address ranges of the proxies in front of LiteLLM, " + "or to an empty list when clients connect directly, to also limit each source address across usernames.", TRUSTED_PROXY_RANGES_KEY, ) @@ -131,13 +131,16 @@ def declared_proxy_ranges(settings: Mapping[str, object]) -> tuple[str, ...] | N """What the operator says fronts LiteLLM: the proxy ranges, an empty tuple for none, None when unsaid. Only a declared topology makes the source address trustworthy enough to limit across usernames. - An unset key, or a value that is not a list of ranges, leaves it unknown and the source scope off. + An unset key, a value that is not a list of ranges, or a list with an entry that is not an address + or range leaves it unknown and the source scope off. """ raw_ranges: Final = settings.get(TRUSTED_PROXY_RANGES_KEY) if isinstance(raw_ranges, (list, tuple, set)) and not raw_ranges: return () cidrs: Final = tuple(normalize_cidr_ranges(raw_ranges, setting_name=TRUSTED_PROXY_RANGES_KEY)) - return cidrs or None + if not cidrs or any(_parse_network(cidr, TRUSTED_PROXY_RANGES_KEY) is None for cidr in cidrs): + return None + return cidrs def _positive_int(raw: object, key: str, default: int) -> int: @@ -176,13 +179,11 @@ def _parse_address(client_ip: str) -> ipaddress.IPv4Address | ipaddress.IPv6Addr return address -def _parse_network(raw_range: str) -> _Network | None: +def _parse_network(raw_range: str, setting_name: str = SOURCE_LIMIT_OVERRIDES_KEY) -> _Network | None: try: return ipaddress.ip_network(raw_range.strip(), strict=False) except ValueError: - verbose_proxy_logger.warning( - "Invalid address or range %r in %s; skipping", raw_range, SOURCE_LIMIT_OVERRIDES_KEY - ) + verbose_proxy_logger.warning("Invalid address or range %r in %s; skipping", raw_range, setting_name) return None diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 0cac0364219..90186f4b8c0 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -934,10 +934,23 @@ async def test_an_empty_trusted_proxy_ranges_means_the_peer_is_the_client_and_th assert await _fail(throttle, username="user-99@corp.com") == "429", "the spray is stopped by the source limit" -@pytest.mark.parametrize("configured", [None, 5, {"10.0.0.0/8": True}, ["", " "]]) +@pytest.mark.parametrize( + "configured", + [ + None, + 5, + {"10.0.0.0/8": True}, + ["", " "], + ["not-a-range"], + ["10.0.0.0/8, 172.16.0.0/12"], + ["10.0.0.0/8", "10.0.0.0/33"], + "10.0.0.0/8;172.16.0.0/12", + ], +) def test_a_trusted_proxy_ranges_value_that_names_no_ranges_leaves_the_topology_unknown(configured): - """Only a real list of ranges or an explicit empty list counts as a declaration; anything else is the same - as unset, so a typo cannot switch the source-wide block on behind a shared ingress.""" + """Only a list of valid ranges or an explicit empty list counts as a declaration; anything else, including a + list with one bad entry, is the same as unset, so a typo cannot switch the source-wide block on against + the shared ingress address and lock out everyone behind it.""" from litellm.proxy.auth.login_throttle import LoginThrottle, declared_proxy_ranges settings = {"trusted_proxy_ranges": configured} if configured is not None else {} diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 5b1532c3629..9cd578ea0a4 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -26751,7 +26751,7 @@ export interface components { supported_db_objects?: components["schemas"]["SupportedDBObjectType"][] | null; /** * Trusted Proxy Ranges - * @description CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler, and whose X-Forwarded-For is used to attribute Admin UI sign-in attempts to a source address. Set it to an empty list when clients connect directly, so the peer address is the source. Left unset, the per-source sign-in limit is off. + * @description CIDR ranges of trusted reverse proxies allowed to provide identity headers for header-based auth paths such as enable_oauth2_proxy_auth and custom_ui_sso_sign_in_handler, and whose X-Forwarded-For is used to attribute Admin UI sign-in attempts to a source address. Set it to an empty list when clients connect directly, so the peer address is the source. Left unset, or containing an entry that is not an address or CIDR range, the per-source sign-in limit is off. */ trusted_proxy_ranges?: string[] | null; /** From 8fc74a78bc0137b07af8a5c494b62cf77acc44ef Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 11:50:30 +0000 Subject: [PATCH 093/141] fix(proxy): reject blank trusted_proxy_ranges entries before they are dropped Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/login_throttle.py | 29 ++++++++++++++----- .../proxy/auth/test_login_utils.py | 5 ++++ 2 files changed, 27 insertions(+), 7 deletions(-) diff --git a/litellm/proxy/auth/login_throttle.py b/litellm/proxy/auth/login_throttle.py index 49eb7a5f20d..b7f7eaceff4 100644 --- a/litellm/proxy/auth/login_throttle.py +++ b/litellm/proxy/auth/login_throttle.py @@ -35,7 +35,7 @@ from litellm.constants import ( LOGIN_THROTTLE_UNKNOWN_SOURCE, ) from litellm.proxy._types import ProxyErrorTypes, ProxyException -from litellm.proxy.auth.network import TrustedProxyConfig, normalize_cidr_ranges, resolve_client_ip +from litellm.proxy.auth.network import TrustedProxyConfig, resolve_client_ip from litellm.secret_managers.main import get_secret_bool DEFAULT_MAX_FAILED_LOGIN_ATTEMPTS_PER_SOURCE: Final = 10 @@ -54,6 +54,7 @@ TRUSTED_PROXY_RANGES_KEY: Final = "trusted_proxy_ranges" _REDIS_FAILURES: Final = (RedisError, RedisCircuitBreakerOpenError, OSError, asyncio.TimeoutError) _LOCAL_BLOCK_EXPIRY: Final = TypeAdapter[float | None](float | None) _SOURCE_LIMIT_OVERRIDES: Final = TypeAdapter[Mapping[str, object]](Mapping[str, object]) +_RANGE_ENTRIES: Final = TypeAdapter[tuple[object, ...]](tuple[object, ...]) Scope: TypeAlias = Literal["user", "source"] @@ -134,13 +135,27 @@ def declared_proxy_ranges(settings: Mapping[str, object]) -> tuple[str, ...] | N An unset key, a value that is not a list of ranges, or a list with an entry that is not an address or range leaves it unknown and the source scope off. """ - raw_ranges: Final = settings.get(TRUSTED_PROXY_RANGES_KEY) - if isinstance(raw_ranges, (list, tuple, set)) and not raw_ranges: - return () - cidrs: Final = tuple(normalize_cidr_ranges(raw_ranges, setting_name=TRUSTED_PROXY_RANGES_KEY)) - if not cidrs or any(_parse_network(cidr, TRUSTED_PROXY_RANGES_KEY) is None for cidr in cidrs): + entries: Final = _configured_range_entries(settings.get(TRUSTED_PROXY_RANGES_KEY)) + if entries is None or any(_parse_network(entry, TRUSTED_PROXY_RANGES_KEY) is None for entry in entries): + return None + return entries + + +def _configured_range_entries(raw_ranges: object) -> tuple[str, ...] | None: + """Every configured entry, blanks included, so a stray empty string fails validation like any other typo.""" + if raw_ranges is None: + return None + if isinstance(raw_ranges, str): + return tuple(part.strip() for part in raw_ranges.split(",")) + try: + return tuple(str(entry).strip() for entry in _RANGE_ENTRIES.validate_python(raw_ranges)) + except ValidationError: + verbose_proxy_logger.warning( + "Invalid %s value: expected a list of address ranges, got %s", + TRUSTED_PROXY_RANGES_KEY, + type(raw_ranges).__name__, + ) return None - return cidrs def _positive_int(raw: object, key: str, default: int) -> int: diff --git a/tests/test_litellm/proxy/auth/test_login_utils.py b/tests/test_litellm/proxy/auth/test_login_utils.py index 90186f4b8c0..55ece36252d 100644 --- a/tests/test_litellm/proxy/auth/test_login_utils.py +++ b/tests/test_litellm/proxy/auth/test_login_utils.py @@ -944,7 +944,12 @@ async def test_an_empty_trusted_proxy_ranges_means_the_peer_is_the_client_and_th ["not-a-range"], ["10.0.0.0/8, 172.16.0.0/12"], ["10.0.0.0/8", "10.0.0.0/33"], + ["10.0.0.0/8", " "], + ["10.0.0.0/8", ""], + ["10.0.0.0/8", None], "10.0.0.0/8;172.16.0.0/12", + "10.0.0.0/8,", + "", ], ) def test_a_trusted_proxy_ranges_value_that_names_no_ranges_leaves_the_topology_unknown(configured): From ca405879a2ba9915b4fec8a35efc48360a719d1e Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 13:22:57 +0000 Subject: [PATCH 094/141] fix(cohere): set embed v3 context length to 512 tokens Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../model_prices_and_context_window_backup.json | 16 ++++++++-------- model_prices_and_context_window.json | 16 ++++++++-------- 2 files changed, 16 insertions(+), 16 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 92df776adf4..60f48ec9679 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -22170,8 +22170,8 @@ "embed-english-light-v3.0": { "input_cost_per_token": 1e-07, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "mode": "embedding", "output_cost_per_token": 0.0 }, @@ -22188,8 +22188,8 @@ "input_cost_per_image": 0.0001, "input_cost_per_token": 1e-07, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "metadata": { "notes": "'supports_image_input' is a deprecated field. Use 'supports_embedding_image_input' instead." }, @@ -22210,8 +22210,8 @@ "embed-multilingual-v3.0": { "input_cost_per_token": 1e-07, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "mode": "embedding", "output_cost_per_token": 0.0, "supports_embedding_image_input": true @@ -22219,8 +22219,8 @@ "embed-multilingual-light-v3.0": { "input_cost_per_token": 0.0001, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "mode": "embedding", "output_cost_per_token": 0.0, "supports_embedding_image_input": true diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 92df776adf4..60f48ec9679 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -22170,8 +22170,8 @@ "embed-english-light-v3.0": { "input_cost_per_token": 1e-07, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "mode": "embedding", "output_cost_per_token": 0.0 }, @@ -22188,8 +22188,8 @@ "input_cost_per_image": 0.0001, "input_cost_per_token": 1e-07, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "metadata": { "notes": "'supports_image_input' is a deprecated field. Use 'supports_embedding_image_input' instead." }, @@ -22210,8 +22210,8 @@ "embed-multilingual-v3.0": { "input_cost_per_token": 1e-07, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "mode": "embedding", "output_cost_per_token": 0.0, "supports_embedding_image_input": true @@ -22219,8 +22219,8 @@ "embed-multilingual-light-v3.0": { "input_cost_per_token": 0.0001, "litellm_provider": "cohere", - "max_input_tokens": 1024, - "max_tokens": 1024, + "max_input_tokens": 512, + "max_tokens": 512, "mode": "embedding", "output_cost_per_token": 0.0, "supports_embedding_image_input": true From b6410d563bd36caf13a5a629aa2d5a8ceb858aa8 Mon Sep 17 00:00:00 2001 From: ryan Date: Fri, 18 Sep 2026 15:46:33 +0000 Subject: [PATCH 095/141] fix(scim): accept entitlements and roles entries without a value on SCIM user PUT SCIMMultiValuedAttribute required value, so a PUT /scim/v2/Users/{id} that carried an IdP-specific entitlements entry such as {"groups": [...]} failed body validation with 422 and the suspend (active: false) never reached update_user. value is now optional and unknown members are kept, so the suspend is applied, keys are blocked, and the entries are stored as sent Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/_lazy_openapi_snapshot.json | 15 ++-- .../management_endpoints/scim/scim_v2.py | 2 +- .../proxy/management_endpoints/scim_v2.py | 4 +- .../scim/test_scim_patch_user.py | 18 ++++- .../scim/test_scim_v2_endpoints.py | 72 +++++++++++++++++++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 4 +- 6 files changed, 106 insertions(+), 9 deletions(-) diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index b244678e201..e889daf6c67 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -38245,6 +38245,7 @@ "type": "object" }, "SCIMMultiValuedAttribute": { + "additionalProperties": true, "properties": { "display": { "anyOf": [ @@ -38280,13 +38281,17 @@ "title": "Type" }, "value": { - "title": "Value", - "type": "string" + "anyOf": [ + { + "type": "string" + }, + { + "type": "null" + } + ], + "title": "Value" } }, - "required": [ - "value" - ], "title": "SCIMMultiValuedAttribute", "type": "object" }, diff --git a/litellm/proxy/management_endpoints/scim/scim_v2.py b/litellm/proxy/management_endpoints/scim/scim_v2.py index 34c1ad42435..2b74dc1e838 100644 --- a/litellm/proxy/management_endpoints/scim/scim_v2.py +++ b/litellm/proxy/management_endpoints/scim/scim_v2.py @@ -2107,7 +2107,7 @@ def _handle_multi_valued_attribute_update(path: str, op_type: str, value: object except ValidationError: raise HTTPException( status_code=400, - detail={"error": f"Invalid value for {base}: expected a list of objects with a 'value' sub-attribute"}, + detail={"error": f"Invalid value for {base}: expected a list of objects or strings"}, ) dumped: Final = [attr.model_dump(exclude_none=True) for attr in attrs] diff --git a/litellm/types/proxy/management_endpoints/scim_v2.py b/litellm/types/proxy/management_endpoints/scim_v2.py index 61fd5c36b16..6f2c48ab283 100644 --- a/litellm/types/proxy/management_endpoints/scim_v2.py +++ b/litellm/types/proxy/management_endpoints/scim_v2.py @@ -61,7 +61,9 @@ class SCIMUserGroup(BaseModel): class SCIMMultiValuedAttribute(BaseModel): - value: str + model_config = ConfigDict(extra="allow") + + value: str | None = None display: str | None = None type: str | None = None primary: bool | None = None diff --git a/tests/test_litellm/proxy/management_endpoints/scim/test_scim_patch_user.py b/tests/test_litellm/proxy/management_endpoints/scim/test_scim_patch_user.py index c3af4208d37..edcce16ab41 100644 --- a/tests/test_litellm/proxy/management_endpoints/scim/test_scim_patch_user.py +++ b/tests/test_litellm/proxy/management_endpoints/scim/test_scim_patch_user.py @@ -422,7 +422,7 @@ def test_apply_patch_ops_invalid_entitlements_value_raises_400(): patch_ops = SCIMPatchOp( Operations=[ SCIMPatchOperation( - op="replace", path="entitlements", value=[{"display": "no value"}] + op="replace", path="entitlements", value=[42] ) ] ) @@ -433,6 +433,22 @@ def test_apply_patch_ops_invalid_entitlements_value_raises_400(): assert exc_info.value.status_code == 400 +def test_apply_patch_ops_replace_entitlements_without_value_member_is_stored_as_sent(): + patch_ops = SCIMPatchOp( + Operations=[ + SCIMPatchOperation( + op="replace", path="entitlements", value=[{"groups": ["S0506MKA55L"]}] + ) + ] + ) + + update_data, _ = _apply_patch_ops( + existing_user=_user_with_metadata({}), patch_ops=patch_ops + ) + + assert update_data["metadata"]["scim_entitlements"] == [{"groups": ["S0506MKA55L"]}] + + def test_apply_patch_ops_add_without_value_raises_400_naming_value_member(): patch_ops = SCIMPatchOp( Operations=[SCIMPatchOperation(op="add", path="entitlements")] diff --git a/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py b/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py index 364ec4aad61..6d85691d8ac 100644 --- a/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py @@ -1,3 +1,4 @@ +import json import logging import time from collections.abc import Callable, Mapping, Sequence @@ -1303,6 +1304,77 @@ async def test_update_user_success(mocker): assert call_args[1]["data"]["teams"] == ["new-team"] +@pytest.mark.asyncio +async def test_update_user_put_with_valueless_entitlements_deactivates_user(scim_test_client, mocker): + """A suspend PUT whose entitlements entries carry no `value` member (an IdP-specific shape) + must not be rejected by body validation: the user is deactivated and the entries are stored as sent""" + existing_user = mocker.MagicMock() + existing_user.teams = [] + existing_user.metadata = {"scim_active": True} + + updated_user = { + "user_id": "suspend-me", + "user_email": "suspend@example.com", + "user_alias": None, + "teams": [], + "metadata": "{}", + } + response_scim_user = SCIMUser( + schemas=["urn:ietf:params:scim:schemas:core:2.0:User"], + id="suspend-me", + userName="suspend-me", + active=False, + ) + + mock_prisma_client = mocker.MagicMock() + mock_prisma_client.db = mocker.MagicMock() + mock_prisma_client.db.litellm_usertable = mocker.MagicMock() + mock_prisma_client.db.litellm_usertable.update = AsyncMock(return_value=updated_user) + + mocker.patch( # test-quality-ok: endpoint collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2._get_prisma_client_or_raise_exception", + AsyncMock(return_value=mock_prisma_client), + ) + mocker.patch( # test-quality-ok: endpoint collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2._check_user_exists", + AsyncMock(return_value=existing_user), + ) + mocker.patch( # test-quality-ok: endpoint collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2._handle_team_membership_changes", + AsyncMock(), + ) + set_keys_blocked_mock = mocker.patch( # test-quality-ok: endpoint collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2._set_user_keys_blocked", + AsyncMock(return_value=1), + ) + mocker.patch( # test-quality-ok: endpoint collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2.ScimTransformations.transform_litellm_user_to_scim_user", + AsyncMock(return_value=response_scim_user), + ) + + async with scim_test_client as client: + response = await client.put( + "/scim/v2/Users/suspend-me", + json={ + "schemas": ["urn:ietf:params:scim:schemas:core:2.0:User"], + "userName": "suspend-me", + "emails": [{"value": "suspend@example.com", "primary": True}], + "entitlements": [{"groups": ["S0506MKA55L", "S0506MKA56M"]}], + "roles": [{"display": "Viewer"}], + "active": False, + }, + ) + + assert response.status_code == 200, response.text + assert response.json()["active"] is False + + written_metadata = json.loads(mock_prisma_client.db.litellm_usertable.update.call_args.kwargs["data"]["metadata"]) + assert written_metadata["scim_active"] is False + assert written_metadata["scim_entitlements"] == [{"groups": ["S0506MKA55L", "S0506MKA56M"]}] + assert written_metadata["scim_roles"] == [{"display": "Viewer"}] + set_keys_blocked_mock.assert_awaited_once_with(user_id="suspend-me", blocked=True) + + @pytest.mark.asyncio @pytest.mark.parametrize("groups", [None, []], ids=["groups-omitted", "groups-empty"]) async def test_update_user_without_groups_preserves_memberships_and_role(mocker, monkeypatch, groups): diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index fd882937e79..4883611c3b1 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -36612,7 +36612,9 @@ export interface components { /** Type */ type?: string | null; /** Value */ - value: string; + value?: string | null; + } & { + [key: string]: unknown; }; /** SCIMPatchOp */ SCIMPatchOp: { From 817cdcef6422df4eee08b6022909c941b4750264 Mon Sep 17 00:00:00 2001 From: ryan Date: Fri, 18 Sep 2026 16:00:54 +0000 Subject: [PATCH 096/141] test(scim): drop redundant docstring from the valueless entitlements PUT test Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../proxy/management_endpoints/scim/test_scim_v2_endpoints.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py b/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py index 6d85691d8ac..dbcf622bbb1 100644 --- a/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/scim/test_scim_v2_endpoints.py @@ -1306,8 +1306,6 @@ async def test_update_user_success(mocker): @pytest.mark.asyncio async def test_update_user_put_with_valueless_entitlements_deactivates_user(scim_test_client, mocker): - """A suspend PUT whose entitlements entries carry no `value` member (an IdP-specific shape) - must not be rejected by body validation: the user is deactivated and the entries are stored as sent""" existing_user = mocker.MagicMock() existing_user.teams = [] existing_user.metadata = {"scim_active": True} From 639c71e5bb03b7c4307fece027db7d87e4835ab5 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 18:09:20 +0000 Subject: [PATCH 097/141] chore(deps): bump anyio from 4.13.0 to 4.14.2 Bumps [anyio](https://github.com/agronholm/anyio) from 4.13.0 to 4.14.2. - [Release notes](https://github.com/agronholm/anyio/releases) - [Commits](https://github.com/agronholm/anyio/compare/4.13.0...4.14.2) --- updated-dependencies: - dependency-name: anyio dependency-version: 4.14.2 dependency-type: indirect ... Signed-off-by: dependabot[bot] --- uv.lock | 388 ++++++++++++++++++++++++++++---------------------------- 1 file changed, 194 insertions(+), 194 deletions(-) diff --git a/uv.lock b/uv.lock index a5e60c68515..f04fa5a17c1 100644 --- a/uv.lock +++ b/uv.lock @@ -10,7 +10,7 @@ resolution-markers = [ ] [options] -exclude-newer = "2026-09-14T23:55:55.024292355Z" +exclude-newer = "0001-01-01T00:00:00Z" # This has no effect and is included for backwards compatibility when using relative exclude-newer values. exclude-newer-span = "P3D" [manifest] @@ -225,9 +225,9 @@ name = "aiologic" version = "0.17.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "sniffio", marker = "python_full_version < '3.13'" }, - { name = "typing-extensions", marker = "python_full_version < '3.13'" }, - { name = "wrapt", marker = "python_full_version < '3.13'" }, + { name = "sniffio" }, + { name = "typing-extensions" }, + { name = "wrapt" }, ] sdist = { url = "https://files.pythonhosted.org/packages/53/a7/809482759f40079f4c4328c7318bf569ae25d457f5017aad30a1b9aafedc/aiologic-0.17.0.tar.gz", hash = "sha256:65aa058e858c94cd208badb188e7f00b54dcabb3ba85b34f794db98074d108b9", size = 251625, upload-time = "2026-06-14T12:24:35.367Z" } wheels = [ @@ -315,16 +315,16 @@ vertex = [ [[package]] name = "anyio" -version = "4.13.0" +version = "4.14.2" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "exceptiongroup", marker = "python_full_version < '3.11'" }, { name = "idna" }, { name = "typing-extensions", marker = "python_full_version < '3.13'" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/19/14/2c5dd9f512b66549ae92767a9c7b330ae88e1932ca57876909410251fe13/anyio-4.13.0.tar.gz", hash = "sha256:334b70e641fd2221c1505b3890c69882fe4a2df910cba14d97019b90b24439dc", size = 231622, upload-time = "2026-03-24T12:59:09.671Z" } +sdist = { url = "https://files.pythonhosted.org/packages/61/cc/a381afa6efea9f496eff839d4a6a1aed3bfafc7b3ab4b0d1b243a12573dd/anyio-4.14.2.tar.gz", hash = "sha256:cfa139f3ed1a23ee8f88a145ddb5ac7605b8bbfd8592baacd7ce3d8bb4313c7f", size = 260176, upload-time = "2026-07-12T20:29:07.082Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/da/42/e921fccf5015463e32a3cf6ee7f980a6ed0f395ceeaa45060b61d86486c2/anyio-4.13.0-py3-none-any.whl", hash = "sha256:08b310f9e24a9594186fd75b4f73f4a4152069e3853f1ed8bfbf58369f4ad708", size = 114353, upload-time = "2026-03-24T12:59:08.246Z" }, + { url = "https://files.pythonhosted.org/packages/da/35/f2287558c17e29fafc8ef3daf819bb9834061cfa43bff8014f7df7f63bdc/anyio-4.14.2-py3-none-any.whl", hash = "sha256:9f505dda5ac9f0c8309b5e8bd445a8c2bf7246f3ce950121e45ea15bc41d1494", size = 125813, upload-time = "2026-07-12T20:29:05.763Z" }, ] [[package]] @@ -519,14 +519,14 @@ name = "aurelio-sdk" version = "0.0.19" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "aiofiles", marker = "python_full_version < '3.14'" }, - { name = "aiohttp", marker = "python_full_version < '3.14'" }, - { name = "colorlog", marker = "python_full_version < '3.14'" }, - { name = "pydantic", marker = "python_full_version < '3.14'" }, - { name = "python-dotenv", marker = "python_full_version < '3.14'" }, - { name = "requests", marker = "python_full_version < '3.14'" }, - { name = "requests-toolbelt", marker = "python_full_version < '3.14'" }, - { name = "tornado", marker = "python_full_version < '3.14'" }, + { name = "aiofiles" }, + { name = "aiohttp" }, + { name = "colorlog" }, + { name = "pydantic" }, + { name = "python-dotenv" }, + { name = "requests" }, + { name = "requests-toolbelt" }, + { name = "tornado" }, ] sdist = { url = "https://files.pythonhosted.org/packages/27/0e/c2e369ad173fb3d76448e46d10beb3dcc53388318933ddf8169a3f21a810/aurelio_sdk-0.0.19.tar.gz", hash = "sha256:14107e7440ff2efd0b4a08c52fb595e7680bd4bc973a0ddfb3b64157c6666b91", size = 15258, upload-time = "2025-03-24T14:37:32.203Z" } wheels = [ @@ -538,9 +538,9 @@ name = "aws-sdk-bedrock-runtime" version = "0.11.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "smithy-aws-core", extra = ["eventstream", "json"], marker = "python_full_version >= '3.12'" }, - { name = "smithy-core", marker = "python_full_version >= '3.12'" }, - { name = "smithy-http", extra = ["aiohttp"], marker = "python_full_version >= '3.12'" }, + { name = "smithy-aws-core", extra = ["eventstream", "json"] }, + { name = "smithy-core" }, + { name = "smithy-http", extra = ["aiohttp"] }, ] sdist = { url = "https://files.pythonhosted.org/packages/8e/b3/9c225cbfe9f17ea2e3d75a0fdd0b325ef79839b9c09a376bda63a7bf3bb3/aws_sdk_bedrock_runtime-0.11.0.tar.gz", hash = "sha256:f2c45d34625bf6a7b56375e29a53a16b376880bda771e4bbf7d84491622eb193", size = 173854, upload-time = "2026-08-24T21:17:16.304Z" } wheels = [ @@ -549,7 +549,7 @@ wheels = [ [package.optional-dependencies] awscrt = [ - { name = "smithy-http", extra = ["awscrt"], marker = "python_full_version >= '3.12'" }, + { name = "smithy-http", extra = ["awscrt"] }, ] [[package]] @@ -1207,7 +1207,7 @@ name = "colorlog" version = "6.10.1" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "colorama", marker = "python_full_version < '3.14' and sys_platform == 'win32'" }, + { name = "colorama", marker = "sys_platform == 'win32'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/a2/61/f083b5ac52e505dfc1c624eafbf8c7589a0d7f32daa398d2e7590efa5fda/colorlog-6.10.1.tar.gz", hash = "sha256:eb4ae5cb65fe7fec7773c2306061a8e63e02efc2c72eba9d27b0fa23c94f1321", size = 17162, upload-time = "2025-10-16T16:14:11.978Z" } wheels = [ @@ -1231,7 +1231,7 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" } }, ] sdist = { url = "https://files.pythonhosted.org/packages/66/54/eb9bfc647b19f2009dd5c7f5ec51c4e6ca831725f1aea7a993034f483147/contourpy-1.3.2.tar.gz", hash = "sha256:b6945942715a034c671b7fc54f9588126b0b8bf23db2696e3ca8328f3ff0ab54", size = 13466130, upload-time = "2025-04-15T17:47:53.79Z" } wheels = [ @@ -1304,7 +1304,7 @@ resolution-markers = [ "python_full_version == '3.11.*'", ] dependencies = [ - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" }, { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/58/01/1253e6698a07380cd31a736d248a3f2a50a7c88779a1813da27503cadc2a/contourpy-1.3.3.tar.gz", hash = "sha256:083e12155b210502d0bca491432bb04d56dc3432f95a979b429f2848c3dbe880", size = 13466174, upload-time = "2025-07-26T12:03:12.549Z" } @@ -1574,8 +1574,8 @@ name = "culsans" version = "0.11.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "aiologic", marker = "python_full_version < '3.13'" }, - { name = "typing-extensions", marker = "python_full_version < '3.13'" }, + { name = "aiologic" }, + { name = "typing-extensions" }, ] sdist = { url = "https://files.pythonhosted.org/packages/d9/e3/49afa1bc180e0d28008ec6bcdf82a4072d1c7a41032b5b759b60814ca4b0/culsans-0.11.0.tar.gz", hash = "sha256:0b43d0d05dce6106293d114c86e3fb4bfc63088cfe8ff08ed3fe36891447fe33", size = 107546, upload-time = "2025-12-31T23:15:38.196Z" } wheels = [ @@ -1829,7 +1829,7 @@ name = "exceptiongroup" version = "1.3.1" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "typing-extensions", marker = "python_full_version < '3.11'" }, + { name = "typing-extensions" }, ] sdist = { url = "https://files.pythonhosted.org/packages/50/79/66800aadf48771f6b62f7eb014e352e5d06856655206165d775e675a02c9/exceptiongroup-1.3.1.tar.gz", hash = "sha256:8b412432c6055b0b7d14c310000ae93352ed6754f70fa8f7c34141f91c4e3219", size = 30371, upload-time = "2025-11-21T23:01:54.787Z" } wheels = [ @@ -2412,11 +2412,11 @@ resolution-markers = [ "python_full_version >= '3.14'", ] dependencies = [ - { name = "google-auth", marker = "python_full_version >= '3.14'" }, - { name = "googleapis-common-protos", marker = "python_full_version >= '3.14'" }, - { name = "proto-plus", marker = "python_full_version >= '3.14'" }, - { name = "protobuf", marker = "python_full_version >= '3.14'" }, - { name = "requests", marker = "python_full_version >= '3.14'" }, + { name = "google-auth" }, + { name = "googleapis-common-protos" }, + { name = "proto-plus" }, + { name = "protobuf" }, + { name = "requests" }, ] sdist = { url = "https://files.pythonhosted.org/packages/09/cd/63f1557235c2440fe0577acdbc32577c5c002684c58c7f4d770a92366a24/google_api_core-2.25.2.tar.gz", hash = "sha256:1c63aa6af0d0d5e37966f157a77f9396d820fba59f9e43e9415bc3dc5baff300", size = 166266, upload-time = "2025-10-03T00:07:34.778Z" } wheels = [ @@ -2425,8 +2425,8 @@ wheels = [ [package.optional-dependencies] grpc = [ - { name = "grpcio", marker = "python_full_version >= '3.14'" }, - { name = "grpcio-status", marker = "python_full_version >= '3.14'" }, + { name = "grpcio" }, + { name = "grpcio-status" }, ] [[package]] @@ -2440,11 +2440,11 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "google-auth", marker = "python_full_version < '3.14'" }, - { name = "googleapis-common-protos", marker = "python_full_version < '3.14'" }, - { name = "proto-plus", marker = "python_full_version < '3.14'" }, - { name = "protobuf", marker = "python_full_version < '3.14'" }, - { name = "requests", marker = "python_full_version < '3.14'" }, + { name = "google-auth" }, + { name = "googleapis-common-protos" }, + { name = "proto-plus" }, + { name = "protobuf" }, + { name = "requests" }, ] sdist = { url = "https://files.pythonhosted.org/packages/16/ce/502a57fb0ec752026d24df1280b162294b22a0afb98a326084f9a979138b/google_api_core-2.30.3.tar.gz", hash = "sha256:e601a37f148585319b26db36e219df68c5d07b6382cff2d580e83404e44d641b", size = 177001, upload-time = "2026-04-10T00:41:28.035Z" } wheels = [ @@ -2453,8 +2453,8 @@ wheels = [ [package.optional-dependencies] grpc = [ - { name = "grpcio", marker = "python_full_version < '3.14'" }, - { name = "grpcio-status", marker = "python_full_version < '3.14'" }, + { name = "grpcio" }, + { name = "grpcio-status" }, ] [[package]] @@ -2623,12 +2623,12 @@ resolution-markers = [ "python_full_version >= '3.14'", ] dependencies = [ - { name = "google-api-core", version = "2.25.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.14'" }, - { name = "google-auth", marker = "python_full_version >= '3.14'" }, - { name = "google-cloud-core", marker = "python_full_version >= '3.14'" }, - { name = "google-crc32c", marker = "python_full_version >= '3.14'" }, - { name = "google-resumable-media", marker = "python_full_version >= '3.14'" }, - { name = "requests", marker = "python_full_version >= '3.14'" }, + { name = "google-api-core", version = "2.25.2", source = { registry = "https://pypi.org/simple" } }, + { name = "google-auth" }, + { name = "google-cloud-core" }, + { name = "google-crc32c" }, + { name = "google-resumable-media" }, + { name = "requests" }, ] sdist = { url = "https://files.pythonhosted.org/packages/bd/ef/7cefdca67a6c8b3af0ec38612f9e78e5a9f6179dd91352772ae1a9849246/google_cloud_storage-3.4.1.tar.gz", hash = "sha256:6f041a297e23a4b485fad8c305a7a6e6831855c208bcbe74d00332a909f82268", size = 17238203, upload-time = "2025-10-08T18:43:39.665Z" } wheels = [ @@ -2646,12 +2646,12 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "google-api-core", version = "2.30.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" }, - { name = "google-auth", marker = "python_full_version < '3.14'" }, - { name = "google-cloud-core", marker = "python_full_version < '3.14'" }, - { name = "google-crc32c", marker = "python_full_version < '3.14'" }, - { name = "google-resumable-media", marker = "python_full_version < '3.14'" }, - { name = "requests", marker = "python_full_version < '3.14'" }, + { name = "google-api-core", version = "2.30.3", source = { registry = "https://pypi.org/simple" } }, + { name = "google-auth" }, + { name = "google-cloud-core" }, + { name = "google-crc32c" }, + { name = "google-resumable-media" }, + { name = "requests" }, ] sdist = { url = "https://files.pythonhosted.org/packages/4c/47/205eb8e9a1739b5345843e5a425775cbdc472cc38e7eda082ba5b8d02450/google_cloud_storage-3.10.1.tar.gz", hash = "sha256:97db9aa4460727982040edd2bd13ff3d5e2260b5331ad22895802da1fc2a5286", size = 17309950, upload-time = "2026-03-23T09:35:23.409Z" } wheels = [ @@ -4081,13 +4081,13 @@ name = "langchain-classic" version = "1.0.7" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "langchain-core", marker = "python_full_version >= '3.11'" }, - { name = "langchain-text-splitters", marker = "python_full_version >= '3.11'" }, - { name = "langsmith", marker = "python_full_version >= '3.11'" }, - { name = "pydantic", marker = "python_full_version >= '3.11'" }, - { name = "pyyaml", marker = "python_full_version >= '3.11'" }, - { name = "requests", marker = "python_full_version >= '3.11'" }, - { name = "sqlalchemy", marker = "python_full_version >= '3.11'" }, + { name = "langchain-core" }, + { name = "langchain-text-splitters" }, + { name = "langsmith" }, + { name = "pydantic" }, + { name = "pyyaml" }, + { name = "requests" }, + { name = "sqlalchemy" }, ] sdist = { url = "https://files.pythonhosted.org/packages/9b/78/84b5065816f348c39fefa4316f209f0135e8410216340a953bec17d9e4e4/langchain_classic-1.0.7.tar.gz", hash = "sha256:debbec8065e69b95108d2652e8d5c44f4516e19aa8d716c02ed2211c3aee099d", size = 10554118, upload-time = "2026-05-07T15:46:56.8Z" } wheels = [ @@ -4102,18 +4102,18 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "aiohttp", marker = "python_full_version < '3.11'" }, - { name = "dataclasses-json", marker = "python_full_version < '3.11'" }, - { name = "httpx-sse", marker = "python_full_version < '3.11'" }, - { name = "langchain", marker = "python_full_version < '3.11'" }, - { name = "langchain-core", marker = "python_full_version < '3.11'" }, - { name = "langsmith", marker = "python_full_version < '3.11'" }, - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, - { name = "pydantic-settings", marker = "python_full_version < '3.11'" }, - { name = "pyyaml", marker = "python_full_version < '3.11'" }, - { name = "requests", marker = "python_full_version < '3.11'" }, - { name = "sqlalchemy", marker = "python_full_version < '3.11'" }, - { name = "tenacity", marker = "python_full_version < '3.11'" }, + { name = "aiohttp" }, + { name = "dataclasses-json" }, + { name = "httpx-sse" }, + { name = "langchain" }, + { name = "langchain-core" }, + { name = "langsmith" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" } }, + { name = "pydantic-settings" }, + { name = "pyyaml" }, + { name = "requests" }, + { name = "sqlalchemy" }, + { name = "tenacity" }, ] sdist = { url = "https://files.pythonhosted.org/packages/83/49/2ff5354273809e9811392bc24bcffda545a196070666aef27bc6aacf1c21/langchain_community-0.3.31.tar.gz", hash = "sha256:250e4c1041539130f6d6ac6f9386cb018354eafccd917b01a4cff1950b80fd81", size = 33241237, upload-time = "2025-10-07T20:17:57.857Z" } wheels = [ @@ -4131,19 +4131,19 @@ resolution-markers = [ "python_full_version == '3.11.*'", ] dependencies = [ - { name = "aiohttp", marker = "python_full_version >= '3.11'" }, - { name = "dataclasses-json", marker = "python_full_version >= '3.11'" }, - { name = "httpx-sse", marker = "python_full_version >= '3.11'" }, - { name = "langchain-classic", marker = "python_full_version >= '3.11'" }, - { name = "langchain-core", marker = "python_full_version >= '3.11'" }, - { name = "langsmith", marker = "python_full_version >= '3.11'" }, - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" }, + { name = "aiohttp" }, + { name = "dataclasses-json" }, + { name = "httpx-sse" }, + { name = "langchain-classic" }, + { name = "langchain-core" }, + { name = "langsmith" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" }, { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" }, - { name = "pydantic-settings", marker = "python_full_version >= '3.11'" }, - { name = "pyyaml", marker = "python_full_version >= '3.11'" }, - { name = "requests", marker = "python_full_version >= '3.11'" }, - { name = "sqlalchemy", marker = "python_full_version >= '3.11'" }, - { name = "tenacity", marker = "python_full_version >= '3.11'" }, + { name = "pydantic-settings" }, + { name = "pyyaml" }, + { name = "requests" }, + { name = "sqlalchemy" }, + { name = "tenacity" }, ] sdist = { url = "https://files.pythonhosted.org/packages/53/97/a03585d42b9bdb6fbd935282d6e3348b10322a24e6ce12d0c99eb461d9af/langchain_community-0.4.1.tar.gz", hash = "sha256:f3b211832728ee89f169ddce8579b80a085222ddb4f4ed445a46e977d17b1e85", size = 33241144, upload-time = "2025-10-27T15:20:32.504Z" } wheels = [ @@ -4215,7 +4215,7 @@ name = "langchain-text-splitters" version = "1.1.2" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "langchain-core", marker = "python_full_version >= '3.11'" }, + { name = "langchain-core" }, ] sdist = { url = "https://files.pythonhosted.org/packages/26/9f/6c545900fefb7b00ddfa3f16b80d61338a0ec68c31c5451eeeab99082760/langchain_text_splitters-1.1.2.tar.gz", hash = "sha256:782a723db0a4746ac91e251c7c1d57fd23636e4f38ed733074e28d7a86f41627", size = 293580, upload-time = "2026-04-16T14:20:39.162Z" } wheels = [ @@ -4959,16 +4959,16 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "aiohttp", marker = "python_full_version < '3.11'" }, - { name = "chevron", marker = "python_full_version < '3.11'" }, - { name = "jsonpickle", marker = "python_full_version < '3.11'" }, - { name = "langchain-community", version = "0.3.31", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, - { name = "packaging", marker = "python_full_version < '3.11'" }, - { name = "pydantic", marker = "python_full_version < '3.11'" }, - { name = "pyhumps", marker = "python_full_version < '3.11'" }, - { name = "requests", marker = "python_full_version < '3.11'" }, - { name = "setuptools", marker = "python_full_version < '3.11'" }, - { name = "tenacity", marker = "python_full_version < '3.11'" }, + { name = "aiohttp" }, + { name = "chevron" }, + { name = "jsonpickle" }, + { name = "langchain-community", version = "0.3.31", source = { registry = "https://pypi.org/simple" } }, + { name = "packaging" }, + { name = "pydantic" }, + { name = "pyhumps" }, + { name = "requests" }, + { name = "setuptools" }, + { name = "tenacity" }, ] sdist = { url = "https://files.pythonhosted.org/packages/4a/6f/9ca1acf766848aaf5f0ac4140c34c91ad0dbfad2654359699644be3352c9/lunary-1.4.36.tar.gz", hash = "sha256:53f002f385c83d9c0e6368e7999923acffbde987f53c5205c2c249c38ee2d75c", size = 20253, upload-time = "2026-02-09T20:49:30.56Z" } wheels = [ @@ -4986,16 +4986,16 @@ resolution-markers = [ "python_full_version == '3.11.*'", ] dependencies = [ - { name = "aiohttp", marker = "python_full_version >= '3.11'" }, - { name = "chevron", marker = "python_full_version >= '3.11'" }, - { name = "jsonpickle", marker = "python_full_version >= '3.11'" }, - { name = "langchain-community", version = "0.4.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" }, - { name = "packaging", marker = "python_full_version >= '3.11'" }, - { name = "pydantic", marker = "python_full_version >= '3.11'" }, - { name = "pyhumps", marker = "python_full_version >= '3.11'" }, - { name = "requests", marker = "python_full_version >= '3.11'" }, - { name = "setuptools", marker = "python_full_version >= '3.11'" }, - { name = "tenacity", marker = "python_full_version >= '3.11'" }, + { name = "aiohttp" }, + { name = "chevron" }, + { name = "jsonpickle" }, + { name = "langchain-community", version = "0.4.1", source = { registry = "https://pypi.org/simple" } }, + { name = "packaging" }, + { name = "pydantic" }, + { name = "pyhumps" }, + { name = "requests" }, + { name = "setuptools" }, + { name = "tenacity" }, ] sdist = { url = "https://files.pythonhosted.org/packages/37/ef/1acbc6957585cc0110e648d787663871717ced3df27fcd3cb5e18fa418f3/lunary-1.4.37.tar.gz", hash = "sha256:1781091e9dceffcc28ebc4be7e085c9fec4102d98d7ca945ed0021e9ce03c36f", size = 20248, upload-time = "2026-02-12T08:15:02.091Z" } wheels = [ @@ -8787,10 +8787,10 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "joblib", marker = "python_full_version < '3.11'" }, - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, - { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, - { name = "threadpoolctl", marker = "python_full_version < '3.11'" }, + { name = "joblib" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" } }, + { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" } }, + { name = "threadpoolctl" }, ] sdist = { url = "https://files.pythonhosted.org/packages/98/c2/a7855e41c9d285dfe86dc50b250978105dce513d6e459ea66a6aeb0e1e0c/scikit_learn-1.7.2.tar.gz", hash = "sha256:20e9e49ecd130598f1ca38a1d85090e1a600147b9c02fa6f15d69cb53d968fda", size = 7193136, upload-time = "2025-09-09T08:21:29.075Z" } wheels = [ @@ -8837,11 +8837,11 @@ resolution-markers = [ "python_full_version == '3.11.*'", ] dependencies = [ - { name = "joblib", marker = "python_full_version >= '3.11'" }, - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" }, + { name = "joblib" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" }, { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" }, - { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" }, - { name = "threadpoolctl", marker = "python_full_version >= '3.11'" }, + { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" } }, + { name = "threadpoolctl" }, ] sdist = { url = "https://files.pythonhosted.org/packages/0e/d4/40988bf3b8e34feec1d0e6a051446b1f66225f8529b9309becaeef62b6c4/scikit_learn-1.8.0.tar.gz", hash = "sha256:9bccbb3b40e3de10351f8f5068e105d0f4083b1a65fa07b6634fbc401a6287fd", size = 7335585, upload-time = "2025-12-10T07:08:53.618Z" } wheels = [ @@ -8891,7 +8891,7 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" } }, ] sdist = { url = "https://files.pythonhosted.org/packages/0f/37/6964b830433e654ec7485e45a00fc9a27cf868d622838f6b6d9c5ec0d532/scipy-1.15.3.tar.gz", hash = "sha256:eae3cf522bc7df64b42cad3925c876e1b0b6c35c1337c93e12c0f366f55b0eaf", size = 59419214, upload-time = "2025-05-08T16:13:05.955Z" } wheels = [ @@ -8953,7 +8953,7 @@ resolution-markers = [ "python_full_version == '3.11.*'", ] dependencies = [ - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" }, { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/7a/97/5a3609c4f8d58b039179648e62dd220f89864f56f7357f5d4f45c29eb2cc/scipy-1.17.1.tar.gz", hash = "sha256:95d8e012d8cb8816c226aef832200b1d45109ed4464303e997c5b13122b297c0", size = 30573822, upload-time = "2026-02-23T00:26:24.851Z" } @@ -9038,20 +9038,20 @@ name = "semantic-router" version = "0.1.15" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "aiohttp", marker = "python_full_version < '3.14'" }, - { name = "aurelio-sdk", marker = "python_full_version < '3.14'" }, - { name = "colorama", marker = "python_full_version < '3.14'" }, - { name = "colorlog", marker = "python_full_version < '3.14'" }, - { name = "litellm", marker = "python_full_version < '3.14'" }, - { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" }, + { name = "aiohttp" }, + { name = "aurelio-sdk" }, + { name = "colorama" }, + { name = "colorlog" }, + { name = "litellm" }, + { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12' or python_full_version >= '3.14'" }, { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12' and python_full_version < '3.14'" }, - { name = "openai", marker = "python_full_version < '3.14'" }, - { name = "pydantic", marker = "python_full_version < '3.14'" }, - { name = "pyyaml", marker = "python_full_version < '3.14'" }, - { name = "regex", marker = "python_full_version < '3.14'" }, - { name = "tiktoken", marker = "python_full_version < '3.14'" }, - { name = "tornado", marker = "python_full_version < '3.14'" }, - { name = "urllib3", marker = "python_full_version < '3.14'" }, + { name = "openai" }, + { name = "pydantic" }, + { name = "pyyaml" }, + { name = "regex" }, + { name = "tiktoken" }, + { name = "tornado" }, + { name = "urllib3" }, ] sdist = { url = "https://files.pythonhosted.org/packages/dc/a9/1a689e916e8b280f1fd8fb335cc059be626a22fe4533baa045d32fcd6de5/semantic_router-0.1.15.tar.gz", hash = "sha256:328256ddc3c2b713101ec69561d6585aecbf1198ea3461e1486289d8c3a35288", size = 95605, upload-time = "2026-05-23T12:58:15.444Z" } wheels = [ @@ -9134,9 +9134,9 @@ name = "smithy-aws-core" version = "0.11.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "aws-sdk-signers", marker = "python_full_version >= '3.12'" }, - { name = "smithy-core", marker = "python_full_version >= '3.12'" }, - { name = "smithy-http", marker = "python_full_version >= '3.12'" }, + { name = "aws-sdk-signers" }, + { name = "smithy-core" }, + { name = "smithy-http" }, ] sdist = { url = "https://files.pythonhosted.org/packages/7d/d3/501c0023548173416109ac42298ca33b708469dc922005770811a597949f/smithy_aws_core-0.11.0.tar.gz", hash = "sha256:29ee89976a520a87e3db557e03e115fdc21a0a60b81161e95174395a1b064da1", size = 38791, upload-time = "2026-08-24T21:16:59.631Z" } wheels = [ @@ -9145,10 +9145,10 @@ wheels = [ [package.optional-dependencies] eventstream = [ - { name = "smithy-aws-event-stream", marker = "python_full_version >= '3.12'" }, + { name = "smithy-aws-event-stream" }, ] json = [ - { name = "smithy-json", marker = "python_full_version >= '3.12'" }, + { name = "smithy-json" }, ] [[package]] @@ -9156,7 +9156,7 @@ name = "smithy-aws-event-stream" version = "0.3.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "smithy-core", marker = "python_full_version >= '3.12'" }, + { name = "smithy-core" }, ] sdist = { url = "https://files.pythonhosted.org/packages/38/0e/6efb3a4ed92c0f1ada6de060ac92e7115a1e34d0ab1fb99a6056734a88ea/smithy_aws_event_stream-0.3.0.tar.gz", hash = "sha256:a0e227367a973144e205a075d0a424f95c92f26656a1018d08900da2ae547c49", size = 12818, upload-time = "2026-05-05T18:04:14.317Z" } wheels = [ @@ -9177,7 +9177,7 @@ name = "smithy-http" version = "0.5.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "smithy-core", marker = "python_full_version >= '3.12'" }, + { name = "smithy-core" }, ] sdist = { url = "https://files.pythonhosted.org/packages/98/78/b5f3113d6c8f0bc1f9777a7f5ca84b892d29efac05850e14f7d4f7e645b5/smithy_http-0.5.0.tar.gz", hash = "sha256:bb4a19672f7c7eeb872a308f777eb505281a5bafb1ee3d1ea9c760c06c352510", size = 31122, upload-time = "2026-08-24T21:16:56.488Z" } wheels = [ @@ -9186,11 +9186,11 @@ wheels = [ [package.optional-dependencies] aiohttp = [ - { name = "aiohttp", marker = "python_full_version >= '3.12'" }, - { name = "yarl", marker = "python_full_version >= '3.12'" }, + { name = "aiohttp" }, + { name = "yarl" }, ] awscrt = [ - { name = "awscrt", marker = "python_full_version >= '3.12'" }, + { name = "awscrt" }, ] [[package]] @@ -9198,8 +9198,8 @@ name = "smithy-json" version = "0.3.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "ijson", marker = "python_full_version >= '3.12'" }, - { name = "smithy-core", marker = "python_full_version >= '3.12'" }, + { name = "ijson" }, + { name = "smithy-core" }, ] sdist = { url = "https://files.pythonhosted.org/packages/c7/ac/04164eefb3da7479f52f6535b4b39cc8384c292cb2bb74279f2acc4f4b4d/smithy_json-0.3.0.tar.gz", hash = "sha256:c81c7034587e01bc64767cbbecb05a7d65ca9070612fd94e8a03e80540290a22", size = 7956, upload-time = "2026-08-20T17:55:32.177Z" } wheels = [ @@ -9277,23 +9277,23 @@ resolution-markers = [ "python_full_version < '3.11'", ] dependencies = [ - { name = "alabaster", marker = "python_full_version < '3.11'" }, - { name = "babel", marker = "python_full_version < '3.11'" }, - { name = "colorama", marker = "python_full_version < '3.11' and sys_platform == 'win32'" }, - { name = "docutils", version = "0.21.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, - { name = "imagesize", marker = "python_full_version < '3.11'" }, - { name = "jinja2", marker = "python_full_version < '3.11'" }, - { name = "packaging", marker = "python_full_version < '3.11'" }, - { name = "pygments", marker = "python_full_version < '3.11'" }, - { name = "requests", marker = "python_full_version < '3.11'" }, - { name = "snowballstemmer", marker = "python_full_version < '3.11'" }, - { name = "sphinxcontrib-applehelp", marker = "python_full_version < '3.11'" }, - { name = "sphinxcontrib-devhelp", marker = "python_full_version < '3.11'" }, - { name = "sphinxcontrib-htmlhelp", marker = "python_full_version < '3.11'" }, - { name = "sphinxcontrib-jsmath", marker = "python_full_version < '3.11'" }, - { name = "sphinxcontrib-qthelp", marker = "python_full_version < '3.11'" }, - { name = "sphinxcontrib-serializinghtml", marker = "python_full_version < '3.11'" }, - { name = "tomli", marker = "python_full_version < '3.11'" }, + { name = "alabaster" }, + { name = "babel" }, + { name = "colorama", marker = "sys_platform == 'win32'" }, + { name = "docutils", version = "0.21.2", source = { registry = "https://pypi.org/simple" } }, + { name = "imagesize" }, + { name = "jinja2" }, + { name = "packaging" }, + { name = "pygments" }, + { name = "requests" }, + { name = "snowballstemmer" }, + { name = "sphinxcontrib-applehelp" }, + { name = "sphinxcontrib-devhelp" }, + { name = "sphinxcontrib-htmlhelp" }, + { name = "sphinxcontrib-jsmath" }, + { name = "sphinxcontrib-qthelp" }, + { name = "sphinxcontrib-serializinghtml" }, + { name = "tomli" }, ] sdist = { url = "https://files.pythonhosted.org/packages/6f/6d/be0b61178fe2cdcb67e2a92fc9ebb488e3c51c4f74a36a7824c0adf23425/sphinx-8.1.3.tar.gz", hash = "sha256:43c1911eecb0d3e161ad78611bc905d1ad0e523e4ddc202a58a821773dc4c927", size = 8184611, upload-time = "2024-10-13T20:27:13.93Z" } wheels = [ @@ -9308,23 +9308,23 @@ resolution-markers = [ "python_full_version == '3.11.*'", ] dependencies = [ - { name = "alabaster", marker = "python_full_version == '3.11.*'" }, - { name = "babel", marker = "python_full_version == '3.11.*'" }, - { name = "colorama", marker = "python_full_version == '3.11.*' and sys_platform == 'win32'" }, - { name = "docutils", version = "0.22.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" }, - { name = "imagesize", marker = "python_full_version == '3.11.*'" }, - { name = "jinja2", marker = "python_full_version == '3.11.*'" }, - { name = "packaging", marker = "python_full_version == '3.11.*'" }, - { name = "pygments", marker = "python_full_version == '3.11.*'" }, - { name = "requests", marker = "python_full_version == '3.11.*'" }, - { name = "roman-numerals", marker = "python_full_version == '3.11.*'" }, - { name = "snowballstemmer", marker = "python_full_version == '3.11.*'" }, - { name = "sphinxcontrib-applehelp", marker = "python_full_version == '3.11.*'" }, - { name = "sphinxcontrib-devhelp", marker = "python_full_version == '3.11.*'" }, - { name = "sphinxcontrib-htmlhelp", marker = "python_full_version == '3.11.*'" }, - { name = "sphinxcontrib-jsmath", marker = "python_full_version == '3.11.*'" }, - { name = "sphinxcontrib-qthelp", marker = "python_full_version == '3.11.*'" }, - { name = "sphinxcontrib-serializinghtml", marker = "python_full_version == '3.11.*'" }, + { name = "alabaster" }, + { name = "babel" }, + { name = "colorama", marker = "sys_platform == 'win32'" }, + { name = "docutils", version = "0.22.4", source = { registry = "https://pypi.org/simple" } }, + { name = "imagesize" }, + { name = "jinja2" }, + { name = "packaging" }, + { name = "pygments" }, + { name = "requests" }, + { name = "roman-numerals" }, + { name = "snowballstemmer" }, + { name = "sphinxcontrib-applehelp" }, + { name = "sphinxcontrib-devhelp" }, + { name = "sphinxcontrib-htmlhelp" }, + { name = "sphinxcontrib-jsmath" }, + { name = "sphinxcontrib-qthelp" }, + { name = "sphinxcontrib-serializinghtml" }, ] sdist = { url = "https://files.pythonhosted.org/packages/42/50/a8c6ccc36d5eacdfd7913ddccd15a9cee03ecafc5ee2bc40e1f168d85022/sphinx-9.0.4.tar.gz", hash = "sha256:594ef59d042972abbc581d8baa577404abe4e6c3b04ef61bd7fc2acbd51f3fa3", size = 8710502, upload-time = "2025-12-04T07:45:27.343Z" } wheels = [ @@ -9341,23 +9341,23 @@ resolution-markers = [ "python_full_version == '3.12.*'", ] dependencies = [ - { name = "alabaster", marker = "python_full_version >= '3.12'" }, - { name = "babel", marker = "python_full_version >= '3.12'" }, - { name = "colorama", marker = "python_full_version >= '3.12' and sys_platform == 'win32'" }, - { name = "docutils", version = "0.22.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" }, - { name = "imagesize", marker = "python_full_version >= '3.12'" }, - { name = "jinja2", marker = "python_full_version >= '3.12'" }, - { name = "packaging", marker = "python_full_version >= '3.12'" }, - { name = "pygments", marker = "python_full_version >= '3.12'" }, - { name = "requests", marker = "python_full_version >= '3.12'" }, - { name = "roman-numerals", marker = "python_full_version >= '3.12'" }, - { name = "snowballstemmer", marker = "python_full_version >= '3.12'" }, - { name = "sphinxcontrib-applehelp", marker = "python_full_version >= '3.12'" }, - { name = "sphinxcontrib-devhelp", marker = "python_full_version >= '3.12'" }, - { name = "sphinxcontrib-htmlhelp", marker = "python_full_version >= '3.12'" }, - { name = "sphinxcontrib-jsmath", marker = "python_full_version >= '3.12'" }, - { name = "sphinxcontrib-qthelp", marker = "python_full_version >= '3.12'" }, - { name = "sphinxcontrib-serializinghtml", marker = "python_full_version >= '3.12'" }, + { name = "alabaster" }, + { name = "babel" }, + { name = "colorama", marker = "sys_platform == 'win32'" }, + { name = "docutils", version = "0.22.4", source = { registry = "https://pypi.org/simple" } }, + { name = "imagesize" }, + { name = "jinja2" }, + { name = "packaging" }, + { name = "pygments" }, + { name = "requests" }, + { name = "roman-numerals" }, + { name = "snowballstemmer" }, + { name = "sphinxcontrib-applehelp" }, + { name = "sphinxcontrib-devhelp" }, + { name = "sphinxcontrib-htmlhelp" }, + { name = "sphinxcontrib-jsmath" }, + { name = "sphinxcontrib-qthelp" }, + { name = "sphinxcontrib-serializinghtml" }, ] sdist = { url = "https://files.pythonhosted.org/packages/cd/bd/f08eb0f4eed5c83f1ba2a3bd18f7745a2b1525fad70660a1c00224ec468a/sphinx-9.1.0.tar.gz", hash = "sha256:7741722357dd75f8190766926071fed3bdc211c74dd2d7d4df5404da95930ddb", size = 8718324, upload-time = "2025-12-31T15:09:27.646Z" } wheels = [ @@ -9505,8 +9505,8 @@ name = "standard-aifc" version = "3.13.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "audioop-lts", marker = "python_full_version >= '3.13'" }, - { name = "standard-chunk", marker = "python_full_version >= '3.13'" }, + { name = "audioop-lts" }, + { name = "standard-chunk" }, ] sdist = { url = "https://files.pythonhosted.org/packages/c4/53/6050dc3dde1671eb3db592c13b55a8005e5040131f7509cef0215212cb84/standard_aifc-3.13.0.tar.gz", hash = "sha256:64e249c7cb4b3daf2fdba4e95721f811bde8bdfc43ad9f936589b7bb2fae2e43", size = 15240, upload-time = "2024-10-30T16:01:31.772Z" } wheels = [ @@ -9527,7 +9527,7 @@ name = "standard-sunau" version = "3.13.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "audioop-lts", marker = "python_full_version >= '3.13'" }, + { name = "audioop-lts" }, ] sdist = { url = "https://files.pythonhosted.org/packages/66/e3/ce8d38cb2d70e05ffeddc28bb09bad77cfef979eb0a299c9117f7ed4e6a9/standard_sunau-3.13.0.tar.gz", hash = "sha256:b319a1ac95a09a2378a8442f403c66f4fd4b36616d6df6ae82b8e536ee790908", size = 9368, upload-time = "2024-10-30T16:01:41.626Z" } wheels = [ @@ -9561,8 +9561,8 @@ name = "taskgroup" version = "0.2.2" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "exceptiongroup", marker = "python_full_version < '3.11'" }, - { name = "typing-extensions", marker = "python_full_version < '3.11'" }, + { name = "exceptiongroup" }, + { name = "typing-extensions" }, ] sdist = { url = "https://files.pythonhosted.org/packages/f0/8d/e218e0160cc1b692e6e0e5ba34e8865dbb171efeb5fc9a704544b3020605/taskgroup-0.2.2.tar.gz", hash = "sha256:078483ac3e78f2e3f973e2edbf6941374fbea81b9c5d0a96f51d297717f4752d", size = 11504, upload-time = "2025-01-03T09:24:13.761Z" } wheels = [ From cf08440557bd0a377f1ce082497d037294c6aca0 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 11:42:52 -0700 Subject: [PATCH 098/141] fix(azure-ai): price FLUX.2 Flex by its resolved cost map key --- .../azure_ai/image_generation/cost_calculator.py | 5 +---- .../test_azure_ai_flux2_image_generation.py | 13 +++++++++++++ 2 files changed, 14 insertions(+), 4 deletions(-) diff --git a/litellm/llms/azure_ai/image_generation/cost_calculator.py b/litellm/llms/azure_ai/image_generation/cost_calculator.py index 086293f26c0..35d0f4fb6c3 100644 --- a/litellm/llms/azure_ai/image_generation/cost_calculator.py +++ b/litellm/llms/azure_ai/image_generation/cost_calculator.py @@ -42,9 +42,6 @@ def cost_calculator( if input_cost_per_pixel: from litellm.cost_calculator import default_image_cost_calculator - cost_model: Final = ( - model if model.startswith(f"{litellm.LlmProviders.AZURE_AI.value}/") else f"azure_ai/{model}" - ) width: Final = optional_params.get("width") if optional_params else None height: Final = optional_params.get("height") if optional_params else None pixel_size: Final = ( @@ -53,7 +50,7 @@ def cost_calculator( else size or image_response.size ) return default_image_cost_calculator( - model=cost_model, + model=_model_info["key"], custom_llm_provider=litellm.LlmProviders.AZURE_AI.value, size=pixel_size, n=num_images, diff --git a/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py index c1d46b7e919..388cca9f054 100644 --- a/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py +++ b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py @@ -172,6 +172,19 @@ def test_flux2_cost_uses_mapped_dimensions_after_response_transformation(dimensi ) == pytest.approx(5e-08 * 2048 * 1024 * 2) +def test_flux2_flex_cost_accepts_lowercase_model_spelling(): + response: Final = ImageResponse(data=[ImageObject(b64_json="aW1n"), ImageObject(b64_json="aW1n")]) + + cost: Final = litellm.completion_cost( + model="azure_ai/flux.2-flex", + completion_response=response, + optional_params={"width": 1536, "height": 1024, "num_images": 2}, + call_type="image_generation", + ) + + assert cost == pytest.approx(5e-08 * 1536 * 1024 * 2) + + def test_flux2_response_preserves_mapped_dimensions(): config = AzureFoundryFluxImageGenerationConfig() params = config.map_openai_params( From 307df09792c1088a2d7b63e424c8040fa7746595 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 18:51:43 +0000 Subject: [PATCH 099/141] test(mcp): assert the self-revoke response instead of echoing the delete mock Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../test_mcp_management_endpoints.py | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-) diff --git a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py index a69cfe0ac7d..afadd6f3d19 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py @@ -24,6 +24,7 @@ from litellm.proxy._types import ( LiteLLM_MCPServerTable, LitellmUserRoles, MCPTransport, + MCPUserCredentialResponse, NewMCPServerRequest, UpdateMCPServerRequest, UserAPIKeyAuth, @@ -5220,7 +5221,11 @@ async def test_user_naming_themselves_still_deletes_own_byok_credential(): delete_mcp_user_credential, ) - delete_mock = AsyncMock(return_value=None) + deleted_rows: list[tuple[str, str]] = [] # mutable-ok: test-local recorder for the fake delete boundary + + async def _fake_delete_user_credential(_prisma_client: object, user_id: str, server_id: str) -> None: + deleted_rows.append((user_id, server_id)) + with ( patch( # test-quality-ok: endpoint test stubs the Prisma client lookup "litellm.proxy.management_endpoints.mcp_management_endpoints.get_prisma_client_or_throw", @@ -5228,19 +5233,20 @@ async def test_user_naming_themselves_still_deletes_own_byok_credential(): ), patch( # test-quality-ok: endpoint test stubs the credential row delete "litellm.proxy.management_endpoints.mcp_management_endpoints.delete_user_credential", - new=delete_mock, + new=_fake_delete_user_credential, ), patch.object( # test-quality-ok: the cache invalidator is module scoped; the suite's only seam mcp_server, "_invalidate_byok_cred_cache", new=AsyncMock() ), ): - await delete_mcp_user_credential( + result = await delete_mcp_user_credential( server_id="srv-byok-self", user_api_key_dict=_make_user_auth("user-self"), user_id="user-self", ) - assert delete_mock.await_args.args[1:] == ("user-self", "srv-byok-self") + assert deleted_rows == [("user-self", "srv-byok-self")] + assert result == MCPUserCredentialResponse(server_id="srv-byok-self", has_credential=False) @pytest.mark.asyncio From 73fddb999e3849ffa31897426a82353ac3705521 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 11:53:26 -0700 Subject: [PATCH 100/141] fix(router): resolve stream_timeout before generic timeouts on the passthrough route --- litellm/passthrough/timeout_utils.py | 56 +++++++++-------- litellm/router.py | 4 +- .../test_pass_through_endpoints.py | 38 ++++++------ tests/test_litellm/test_router.py | 61 +++++++++---------- 4 files changed, 80 insertions(+), 79 deletions(-) diff --git a/litellm/passthrough/timeout_utils.py b/litellm/passthrough/timeout_utils.py index fb649a9eeaf..0170d93c156 100644 --- a/litellm/passthrough/timeout_utils.py +++ b/litellm/passthrough/timeout_utils.py @@ -1,9 +1,20 @@ import sys from typing import Final +from pydantic import BaseModel, ConfigDict + DEFAULT_PASS_THROUGH_REQUEST_TIMEOUT_SECONDS: Final = 600.0 +class _TimeoutFields(BaseModel): + model_config = ConfigDict(frozen=True) + + stream: bool = False + stream_timeout: float | None = None + timeout: float | None = None + request_timeout: float | None = None + + def resolve_pass_through_request_timeout( endpoint_timeout: float | None = None, ) -> float: @@ -33,8 +44,8 @@ def resolve_pass_through_request_timeout( def resolve_llm_passthrough_timeout( kwargs: dict | None = None, litellm_params: dict | None = None, - router_timeout: float | None = None, - router_stream_timeout: float | None = None, + router_timeout: float | str | None = None, + router_stream_timeout: float | str | None = None, ) -> float: """ Resolve upstream httpx timeout for SDK native passthrough (e.g. Bedrock /converse, @@ -44,28 +55,23 @@ def resolve_llm_passthrough_timeout( timeout/request_timeout -> router_timeout -> general_settings.pass_through_request_timeout -> 600s default. - Streaming (``kwargs["stream"]`` truthy) additionally consults ``stream_timeout`` at each - level before the non-streaming key, matching ``Router._get_stream_timeout`` on the - completion route: kwargs stream_timeout -> kwargs timeout/request_timeout -> - litellm_params stream_timeout -> litellm_params timeout/request_timeout -> - router_stream_timeout -> router_timeout -> pass_through_request_timeout -> 600s. + Streaming (``kwargs["stream"]`` truthy) resolves ``stream_timeout`` at every level before + any generic timeout, matching ``Router._get_stream_timeout`` on the completion route: + kwargs stream_timeout -> litellm_params stream_timeout -> router_stream_timeout, then the + non-streaming chain above. """ - kwargs = kwargs or {} - litellm_params = litellm_params or {} - is_stream: Final[bool] = bool(kwargs.get("stream", False)) - - keys: Final[tuple[str, ...]] = ( - ("stream_timeout", "timeout", "request_timeout") if is_stream else ("timeout", "request_timeout") + request: Final = _TimeoutFields.model_validate(kwargs or {}) + deployment: Final = _TimeoutFields.model_validate(litellm_params or {}) + stream_candidates: Final = ( + (request.stream_timeout, deployment.stream_timeout, router_stream_timeout) if request.stream else () ) - for source in (kwargs, litellm_params): - for key in keys: - val = source.get(key) - if val is not None: - return float(val) - - if is_stream and router_stream_timeout is not None: - return float(router_stream_timeout) - if router_timeout is not None: - return float(router_timeout) - - return resolve_pass_through_request_timeout() + candidates: Final = ( + *stream_candidates, + request.timeout, + request.request_timeout, + deployment.timeout, + deployment.request_timeout, + router_timeout, + ) + resolved: Final = next((float(val) for val in candidates if val is not None), None) + return resolved if resolved is not None else resolve_pass_through_request_timeout() diff --git a/litellm/router.py b/litellm/router.py index 7ce7ba30502..a5523d7af79 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -3880,7 +3880,9 @@ class Router: float(self._explicit_timeout) if isinstance(self._explicit_timeout, (int, float)) else None ) _router_stream_timeout: Final = ( - float(self.stream_timeout) if isinstance(self.stream_timeout, (int, float)) else None + self.stream_timeout + if self.stream_timeout is not None + else self.default_litellm_params.get("stream_timeout") ) kwargs["timeout"] = resolve_llm_passthrough_timeout( kwargs=kwargs, diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index 7f8663ea860..54336800db4 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -1120,7 +1120,6 @@ def test_resolve_llm_passthrough_timeout_precedence(): def test_resolve_llm_passthrough_timeout_stream_timeout_precedence(): - # streaming: stream_timeout wins at each level, then falls through to the non-stream keys assert ( resolve_llm_passthrough_timeout( kwargs={"stream": True, "stream_timeout": 1800, "timeout": 45}, @@ -1129,36 +1128,35 @@ def test_resolve_llm_passthrough_timeout_stream_timeout_precedence(): ) assert ( resolve_llm_passthrough_timeout( - kwargs={"stream": True}, + kwargs={"stream": True, "timeout": 45}, litellm_params={"stream_timeout": 1800, "timeout": 90}, ) == 1800.0 ) + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": True, "timeout": 45}, + litellm_params={"timeout": 90}, + router_timeout=120, + router_stream_timeout=1800, + ) + == 1800.0 + ) + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": True}, + router_stream_timeout="1800", + ) + == 1800.0 + ) assert ( resolve_llm_passthrough_timeout( kwargs={"stream": True}, litellm_params={"timeout": 90}, - router_stream_timeout=1800, + router_timeout=120, ) == 90.0 ) - assert ( - resolve_llm_passthrough_timeout( - kwargs={"stream": True}, - router_timeout=120, - router_stream_timeout=1800, - ) - == 1800.0 - ) - assert ( - resolve_llm_passthrough_timeout( - kwargs={"stream": True}, - router_timeout=120, - ) - == 120.0 - ) - - # non-streaming: stream_timeout is ignored everywhere assert ( resolve_llm_passthrough_timeout( kwargs={"stream": False, "stream_timeout": 1800}, diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 4c39f8ba4e4..e92e5de23dc 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -5480,13 +5480,17 @@ def test_update_kwargs_with_deployment_uses_pass_through_request_timeout(): assert kwargs["timeout"] == 6.0 +def _passthrough_timeout(router: litellm.Router, deployment: dict, stream: bool) -> float: + kwargs: Final[dict] = {"stream": stream} + router._update_kwargs_with_deployment( + deployment=deployment, + kwargs=kwargs, + function_name="_ageneric_api_call_with_fallbacks", + ) + return kwargs["timeout"] + + def test_update_kwargs_with_deployment_passthrough_honors_stream_timeout(): - """ - The SDK-native passthrough route (anthropic /v1/messages, bedrock /converse) resolves - its upstream timeout separately from the completion route. A streaming call must get - stream_timeout (deployment litellm_params first, then router_settings), while a - non-streaming call on the same deployment keeps the non-stream resolution. - """ router = litellm.Router( model_list=[ { @@ -5503,6 +5507,7 @@ def test_update_kwargs_with_deployment_passthrough_honors_stream_timeout(): "litellm_params": { "model": "anthropic/claude-sonnet-4-5", "api_key": "fake-key", + "timeout": 60, }, }, ], @@ -5511,37 +5516,27 @@ def test_update_kwargs_with_deployment_passthrough_honors_stream_timeout(): ) per_deployment, router_default = router.model_list - kwargs: dict = {"stream": True} - router._update_kwargs_with_deployment( - deployment=per_deployment, - kwargs=kwargs, - function_name="_ageneric_api_call_with_fallbacks", - ) - assert kwargs["timeout"] == 1800.0 + assert _passthrough_timeout(router, per_deployment, stream=True) == 1800.0 + assert _passthrough_timeout(router, router_default, stream=True) == 900.0 + assert _passthrough_timeout(router, per_deployment, stream=False) == 60.0 + assert _passthrough_timeout(router, router_default, stream=False) == 60.0 - kwargs = {"stream": True} - router._update_kwargs_with_deployment( - deployment=router_default, - kwargs=kwargs, - function_name="_ageneric_api_call_with_fallbacks", - ) - assert kwargs["timeout"] == 900.0 - kwargs = {"stream": False} - router._update_kwargs_with_deployment( - deployment=per_deployment, - kwargs=kwargs, - function_name="_ageneric_api_call_with_fallbacks", +def test_update_kwargs_with_deployment_passthrough_router_stream_timeout_sources(): + deployment: Final[dict] = { + "model_name": "anthropic-router-default", + "litellm_params": {"model": "anthropic/claude-sonnet-4-5", "api_key": "fake-key"}, + } + string_router = litellm.Router(model_list=[deployment], timeout=120, stream_timeout="900") + default_router = litellm.Router( + model_list=[deployment], + timeout=120, + default_litellm_params={"stream_timeout": 700}, ) - assert kwargs["timeout"] == 60.0 - kwargs = {"stream": False} - router._update_kwargs_with_deployment( - deployment=router_default, - kwargs=kwargs, - function_name="_ageneric_api_call_with_fallbacks", - ) - assert kwargs["timeout"] == 120.0 + assert _passthrough_timeout(string_router, string_router.model_list[0], stream=True) == 900.0 + assert _passthrough_timeout(default_router, default_router.model_list[0], stream=True) == 700.0 + assert _passthrough_timeout(default_router, default_router.model_list[0], stream=False) == 120.0 @pytest.mark.asyncio From afa8b8a9037c17dc28c2d06649464582dffe1f67 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 19:11:39 +0000 Subject: [PATCH 101/141] test(docs): read only the first column of the router_settings reference table Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/documentation_tests/test_router_settings.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/tests/documentation_tests/test_router_settings.py b/tests/documentation_tests/test_router_settings.py index 75032f80dfa..7e3d0c07459 100644 --- a/tests/documentation_tests/test_router_settings.py +++ b/tests/documentation_tests/test_router_settings.py @@ -51,9 +51,7 @@ try: if general_settings_section: # Extract the table rows, which contain the documented keys table_content = general_settings_section.group(1) - doc_key_pattern = re.compile( - r"\|\s*([^\|]+?)\s*\|" - ) # Capture the key from each row of the table + doc_key_pattern = re.compile(r"^\|\s*([^\|]+?)\s*\|", re.MULTILINE) documented_keys.update(doc_key_pattern.findall(table_content)) except Exception as e: raise Exception( From 2299846b48e4c29868bdbd7a5e145493e4de067a Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 19:18:02 +0000 Subject: [PATCH 102/141] fix(bedrock_mantle): align gpt-5.6-sol pricing with the AWS in-region model card Co-authored-by: kusumakarb Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../model_prices_and_context_window_backup.json | 16 ++++++++-------- model_prices_and_context_window.json | 16 ++++++++-------- ...st_bedrock_mantle_responses_transformation.py | 16 ++++++++++++++++ 3 files changed, 32 insertions(+), 16 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 94e20fa0213..b91c7274a1d 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -57990,14 +57990,14 @@ "supports_tool_choice": true }, "bedrock_mantle/openai.gpt-5.6-sol": { - "input_cost_per_token": 5.5e-06, - "input_cost_per_token_above_272k_tokens": 1.1e-05, - "cache_creation_input_token_cost": 6.875e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, - "cache_read_input_token_cost": 5.5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "output_cost_per_token": 3.3e-05, - "output_cost_per_token_above_272k_tokens": 4.95e-05, + "input_cost_per_token": 4.4e-06, + "input_cost_per_token_above_272k_tokens": 8.8e-06, + "cache_creation_input_token_cost": 5.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05, + "cache_read_input_token_cost": 4.4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8.8e-07, + "output_cost_per_token": 2.2e-05, + "output_cost_per_token_above_272k_tokens": 3.3e-05, "search_context_cost_per_query": { "search_context_size_high": 0.012, "search_context_size_low": 0.012, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 94e20fa0213..b91c7274a1d 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -57990,14 +57990,14 @@ "supports_tool_choice": true }, "bedrock_mantle/openai.gpt-5.6-sol": { - "input_cost_per_token": 5.5e-06, - "input_cost_per_token_above_272k_tokens": 1.1e-05, - "cache_creation_input_token_cost": 6.875e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, - "cache_read_input_token_cost": 5.5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "output_cost_per_token": 3.3e-05, - "output_cost_per_token_above_272k_tokens": 4.95e-05, + "input_cost_per_token": 4.4e-06, + "input_cost_per_token_above_272k_tokens": 8.8e-06, + "cache_creation_input_token_cost": 5.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05, + "cache_read_input_token_cost": 4.4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8.8e-07, + "output_cost_per_token": 2.2e-05, + "output_cost_per_token_above_272k_tokens": 3.3e-05, "search_context_cost_per_query": { "search_context_size_high": 0.012, "search_context_size_low": 0.012, diff --git a/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py b/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py index 901c005f5a3..81c39f736a5 100644 --- a/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py +++ b/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py @@ -1839,6 +1839,22 @@ class TestBedrockMantleResponsesSigV4: class TestBedrockMantleResponsesPricing: + @pytest.mark.parametrize( + "model", + ["openai.gpt-5.6-sol", "openai.gpt-5.6-terra", "openai.gpt-5.6-luna"], + ) + def test_mantle_matches_in_region_converse_pricing(self, local_cost_map, model): + mantle = litellm.model_cost[f"bedrock_mantle/{model}"] + converse = litellm.model_cost[f"us.{model}"] + + cost_fields = [k for k in converse if "cost" in k and k != "search_context_cost_per_query"] + assert cost_fields, "expected cost fields on the converse entry" + for field in cost_fields: + assert mantle.get(field) == pytest.approx(converse[field]), ( + f"{model}: {field} is {mantle.get(field)} on bedrock_mantle " + f"but {converse[field]} on us. (bedrock_converse)" + ) + def test_models_registered(self, local_cost_map): assert "bedrock_mantle/openai.gpt-5.5" in litellm.bedrock_mantle_models assert "bedrock_mantle/openai.gpt-5.4" in litellm.bedrock_mantle_models From 31f1c5d33107d6cd5179c3ec8c6074a8d5742f87 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 19:18:14 +0000 Subject: [PATCH 103/141] fix(openrouter): refresh glm-5.3, glm-latest, deepseek-flash-latest and hy3 pricing Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 22 +++++++++---------- model_prices_and_context_window.json | 22 +++++++++---------- 2 files changed, 22 insertions(+), 22 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index b91c7274a1d..f8a537fb922 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -65896,9 +65896,9 @@ "supports_web_search": false }, "openrouter/z-ai/glm-5.3": { - "input_cost_per_token": 1.4e-06, - "output_cost_per_token": 4.4e-06, - "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 9.1e-07, + "output_cost_per_token": 2.86e-06, + "cache_read_input_token_cost": 1.69e-07, "litellm_provider": "openrouter", "max_input_tokens": 1310720, "max_output_tokens": 943717, @@ -70619,14 +70619,14 @@ "supports_web_search": false }, "openrouter/~deepseek/deepseek-flash-latest": { - "cache_read_input_token_cost": 1.5e-08, - "input_cost_per_token": 1.5e-07, + "cache_read_input_token_cost": 4.2e-09, + "input_cost_per_token": 1.4e-07, "litellm_provider": "openrouter", "max_input_tokens": 1048576, "max_output_tokens": 393216, "max_tokens": 393216, "mode": "chat", - "output_cost_per_token": 6e-07, + "output_cost_per_token": 4.2e-07, "source": "https://openrouter.ai/api/v1/models", "supports_audio_input": false, "supports_function_calling": true, @@ -70911,14 +70911,14 @@ "supports_web_search": false }, "openrouter/~z-ai/glm-latest": { - "cache_read_input_token_cost": 1.5e-07, + "cache_read_input_token_cost": 1.46625e-07, "input_cost_per_token": 9e-07, "litellm_provider": "openrouter", "max_input_tokens": 1310720, "max_output_tokens": 235929, "max_tokens": 235929, "mode": "chat", - "output_cost_per_token": 3e-06, + "output_cost_per_token": 2.805e-06, "source": "https://openrouter.ai/api/v1/models", "supports_audio_input": false, "supports_function_calling": true, @@ -74062,14 +74062,14 @@ "supports_web_search": false }, "openrouter/tencent/hy3": { - "cache_read_input_token_cost": 3.3e-08, - "input_cost_per_token": 1.32e-07, + "cache_read_input_token_cost": 2.0625e-08, + "input_cost_per_token": 8.25e-08, "litellm_provider": "openrouter", "max_input_tokens": 262144, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "chat", - "output_cost_per_token": 5.28e-07, + "output_cost_per_token": 3.3e-07, "source": "https://openrouter.ai/api/v1/models", "supports_audio_input": false, "supports_function_calling": true, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index b91c7274a1d..f8a537fb922 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -65896,9 +65896,9 @@ "supports_web_search": false }, "openrouter/z-ai/glm-5.3": { - "input_cost_per_token": 1.4e-06, - "output_cost_per_token": 4.4e-06, - "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 9.1e-07, + "output_cost_per_token": 2.86e-06, + "cache_read_input_token_cost": 1.69e-07, "litellm_provider": "openrouter", "max_input_tokens": 1310720, "max_output_tokens": 943717, @@ -70619,14 +70619,14 @@ "supports_web_search": false }, "openrouter/~deepseek/deepseek-flash-latest": { - "cache_read_input_token_cost": 1.5e-08, - "input_cost_per_token": 1.5e-07, + "cache_read_input_token_cost": 4.2e-09, + "input_cost_per_token": 1.4e-07, "litellm_provider": "openrouter", "max_input_tokens": 1048576, "max_output_tokens": 393216, "max_tokens": 393216, "mode": "chat", - "output_cost_per_token": 6e-07, + "output_cost_per_token": 4.2e-07, "source": "https://openrouter.ai/api/v1/models", "supports_audio_input": false, "supports_function_calling": true, @@ -70911,14 +70911,14 @@ "supports_web_search": false }, "openrouter/~z-ai/glm-latest": { - "cache_read_input_token_cost": 1.5e-07, + "cache_read_input_token_cost": 1.46625e-07, "input_cost_per_token": 9e-07, "litellm_provider": "openrouter", "max_input_tokens": 1310720, "max_output_tokens": 235929, "max_tokens": 235929, "mode": "chat", - "output_cost_per_token": 3e-06, + "output_cost_per_token": 2.805e-06, "source": "https://openrouter.ai/api/v1/models", "supports_audio_input": false, "supports_function_calling": true, @@ -74062,14 +74062,14 @@ "supports_web_search": false }, "openrouter/tencent/hy3": { - "cache_read_input_token_cost": 3.3e-08, - "input_cost_per_token": 1.32e-07, + "cache_read_input_token_cost": 2.0625e-08, + "input_cost_per_token": 8.25e-08, "litellm_provider": "openrouter", "max_input_tokens": 262144, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "chat", - "output_cost_per_token": 5.28e-07, + "output_cost_per_token": 3.3e-07, "source": "https://openrouter.ai/api/v1/models", "supports_audio_input": false, "supports_function_calling": true, From a774e71175177641c2914dce33ac5d8d106ca131 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 19:18:45 +0000 Subject: [PATCH 104/141] test(bedrock_mantle): document mantle in-region pricing invariant Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../test_bedrock_mantle_responses_transformation.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py b/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py index 81c39f736a5..951911ac066 100644 --- a/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py +++ b/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py @@ -1844,6 +1844,11 @@ class TestBedrockMantleResponsesPricing: ["openai.gpt-5.6-sol", "openai.gpt-5.6-terra", "openai.gpt-5.6-luna"], ) def test_mantle_matches_in_region_converse_pricing(self, local_cost_map, model): + """bedrock-mantle serves these models In-Region only, and the AWS model + cards price In-Region and Geo CRIS identically -- so every cost field on + the mantle key must equal the `us.` converse key. A price change applied + to one namespace but not the other shows up here. + """ mantle = litellm.model_cost[f"bedrock_mantle/{model}"] converse = litellm.model_cost[f"us.{model}"] From 1e7c5400fd792c2b7f5d9915ac832c4147f256dc Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 19:27:25 +0000 Subject: [PATCH 105/141] fix(proxy): canonicalize azure speech paths and bill uploaded short audio Resolve dot segments in the /azure_speech endpoint path before the endpoint family and the admin-only batch guard are decided, so the guard and the forwarded upstream path agree. Bill short-audio requests for the longer of the uploaded audio duration and the recognized duration, so a NoMatch or silence response still charges for the audio Azure processed Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../llm_passthrough_endpoints.py | 16 ++- ...zure_speech_passthrough_logging_handler.py | 40 +++++++- ...zure_speech_passthrough_logging_handler.py | 99 ++++++++++++++++--- .../test_llm_pass_through_endpoints.py | 75 ++++++++++++++ 4 files changed, 207 insertions(+), 23 deletions(-) diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index ab4b636cb60..ed70369506d 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -12,6 +12,7 @@ import hmac import inspect import json import os +import posixpath import re from collections.abc import AsyncGenerator, Callable, Mapping, Sequence from dataclasses import dataclass @@ -1408,6 +1409,18 @@ def azure_speech_path_manages_shared_resources(endpoint_path: str) -> bool: ) +def canonical_azure_speech_endpoint_path(endpoint: str) -> str: + """ + The path Azure will actually serve, with ``.`` and ``..`` segments resolved, so the + endpoint family and the admin guard are decided on the same path the upstream request uses. + """ + raw_path: Final = httpx.URL(endpoint).path + resolved_path: Final = posixpath.normpath(f"/{raw_path.lstrip('/')}") + if raw_path.endswith("/") and resolved_path != "/": + return f"{resolved_path}/" + return resolved_path + + @router.api_route( f"{AZURE_SPEECH_PASS_THROUGH_ROUTE_PREFIX}/{{endpoint:path}}", methods=["GET", "POST", "PUT", "DELETE", "PATCH"], # mutable-ok: fastapi route methods must be a list @@ -1430,8 +1443,7 @@ async def azure_speech_proxy_route( [Docs](https://docs.litellm.ai/docs/pass_through/azure_speech) """ - endpoint_path: Final = httpx.URL(endpoint).path - normalized_endpoint_path: Final = endpoint_path if endpoint_path.startswith("/") else f"/{endpoint_path}" + normalized_endpoint_path: Final = canonical_azure_speech_endpoint_path(endpoint) base_url: Final = resolve_azure_speech_base_url( endpoint_path=normalized_endpoint_path, api_base=get_secret_str(secret_name="AZURE_SPEECH_API_BASE"), diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py index 588b7cc8e56..8cd1b137de8 100644 --- a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/azure_speech_passthrough_logging_handler.py @@ -18,6 +18,7 @@ from litellm.constants import ( AZURE_SPEECH_TICKS_PER_SECOND, ) from litellm.cost_calculator import transcription_cost +from litellm.litellm_core_utils.audio_utils.utils import calculate_request_duration from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.litellm_logging import ( get_standard_logging_object_payload, @@ -53,6 +54,23 @@ class AzureSpeechPassthroughLoggingHandler: return 0.0 return (offset + duration) / AZURE_SPEECH_TICKS_PER_SECOND + @staticmethod + def _uploaded_audio_seconds(httpx_response: httpx.Response) -> float: + try: + uploaded_audio: Final = httpx_response.request.content + except RuntimeError: + return 0.0 + return calculate_request_duration(uploaded_audio) or 0.0 + + @staticmethod + def _short_audio_seconds( + httpx_response: httpx.Response, response_body: Mapping[str, object] | Sequence[object] | None + ) -> float: + return max( + AzureSpeechPassthroughLoggingHandler._uploaded_audio_seconds(httpx_response), + AzureSpeechPassthroughLoggingHandler._recognized_audio_seconds(response_body), + ) + @staticmethod def _fast_transcription_audio_seconds(response_body: Mapping[str, object] | Sequence[object] | None) -> float: if not isinstance(response_body, Mapping): @@ -63,16 +81,26 @@ class AzureSpeechPassthroughLoggingHandler: return duration_milliseconds / AZURE_SPEECH_MILLISECONDS_PER_SECOND @staticmethod - def _billed_audio_seconds(url_route: str, response_body: Mapping[str, object] | Sequence[object] | None) -> float: + def _billed_audio_seconds( + url_route: str, + httpx_response: httpx.Response, + response_body: Mapping[str, object] | Sequence[object] | None, + ) -> float: if AzureSpeechPassthroughLoggingHandler._is_short_audio_route(url_route): - return AzureSpeechPassthroughLoggingHandler._recognized_audio_seconds(response_body) + return AzureSpeechPassthroughLoggingHandler._short_audio_seconds(httpx_response, response_body) if AzureSpeechPassthroughLoggingHandler._is_fast_transcription_route(url_route): return AzureSpeechPassthroughLoggingHandler._fast_transcription_audio_seconds(response_body) return 0.0 @staticmethod - def _response_cost(url_route: str, response_body: Mapping[str, object] | Sequence[object] | None) -> float: - audio_seconds: Final = AzureSpeechPassthroughLoggingHandler._billed_audio_seconds(url_route, response_body) + def _response_cost( + url_route: str, + httpx_response: httpx.Response, + response_body: Mapping[str, object] | Sequence[object] | None, + ) -> float: + audio_seconds: Final = AzureSpeechPassthroughLoggingHandler._billed_audio_seconds( + url_route, httpx_response, response_body + ) if audio_seconds <= 0.0: return 0.0 try: @@ -103,7 +131,9 @@ class AzureSpeechPassthroughLoggingHandler: ) -> PassThroughEndpointLoggingTypedDict: try: model_name: Final = AzureSpeechPassthroughLoggingHandler._model_from_url_route(url_route) - response_cost: Final = AzureSpeechPassthroughLoggingHandler._response_cost(url_route, response_body) + response_cost: Final = AzureSpeechPassthroughLoggingHandler._response_cost( + url_route, httpx_response, response_body + ) updated_kwargs: Final = { # mutable-ok: the logging pipeline requires a plain kwargs dict **kwargs, diff --git a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py index 50d3de64f72..670f8e65823 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_azure_speech_passthrough_logging_handler.py @@ -1,5 +1,8 @@ +import io import json +import wave from datetime import datetime +from typing import Final from unittest.mock import MagicMock import httpx @@ -29,6 +32,25 @@ TRANSCRIPT_BODY = { TRANSCRIPT = json.dumps(TRANSCRIPT_BODY) TRANSCRIPT_AUDIO_SECONDS = 3.0 PRICE_PER_SECOND = 0.5 +WAV_SAMPLE_RATE: Final = 16000 +UNRECOGNIZED_BODIES: Final = ( + {"RecognitionStatus": "NoMatch", "Offset": 0, "Duration": 0}, + {"RecognitionStatus": "InitialSilenceTimeout"}, + {"Offset": "5000000", "Duration": "25000000"}, + {}, + [], + None, +) + + +def _pcm16_wav(seconds: float) -> bytes: + buffer: Final = io.BytesIO() + with wave.open(buffer, "wb") as wav: + wav.setnchannels(1) + wav.setsampwidth(2) + wav.setframerate(WAV_SAMPLE_RATE) + wav.writeframes(b"\x00\x00" * int(seconds * WAV_SAMPLE_RATE)) + return buffer.getvalue() @pytest.fixture(autouse=True) @@ -45,8 +67,8 @@ def azure_stt_price(monkeypatch: pytest.MonkeyPatch): ) -def _make_response(url: str) -> httpx.Response: - request = httpx.Request("POST", url, headers={"Ocp-Apim-Subscription-Key": "server-secret"}) +def _make_response(url: str, uploaded: bytes = b"") -> httpx.Response: + request = httpx.Request("POST", url, headers={"Ocp-Apim-Subscription-Key": "server-secret"}, content=uploaded) return httpx.Response(200, request=request, text=TRANSCRIPT) @@ -92,22 +114,13 @@ class TestAzureSpeechPassthroughHandler: assert logging_obj.model_call_details["custom_llm_provider"] == "azure_speech" assert logging_obj.model_call_details["response_cost"] == pytest.approx(expected_cost) - @pytest.mark.parametrize( - "response_body", - [ - {"RecognitionStatus": "NoMatch", "Offset": 0, "Duration": 0}, - {"RecognitionStatus": "InitialSilenceTimeout"}, - {"Offset": "5000000", "Duration": "25000000"}, - {}, - [], - None, - ], - ) - def test_short_audio_without_recognized_duration_logs_zero_cost( - self, response_body: dict[str, object] | list[dict[str, object]] | None + @pytest.mark.parametrize("response_body", UNRECOGNIZED_BODIES) + @pytest.mark.parametrize("uploaded", [b"", b"not audio at all"]) + def test_short_audio_with_neither_recognized_nor_decodable_audio_logs_zero_cost( + self, response_body: dict[str, object] | list[dict[str, object]] | None, uploaded: bytes ): handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( - httpx_response=_make_response(SHORT_AUDIO_URL), + httpx_response=_make_response(SHORT_AUDIO_URL, uploaded), response_body=response_body, logging_obj=_make_logging_obj(), url_route=SHORT_AUDIO_URL, @@ -121,6 +134,60 @@ class TestAzureSpeechPassthroughHandler: assert handler_result["kwargs"]["model"] == "azure_speech/short-audio" assert handler_result["kwargs"]["response_cost"] == 0.0 + @pytest.mark.parametrize("response_body", UNRECOGNIZED_BODIES) + def test_short_audio_bills_the_uploaded_audio_when_nothing_was_recognized( + self, response_body: dict[str, object] | list[dict[str, object]] | None + ): + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(SHORT_AUDIO_URL, _pcm16_wav(seconds=2.0)), + response_body=response_body, + logging_obj=_make_logging_obj(), + url_route=SHORT_AUDIO_URL, + result="", + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert handler_result["kwargs"]["response_cost"] == pytest.approx(2.0 * PRICE_PER_SECOND) + + @pytest.mark.parametrize( + "uploaded_seconds,expected_seconds", + [(1.0, TRANSCRIPT_AUDIO_SECONDS), (TRANSCRIPT_AUDIO_SECONDS + 2.0, TRANSCRIPT_AUDIO_SECONDS + 2.0)], + ) + def test_short_audio_bills_the_longer_of_uploaded_and_recognized_audio( + self, uploaded_seconds: float, expected_seconds: float + ): + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(SHORT_AUDIO_URL, _pcm16_wav(seconds=uploaded_seconds)), + response_body=TRANSCRIPT_BODY, + logging_obj=_make_logging_obj(), + url_route=SHORT_AUDIO_URL, + result=TRANSCRIPT, + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert handler_result["kwargs"]["response_cost"] == pytest.approx(expected_seconds * PRICE_PER_SECOND) + + def test_fast_transcription_ignores_the_uploaded_multipart_body(self): + handler_result = AzureSpeechPassthroughLoggingHandler.azure_speech_passthrough_handler( + httpx_response=_make_response(FAST_URL, _pcm16_wav(seconds=30.0)), + response_body=FAST_BODY, + logging_obj=_make_logging_obj(), + url_route=FAST_URL, + result=json.dumps(FAST_BODY), + start_time=datetime.now(), + end_time=datetime.now(), + cache_hit=False, + request_body={}, + ) + + assert handler_result["kwargs"]["response_cost"] == pytest.approx(FAST_AUDIO_SECONDS * PRICE_PER_SECOND) + @pytest.mark.parametrize( "response_body", [{"durationMilliseconds": 0}, {"durationMilliseconds": "5061"}, {"duration": 5061}, {}, [], None], diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index 9d210ac2f4f..636980eb6e3 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -33,6 +33,7 @@ from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import ( _proxy_general_settings, anthropic_proxy_route, azure_proxy_route, + azure_speech_proxy_route, bedrock_llm_proxy_route, bedrock_proxy_route, create_pass_through_route, @@ -6443,6 +6444,7 @@ AZURE_SPEECH_PCM16_HEADER: Final = ( b"RIFF\x24\x0c\x00\x00WAVEfmt \x10\x00\x00\x00\x01\x00\x01\x00\x80\x3e\x00\x00\x00\x7d\x00\x00\x02\x00\x10\x00data\x00\x0c\x00\x00" ) AZURE_SPEECH_WAV_BYTES: Final = AZURE_SPEECH_PCM16_HEADER + b"\x00" * 3072 +AZURE_SPEECH_WAV_SECONDS: Final = 3072 / (16000 * 2) AZURE_SPEECH_NON_UTF8_WAV_BYTES: Final = AZURE_SPEECH_PCM16_HEADER + bytes(range(256)) * 12 AZURE_SPEECH_TRANSCRIPT: Final = {"RecognitionStatus": "Success", "DisplayText": "The eagle has landed."} @@ -6826,6 +6828,79 @@ class TestAzureSpeechProxyRoute: assert "/azure_speech" in LiteLLMRoutes.mapped_pass_through_routes.value + def test_short_audio_with_no_recognized_speech_is_billed_for_the_uploaded_audio( + self, azure_speech_client: TestClient, monkeypatch: pytest.MonkeyPatch + ) -> None: + from litellm.integrations.custom_logger import CustomLogger + + class _Recorder(CustomLogger): + def __init__(self) -> None: + super().__init__() + self.payloads: list[dict[str, object]] = [] # mutable-ok: test recorder accumulates callback payloads + + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time) -> None: + self.payloads.append(kwargs["standard_logging_object"]) + + recorder: Final = _Recorder() + monkeypatch.setattr(litellm, "_async_success_callback", [*litellm._async_success_callback, recorder]) + monkeypatch.setitem( + litellm.model_cost, + "azure/speech/azure-stt", + { + "litellm_provider": "azure", + "mode": "audio_transcription", + "input_cost_per_second": 0.25, + "output_cost_per_second": 0.0, + }, + ) + with respx.mock(assert_all_called=True) as upstream: + upstream.post(f"https://eastus.stt.speech.microsoft.com{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}").mock( + return_value=httpx.Response(200, json={"RecognitionStatus": "NoMatch", "Offset": 0, "Duration": 0}) + ) + + response = azure_speech_client.post( + f"/azure_speech{AZURE_SPEECH_SHORT_AUDIO_ENDPOINT}", + content=AZURE_SPEECH_WAV_BYTES, + headers={"Content-Type": "audio/wav", "Authorization": "Bearer sk-virtual"}, + ) + + assert response.status_code == 200 + assert [p["model"] for p in recorder.payloads] == ["azure_speech/short-audio"] + assert recorder.payloads[0]["response_cost"] == pytest.approx(AZURE_SPEECH_WAV_SECONDS * 0.25) + + +class TestAzureSpeechProxyRoutePathTraversal: + """Calls the route function directly because httpx clients resolve dot segments before sending.""" + + @pytest.mark.parametrize( + "endpoint", + [ + f"speech/..{AZURE_SPEECH_BATCH_ENDPOINT}", + f"speech/recognition/../..{AZURE_SPEECH_BATCH_ENDPOINT}/", + f"speech/./..{AZURE_SPEECH_BATCH_ENDPOINT}/8a5d3f2c-0b1e-4c7d-9e6f-1234567890ab", + ], + ) + @pytest.mark.asyncio + async def test_dot_segments_cannot_reach_shared_batch_resources_with_a_non_admin_key( + self, monkeypatch: pytest.MonkeyPatch, endpoint: str + ) -> None: + monkeypatch.setenv("AZURE_SPEECH_API_KEY", "server-subscription-key") + monkeypatch.setenv("AZURE_SPEECH_REGION", "eastus") + monkeypatch.delenv("AZURE_SPEECH_API_BASE", raising=False) + request: Final = MagicMock(spec=Request) + request.method = "GET" + + with pytest.raises(HTTPException) as denied: + await azure_speech_proxy_route( + endpoint=endpoint, + request=request, + fastapi_response=MagicMock(spec=Response), + user_api_key_dict=UserAPIKeyAuth(api_key="sk-virtual"), + ) + + assert denied.value.status_code == 403 + assert AZURE_SPEECH_FAST_ENDPOINT in str(denied.value.detail) + def _azure_speech_real_auth_attrs() -> dict[str, object]: from litellm.caching.caching import DualCache From d4ee62eb8c3083a48fb21f72da49c14a578dc993 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 12:29:32 -0700 Subject: [PATCH 106/141] refactor(passthrough): type the timeout resolver mapping parameters --- litellm/passthrough/timeout_utils.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/litellm/passthrough/timeout_utils.py b/litellm/passthrough/timeout_utils.py index 0170d93c156..829277105e3 100644 --- a/litellm/passthrough/timeout_utils.py +++ b/litellm/passthrough/timeout_utils.py @@ -1,4 +1,5 @@ import sys +from collections.abc import Mapping from typing import Final from pydantic import BaseModel, ConfigDict @@ -42,8 +43,8 @@ def resolve_pass_through_request_timeout( def resolve_llm_passthrough_timeout( - kwargs: dict | None = None, - litellm_params: dict | None = None, + kwargs: Mapping[str, object] | None = None, + litellm_params: Mapping[str, object] | None = None, router_timeout: float | str | None = None, router_stream_timeout: float | str | None = None, ) -> float: From 12b7268ae986bde9dc08c19fff8b188eec084fa7 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 12:29:55 -0700 Subject: [PATCH 107/141] feat(vscode): add LiteLLM language model provider extension --- .github/workflows/test-vscode-extension.yml | 65 + vscode-extension/.gitignore | 3 + vscode-extension/.vscodeignore | 10 + vscode-extension/LICENSE | 21 + vscode-extension/README.md | 31 + vscode-extension/package-lock.json | 3570 +++++++++++++++++++ vscode-extension/package.json | 87 + vscode-extension/src/extension.ts | 17 + vscode-extension/src/gateway.ts | 114 + vscode-extension/src/messages.ts | 188 + vscode-extension/src/models.ts | 170 + vscode-extension/src/provider.ts | 136 + vscode-extension/src/stream.ts | 90 + vscode-extension/src/vscode.proposed.d.ts | 15 + vscode-extension/test/gateway.test.ts | 205 ++ vscode-extension/test/messages.test.ts | 168 + vscode-extension/test/models.test.ts | 185 + vscode-extension/test/provider.test.ts | 258 ++ vscode-extension/test/stream.test.ts | 93 + vscode-extension/test/vscode-mock.ts | 82 + vscode-extension/tsconfig.json | 19 + vscode-extension/vitest.config.mts | 13 + 22 files changed, 5540 insertions(+) create mode 100644 .github/workflows/test-vscode-extension.yml create mode 100644 vscode-extension/.gitignore create mode 100644 vscode-extension/.vscodeignore create mode 100644 vscode-extension/LICENSE create mode 100644 vscode-extension/README.md create mode 100644 vscode-extension/package-lock.json create mode 100644 vscode-extension/package.json create mode 100644 vscode-extension/src/extension.ts create mode 100644 vscode-extension/src/gateway.ts create mode 100644 vscode-extension/src/messages.ts create mode 100644 vscode-extension/src/models.ts create mode 100644 vscode-extension/src/provider.ts create mode 100644 vscode-extension/src/stream.ts create mode 100644 vscode-extension/src/vscode.proposed.d.ts create mode 100644 vscode-extension/test/gateway.test.ts create mode 100644 vscode-extension/test/messages.test.ts create mode 100644 vscode-extension/test/models.test.ts create mode 100644 vscode-extension/test/provider.test.ts create mode 100644 vscode-extension/test/stream.test.ts create mode 100644 vscode-extension/test/vscode-mock.ts create mode 100644 vscode-extension/tsconfig.json create mode 100644 vscode-extension/vitest.config.mts diff --git a/.github/workflows/test-vscode-extension.yml b/.github/workflows/test-vscode-extension.yml new file mode 100644 index 00000000000..886268d9e2c --- /dev/null +++ b/.github/workflows/test-vscode-extension.yml @@ -0,0 +1,65 @@ +name: VS Code Extension +permissions: + contents: read + +on: + pull_request: + branches: + - main + - litellm_internal_staging + - litellm_oss_staging + - "litellm_**" + paths: + - "vscode-extension/**" + - ".github/workflows/test-vscode-extension.yml" + push: + branches: + - main + paths: + - "vscode-extension/**" + - ".github/workflows/test-vscode-extension.yml" + +concurrency: + group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }} + cancel-in-progress: ${{ github.event_name == 'pull_request' }} + +jobs: + vscode-extension: + runs-on: ubuntu-latest + timeout-minutes: 10 + defaults: + run: + working-directory: vscode-extension + + steps: + - name: Checkout repository + uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0 + with: + fetch-depth: 1 + persist-credentials: false + + - name: Set up Node.js + uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0 + with: + node-version: "24" + cache: npm + cache-dependency-path: vscode-extension/package-lock.json + + - name: Install dependencies + run: npm ci + + - name: Typecheck + run: npm run typecheck + + - name: Unit tests + run: npm test + + - name: Package extension + run: npm run package + + - name: Upload VSIX + uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 # v4.6.1 + with: + name: litellm-vscode + path: vscode-extension/*.vsix + if-no-files-found: error diff --git a/vscode-extension/.gitignore b/vscode-extension/.gitignore new file mode 100644 index 00000000000..a08e1da2de7 --- /dev/null +++ b/vscode-extension/.gitignore @@ -0,0 +1,3 @@ +node_modules/ +dist/ +*.vsix diff --git a/vscode-extension/.vscodeignore b/vscode-extension/.vscodeignore new file mode 100644 index 00000000000..dc7c667ce84 --- /dev/null +++ b/vscode-extension/.vscodeignore @@ -0,0 +1,10 @@ +.gitignore +.vscodeignore +node_modules/** +src/** +test/** +tsconfig.json +package-lock.json +**/*.map +**/*.vsix +vitest.config.mts diff --git a/vscode-extension/LICENSE b/vscode-extension/LICENSE new file mode 100644 index 00000000000..dd11dc52350 --- /dev/null +++ b/vscode-extension/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Berri AI + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/vscode-extension/README.md b/vscode-extension/README.md new file mode 100644 index 00000000000..47c1a3afd27 --- /dev/null +++ b/vscode-extension/README.md @@ -0,0 +1,31 @@ +# LiteLLM for VS Code + +Chat in VS Code with every model your [LiteLLM AI Gateway](https://docs.litellm.ai) exposes. The extension registers LiteLLM as a language model provider, so the gateway's models show up in the chat model picker next to the built-in ones, with the price and reasoning effort controls the gateway reports for each of them + +## What you get + +The model list comes from the gateway's `GET /model_group/info` endpoint, scoped to the virtual key you configure, so the picker shows exactly the chat models that key can use. Each model carries its input and output price per 1M tokens in the picker and in the Language Models editor, and its context limits come from the gateway too, so VS Code sizes prompts correctly. A model whose gateway entry lists `supported_reasoning_efforts` gets a Reasoning Effort submenu in the picker's Configure Model menu, and the chosen effort is sent as `reasoning_effort` on every request to that model. Requests go to `POST /v1/chat/completions` on the gateway as streaming chat completions with tools and images passed through, so routing, fallbacks, guardrails, and spend tracking all apply as usual + +## Setup + +1. Install the extension +2. Run `Chat: Manage Language Models` from the Command Palette and pick `LiteLLM` +3. Enter a name for the connection, the gateway URL (for example `https://litellm.example.com`), and a LiteLLM virtual key. The key is stored in VS Code's secret storage +4. Open the chat model picker. The gateway's chat models are listed under the name you chose, each with its price + +Add the same provider again with another name to reach a second gateway or a second key. Run `LiteLLM: Refresh Models` after the gateway's model list changes. To change the key of an existing connection or to drop it, use the gear on its row in the Language Models editor (`Update API Key`, `Delete`); to change the URL, open its entry with `Open in Language Models (JSON)` from the same menu. If the stored key is ever lost the editor shows a `missing its API key` row for that connection until you update the key + +## Requirements + +VS Code 1.109 or newer and a LiteLLM AI Gateway the key can reach. The key needs access to at least one model group whose mode is `chat` + +## Development + +``` +npm ci +npm run typecheck +npm test +npm run package +``` + +`npm run package` writes a `.vsix` you can install with `code --install-extension litellm-vscode-.vsix` diff --git a/vscode-extension/package-lock.json b/vscode-extension/package-lock.json new file mode 100644 index 00000000000..378579dc024 --- /dev/null +++ b/vscode-extension/package-lock.json @@ -0,0 +1,3570 @@ +{ + "name": "litellm-vscode", + "version": "0.1.0", + "lockfileVersion": 3, + "requires": true, + "packages": { + "": { + "name": "litellm-vscode", + "version": "0.1.0", + "license": "MIT", + "dependencies": { + "openai": "^7.18.0" + }, + "devDependencies": { + "@types/node": "^22.20.3", + "@types/vscode": "^1.109.0", + "@vscode/vsce": "^4.0.0", + "esbuild": "^0.28.2", + "typescript": "^5.9.3", + "vitest": "^4.1.11" + }, + "engines": { + "vscode": "^1.109.0" + } + }, + "node_modules/@azure/abort-controller": { + "version": "2.2.0", + "resolved": "https://registry.npmjs.org/@azure/abort-controller/-/abort-controller-2.2.0.tgz", + "integrity": "sha512-fNAjWnA/nZ2jz31kxR/AqRaUT8ewHBw/WuBIosK0moMy1C9e5ValbDfFdIxJzVOOYaYkV/b2F1S4H/aHiqfVQg==", + "dev": true, + "license": "MIT", + "dependencies": { + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/core-auth": { + "version": "1.11.0", + "resolved": "https://registry.npmjs.org/@azure/core-auth/-/core-auth-1.11.0.tgz", + "integrity": "sha512-IUZydyTUkDnYdstOW9pFOOUQlBjAepK5teihDE3x6yxsPJs/hsAaaYpeGxdxrgtOiJbBKSjKW7MDk7AEhb4LRg==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/abort-controller": "^2.1.2", + "@azure/core-util": "^1.13.0", + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/core-client": { + "version": "1.11.1", + "resolved": "https://registry.npmjs.org/@azure/core-client/-/core-client-1.11.1.tgz", + "integrity": "sha512-2QygG2F76ZpMP2eMztiJvAiFMu71M9rDeU7vO/QKg5Css7MgM4frUOslFjhVjRhbGaCNPtz/S8M6y46/fFKVuQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/abort-controller": "^2.1.2", + "@azure/core-auth": "^1.10.0", + "@azure/core-rest-pipeline": "^1.22.0", + "@azure/core-tracing": "^1.3.0", + "@azure/core-util": "^1.13.0", + "@azure/logger": "^1.3.0", + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/core-process": { + "version": "1.0.0", + "resolved": "https://registry.npmjs.org/@azure/core-process/-/core-process-1.0.0.tgz", + "integrity": "sha512-/shnJ+ooO8WPxDhPEeI/2oRQuubn16gZ6CvlbpWbEswZfzwI9tI/sMAHmF3x1LuQ9yZYXfLW3TjzGMLEC5blKg==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/core-rest-pipeline": { + "version": "1.25.0", + "resolved": "https://registry.npmjs.org/@azure/core-rest-pipeline/-/core-rest-pipeline-1.25.0.tgz", + "integrity": "sha512-bMs8ekJLjX8wPV+9IPBges1SLPyuDtE9g5gLDWOpxzKcoOFQnpLGkbcT1tdw3FaAmDS1gnPmMmJ6y/T5B96kIA==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/abort-controller": "^2.1.2", + "@azure/core-auth": "^1.10.0", + "@azure/core-tracing": "^1.3.0", + "@azure/core-util": "^1.13.0", + "@azure/logger": "^1.3.0", + "@typespec/ts-http-runtime": "^0.3.4", + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/core-tracing": { + "version": "1.4.0", + "resolved": "https://registry.npmjs.org/@azure/core-tracing/-/core-tracing-1.4.0.tgz", + "integrity": "sha512-eGwxD0AtncrxeBM4tG8R55Pc3rdX1hNW2WibJAgYpCVA6E93mvvVH+LcssoVjOBrSKWS55yEIHsk0X8ctHmfOQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/core-util": { + "version": "1.14.0", + "resolved": "https://registry.npmjs.org/@azure/core-util/-/core-util-1.14.0.tgz", + "integrity": "sha512-9n2pWK61veAuN0V20t9lOuoV4CFMdyAZ1ygZzvBGk/pBBJRib/PjL9PLXa/aI2CcPpyHfqVsxxqLCYl6uZlfDw==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/abort-controller": "^2.1.2", + "@typespec/ts-http-runtime": "^0.3.0", + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/identity": { + "version": "4.13.3", + "resolved": "https://registry.npmjs.org/@azure/identity/-/identity-4.13.3.tgz", + "integrity": "sha512-zGQPtqvXPgSA8yfV2CkIQ1qirqk0p9AIVpC5uEkdXQYcKl07QHvyaGYRnZOk0AsQUmxNb4wfkcwY5di8Z5xa9A==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/abort-controller": "^2.0.0", + "@azure/core-auth": "^1.9.0", + "@azure/core-client": "^1.9.2", + "@azure/core-process": "^1.0.0", + "@azure/core-rest-pipeline": "^1.17.0", + "@azure/core-tracing": "^1.0.0", + "@azure/core-util": "^1.11.0", + "@azure/logger": "^1.0.0", + "@azure/msal-browser": "^5.5.0", + "@azure/msal-node": "^6.0.0", + "open": "^10.1.0", + "tslib": "^2.2.0" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/logger": { + "version": "1.4.0", + "resolved": "https://registry.npmjs.org/@azure/logger/-/logger-1.4.0.tgz", + "integrity": "sha512-rbAE25KUfjU/s3XHUdJgceoCP5dEOpMx85J04kF+QMdta73XkuG9JGHHinch+XIoKpBdqljin+KqURpJriSzLA==", + "dev": true, + "license": "MIT", + "dependencies": { + "@typespec/ts-http-runtime": "^0.3.0", + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@azure/msal-browser": { + "version": "5.22.0", + "resolved": "https://registry.npmjs.org/@azure/msal-browser/-/msal-browser-5.22.0.tgz", + "integrity": "sha512-5kgu9xeEKgGc2JeidxAtU15NJTqiH/CMCRRQAJ4Rac56kB7KVg91vbNmn+z3RO1vNomPN69UvjKG9h1Pghx6dQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/msal-common": "16.14.1" + }, + "engines": { + "node": ">=0.8.0" + } + }, + "node_modules/@azure/msal-common": { + "version": "16.14.1", + "resolved": "https://registry.npmjs.org/@azure/msal-common/-/msal-common-16.14.1.tgz", + "integrity": "sha512-Or6xhPNyi4zHW25158yxBoyxuCqNSPa5YBVqfF1J5Ks4MJWBo/USXdp05DQIPu1Zli00YZu6t0+h6KvHJealxQ==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=0.8.0" + } + }, + "node_modules/@azure/msal-node": { + "version": "6.0.1", + "resolved": "https://registry.npmjs.org/@azure/msal-node/-/msal-node-6.0.1.tgz", + "integrity": "sha512-ixSO1Y/kCVRthRs+hSx/5qkwaunX1/RAePhlMN0wIpIQ4WEZ6AREGGnGd1AP0qspHVsAwzWQKGubpjh50JyJIQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/msal-common": "16.14.1", + "jsonwebtoken": "^9.0.0" + }, + "engines": { + "node": ">=20" + } + }, + "node_modules/@esbuild/aix-ppc64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/aix-ppc64/-/aix-ppc64-0.28.2.tgz", + "integrity": "sha512-XExcO+dvLKvVtNTibSTBej1NCAbaGhWn9Ww1ZPx80qsahhPFe/8jgWP0IchNe0F3HwkU7n8ejhH8bjonqht8mQ==", + "cpu": [ + "ppc64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "aix" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/android-arm": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/android-arm/-/android-arm-0.28.2.tgz", + "integrity": "sha512-kXXoiPVVGQcnIYGOeaovwOURpniDBpSq4A03qkQ+BMQqtGG6HYap3xne9C1O1yo4TR3qxlCX5IqqmX6fFo2Lqg==", + "cpu": [ + "arm" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "android" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/android-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/android-arm64/-/android-arm64-0.28.2.tgz", + "integrity": "sha512-5YfKeeI8qWfBZIX+u2xZC3Zlb3Os/gLS2sbEKM+I4ZOcsWmHS2WLysCcQZDAFRslDUU5Oiq44gf6PYN1vGwG5A==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "android" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/android-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/android-x64/-/android-x64-0.28.2.tgz", + "integrity": "sha512-O387ite7SzUyCcy3JQX4P4bLtEA7bLLkx+esve5JHnyYfNTxcVpXZo9jhdB0lTKN44gztELTdU7nS8Nr16Fs1Q==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "android" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/darwin-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/darwin-arm64/-/darwin-arm64-0.28.2.tgz", + "integrity": "sha512-n4KqkOQrraxHJcgjM1RvwbigfQKIKJVpM7xp+KsxiyUSrRdIXnt73VhrPAx0fV44hgfmIVKjxMN9J1t5jySVkw==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/darwin-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/darwin-x64/-/darwin-x64-0.28.2.tgz", + "integrity": "sha512-uq6suIWYP37qzGddBKPw5QEQPi6HiLGsO7UmkpfyaYNQ3D+rN6w6WfwH+nuqcGXWvawGwxOEroO4YGnFh95azw==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/freebsd-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/freebsd-arm64/-/freebsd-arm64-0.28.2.tgz", + "integrity": "sha512-n+I0BTSRIoy+d6RPKnEVwql5UwBJolytvY4mAOIEJorKlqgPII8ix6slVVrfZ5Tnj7glIZvloylbB/EJPMWEXw==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "freebsd" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/freebsd-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/freebsd-x64/-/freebsd-x64-0.28.2.tgz", + "integrity": "sha512-78XJTJkvPs0kz2w61301PJjXl4g7q3JqiYMZ/M/yVI73EHBrCRTgkhu9oqG7vPqq+a/yadEW8aD+agKlk5xrmg==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "freebsd" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-arm": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-arm/-/linux-arm-0.28.2.tgz", + "integrity": "sha512-XlDnu2q5yoqems+xay6wSAcg9DDD7K9RLKZEBOMZm3ckNpJBvOX20tSfby8KfrrhINDyv9V2YVZKY/SpoGJI8w==", + "cpu": [ + "arm" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-arm64/-/linux-arm64-0.28.2.tgz", + "integrity": "sha512-pW4AC0P3it8c7do9MVM4p51FzHzdM/TZrerurgRcHJ2WTa1VQ1CIq18xncfpBJw4ojkiZZrKW2yIBWBP92j6Ug==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-ia32": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-ia32/-/linux-ia32-0.28.2.tgz", + "integrity": "sha512-CYbnj78HsIeA+DhgUKgFCfvNsTHFhMMrinUrMZpDXJXKN8T3XViTZ/+wtHeVxEWY8ewSzTFN+nRmSwO2tZaLUQ==", + "cpu": [ + "ia32" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-loong64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-loong64/-/linux-loong64-0.28.2.tgz", + "integrity": "sha512-buwkd8nsph4R+ajRvw0qM5Hja/TXQow3ptzWO2EbG/cqcIkHloRrdlBtQlshyYGTNFvfkfJ5tpPLVkY4DtsPfQ==", + "cpu": [ + "loong64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-mips64el": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-mips64el/-/linux-mips64el-0.28.2.tgz", + "integrity": "sha512-ZVykbDyk7519VwiNb9Lcj9m8XM6v5V9uKPvrEMkkEedVewf+0itkhahp4HDpgERXhwLRpWFypsGbG/J8s0QjJA==", + "cpu": [ + "mips64el" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-ppc64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-ppc64/-/linux-ppc64-0.28.2.tgz", + "integrity": "sha512-CAXl+Dtd9UUuJd8pKKdwh6MLm3MUMiqMPmhZ3tTSXPqfyQ3vDl6R5hZdZ/kYojK4ofXtdfSv1tFq8XzWx3heNQ==", + "cpu": [ + "ppc64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-riscv64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-riscv64/-/linux-riscv64-0.28.2.tgz", + "integrity": "sha512-GeXCej4IQtU1B+QlDV8W/RRvbzI3O/Stss+/bCXv4lZls5WGRtu2a+3JkA3i4qIUlMXpcHebWpF8AkJhATowuA==", + "cpu": [ + "riscv64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-s390x": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-s390x/-/linux-s390x-0.28.2.tgz", + "integrity": "sha512-3H1weTYZPxt/WOhByszQZybS9w5lKzUn1FDMsgEChbHWQwHYQQRfBxgCcZvPhjHfKyJjIievvMmEUawJrdY9Dg==", + "cpu": [ + "s390x" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/linux-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/linux-x64/-/linux-x64-0.28.2.tgz", + "integrity": "sha512-4xTZr1FUmSoQW4XIWmit3tzQrUTZM+N3P0XV8xROKYF50XfI7xeO90+1bZvNwxIufQ9hDQVRJH5YhgPVF8A/HQ==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/netbsd-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/netbsd-arm64/-/netbsd-arm64-0.28.2.tgz", + "integrity": "sha512-sSATRjPeDBg3pdgHoQfoYBob11Kk1FGa9lui5RIHZCoCkJa9QKlvl3/vKz2usCmYYjs7ymJR/2Nnsqe+Hjt5nw==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "netbsd" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/netbsd-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/netbsd-x64/-/netbsd-x64-0.28.2.tgz", + "integrity": "sha512-lqnzCV+mM0gIADaKihiCg6ifgfU2L3h5E33rNQBN1Y4MaVGnzryzmvvf7UHxprpQdE8hpqLolJ9Rl+SkIRDpyw==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "netbsd" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/openbsd-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/openbsd-arm64/-/openbsd-arm64-0.28.2.tgz", + "integrity": "sha512-AL2qJILH7lNjrDmCQDvdxMfAUIv8KMNZOvrwAQ8i8//ntL9FflhOyMJ8OZSMBb8/AWXe3/5v5S20y3zCoZWKoQ==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "openbsd" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/openbsd-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/openbsd-x64/-/openbsd-x64-0.28.2.tgz", + "integrity": "sha512-QtiuPytchRyC4rwUKhexJdQKvDuZ6hWloi3igqPQNUJCS1/v9EiO3UTOXR6A3FoMo4fnAKbWJdqaIwhOzh8qEw==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "openbsd" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/openharmony-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/openharmony-arm64/-/openharmony-arm64-0.28.2.tgz", + "integrity": "sha512-WkhYDmpTjLvGlScA1rwjRUmhl4k8oXR3cIbtqWmELgU/dFeHHlEllxDvdWcNJV9rbzCexB5vz8gtNewWLgCT7Q==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "openharmony" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/sunos-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/sunos-x64/-/sunos-x64-0.28.2.tgz", + "integrity": "sha512-GPMSkTOtMnv2U2F8gxe4Io6qmVs+YKyp832Etqqxr0hFngmXQ3rzwytelm3GIn7T4VviRUlf3sOgBOiTdvaf7g==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "sunos" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/win32-arm64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/win32-arm64/-/win32-arm64-0.28.2.tgz", + "integrity": "sha512-PIhhEkE9uPBleRBrQEJpUn7MBnibZzbGzYWPmY3x+YoVg/95zbjB4CxPPOQ8l5tYYM4mMaCthF8/1DIfBQQyWQ==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/win32-ia32": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/win32-ia32/-/win32-ia32-0.28.2.tgz", + "integrity": "sha512-YmJbfTlvU7Sdn9BB+4PRES4oB6pxgS37MAONj+hBr/cpXS1aBPKXxNnDbu+QCWPj0o9dgyxeq79g6c5P8KeuYA==", + "cpu": [ + "ia32" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@esbuild/win32-x64": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/@esbuild/win32-x64/-/win32-x64-0.28.2.tgz", + "integrity": "sha512-5ebpxr3nWMzrL/rnUI755Jkuee0bHL/Gq0WTF9lvcpv73wAp5eu8MfBUgWK9bhWvZjj7yX8etf/8tI8Ney695g==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">=18" + } + }, + "node_modules/@jridgewell/sourcemap-codec": { + "version": "1.6.0", + "resolved": "https://registry.npmjs.org/@jridgewell/sourcemap-codec/-/sourcemap-codec-1.6.0.tgz", + "integrity": "sha512-T7jf+5zgsZHwNJ4lvQ7/aezbyk0nNX+zJVWpmHA7VYsEx7a7qr5Rg5IbtJFqkgze5Y2sruq1RUY8Q837Od7iFw==", + "dev": true, + "license": "MIT" + }, + "node_modules/@napi-rs/keyring": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring/-/keyring-1.3.0.tgz", + "integrity": "sha512-WrOw/bcXm0f9qHkumlT1QlArXSTWqaY9sunsDpOk+yCCorCKMxvWT/a3xko4EYHVdeZoh00yI2TydXn6eyICDA==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 10" + }, + "funding": { + "type": "github", + "url": "https://github.com/sponsors/Brooooooklyn" + }, + "optionalDependencies": { + "@napi-rs/keyring-darwin-arm64": "1.3.0", + "@napi-rs/keyring-darwin-x64": "1.3.0", + "@napi-rs/keyring-freebsd-x64": "1.3.0", + "@napi-rs/keyring-linux-arm-gnueabihf": "1.3.0", + "@napi-rs/keyring-linux-arm64-gnu": "1.3.0", + "@napi-rs/keyring-linux-arm64-musl": "1.3.0", + "@napi-rs/keyring-linux-riscv64-gnu": "1.3.0", + "@napi-rs/keyring-linux-x64-gnu": "1.3.0", + "@napi-rs/keyring-linux-x64-musl": "1.3.0", + "@napi-rs/keyring-win32-arm64-msvc": "1.3.0", + "@napi-rs/keyring-win32-ia32-msvc": "1.3.0", + "@napi-rs/keyring-win32-x64-msvc": "1.3.0" + } + }, + "node_modules/@napi-rs/keyring-darwin-arm64": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-darwin-arm64/-/keyring-darwin-arm64-1.3.0.tgz", + "integrity": "sha512-pl76hJvdYUBn6I24bXiOBMA9nbDapo3I5B+f3OorjDU4dUMSypXeKbOVehJe8fhgTiH24flMyTS3aAIy43xegQ==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-darwin-x64": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-darwin-x64/-/keyring-darwin-x64-1.3.0.tgz", + "integrity": "sha512-YcJtEV5LA3cvA4z3BurgxH5IhTsW1JfIvcAAcqcecwk06Si9F9NqkxbZVIfDwQ8oRHgaBmT3zZJnLAotCrVahw==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-freebsd-x64": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-freebsd-x64/-/keyring-freebsd-x64-1.3.0.tgz", + "integrity": "sha512-vlLf31TGhfRAaxLDBhg8b89ss0HHD/lyNmL5F3UjSaz5CUXElsJmKYq9fqA/B+cZKUEUcLHHGhF0I/CqcFdaVw==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "freebsd" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-linux-arm-gnueabihf": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-linux-arm-gnueabihf/-/keyring-linux-arm-gnueabihf-1.3.0.tgz", + "integrity": "sha512-KiWdMMu/Inz/bHHIAGrnF7r54FZDYXuHO6UFF/rhIrshUsxbMG1Rl9lEymNtqqsVo927G0VYcb02FzWQ3iBQRQ==", + "cpu": [ + "arm" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-linux-arm64-gnu": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-linux-arm64-gnu/-/keyring-linux-arm64-gnu-1.3.0.tgz", + "integrity": "sha512-eyKGpY40lm9Jvs1aD294XRH4y7+TlJM0YVAryZeXA6TX0mb4gMkxVXwSQv7MCwgah7raeUd0dKUb4BPAYIgcMg==", + "cpu": [ + "arm64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-linux-arm64-musl": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-linux-arm64-musl/-/keyring-linux-arm64-musl-1.3.0.tgz", + "integrity": "sha512-iIK6JWHXAJqDrEyLY3TmswwloVyt2vj+04TZnew+uSJ9gnDO8EwRbp3/iw3LpWaXiDO7VomGO6y8I0Id8uBZSw==", + "cpu": [ + "arm64" + ], + "dev": true, + "libc": [ + "musl" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-linux-riscv64-gnu": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-linux-riscv64-gnu/-/keyring-linux-riscv64-gnu-1.3.0.tgz", + "integrity": "sha512-/PGqrwn6EwgtK6vccASSXJRfOSP4vN1F4ASsIQ+7MdrK6hNvAJ1FZPrIuD5gGGdxezo3F++To2Wq7DbuGIeuNQ==", + "cpu": [ + "riscv64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-linux-x64-gnu": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-linux-x64-gnu/-/keyring-linux-x64-gnu-1.3.0.tgz", + "integrity": "sha512-2PDK1WKWTu9lBGq9VvNEkSlQD3O7YwVpmnyN2M3cy4v7NJ/8gDMd9GXv3G+FVXN13uhp4gnnPBS+ScefmEeD2A==", + "cpu": [ + "x64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-linux-x64-musl": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-linux-x64-musl/-/keyring-linux-x64-musl-1.3.0.tgz", + "integrity": "sha512-oJ2HkX8YUo46QBkn0pG+HuIKQNqr523q6vBobCn+P95s4C4K6/kLBqHY/1bg5J4ap31DzsznhnFKcfBNBsjCnw==", + "cpu": [ + "x64" + ], + "dev": true, + "libc": [ + "musl" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-win32-arm64-msvc": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-win32-arm64-msvc/-/keyring-win32-arm64-msvc-1.3.0.tgz", + "integrity": "sha512-tOd3c/uAaeoE4ycVlmAdSvygz0Zt3zdca6Y7gokBeIbaRDWpjDIUOpU3MvML59XAaqyuKGsVVu0F/DZb1lHPmw==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-win32-ia32-msvc": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-win32-ia32-msvc/-/keyring-win32-ia32-msvc-1.3.0.tgz", + "integrity": "sha512-sPSqeAFZMGqP1R++M2JTza7GQJJ/TpCo6JU6Vcd4jnebvOaEDs9b7eipakU1PJdSvhpC2yXMCNRk9gXfrhuwHQ==", + "cpu": [ + "ia32" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@napi-rs/keyring-win32-x64-msvc": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/@napi-rs/keyring-win32-x64-msvc/-/keyring-win32-x64-msvc-1.3.0.tgz", + "integrity": "sha512-4DnCWXwDc0HRKwyRlG5y0VhKZW2tNRQfKKfyj6IX/KWfDNyq9hn4n+GL1auyDcOO/v8PwnhmYo2+rOOqCkvvOg==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@oxc-project/types": { + "version": "0.150.0", + "resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.150.0.tgz", + "integrity": "sha512-rDS5/31E9HfPl/CIzGrn0DOlvBbXFseQ5URJ9sYMfstbKLD/c6Gm9vmRzRGDdAXyOIL4zmO37lc9RIwYqVruZw==", + "dev": true, + "license": "MIT", + "funding": { + "url": "https://github.com/sponsors/oxc-project" + } + }, + "node_modules/@rolldown/binding-android-arm-eabi": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm-eabi/-/binding-android-arm-eabi-1.2.9.tgz", + "integrity": "sha512-tNISae1QEf/vkb3xkRcjV5SEdzPE97We5IVaa2Z8jSszQPZ8U60B/YCYpw4QI7VidYsBtKavczXf+DyDs9WGxw==", + "cpu": [ + "arm" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "android" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-android-arm64": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.2.9.tgz", + "integrity": "sha512-YC8YsI30o606GTZi0VyzYlsDKFP8W61i/QzayHDkLbNEz/IShqAmTa+hsJRj13xTHA0H+6fk4b2UmGn+Q/cMlg==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "android" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-darwin-arm64": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.2.9.tgz", + "integrity": "sha512-IwhlH3qK5urrY8hZiEgGkHKEFN901p/p2bjxCxJlr4GyNnF7wYpUvK+Y43uaRYuC4hpfjzbR3SJC3arX1jGvmw==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-darwin-x64": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.2.9.tgz", + "integrity": "sha512-XxpJfVzFh+jilRxIXUqcfYAYcunIc/XEzIizsOL1fcJee5Sf7H3mH8WlLmfHfluz5amqR88QQo9izKtmMlavAw==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-freebsd-x64": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.2.9.tgz", + "integrity": "sha512-kSfvhmgeWyfkbT3p/1s5vSgboogoah2zkm9fX2zjg2hHxSV7T4KhMWRUUaRk4OXNqoD3QAUeRqLcs1aZOK4U1g==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "freebsd" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-linux-arm-gnueabihf": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.2.9.tgz", + "integrity": "sha512-1RVzG17pxqbTfYLC352JlLt6kKLG+6Hr30n8DlIJqsnV5luUDd2Qdx9Ayw1Cabfyb1K9k0jXEZ7evxkRoT+uiw==", + "cpu": [ + "arm" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-linux-arm64-gnu": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.2.9.tgz", + "integrity": "sha512-BXqPvZ2drqVD+/Z8UpKwcs4Mp7grM+eGFku4CAEKrEtcbAsUpzREphK1sogCRZGreVPiMkiiBtw0n3TPteuqvw==", + "cpu": [ + "arm64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-linux-arm64-musl": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.2.9.tgz", + "integrity": "sha512-11vWvo8YDwLzukt27J3aYDWU+gg2P7J+ZOmiJ0hkF5BXZDW7pVya7r40MXDy6ya0i9KamoENSVKIugvJNgFXIA==", + "cpu": [ + "arm64" + ], + "dev": true, + "libc": [ + "musl" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-linux-ppc64-gnu": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.2.9.tgz", + "integrity": "sha512-a1tijMkdwsIARtc0F39ApURROkf3NwqinI6TOiSSWCTR7dT96dffNvMUtDHnq64wKNTIZOIlzKrFvvFUznJiyw==", + "cpu": [ + "ppc64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-linux-s390x-gnu": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.2.9.tgz", + "integrity": "sha512-x6SQNdAvv4c3hWqTMaWuawzMX9myaCs/yEmlGsxJzkdClnHW7FbrjQuSiRDhuSYzEYoEMhsaJy9qHG/XNemJPQ==", + "cpu": [ + "s390x" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-linux-x64-gnu": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.2.9.tgz", + "integrity": "sha512-9s0AZ8BFK5/n7B/TBoa2yJE3gI3KURrbXcPBlsAsvjU4VeJKgE90y1YtNxyEUIcHPQkg6/yfF3qihUrcM/Kf0Q==", + "cpu": [ + "x64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-linux-x64-musl": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.2.9.tgz", + "integrity": "sha512-P7VWAmV+WdJluH7ovnRGoiv2i8To7GAZ+kGzfGup635cyL7SyYl3lSUaA3Gp5THf0n/Co5EyEqb2zbqq+nMOHQ==", + "cpu": [ + "x64" + ], + "dev": true, + "libc": [ + "musl" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-openharmony-arm64": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.2.9.tgz", + "integrity": "sha512-1qixtsE4BK8h+yS3BfmZ09UhA7O/N4IACva6YBr7EBvCJraByTuRcgOTaiA62Tm0vey3UcKXLOaoGHtYmNGEVg==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "openharmony" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-win32-arm64-msvc": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.2.9.tgz", + "integrity": "sha512-ok8IQjcEPs1AKZfuEUznVBrJw+gK4soq+bx8b1X2XoMqVClarc1q5JDmVtWXY1xfr6ZuHTAsPXHTgTrqKTZeww==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/binding-win32-x64-msvc": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.2.9.tgz", + "integrity": "sha512-Ip2mXoU0hM0boq3Rf+ekuT653OROSo6aSYcPT1VHE4q52KvyxgFkQgrgb/IEsxOuvQ2fZZbs8khJAyCEPM24/g==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": "^20.19.0 || >=22.12.0" + } + }, + "node_modules/@rolldown/pluginutils": { + "version": "1.0.1", + "resolved": "https://registry.npmjs.org/@rolldown/pluginutils/-/pluginutils-1.0.1.tgz", + "integrity": "sha512-2j9bGt5Jh8hj+vPtgzPtl72j0yRxHAyumoo6TNfAjsLB04UtpSvPbPcDcBMxz7n+9CYB0c1GxQFxYRg2jimqGw==", + "dev": true, + "license": "MIT" + }, + "node_modules/@secretlint/core": { + "version": "10.2.2", + "resolved": "https://registry.npmjs.org/@secretlint/core/-/core-10.2.2.tgz", + "integrity": "sha512-6rdwBwLP9+TO3rRjMVW1tX+lQeo5gBbxl1I5F8nh8bgGtKwdlCMhMKsBWzWg1ostxx/tIG7OjZI0/BxsP8bUgw==", + "dev": true, + "license": "MIT", + "dependencies": { + "@secretlint/profiler": "^10.2.2", + "@secretlint/types": "^10.2.2", + "debug": "^4.4.1", + "structured-source": "^4.0.0" + }, + "engines": { + "node": ">=20.0.0" + } + }, + "node_modules/@secretlint/profiler": { + "version": "10.2.2", + "resolved": "https://registry.npmjs.org/@secretlint/profiler/-/profiler-10.2.2.tgz", + "integrity": "sha512-qm9rWfkh/o8OvzMIfY8a5bCmgIniSpltbVlUVl983zDG1bUuQNd1/5lUEeWx5o/WJ99bXxS7yNI4/KIXfHexig==", + "dev": true, + "license": "MIT" + }, + "node_modules/@secretlint/secretlint-rule-no-dotenv": { + "version": "10.2.2", + "resolved": "https://registry.npmjs.org/@secretlint/secretlint-rule-no-dotenv/-/secretlint-rule-no-dotenv-10.2.2.tgz", + "integrity": "sha512-KJRbIShA9DVc5Va3yArtJ6QDzGjg3PRa1uYp9As4RsyKtKSSZjI64jVca57FZ8gbuk4em0/0Jq+uy6485wxIdg==", + "dev": true, + "license": "MIT", + "dependencies": { + "@secretlint/types": "^10.2.2" + }, + "engines": { + "node": ">=20.0.0" + } + }, + "node_modules/@secretlint/secretlint-rule-preset-recommend": { + "version": "10.2.2", + "resolved": "https://registry.npmjs.org/@secretlint/secretlint-rule-preset-recommend/-/secretlint-rule-preset-recommend-10.2.2.tgz", + "integrity": "sha512-K3jPqjva8bQndDKJqctnGfwuAxU2n9XNCPtbXVI5JvC7FnQiNg/yWlQPbMUlBXtBoBGFYp08A94m6fvtc9v+zA==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=20.0.0" + } + }, + "node_modules/@secretlint/source-creator": { + "version": "10.2.2", + "resolved": "https://registry.npmjs.org/@secretlint/source-creator/-/source-creator-10.2.2.tgz", + "integrity": "sha512-h6I87xJfwfUTgQ7irWq7UTdq/Bm1RuQ/fYhA3dtTIAop5BwSFmZyrchph4WcoEvbN460BWKmk4RYSvPElIIvxw==", + "dev": true, + "license": "MIT", + "dependencies": { + "@secretlint/types": "^10.2.2", + "istextorbinary": "^9.5.0" + }, + "engines": { + "node": ">=20.0.0" + } + }, + "node_modules/@secretlint/types": { + "version": "10.2.2", + "resolved": "https://registry.npmjs.org/@secretlint/types/-/types-10.2.2.tgz", + "integrity": "sha512-Nqc90v4lWCXyakD6xNyNACBJNJ0tNCwj2WNk/7ivyacYHxiITVgmLUFXTBOeCdy79iz6HtN9Y31uw/jbLrdOAg==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=20.0.0" + } + }, + "node_modules/@standard-schema/spec": { + "version": "1.1.0", + "resolved": "https://registry.npmjs.org/@standard-schema/spec/-/spec-1.1.0.tgz", + "integrity": "sha512-l2aFy5jALhniG5HgqrD6jXLi/rUWrKvqN/qJx6yoJsgKhblVd+iqqU4RCXavm/jPityDo5TCvKMnpjKnOriy0w==", + "dev": true, + "license": "MIT" + }, + "node_modules/@types/chai": { + "version": "5.2.3", + "resolved": "https://registry.npmjs.org/@types/chai/-/chai-5.2.3.tgz", + "integrity": "sha512-Mw558oeA9fFbv65/y4mHtXDs9bPnFMZAL/jxdPFUpOHHIXX91mcgEHbS5Lahr+pwZFR8A7GQleRWeI6cGFC2UA==", + "dev": true, + "license": "MIT", + "dependencies": { + "@types/deep-eql": "*", + "assertion-error": "^2.0.1" + } + }, + "node_modules/@types/deep-eql": { + "version": "4.0.2", + "resolved": "https://registry.npmjs.org/@types/deep-eql/-/deep-eql-4.0.2.tgz", + "integrity": "sha512-c9h9dVVMigMPc4bwTvC5dxqtqJZwQPePsWjPlpSOnojbor6pGqdk541lfA7AqFQr5pB1BRdq0juY9db81BwyFw==", + "dev": true, + "license": "MIT" + }, + "node_modules/@types/estree": { + "version": "1.0.9", + "resolved": "https://registry.npmjs.org/@types/estree/-/estree-1.0.9.tgz", + "integrity": "sha512-GhdPgy1el4/ImP05X05Uw4cw2/M93BCUmnEvWZNStlCzEKME4Fkk+YpoA5OiHNQmoS7Cafb8Xa3Pya8m1Qrzeg==", + "dev": true, + "license": "MIT" + }, + "node_modules/@types/node": { + "version": "22.20.3", + "resolved": "https://registry.npmjs.org/@types/node/-/node-22.20.3.tgz", + "integrity": "sha512-DZmzkmwHzXrLPAXPyKNDzlIwMMUZCVacoD25ywdy5YTKGbOx/2ld+Q38Im2zJ0vBuZP5Prd3VZutKZyXwkOS8A==", + "dev": true, + "license": "MIT", + "dependencies": { + "undici-types": "~6.21.0" + } + }, + "node_modules/@types/vscode": { + "version": "1.138.0", + "resolved": "https://registry.npmjs.org/@types/vscode/-/vscode-1.138.0.tgz", + "integrity": "sha512-DhlvrucJa8n6EHst7qQcXOcjLKs/VRTQy55plLO+RN74jonYsr/t9dd66zBnAyz1tpDig0okQAxuoPHpVNje2g==", + "dev": true, + "license": "MIT" + }, + "node_modules/@typespec/ts-http-runtime": { + "version": "0.3.9", + "resolved": "https://registry.npmjs.org/@typespec/ts-http-runtime/-/ts-http-runtime-0.3.9.tgz", + "integrity": "sha512-edSdeAqkdxBVzA1yL1LrLCml1YjyCVvPMtMqJpbF+6K609tHe8V6sQUzFQSGcYNhcuhOceZtjvN32+mpIth30A==", + "dev": true, + "license": "MIT", + "dependencies": { + "http-proxy-agent": "^7.0.0", + "https-proxy-agent": "^7.0.0", + "tslib": "^2.6.2" + }, + "engines": { + "node": ">=22.0.0" + } + }, + "node_modules/@vitest/expect": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/@vitest/expect/-/expect-4.1.11.tgz", + "integrity": "sha512-VX2x5vNJXET47KAFzwERI+KRMtTTCSWTfSMKsW7JsUsXV4psq++e3DvZpuTDOpHcxytiDs6p2nhVb2tVDiiUYw==", + "dev": true, + "license": "MIT", + "dependencies": { + "@standard-schema/spec": "^1.1.0", + "@types/chai": "^5.2.2", + "@vitest/spy": "4.1.11", + "@vitest/utils": "4.1.11", + "chai": "^6.2.2", + "tinyrainbow": "^3.1.0" + }, + "funding": { + "url": "https://opencollective.com/vitest" + } + }, + "node_modules/@vitest/mocker": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/@vitest/mocker/-/mocker-4.1.11.tgz", + "integrity": "sha512-2XJVD55d1o5AZous5CCGKS74g/riOj9odEt2bQpCVZeblHyHdnMeFl4jl0XjU21stf4mbjUkew2eXQZt65g5CQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "@vitest/spy": "4.1.11", + "estree-walker": "^3.0.3", + "magic-string": "^0.30.21" + }, + "funding": { + "url": "https://opencollective.com/vitest" + }, + "peerDependencies": { + "msw": "^2.4.9", + "vite": "^6.0.0 || ^7.0.0 || ^8.0.0" + }, + "peerDependenciesMeta": { + "msw": { + "optional": true + }, + "vite": { + "optional": true + } + } + }, + "node_modules/@vitest/pretty-format": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/@vitest/pretty-format/-/pretty-format-4.1.11.tgz", + "integrity": "sha512-yiZzPbGTS9Sr/JpFl8zHrcIkAofNbFV6k21vIgQN/cY/oxZeXhJv5sc/MBJ5jFKWmWs+oJHw0UXLZjmf931+Vw==", + "dev": true, + "license": "MIT", + "dependencies": { + "tinyrainbow": "^3.1.0" + }, + "funding": { + "url": "https://opencollective.com/vitest" + } + }, + "node_modules/@vitest/runner": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/@vitest/runner/-/runner-4.1.11.tgz", + "integrity": "sha512-LztvUgdwMNJMIkj3hQnnxiC2Xy1zNxq928W/xhjCLaNCzqTZOudjwbQf6v9IntZGPw132i2Lq2rgTRZHD3JHNw==", + "dev": true, + "license": "MIT", + "dependencies": { + "@vitest/utils": "4.1.11", + "pathe": "^2.0.3" + }, + "funding": { + "url": "https://opencollective.com/vitest" + } + }, + "node_modules/@vitest/snapshot": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/@vitest/snapshot/-/snapshot-4.1.11.tgz", + "integrity": "sha512-pN7ikn1ON7h8ee4gIAp4AzyK+zBtJPzVbqOgu5LCEh4VaJVbPQcgYQYJIMGQPXVeJJq1fnfazis7a5pFNPahog==", + "dev": true, + "license": "MIT", + "dependencies": { + "@vitest/pretty-format": "4.1.11", + "@vitest/utils": "4.1.11", + "magic-string": "^0.30.21", + "pathe": "^2.0.3" + }, + "funding": { + "url": "https://opencollective.com/vitest" + } + }, + "node_modules/@vitest/spy": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/@vitest/spy/-/spy-4.1.11.tgz", + "integrity": "sha512-apNa/prQy2qCeywhnixOHPRCgGNhvg7T4Dapfl1GahLp/R+uhBm5cPyFoNVyqsNd2h1nJxL6BqqdIjiABL60YA==", + "dev": true, + "license": "MIT", + "funding": { + "url": "https://opencollective.com/vitest" + } + }, + "node_modules/@vitest/utils": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/@vitest/utils/-/utils-4.1.11.tgz", + "integrity": "sha512-zTCVGpyFsGWBhllOyKlTw/vnr6D9qxsfSDyfbyZmTyjHw5N/VuvzHpHoQjm2ZJzn4RJgx5w4r7V0er69CmLgPQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "@vitest/pretty-format": "4.1.11", + "convert-source-map": "^2.0.0", + "tinyrainbow": "^3.1.0" + }, + "funding": { + "url": "https://opencollective.com/vitest" + } + }, + "node_modules/@vscode/vsce": { + "version": "4.0.0", + "resolved": "https://registry.npmjs.org/@vscode/vsce/-/vsce-4.0.0.tgz", + "integrity": "sha512-NImwuLaenMmb5D5Jer9/lzi/F9ZQUBOp8Azhj/BVYcTFgixv8KehFXqEUDjQlD2tAiw2E6dDGyjTuAB//di60A==", + "dev": true, + "license": "MIT", + "dependencies": { + "@azure/identity": "^4.13.2", + "@napi-rs/keyring": "^1.3.0", + "@secretlint/core": "^10.2.2", + "@secretlint/secretlint-rule-no-dotenv": "^10.2.2", + "@secretlint/secretlint-rule-preset-recommend": "^10.2.2", + "@secretlint/source-creator": "^10.2.2", + "@secretlint/types": "^10.2.2", + "@vscode/vsce-sign": "^2.1.0", + "azure-devops-node-api": "^12.5.0", + "cockatiel": "^3.2.1", + "commander": "^12.1.0", + "hosted-git-info": "^4.1.0", + "jsonc-parser": "^3.3.1", + "marked": "^18.0.11", + "mime": "^1.6.0", + "minimatch": "^10.2.6", + "parse5": "^8.0.1", + "proper-lockfile": "^4.1.2", + "read": "^1.0.7", + "semver": "^7.8.5", + "tinyglobby": "^0.2.17", + "typed-rest-client": "^1.8.11", + "url-join": "^4.0.1", + "xml2js": "^0.5.0", + "yauzl": "^3.4.0", + "yazl": "^2.5.1" + }, + "bin": { + "vsce": "vsce" + }, + "engines": { + "node": ">= 22" + } + }, + "node_modules/@vscode/vsce-sign": { + "version": "2.1.0", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign/-/vsce-sign-2.1.0.tgz", + "integrity": "sha512-9AQrqazrBgTgRSuwleLVXUrIUphY02/SFCh2TKYoLV/xifJAdblhdmEmw5gUrYSPQ3sRwNs9iyCMD14sATEE6g==", + "dev": true, + "hasInstallScript": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optionalDependencies": { + "@vscode/vsce-sign-alpine-arm64": "2.0.6", + "@vscode/vsce-sign-alpine-x64": "2.0.6", + "@vscode/vsce-sign-darwin-arm64": "2.0.6", + "@vscode/vsce-sign-darwin-x64": "2.0.6", + "@vscode/vsce-sign-linux-arm": "2.0.6", + "@vscode/vsce-sign-linux-arm64": "2.0.6", + "@vscode/vsce-sign-linux-x64": "2.0.6", + "@vscode/vsce-sign-win32-arm64": "2.0.6", + "@vscode/vsce-sign-win32-x64": "2.0.6" + } + }, + "node_modules/@vscode/vsce-sign-alpine-arm64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-alpine-arm64/-/vsce-sign-alpine-arm64-2.0.6.tgz", + "integrity": "sha512-wKkJBsvKF+f0GfsUuGT0tSW0kZL87QggEiqNqK6/8hvqsXvpx8OsTEc3mnE1kejkh5r+qUyQ7PtF8jZYN0mo8Q==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "alpine" + ] + }, + "node_modules/@vscode/vsce-sign-alpine-x64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-alpine-x64/-/vsce-sign-alpine-x64-2.0.6.tgz", + "integrity": "sha512-YoAGlmdK39vKi9jA18i4ufBbd95OqGJxRvF3n6ZbCyziwy3O+JgOpIUPxv5tjeO6gQfx29qBivQ8ZZTUF2Ba0w==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "alpine" + ] + }, + "node_modules/@vscode/vsce-sign-darwin-arm64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-darwin-arm64/-/vsce-sign-darwin-arm64-2.0.6.tgz", + "integrity": "sha512-5HMHaJRIQuozm/XQIiJiA0W9uhdblwwl2ZNDSSAeXGO9YhB9MH5C4KIHOmvyjUnKy4UCuiP43VKpIxW1VWP4tQ==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "darwin" + ] + }, + "node_modules/@vscode/vsce-sign-darwin-x64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-darwin-x64/-/vsce-sign-darwin-x64-2.0.6.tgz", + "integrity": "sha512-25GsUbTAiNfHSuRItoQafXOIpxlYj+IXb4/qarrXu7kmbH94jlm5sdWSCKrrREs8+GsXF1b+l3OB7VJy5jsykw==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "darwin" + ] + }, + "node_modules/@vscode/vsce-sign-linux-arm": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-linux-arm/-/vsce-sign-linux-arm-2.0.6.tgz", + "integrity": "sha512-UndEc2Xlq4HsuMPnwu7420uqceXjs4yb5W8E2/UkaHBB9OWCwMd3/bRe/1eLe3D8kPpxzcaeTyXiK3RdzS/1CA==", + "cpu": [ + "arm" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "linux" + ] + }, + "node_modules/@vscode/vsce-sign-linux-arm64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-linux-arm64/-/vsce-sign-linux-arm64-2.0.6.tgz", + "integrity": "sha512-cfb1qK7lygtMa4NUl2582nP7aliLYuDEVpAbXJMkDq1qE+olIw/es+C8j1LJwvcRq1I2yWGtSn3EkDp9Dq5FdA==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "linux" + ] + }, + "node_modules/@vscode/vsce-sign-linux-x64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-linux-x64/-/vsce-sign-linux-x64-2.0.6.tgz", + "integrity": "sha512-/olerl1A4sOqdP+hjvJ1sbQjKN07Y3DVnxO4gnbn/ahtQvFrdhUi0G1VsZXDNjfqmXw57DmPi5ASnj/8PGZhAA==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "linux" + ] + }, + "node_modules/@vscode/vsce-sign-win32-arm64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-win32-arm64/-/vsce-sign-win32-arm64-2.0.6.tgz", + "integrity": "sha512-ivM/MiGIY0PJNZBoGtlRBM/xDpwbdlCWomUWuLmIxbi1Cxe/1nooYrEQoaHD8ojVRgzdQEUzMsRbyF5cJJgYOg==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "win32" + ] + }, + "node_modules/@vscode/vsce-sign-win32-x64": { + "version": "2.0.6", + "resolved": "https://registry.npmjs.org/@vscode/vsce-sign-win32-x64/-/vsce-sign-win32-x64-2.0.6.tgz", + "integrity": "sha512-mgth9Kvze+u8CruYMmhHw6Zgy3GRX2S+Ed5oSokDEK5vPEwGGKnmuXua9tmFhomeAnhgJnL4DCna3TiNuGrBTQ==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "SEE LICENSE IN LICENSE.txt", + "optional": true, + "os": [ + "win32" + ] + }, + "node_modules/agent-base": { + "version": "7.1.4", + "resolved": "https://registry.npmjs.org/agent-base/-/agent-base-7.1.4.tgz", + "integrity": "sha512-MnA+YT8fwfJPgBx3m60MNqakm30XOkyIoH1y6huTQvC0PwZG7ki8NacLBcrPbNoo8vEZy7Jpuk7+jMO+CUovTQ==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 14" + } + }, + "node_modules/assertion-error": { + "version": "2.0.1", + "resolved": "https://registry.npmjs.org/assertion-error/-/assertion-error-2.0.1.tgz", + "integrity": "sha512-Izi8RQcffqCeNVgFigKli1ssklIbpHnCYc6AknXGYoB6grJqyeby7jv12JUQgmTAnIDnbck1uxksT4dzN3PWBA==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=12" + } + }, + "node_modules/azure-devops-node-api": { + "version": "12.5.0", + "resolved": "https://registry.npmjs.org/azure-devops-node-api/-/azure-devops-node-api-12.5.0.tgz", + "integrity": "sha512-R5eFskGvOm3U/GzeAuxRkUsAl0hrAwGgWn6zAd2KrZmrEhWZVqLew4OOupbQlXUuojUzpGtq62SmdhJ06N88og==", + "dev": true, + "license": "MIT", + "dependencies": { + "tunnel": "0.0.6", + "typed-rest-client": "^1.8.4" + } + }, + "node_modules/balanced-match": { + "version": "4.0.4", + "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-4.0.4.tgz", + "integrity": "sha512-BLrgEcRTwX2o6gGxGOCNyMvGSp35YofuYzw9h1IMTRmKqttAZZVU67bdb9Pr2vUHA8+j3i2tJfjO6C6+4myGTA==", + "dev": true, + "license": "MIT", + "engines": { + "node": "18 || 20 || >=22" + } + }, + "node_modules/binaryextensions": { + "version": "6.11.0", + "resolved": "https://registry.npmjs.org/binaryextensions/-/binaryextensions-6.11.0.tgz", + "integrity": "sha512-sXnYK/Ij80TO3lcqZVV2YgfKN5QjUWIRk/XSm2J/4bd/lPko3lvk0O4ZppH6m+6hB2/GTu+ptNwVFe1xh+QLQw==", + "dev": true, + "license": "Artistic-2.0", + "dependencies": { + "editions": "^6.21.0" + }, + "engines": { + "node": ">=4" + }, + "funding": { + "url": "https://bevry.me/fund" + } + }, + "node_modules/boundary": { + "version": "2.0.0", + "resolved": "https://registry.npmjs.org/boundary/-/boundary-2.0.0.tgz", + "integrity": "sha512-rJKn5ooC9u8q13IMCrW0RSp31pxBCHE3y9V/tp3TdWSLf8Em3p6Di4NBpfzbJge9YjjFEsD0RtFEjtvHL5VyEA==", + "dev": true, + "license": "BSD-2-Clause" + }, + "node_modules/brace-expansion": { + "version": "5.0.12", + "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.12.tgz", + "integrity": "sha512-YovQ3rzhaLMIrDjNDMkNS01tea93qhEhG5xy8f6+R0l+dw3Ki+5sCoIoI942iuLZTHWogWktgwVDhU09iNEimQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "balanced-match": "^4.0.2" + }, + "engines": { + "node": "20 || >=22" + } + }, + "node_modules/buffer-crc32": { + "version": "0.2.13", + "resolved": "https://registry.npmjs.org/buffer-crc32/-/buffer-crc32-0.2.13.tgz", + "integrity": "sha512-VO9Ht/+p3SN7SKWqcrgEzjGbRSJYTx+Q1pTQC0wrWqHx0vpJraQ6GtHx8tvcg1rlK1byhU5gccxgOgj7B0TDkQ==", + "dev": true, + "license": "MIT", + "engines": { + "node": "*" + } + }, + "node_modules/buffer-equal-constant-time": { + "version": "1.0.1", + "resolved": "https://registry.npmjs.org/buffer-equal-constant-time/-/buffer-equal-constant-time-1.0.1.tgz", + "integrity": "sha512-zRpUiDwd/xk6ADqPMATG8vc9VPrkck7T07OIx0gnjmJAnHnTVXNQG3vfvWNuiZIkwu9KrKdA1iJKfsfTVxE6NA==", + "dev": true, + "license": "BSD-3-Clause" + }, + "node_modules/bundle-name": { + "version": "4.1.0", + "resolved": "https://registry.npmjs.org/bundle-name/-/bundle-name-4.1.0.tgz", + "integrity": "sha512-tjwM5exMg6BGRI+kNmTntNsvdZS1X8BFYS6tnJ2hdH0kVxM6/eVZ2xy+FqStSWvYmtfFMDLIxurorHwDKfDz5Q==", + "dev": true, + "license": "MIT", + "dependencies": { + "run-applescript": "^7.0.0" + }, + "engines": { + "node": ">=18" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/call-bind-apply-helpers": { + "version": "1.0.2", + "resolved": "https://registry.npmjs.org/call-bind-apply-helpers/-/call-bind-apply-helpers-1.0.2.tgz", + "integrity": "sha512-Sp1ablJ0ivDkSzjcaJdxEunN5/XvksFJ2sMBFfq6x0ryhQV/2b/KwFe21cMpmHtPOSij8K99/wSfoEuTObmuMQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "es-errors": "^1.3.0", + "function-bind": "^1.1.2" + }, + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/call-bound": { + "version": "1.0.4", + "resolved": "https://registry.npmjs.org/call-bound/-/call-bound-1.0.4.tgz", + "integrity": "sha512-+ys997U96po4Kx/ABpBCqhA9EuxJaQWDQg7295H4hBphv3IZg0boBKuwYpt4YXp6MZ5AmZQnU/tyMTlRpaSejg==", + "dev": true, + "license": "MIT", + "dependencies": { + "call-bind-apply-helpers": "^1.0.2", + "get-intrinsic": "^1.3.0" + }, + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/chai": { + "version": "6.2.2", + "resolved": "https://registry.npmjs.org/chai/-/chai-6.2.2.tgz", + "integrity": "sha512-NUPRluOfOiTKBKvWPtSD4PhFvWCqOi0BGStNWs57X9js7XGTprSmFoz5F0tWhR4WPjNeR9jXqdC7/UpSJTnlRg==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=18" + } + }, + "node_modules/cockatiel": { + "version": "3.2.1", + "resolved": "https://registry.npmjs.org/cockatiel/-/cockatiel-3.2.1.tgz", + "integrity": "sha512-gfrHV6ZPkquExvMh9IOkKsBzNDk6sDuZ6DdBGUBkvFnTCqCxzpuq48RySgP0AnaqQkw2zynOFj9yly6T1Q2G5Q==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=16" + } + }, + "node_modules/commander": { + "version": "12.1.0", + "resolved": "https://registry.npmjs.org/commander/-/commander-12.1.0.tgz", + "integrity": "sha512-Vw8qHK3bZM9y/P10u3Vib8o/DdkvA2OtPtZvD871QKjy74Wj1WSKFILMPRPSdUSx5RFK1arlJzEtA4PkFgnbuA==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=18" + } + }, + "node_modules/convert-source-map": { + "version": "2.0.0", + "resolved": "https://registry.npmjs.org/convert-source-map/-/convert-source-map-2.0.0.tgz", + "integrity": "sha512-Kvp459HrV2FEJ1CAsi1Ku+MY3kasH19TFykTz2xWmMeq6bk2NU3XXvfJ+Q61m0xktWwt+1HSYf3JZsTms3aRJg==", + "dev": true, + "license": "MIT" + }, + "node_modules/debug": { + "version": "4.4.3", + "resolved": "https://registry.npmjs.org/debug/-/debug-4.4.3.tgz", + "integrity": "sha512-RGwwWnwQvkVfavKVt22FGLw+xYSdzARwm0ru6DhTVA3umU5hZc28V3kO4stgYryrTlLpuvgI9GiijltAjNbcqA==", + "dev": true, + "license": "MIT", + "dependencies": { + "ms": "^2.1.3" + }, + "engines": { + "node": ">=6.0" + }, + "peerDependenciesMeta": { + "supports-color": { + "optional": true + } + } + }, + "node_modules/default-browser": { + "version": "5.5.1", + "resolved": "https://registry.npmjs.org/default-browser/-/default-browser-5.5.1.tgz", + "integrity": "sha512-m1pAzaJgZ/gssEqlOhJkPJp8Xly7QyW6xcrkUa2KKcDeDSEMP7X8xipU3snUcfisTQx0w1AGae+9UtJSfVnXGw==", + "dev": true, + "license": "MIT", + "dependencies": { + "bundle-name": "^4.1.0", + "default-browser-id": "^5.0.0" + }, + "engines": { + "node": ">=18" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/default-browser-id": { + "version": "5.0.1", + "resolved": "https://registry.npmjs.org/default-browser-id/-/default-browser-id-5.0.1.tgz", + "integrity": "sha512-x1VCxdX4t+8wVfd1so/9w+vQ4vx7lKd2Qp5tDRutErwmR85OgmfX7RlLRMWafRMY7hbEiXIbudNrjOAPa/hL8Q==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=18" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/define-lazy-prop": { + "version": "3.0.0", + "resolved": "https://registry.npmjs.org/define-lazy-prop/-/define-lazy-prop-3.0.0.tgz", + "integrity": "sha512-N+MeXYoqr3pOgn8xfyRPREN7gHakLYjhsHhWGT3fWAiL4IkAt0iDw14QiiEm2bE30c5XX5q0FtAA3CK5f9/BUg==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=12" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/detect-libc": { + "version": "2.1.2", + "resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-2.1.2.tgz", + "integrity": "sha512-Btj2BOOO83o3WyH59e8MgXsxEQVcarkUOpEYrubB0urwnN10yQ364rsiByU11nZlqWYZm05i/of7io4mzihBtQ==", + "dev": true, + "license": "Apache-2.0", + "engines": { + "node": ">=8" + } + }, + "node_modules/dunder-proto": { + "version": "1.0.1", + "resolved": "https://registry.npmjs.org/dunder-proto/-/dunder-proto-1.0.1.tgz", + "integrity": "sha512-KIN/nDJBQRcXw0MLVhZE9iQHmG68qAVIBg9CqmUYjmQIhgij9U5MFvrqkUL5FbtyyzZuOeOt0zdeRe4UY7ct+A==", + "dev": true, + "license": "MIT", + "dependencies": { + "call-bind-apply-helpers": "^1.0.1", + "es-errors": "^1.3.0", + "gopd": "^1.2.0" + }, + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/ecdsa-sig-formatter": { + "version": "1.0.11", + "resolved": "https://registry.npmjs.org/ecdsa-sig-formatter/-/ecdsa-sig-formatter-1.0.11.tgz", + "integrity": "sha512-nagl3RYrbNv6kQkeJIpt6NJZy8twLB/2vtz6yN9Z4vRKHN4/QZJIEbqohALSgwKdnksuY3k5Addp5lg8sVoVcQ==", + "dev": true, + "license": "Apache-2.0", + "dependencies": { + "safe-buffer": "^5.0.1" + } + }, + "node_modules/editions": { + "version": "6.22.0", + "resolved": "https://registry.npmjs.org/editions/-/editions-6.22.0.tgz", + "integrity": "sha512-UgGlf8IW75je7HZjNDpJdCv4cGJWIi6yumFdZ0R7A8/CIhQiWUjyGLCxdHpd8bmyD1gnkfUNK0oeOXqUS2cpfQ==", + "dev": true, + "license": "Artistic-2.0", + "dependencies": { + "version-range": "^4.15.0" + }, + "engines": { + "ecmascript": ">= es5", + "node": ">=4" + }, + "funding": { + "url": "https://bevry.me/fund" + } + }, + "node_modules/entities": { + "version": "8.1.0", + "resolved": "https://registry.npmjs.org/entities/-/entities-8.1.0.tgz", + "integrity": "sha512-kxL7msIffSuh9aaFAMD7rxAIuTRMAHMeBtgHW2yUdWw732ZNh4MehkF2gdjvtdmikkaIP9bFDDJOPlsvm7avrA==", + "dev": true, + "license": "BSD-2-Clause", + "engines": { + "node": ">=20.19.0" + }, + "funding": { + "url": "https://github.com/fb55/entities?sponsor=1" + } + }, + "node_modules/es-define-property": { + "version": "1.0.1", + "resolved": "https://registry.npmjs.org/es-define-property/-/es-define-property-1.0.1.tgz", + "integrity": "sha512-e3nRfgfUZ4rNGL232gUgX06QNyyez04KdjFrF+LTRoOXmrOgFKDg4BCdsjW8EnT69eqdYGmRpJwiPVYNrCaW3g==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/es-errors": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/es-errors/-/es-errors-1.3.0.tgz", + "integrity": "sha512-Zf5H2Kxt2xjTvbJvP2ZWLEICxA6j+hAmMzIlypy4xcBg1vKVnx89Wy0GbS+kf5cwCVFFzdCFh2XSCFNULS6csw==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/es-module-lexer": { + "version": "2.3.2", + "resolved": "https://registry.npmjs.org/es-module-lexer/-/es-module-lexer-2.3.2.tgz", + "integrity": "sha512-poHGpORABojJJucnV9KbOavETW8lBVnphkW77ER5/BQ5Fz7oXSoCNek7IH3vR5nRjdsEz926ibFYX8KtLQmdyw==", + "dev": true, + "license": "MIT" + }, + "node_modules/es-object-atoms": { + "version": "1.1.2", + "resolved": "https://registry.npmjs.org/es-object-atoms/-/es-object-atoms-1.1.2.tgz", + "integrity": "sha512-HWcBoN6NileqtSydK2FqHbS/LoDd2pqrnQHLyJzBj4kOp/ky2MWMN694xOfkK8/SnUsW2DH7EfyVlydKCsm1Zw==", + "dev": true, + "license": "MIT", + "dependencies": { + "es-errors": "^1.3.0" + }, + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/esbuild": { + "version": "0.28.2", + "resolved": "https://registry.npmjs.org/esbuild/-/esbuild-0.28.2.tgz", + "integrity": "sha512-HKVLS8dvII+xoKW9kmqxbRKrnWEXfJJr/FZhhJmiqIB0e053QNYFqOBouTMO/k5sID4MvCiUCvv8b9M4h32wIA==", + "dev": true, + "hasInstallScript": true, + "license": "MIT", + "bin": { + "esbuild": "bin/esbuild" + }, + "engines": { + "node": ">=18" + }, + "optionalDependencies": { + "@esbuild/aix-ppc64": "0.28.2", + "@esbuild/android-arm": "0.28.2", + "@esbuild/android-arm64": "0.28.2", + "@esbuild/android-x64": "0.28.2", + "@esbuild/darwin-arm64": "0.28.2", + "@esbuild/darwin-x64": "0.28.2", + "@esbuild/freebsd-arm64": "0.28.2", + "@esbuild/freebsd-x64": "0.28.2", + "@esbuild/linux-arm": "0.28.2", + "@esbuild/linux-arm64": "0.28.2", + "@esbuild/linux-ia32": "0.28.2", + "@esbuild/linux-loong64": "0.28.2", + "@esbuild/linux-mips64el": "0.28.2", + "@esbuild/linux-ppc64": "0.28.2", + "@esbuild/linux-riscv64": "0.28.2", + "@esbuild/linux-s390x": "0.28.2", + "@esbuild/linux-x64": "0.28.2", + "@esbuild/netbsd-arm64": "0.28.2", + "@esbuild/netbsd-x64": "0.28.2", + "@esbuild/openbsd-arm64": "0.28.2", + "@esbuild/openbsd-x64": "0.28.2", + "@esbuild/openharmony-arm64": "0.28.2", + "@esbuild/sunos-x64": "0.28.2", + "@esbuild/win32-arm64": "0.28.2", + "@esbuild/win32-ia32": "0.28.2", + "@esbuild/win32-x64": "0.28.2" + } + }, + "node_modules/estree-walker": { + "version": "3.0.3", + "resolved": "https://registry.npmjs.org/estree-walker/-/estree-walker-3.0.3.tgz", + "integrity": "sha512-7RUKfXgSMMkzt6ZuXmqapOurLGPPfgj6l9uRZ7lRGolvk0y2yocc35LdcxKC5PQZdn2DMqioAQ2NoWcrTKmm6g==", + "dev": true, + "license": "MIT", + "dependencies": { + "@types/estree": "^1.0.0" + } + }, + "node_modules/expect-type": { + "version": "1.4.0", + "resolved": "https://registry.npmjs.org/expect-type/-/expect-type-1.4.0.tgz", + "integrity": "sha512-KfYbmpRm0VbLjEvVa9yGwCi9GI34xvi7A/HXYWQO65CSD2u3MczUJSuwXKFIxlGsgBQizV9q5J9NHj4VG0n+pA==", + "dev": true, + "license": "Apache-2.0", + "engines": { + "node": ">=12.0.0" + } + }, + "node_modules/fdir": { + "version": "6.5.0", + "resolved": "https://registry.npmjs.org/fdir/-/fdir-6.5.0.tgz", + "integrity": "sha512-tIbYtZbucOs0BRGqPJkshJUYdL+SDH7dVM8gjy+ERp3WAUjLEFJE+02kanyHtwjWOnwrKYBiwAmM0p4kLJAnXg==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=12.0.0" + }, + "peerDependencies": { + "picomatch": "^3 || ^4" + }, + "peerDependenciesMeta": { + "picomatch": { + "optional": true + } + } + }, + "node_modules/fsevents": { + "version": "2.3.3", + "resolved": "https://registry.npmjs.org/fsevents/-/fsevents-2.3.3.tgz", + "integrity": "sha512-5xoDfX+fL7faATnagmWPpbFtwh/R77WmMMqqHGS65C3vvB0YHrgF+B1YmZ3441tMj5n63k0212XNoJwzlhffQw==", + "dev": true, + "hasInstallScript": true, + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": "^8.16.0 || ^10.6.0 || >=11.0.0" + } + }, + "node_modules/function-bind": { + "version": "1.1.2", + "resolved": "https://registry.npmjs.org/function-bind/-/function-bind-1.1.2.tgz", + "integrity": "sha512-7XHNxH7qX9xG5mIwxkhumTox/MIRNcOgDrxWsMt2pAr23WHp6MrRlN7FBSFpCpr+oVO0F744iUgR82nJMfG2SA==", + "dev": true, + "license": "MIT", + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/get-intrinsic": { + "version": "1.3.0", + "resolved": "https://registry.npmjs.org/get-intrinsic/-/get-intrinsic-1.3.0.tgz", + "integrity": "sha512-9fSjSaos/fRIVIp+xSJlE6lfwhES7LNtKaCBIamHsjr2na1BiABJPo0mOjjz8GJDURarmCPGqaiVg5mfjb98CQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "call-bind-apply-helpers": "^1.0.2", + "es-define-property": "^1.0.1", + "es-errors": "^1.3.0", + "es-object-atoms": "^1.1.1", + "function-bind": "^1.1.2", + "get-proto": "^1.0.1", + "gopd": "^1.2.0", + "has-symbols": "^1.1.0", + "hasown": "^2.0.2", + "math-intrinsics": "^1.1.0" + }, + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/get-proto": { + "version": "1.0.1", + "resolved": "https://registry.npmjs.org/get-proto/-/get-proto-1.0.1.tgz", + "integrity": "sha512-sTSfBjoXBp89JvIKIefqw7U2CCebsc74kiY6awiGogKtoSGbgjYE/G/+l9sF3MWFPNc9IcoOC4ODfKHfxFmp0g==", + "dev": true, + "license": "MIT", + "dependencies": { + "dunder-proto": "^1.0.1", + "es-object-atoms": "^1.0.0" + }, + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/gopd": { + "version": "1.2.0", + "resolved": "https://registry.npmjs.org/gopd/-/gopd-1.2.0.tgz", + "integrity": "sha512-ZUKRh6/kUFoAiTAtTYPZJ3hw9wNxx+BIBOijnlG9PnrJsCcSjs1wyyD6vJpaYtgnzDrKYRSqf3OO6Rfa93xsRg==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/graceful-fs": { + "version": "4.2.11", + "resolved": "https://registry.npmjs.org/graceful-fs/-/graceful-fs-4.2.11.tgz", + "integrity": "sha512-RbJ5/jmFcNNCcDV5o9eTnBLJ/HszWV0P73bc+Ff4nS/rJj+YaS6IGyiOL0VoBYX+l1Wrl3k63h/KrH+nhJ0XvQ==", + "dev": true, + "license": "ISC" + }, + "node_modules/has-symbols": { + "version": "1.1.0", + "resolved": "https://registry.npmjs.org/has-symbols/-/has-symbols-1.1.0.tgz", + "integrity": "sha512-1cDNdwJ2Jaohmb3sg4OmKaMBwuC48sYni5HUw2DvsC8LjGTLK9h+eb1X6RyuOHe4hT0ULCW68iomhjUoKUqlPQ==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/hasown": { + "version": "2.0.4", + "resolved": "https://registry.npmjs.org/hasown/-/hasown-2.0.4.tgz", + "integrity": "sha512-T2UbfbBEF32wiepXIsMlTW9+dDYC6wMh/t/vYA4tuOMKqWz/n3vr1NFSxQiyP+zk2mXsoMA/i/7qV6LKut1t1A==", + "dev": true, + "license": "MIT", + "dependencies": { + "function-bind": "^1.1.2" + }, + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/hosted-git-info": { + "version": "4.1.0", + "resolved": "https://registry.npmjs.org/hosted-git-info/-/hosted-git-info-4.1.0.tgz", + "integrity": "sha512-kyCuEOWjJqZuDbRHzL8V93NzQhwIB71oFWSyzVo+KPZI+pnQPPxucdkrOZvkLRnrf5URsQM+IJ09Dw29cRALIA==", + "dev": true, + "license": "ISC", + "dependencies": { + "lru-cache": "^6.0.0" + }, + "engines": { + "node": ">=10" + } + }, + "node_modules/http-proxy-agent": { + "version": "7.0.2", + "resolved": "https://registry.npmjs.org/http-proxy-agent/-/http-proxy-agent-7.0.2.tgz", + "integrity": "sha512-T1gkAiYYDWYx3V5Bmyu7HcfcvL7mUrTWiM6yOfa3PIphViJ/gFPbvidQ+veqSOHci/PxBcDabeUNCzpOODJZig==", + "dev": true, + "license": "MIT", + "dependencies": { + "agent-base": "^7.1.0", + "debug": "^4.3.4" + }, + "engines": { + "node": ">= 14" + } + }, + "node_modules/https-proxy-agent": { + "version": "7.0.6", + "resolved": "https://registry.npmjs.org/https-proxy-agent/-/https-proxy-agent-7.0.6.tgz", + "integrity": "sha512-vK9P5/iUfdl95AI+JVyUuIcVtd4ofvtrOr3HNtM2yxC9bnMbEdp3x01OhQNnjb8IJYi38VlTE3mBXwcfvywuSw==", + "dev": true, + "license": "MIT", + "dependencies": { + "agent-base": "^7.1.2", + "debug": "4" + }, + "engines": { + "node": ">= 14" + } + }, + "node_modules/is-docker": { + "version": "3.0.0", + "resolved": "https://registry.npmjs.org/is-docker/-/is-docker-3.0.0.tgz", + "integrity": "sha512-eljcgEDlEns/7AXFosB5K/2nCM4P7FQPkGc/DWLy5rmFEWvZayGrik1d9/QIY5nJ4f9YsVvBkA6kJpHn9rISdQ==", + "dev": true, + "license": "MIT", + "bin": { + "is-docker": "cli.js" + }, + "engines": { + "node": "^12.20.0 || ^14.13.1 || >=16.0.0" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/is-inside-container": { + "version": "1.0.0", + "resolved": "https://registry.npmjs.org/is-inside-container/-/is-inside-container-1.0.0.tgz", + "integrity": "sha512-KIYLCCJghfHZxqjYBE7rEy0OBuTd5xCHS7tHVgvCLkx7StIoaxwNW3hCALgEUjFfeRk+MG/Qxmp/vtETEF3tRA==", + "dev": true, + "license": "MIT", + "dependencies": { + "is-docker": "^3.0.0" + }, + "bin": { + "is-inside-container": "cli.js" + }, + "engines": { + "node": ">=14.16" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/is-wsl": { + "version": "3.1.1", + "resolved": "https://registry.npmjs.org/is-wsl/-/is-wsl-3.1.1.tgz", + "integrity": "sha512-e6rvdUCiQCAuumZslxRJWR/Doq4VpPR82kqclvcS0efgt430SlGIk05vdCN58+VrzgtIcfNODjozVielycD4Sw==", + "dev": true, + "license": "MIT", + "dependencies": { + "is-inside-container": "^1.0.0" + }, + "engines": { + "node": ">=16" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/istextorbinary": { + "version": "9.5.0", + "resolved": "https://registry.npmjs.org/istextorbinary/-/istextorbinary-9.5.0.tgz", + "integrity": "sha512-5mbUj3SiZXCuRf9fT3ibzbSSEWiy63gFfksmGfdOzujPjW3k+z8WvIBxcJHBoQNlaZaiyB25deviif2+osLmLw==", + "dev": true, + "license": "Artistic-2.0", + "dependencies": { + "binaryextensions": "^6.11.0", + "editions": "^6.21.0", + "textextensions": "^6.11.0" + }, + "engines": { + "node": ">=4" + }, + "funding": { + "url": "https://bevry.me/fund" + } + }, + "node_modules/jsonc-parser": { + "version": "3.3.1", + "resolved": "https://registry.npmjs.org/jsonc-parser/-/jsonc-parser-3.3.1.tgz", + "integrity": "sha512-HUgH65KyejrUFPvHFPbqOY0rsFip3Bo5wb4ngvdi1EpCYWUQDC5V+Y7mZws+DLkr4M//zQJoanu1SP+87Dv1oQ==", + "dev": true, + "license": "MIT" + }, + "node_modules/jsonwebtoken": { + "version": "9.0.3", + "resolved": "https://registry.npmjs.org/jsonwebtoken/-/jsonwebtoken-9.0.3.tgz", + "integrity": "sha512-MT/xP0CrubFRNLNKvxJ2BYfy53Zkm++5bX9dtuPbqAeQpTVe0MQTFhao8+Cp//EmJp244xt6Drw/GVEGCUj40g==", + "dev": true, + "license": "MIT", + "dependencies": { + "jws": "^4.0.1", + "lodash.includes": "^4.3.0", + "lodash.isboolean": "^3.0.3", + "lodash.isinteger": "^4.0.4", + "lodash.isnumber": "^3.0.3", + "lodash.isplainobject": "^4.0.6", + "lodash.isstring": "^4.0.1", + "lodash.once": "^4.0.0", + "ms": "^2.1.1", + "semver": "^7.5.4" + }, + "engines": { + "node": ">=12", + "npm": ">=6" + } + }, + "node_modules/jwa": { + "version": "2.0.1", + "resolved": "https://registry.npmjs.org/jwa/-/jwa-2.0.1.tgz", + "integrity": "sha512-hRF04fqJIP8Abbkq5NKGN0Bbr3JxlQ+qhZufXVr0DvujKy93ZCbXZMHDL4EOtodSbCWxOqR8MS1tXA5hwqCXDg==", + "dev": true, + "license": "MIT", + "dependencies": { + "buffer-equal-constant-time": "^1.0.1", + "ecdsa-sig-formatter": "1.0.11", + "safe-buffer": "^5.0.1" + } + }, + "node_modules/jws": { + "version": "4.0.1", + "resolved": "https://registry.npmjs.org/jws/-/jws-4.0.1.tgz", + "integrity": "sha512-EKI/M/yqPncGUUh44xz0PxSidXFr/+r0pA70+gIYhjv+et7yxM+s29Y+VGDkovRofQem0fs7Uvf4+YmAdyRduA==", + "dev": true, + "license": "MIT", + "dependencies": { + "jwa": "^2.0.1", + "safe-buffer": "^5.0.1" + } + }, + "node_modules/lightningcss": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss/-/lightningcss-1.33.0.tgz", + "integrity": "sha512-WkUDrojuJs0xkgGf2udWxa3yGBRxPtxUkB79i6aCZLRgc7PM8fZe9TosfPDcvEpQZbuFASnHYmRLBLUbmLOIIA==", + "dev": true, + "license": "MPL-2.0", + "dependencies": { + "detect-libc": "^2.0.3" + }, + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + }, + "optionalDependencies": { + "lightningcss-android-arm64": "1.33.0", + "lightningcss-darwin-arm64": "1.33.0", + "lightningcss-darwin-x64": "1.33.0", + "lightningcss-freebsd-x64": "1.33.0", + "lightningcss-linux-arm-gnueabihf": "1.33.0", + "lightningcss-linux-arm64-gnu": "1.33.0", + "lightningcss-linux-arm64-musl": "1.33.0", + "lightningcss-linux-x64-gnu": "1.33.0", + "lightningcss-linux-x64-musl": "1.33.0", + "lightningcss-win32-arm64-msvc": "1.33.0", + "lightningcss-win32-x64-msvc": "1.33.0" + } + }, + "node_modules/lightningcss-android-arm64": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-android-arm64/-/lightningcss-android-arm64-1.33.0.tgz", + "integrity": "sha512-gEpRTalKdosp4Bb8qWtc2iOgE5SeIHlpS1up9bFq2wAyYhl1UdTObYiHe98zEM9SQvSoqQZ1IQD0JNpg3Ml5pg==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MPL-2.0", + "optional": true, + "os": [ + "android" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-darwin-arm64": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-darwin-arm64/-/lightningcss-darwin-arm64-1.33.0.tgz", + "integrity": "sha512-Sciaz8eenNTKn9b3t7+xr0ipTp9YxKQY4npwQ3mrRuL0BAVHBLyZxofhaKBAVtzmtRZ/zTyo0/to4B1uWG/Djg==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MPL-2.0", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-darwin-x64": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-darwin-x64/-/lightningcss-darwin-x64-1.33.0.tgz", + "integrity": "sha512-Z5UPAxzrjlWNNyGy6i65cJzzvgJ5D3T6wMvs+gWpY9d7qRhANrxqAp6LhxIgZhWEw18RfJTGcRxjuLIBr+m8XQ==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MPL-2.0", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-freebsd-x64": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-freebsd-x64/-/lightningcss-freebsd-x64-1.33.0.tgz", + "integrity": "sha512-QQM/Ti/hQajJwCY+RiWuCZ9sdtI/XQk7nDK5vC8kkdwixezOlDgvDx7+RT+QjK6FcFT4MpsuoBnHIo/O3StRRg==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MPL-2.0", + "optional": true, + "os": [ + "freebsd" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-arm-gnueabihf": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-linux-arm-gnueabihf/-/lightningcss-linux-arm-gnueabihf-1.33.0.tgz", + "integrity": "sha512-N7FVBe6iS24MlM6R/4RBTxGhQheZGs7tiQ9U32UtF75NzP5Q7xWPRqLBCKxlRQRk3rY1jCIPLzx7WzOhuUIRLQ==", + "cpu": [ + "arm" + ], + "dev": true, + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-arm64-gnu": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-gnu/-/lightningcss-linux-arm64-gnu-1.33.0.tgz", + "integrity": "sha512-j2v/itmy4HlNxlc6voKXYgBqNi0Ng2LShg4z7GufpEgs05P+2suBVyi9I6YHq5uoVFx9ETin3eCEhLVyXGQnKg==", + "cpu": [ + "arm64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-arm64-musl": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-musl/-/lightningcss-linux-arm64-musl-1.33.0.tgz", + "integrity": "sha512-yiO5ROMuYQgXbC60yjZU5CYSFZGKXL0HFATXt9mHJn1+zW55oCtMI9NfcVhYLMFDL7gV7oBPon/EmMMGg2OvtQ==", + "cpu": [ + "arm64" + ], + "dev": true, + "libc": [ + "musl" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-x64-gnu": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-linux-x64-gnu/-/lightningcss-linux-x64-gnu-1.33.0.tgz", + "integrity": "sha512-ar+Ju7LmcN0Jo4FpL4hpFybwNG9/3A/Br5KW2n2jyODg3MEZXaDYADdemoNS+BDNfMgKvylJLj4S5tyRActuAg==", + "cpu": [ + "x64" + ], + "dev": true, + "libc": [ + "glibc" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-x64-musl": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-linux-x64-musl/-/lightningcss-linux-x64-musl-1.33.0.tgz", + "integrity": "sha512-RYiYbkokw0trfKqqzfF55lginwEPrD3OJDfTuJzFs1MK6iFnDenaz1fqLLtX4ITG3OktJQXOeTaw1awrBAlZPw==", + "cpu": [ + "x64" + ], + "dev": true, + "libc": [ + "musl" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-win32-arm64-msvc": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-win32-arm64-msvc/-/lightningcss-win32-arm64-msvc-1.33.0.tgz", + "integrity": "sha512-1K+MPfLSFVpphzpdbfkhlWk6wBrTObBzS2T6db10PNOZgR9GoVsAWzwNyuhUYYbTp23j+4RrncfujZ4uAzXvwA==", + "cpu": [ + "arm64" + ], + "dev": true, + "license": "MPL-2.0", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-win32-x64-msvc": { + "version": "1.33.0", + "resolved": "https://registry.npmjs.org/lightningcss-win32-x64-msvc/-/lightningcss-win32-x64-msvc-1.33.0.tgz", + "integrity": "sha512-OlEICDx/Xl0FqSp4bry8zFnCvGpig3Gl4gCquvYwHuqJKEC1+n9NgDniFvqHGmMv1ZkqDJrDqKKSykTDX+ehuA==", + "cpu": [ + "x64" + ], + "dev": true, + "license": "MPL-2.0", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lodash.includes": { + "version": "4.3.0", + "resolved": "https://registry.npmjs.org/lodash.includes/-/lodash.includes-4.3.0.tgz", + "integrity": "sha512-W3Bx6mdkRTGtlJISOvVD/lbqjTlPPUDTMnlXZFnVwi9NKJ6tiAk6LVdlhZMm17VZisqhKcgzpO5Wz91PCt5b0w==", + "dev": true, + "license": "MIT" + }, + "node_modules/lodash.isboolean": { + "version": "3.0.3", + "resolved": "https://registry.npmjs.org/lodash.isboolean/-/lodash.isboolean-3.0.3.tgz", + "integrity": "sha512-Bz5mupy2SVbPHURB98VAcw+aHh4vRV5IPNhILUCsOzRmsTmSQ17jIuqopAentWoehktxGd9e/hbIXq980/1QJg==", + "dev": true, + "license": "MIT" + }, + "node_modules/lodash.isinteger": { + "version": "4.0.4", + "resolved": "https://registry.npmjs.org/lodash.isinteger/-/lodash.isinteger-4.0.4.tgz", + "integrity": "sha512-DBwtEWN2caHQ9/imiNeEA5ys1JoRtRfY3d7V9wkqtbycnAmTvRRmbHKDV4a0EYc678/dia0jrte4tjYwVBaZUA==", + "dev": true, + "license": "MIT" + }, + "node_modules/lodash.isnumber": { + "version": "3.0.3", + "resolved": "https://registry.npmjs.org/lodash.isnumber/-/lodash.isnumber-3.0.3.tgz", + "integrity": "sha512-QYqzpfwO3/CWf3XP+Z+tkQsfaLL/EnUlXWVkIk5FUPc4sBdTehEqZONuyRt2P67PXAk+NXmTBcc97zw9t1FQrw==", + "dev": true, + "license": "MIT" + }, + "node_modules/lodash.isplainobject": { + "version": "4.0.6", + "resolved": "https://registry.npmjs.org/lodash.isplainobject/-/lodash.isplainobject-4.0.6.tgz", + "integrity": "sha512-oSXzaWypCMHkPC3NvBEaPHf0KsA5mvPrOPgQWDsbg8n7orZ290M0BmC/jgRZ4vcJ6DTAhjrsSYgdsW/F+MFOBA==", + "dev": true, + "license": "MIT" + }, + "node_modules/lodash.isstring": { + "version": "4.0.1", + "resolved": "https://registry.npmjs.org/lodash.isstring/-/lodash.isstring-4.0.1.tgz", + "integrity": "sha512-0wJxfxH1wgO3GrbuP+dTTk7op+6L41QCXbGINEmD+ny/G/eCqGzxyCsh7159S+mgDDcoarnBw6PC1PS5+wUGgw==", + "dev": true, + "license": "MIT" + }, + "node_modules/lodash.once": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/lodash.once/-/lodash.once-4.1.1.tgz", + "integrity": "sha512-Sb487aTOCr9drQVL8pIxOzVhafOjZN9UU54hiN8PU3uAiSV7lx1yYNpbNmex2PK6dSJoNTSJUUswT651yww3Mg==", + "dev": true, + "license": "MIT" + }, + "node_modules/lru-cache": { + "version": "6.0.0", + "resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-6.0.0.tgz", + "integrity": "sha512-Jo6dJ04CmSjuznwJSS3pUeWmd/H0ffTlkXXgwZi+eq1UCmqQwCh+eLsYOYCwY991i2Fah4h1BEMCx4qThGbsiA==", + "dev": true, + "license": "ISC", + "dependencies": { + "yallist": "^4.0.0" + }, + "engines": { + "node": ">=10" + } + }, + "node_modules/magic-string": { + "version": "0.30.21", + "resolved": "https://registry.npmjs.org/magic-string/-/magic-string-0.30.21.tgz", + "integrity": "sha512-vd2F4YUyEXKGcLHoq+TEyCjxueSeHnFxyyjNp80yg0XV4vUhnDer/lvvlqM/arB5bXQN5K2/3oinyCRyx8T2CQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "@jridgewell/sourcemap-codec": "^1.5.5" + } + }, + "node_modules/marked": { + "version": "18.0.13", + "resolved": "https://registry.npmjs.org/marked/-/marked-18.0.13.tgz", + "integrity": "sha512-xTxVzZsBFwunP6HDmtBkabUQEYArnP7/rMDGmPj9SlrKlQ4i8MdYVow+nJL0eOqwpUqhzBoTBRADGN6uYwPyOw==", + "dev": true, + "license": "MIT", + "bin": { + "marked": "bin/marked.js" + }, + "engines": { + "node": ">= 20" + } + }, + "node_modules/math-intrinsics": { + "version": "1.1.0", + "resolved": "https://registry.npmjs.org/math-intrinsics/-/math-intrinsics-1.1.0.tgz", + "integrity": "sha512-/IXtbwEk5HTPyEwyKX6hGkYXxM9nbj64B+ilVJnC/R6B0pH5G4V3b0pVbL7DBj4tkhBAppbQUlf6F6Xl9LHu1g==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 0.4" + } + }, + "node_modules/mime": { + "version": "1.6.0", + "resolved": "https://registry.npmjs.org/mime/-/mime-1.6.0.tgz", + "integrity": "sha512-x0Vn8spI+wuJ1O6S7gnbaQg8Pxh4NNHb7KSINmEWKiPE4RKOplvijn+NkmYmmRgP68mc70j2EbeTFRsrswaQeg==", + "dev": true, + "license": "MIT", + "bin": { + "mime": "cli.js" + }, + "engines": { + "node": ">=4" + } + }, + "node_modules/minimatch": { + "version": "10.2.6", + "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.2.6.tgz", + "integrity": "sha512-vpLQEs+VLCr1nU0BXS07maYoFwlDAH0gngQuuttxIwutDFEMHq2blX+8vpgxDdK3J1PwjCJiep77OitTZ4Ll1A==", + "dev": true, + "license": "BlueOak-1.0.0", + "dependencies": { + "brace-expansion": "^5.0.8" + }, + "engines": { + "node": "18 || 20 || >=22" + }, + "funding": { + "url": "https://github.com/sponsors/isaacs" + } + }, + "node_modules/ms": { + "version": "2.1.3", + "resolved": "https://registry.npmjs.org/ms/-/ms-2.1.3.tgz", + "integrity": "sha512-6FlzubTLZG3J2a/NVCAleEhjzq5oxgHyaCU9yYXvcLsvoVaHJq/s5xXI6/XXP6tz7R9xAOtHnSO/tXtF3WRTlA==", + "dev": true, + "license": "MIT" + }, + "node_modules/mute-stream": { + "version": "0.0.8", + "resolved": "https://registry.npmjs.org/mute-stream/-/mute-stream-0.0.8.tgz", + "integrity": "sha512-nnbWWOkoWyUsTjKrhgD0dcz22mdkSnpYqbEjIm2nhwhuxlSkpywJmBo8h0ZqJdkp73mb90SssHkN4rsRaBAfAA==", + "dev": true, + "license": "ISC" + }, + "node_modules/nanoid": { + "version": "3.3.19", + "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.19.tgz", + "integrity": "sha512-Y2tUNy4ouw6tq5oDSKeQYGOyhkUBhNOcGV/02KC+6kd9eDGqdZd++mjMiIDilrBYvjEnCYvVtsuHCuP+okSfug==", + "dev": true, + "funding": [ + { + "type": "github", + "url": "https://github.com/sponsors/ai" + } + ], + "license": "MIT", + "bin": { + "nanoid": "bin/nanoid.cjs" + }, + "engines": { + "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1" + } + }, + "node_modules/object-inspect": { + "version": "1.13.4", + "resolved": "https://registry.npmjs.org/object-inspect/-/object-inspect-1.13.4.tgz", + "integrity": "sha512-W67iLl4J2EXEGTbfeHCffrjDfitvLANg0UlX3wFUUSTx92KXRFegMHUVgSqE+wvhAbi4WqjGg9czysTV2Epbew==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/obug": { + "version": "2.2.1", + "resolved": "https://registry.npmjs.org/obug/-/obug-2.2.1.tgz", + "integrity": "sha512-XrsrhT5sybtKI6wakr2SPOlGZWWYbUXZ7a0jT8/QOeAPau+1X/bSegNe5YR75oJmEZQbKningirmGOEJCIk61Q==", + "dev": true, + "funding": [ + "https://github.com/sponsors/sxzz", + "https://opencollective.com/debug" + ], + "license": "MIT", + "engines": { + "node": ">=12.20.0" + } + }, + "node_modules/open": { + "version": "10.2.0", + "resolved": "https://registry.npmjs.org/open/-/open-10.2.0.tgz", + "integrity": "sha512-YgBpdJHPyQ2UE5x+hlSXcnejzAvD0b22U2OuAP+8OnlJT+PjWPxtgmGqKKc+RgTM63U9gN0YzrYc71R2WT/hTA==", + "dev": true, + "license": "MIT", + "dependencies": { + "default-browser": "^5.2.1", + "define-lazy-prop": "^3.0.0", + "is-inside-container": "^1.0.0", + "wsl-utils": "^0.1.0" + }, + "engines": { + "node": ">=18" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/openai": { + "version": "7.18.0", + "resolved": "https://registry.npmjs.org/openai/-/openai-7.18.0.tgz", + "integrity": "sha512-S+xxaUf9VzIHEPDUpUFnRgvR2Ho0K1yZEaSJ8gd7xQlCNm4sdOD7/QaKWLRTYK3MK6KhwpD3cOEC0F4OPG698A==", + "license": "Apache-2.0", + "engines": { + "node": ">=22.0.0" + }, + "peerDependencies": { + "@aws-sdk/credential-provider-node": ">=3.972.0 <4", + "@smithy/hash-node": ">=4.3.0 <5", + "@smithy/signature-v4": ">=5.4.0 <6", + "undici": ">=5 <9", + "ws": "^8.21.0", + "zod": "^3.25 || ^4.0" + }, + "peerDependenciesMeta": { + "@aws-sdk/credential-provider-node": { + "optional": true + }, + "@smithy/hash-node": { + "optional": true + }, + "@smithy/signature-v4": { + "optional": true + }, + "undici": { + "optional": true + }, + "ws": { + "optional": true + }, + "zod": { + "optional": true + } + } + }, + "node_modules/parse5": { + "version": "8.0.1", + "resolved": "https://registry.npmjs.org/parse5/-/parse5-8.0.1.tgz", + "integrity": "sha512-z1e/HMG90obSGeidlli3hj7cbocou0/wa5HacvI3ASx34PecNjNQeaHNo5WIZpWofN9kgkqV1q5YvXe3F0FoPw==", + "dev": true, + "license": "MIT", + "dependencies": { + "entities": "^8.0.0" + }, + "funding": { + "url": "https://github.com/inikulin/parse5?sponsor=1" + } + }, + "node_modules/pathe": { + "version": "2.0.3", + "resolved": "https://registry.npmjs.org/pathe/-/pathe-2.0.3.tgz", + "integrity": "sha512-WUjGcAqP1gQacoQe+OBJsFA7Ld4DyXuUIjZ5cc75cLHvJ7dtNsTugphxIADwspS+AraAUePCKrSVtPLFj/F88w==", + "dev": true, + "license": "MIT" + }, + "node_modules/pend": { + "version": "1.2.0", + "resolved": "https://registry.npmjs.org/pend/-/pend-1.2.0.tgz", + "integrity": "sha512-F3asv42UuXchdzt+xXqfW1OGlVBe+mxa2mqI0pg5yAHZPvFmY3Y6drSf/GQ1A86WgWEN9Kzh/WrgKa6iGcHXLg==", + "dev": true, + "license": "MIT" + }, + "node_modules/picocolors": { + "version": "1.1.1", + "resolved": "https://registry.npmjs.org/picocolors/-/picocolors-1.1.1.tgz", + "integrity": "sha512-xceH2snhtb5M9liqDsmEw56le376mTZkEX/jEb/RxNFyegNul7eNslCXP9FDj/Lcu0X8KEyMceP2ntpaHrDEVA==", + "dev": true, + "license": "ISC" + }, + "node_modules/picomatch": { + "version": "4.0.7", + "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.7.tgz", + "integrity": "sha512-qcJu88Q2IWqJsDD529JKMdwGm/dvInW4HvQnRwiH9JtihJvzGOscDtHE3x1pBKeUOTysQ8kVmLnJ2kJu7yhcGA==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=12" + }, + "funding": { + "url": "https://github.com/sponsors/jonschlinkert" + } + }, + "node_modules/postcss": { + "version": "8.5.28", + "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.28.tgz", + "integrity": "sha512-RRuzqDtt5Y9h3quz5hWhK+TPnsmVs6WwSU6LkJMeY4HstUEDuYTG8UJSdawMRzmzAtV+KEoG8N3Qg2qLy5vM/A==", + "dev": true, + "funding": [ + { + "type": "opencollective", + "url": "https://opencollective.com/postcss/" + }, + { + "type": "tidelift", + "url": "https://tidelift.com/funding/github/npm/postcss" + }, + { + "type": "github", + "url": "https://github.com/sponsors/ai" + } + ], + "license": "MIT", + "dependencies": { + "nanoid": "^3.3.18", + "picocolors": "^1.1.1", + "source-map-js": "^1.2.1" + }, + "engines": { + "node": "^10 || ^12 || >=14" + } + }, + "node_modules/proper-lockfile": { + "version": "4.1.2", + "resolved": "https://registry.npmjs.org/proper-lockfile/-/proper-lockfile-4.1.2.tgz", + "integrity": "sha512-TjNPblN4BwAWMXU8s9AEz4JmQxnD1NNL7bNOY/AKUzyamc379FWASUhc/K1pL2noVb+XmZKLL68cjzLsiOAMaA==", + "dev": true, + "license": "MIT", + "dependencies": { + "graceful-fs": "^4.2.4", + "retry": "^0.12.0", + "signal-exit": "^3.0.2" + } + }, + "node_modules/qs": { + "version": "6.16.0", + "resolved": "https://registry.npmjs.org/qs/-/qs-6.16.0.tgz", + "integrity": "sha512-h6fhOIaRrID2CbEY2fqs+7t+UXZo+MLAnU5gRIq85uFtdiUPCdsApMlHhXogKVM4HM2DVbIjGNTTYH2OcmP1vA==", + "dev": true, + "license": "BSD-3-Clause", + "dependencies": { + "es-define-property": "^1.0.1", + "side-channel": "^1.1.1" + }, + "engines": { + "node": ">=0.6" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/read": { + "version": "1.0.7", + "resolved": "https://registry.npmjs.org/read/-/read-1.0.7.tgz", + "integrity": "sha512-rSOKNYUmaxy0om1BNjMN4ezNT6VKK+2xF4GBhc81mkH7L60i6dp8qPYrkndNLT3QPphoII3maL9PVC9XmhHwVQ==", + "dev": true, + "license": "ISC", + "dependencies": { + "mute-stream": "~0.0.4" + }, + "engines": { + "node": ">=0.8" + } + }, + "node_modules/retry": { + "version": "0.12.0", + "resolved": "https://registry.npmjs.org/retry/-/retry-0.12.0.tgz", + "integrity": "sha512-9LkiTwjUh6rT555DtE9rTX+BKByPfrMzEAtnlEtdEwr3Nkffwiihqe2bWADg+OQRjt9gl6ICdmB/ZFDCGAtSow==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">= 4" + } + }, + "node_modules/rolldown": { + "version": "1.2.9", + "resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.2.9.tgz", + "integrity": "sha512-hx/Pv0N1haXRb11qkfnK5MXB/iqr7i0yjWQqmO9uHqZpBgQSqzc8UsSnEpalsh+j1I8qQ2CkXAkJC8Br3dKSlg==", + "dev": true, + "license": "MIT", + "dependencies": { + "@oxc-project/types": "=0.150.0", + "@rolldown/pluginutils": "^1.0.0" + }, + "bin": { + "rolldown": "bin/cli.mjs" + }, + "engines": { + "node": "^20.19.0 || >=22.12.0" + }, + "optionalDependencies": { + "@rolldown/binding-android-arm-eabi": "1.2.9", + "@rolldown/binding-android-arm64": "1.2.9", + "@rolldown/binding-darwin-arm64": "1.2.9", + "@rolldown/binding-darwin-x64": "1.2.9", + "@rolldown/binding-freebsd-x64": "1.2.9", + "@rolldown/binding-linux-arm-gnueabihf": "1.2.9", + "@rolldown/binding-linux-arm64-gnu": "1.2.9", + "@rolldown/binding-linux-arm64-musl": "1.2.9", + "@rolldown/binding-linux-ppc64-gnu": "1.2.9", + "@rolldown/binding-linux-s390x-gnu": "1.2.9", + "@rolldown/binding-linux-x64-gnu": "1.2.9", + "@rolldown/binding-linux-x64-musl": "1.2.9", + "@rolldown/binding-openharmony-arm64": "1.2.9", + "@rolldown/binding-win32-arm64-msvc": "1.2.9", + "@rolldown/binding-win32-x64-msvc": "1.2.9" + } + }, + "node_modules/run-applescript": { + "version": "7.1.0", + "resolved": "https://registry.npmjs.org/run-applescript/-/run-applescript-7.1.0.tgz", + "integrity": "sha512-DPe5pVFaAsinSaV6QjQ6gdiedWDcRCbUuiQfQa2wmWV7+xC9bGulGI8+TdRmoFkAPaBXk8CrAbnlY2ISniJ47Q==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=18" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/safe-buffer": { + "version": "5.2.1", + "resolved": "https://registry.npmjs.org/safe-buffer/-/safe-buffer-5.2.1.tgz", + "integrity": "sha512-rp3So07KcdmmKbGvgaNxQSJr7bGVSVk5S9Eq1F+ppbRo70+YeaDxkw5Dd8NPN+GD6bjnYm2VuPuCXmpuYvmCXQ==", + "dev": true, + "funding": [ + { + "type": "github", + "url": "https://github.com/sponsors/feross" + }, + { + "type": "patreon", + "url": "https://www.patreon.com/feross" + }, + { + "type": "consulting", + "url": "https://feross.org/support" + } + ], + "license": "MIT" + }, + "node_modules/sax": { + "version": "1.6.1", + "resolved": "https://registry.npmjs.org/sax/-/sax-1.6.1.tgz", + "integrity": "sha512-42tBVwLWnaQvW5zc4HbZrTuWccECCZfBi92FDuwtqxasH+JbPB3/FOKb1m222K42R4WxuxzzMsTswfzgtSu64Q==", + "dev": true, + "license": "BlueOak-1.0.0", + "engines": { + "node": ">=11.0.0" + } + }, + "node_modules/semver": { + "version": "7.8.5", + "resolved": "https://registry.npmjs.org/semver/-/semver-7.8.5.tgz", + "integrity": "sha512-Y7/KDsb8LjooZpwaqGyulO6DQlksgCncchHGk+sZIY4SBvUocMBEFH5Ur1fI4dV+Jvl0w6cjvucaIi40puRioA==", + "dev": true, + "license": "ISC", + "bin": { + "semver": "bin/semver.js" + }, + "engines": { + "node": ">=10" + } + }, + "node_modules/side-channel": { + "version": "1.1.1", + "resolved": "https://registry.npmjs.org/side-channel/-/side-channel-1.1.1.tgz", + "integrity": "sha512-6x6dK6zJdpTzF4sQeNYxwtvBzf6Eg4GtlesS94HOvTudUeyK2WXAaIfmDgsyslYrRBeFIlsi54AYsFGUuhmvrQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "es-errors": "^1.3.0", + "object-inspect": "^1.13.4", + "side-channel-list": "^1.0.1", + "side-channel-map": "^1.0.1", + "side-channel-weakmap": "^1.0.2" + }, + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/side-channel-list": { + "version": "1.0.1", + "resolved": "https://registry.npmjs.org/side-channel-list/-/side-channel-list-1.0.1.tgz", + "integrity": "sha512-mjn/0bi/oUURjc5Xl7IaWi/OJJJumuoJFQJfDDyO46+hBWsfaVM65TBHq2eoZBhzl9EchxOijpkbRC8SVBQU0w==", + "dev": true, + "license": "MIT", + "dependencies": { + "es-errors": "^1.3.0", + "object-inspect": "^1.13.4" + }, + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/side-channel-map": { + "version": "1.0.1", + "resolved": "https://registry.npmjs.org/side-channel-map/-/side-channel-map-1.0.1.tgz", + "integrity": "sha512-VCjCNfgMsby3tTdo02nbjtM/ewra6jPHmpThenkTYh8pG9ucZ/1P8So4u4FGBek/BjpOVsDCMoLA/iuBKIFXRA==", + "dev": true, + "license": "MIT", + "dependencies": { + "call-bound": "^1.0.2", + "es-errors": "^1.3.0", + "get-intrinsic": "^1.2.5", + "object-inspect": "^1.13.3" + }, + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/side-channel-weakmap": { + "version": "1.0.2", + "resolved": "https://registry.npmjs.org/side-channel-weakmap/-/side-channel-weakmap-1.0.2.tgz", + "integrity": "sha512-WPS/HvHQTYnHisLo9McqBHOJk2FkHO/tlpvldyrnem4aeQp4hai3gythswg6p01oSoTl58rcpiFAjF2br2Ak2A==", + "dev": true, + "license": "MIT", + "dependencies": { + "call-bound": "^1.0.2", + "es-errors": "^1.3.0", + "get-intrinsic": "^1.2.5", + "object-inspect": "^1.13.3", + "side-channel-map": "^1.0.1" + }, + "engines": { + "node": ">= 0.4" + }, + "funding": { + "url": "https://github.com/sponsors/ljharb" + } + }, + "node_modules/siginfo": { + "version": "2.0.0", + "resolved": "https://registry.npmjs.org/siginfo/-/siginfo-2.0.0.tgz", + "integrity": "sha512-ybx0WO1/8bSBLEWXZvEd7gMW3Sn3JFlW3TvX1nREbDLRNQNaeNN8WK0meBwPdAaOI7TtRRRJn/Es1zhrrCHu7g==", + "dev": true, + "license": "ISC" + }, + "node_modules/signal-exit": { + "version": "3.0.7", + "resolved": "https://registry.npmjs.org/signal-exit/-/signal-exit-3.0.7.tgz", + "integrity": "sha512-wnD2ZE+l+SPC/uoS0vXeE9L1+0wuaMqKlfz9AMUo38JsyLSBWSFcHR1Rri62LZc12vLr1gb3jl7iwQhgwpAbGQ==", + "dev": true, + "license": "ISC" + }, + "node_modules/source-map-js": { + "version": "1.2.1", + "resolved": "https://registry.npmjs.org/source-map-js/-/source-map-js-1.2.1.tgz", + "integrity": "sha512-UXWMKhLOwVKb728IUtQPXxfYU+usdybtUrK/8uGE8CQMvrhOpwvzDBwj0QhSL7MQc7vIsISBG8VQ8+IDQxpfQA==", + "dev": true, + "license": "BSD-3-Clause", + "engines": { + "node": ">=0.10.0" + } + }, + "node_modules/stackback": { + "version": "0.0.2", + "resolved": "https://registry.npmjs.org/stackback/-/stackback-0.0.2.tgz", + "integrity": "sha512-1XMJE5fQo1jGH6Y/7ebnwPOBEkIEnT4QF32d5R1+VXdXveM0IBMJt8zfaxX1P3QhVwrYe+576+jkANtSS2mBbw==", + "dev": true, + "license": "MIT" + }, + "node_modules/std-env": { + "version": "4.2.0", + "resolved": "https://registry.npmjs.org/std-env/-/std-env-4.2.0.tgz", + "integrity": "sha512-oCUKSupKTHX53EyjDtuZQ64pjLJ6yYCtpmEw0goYxtjG9KpbRe8KAsl2tBUGU9DyMcJ0RwJ8GqJAFzMXcXW1Rw==", + "dev": true, + "license": "MIT" + }, + "node_modules/structured-source": { + "version": "4.0.0", + "resolved": "https://registry.npmjs.org/structured-source/-/structured-source-4.0.0.tgz", + "integrity": "sha512-qGzRFNJDjFieQkl/sVOI2dUjHKRyL9dAJi2gCPGJLbJHBIkyOHxjuocpIEfbLioX+qSJpvbYdT49/YCdMznKxA==", + "dev": true, + "license": "BSD-2-Clause", + "dependencies": { + "boundary": "^2.0.0" + } + }, + "node_modules/textextensions": { + "version": "6.11.0", + "resolved": "https://registry.npmjs.org/textextensions/-/textextensions-6.11.0.tgz", + "integrity": "sha512-tXJwSr9355kFJI3lbCkPpUH5cP8/M0GGy2xLO34aZCjMXBaK3SoPnZwr/oWmo1FdCnELcs4npdCIOFtq9W3ruQ==", + "dev": true, + "license": "Artistic-2.0", + "dependencies": { + "editions": "^6.21.0" + }, + "engines": { + "node": ">=4" + }, + "funding": { + "url": "https://bevry.me/fund" + } + }, + "node_modules/tinybench": { + "version": "2.9.0", + "resolved": "https://registry.npmjs.org/tinybench/-/tinybench-2.9.0.tgz", + "integrity": "sha512-0+DUvqWMValLmha6lr4kD8iAMK1HzV0/aKnCtWb9v9641TnP/MFb7Pc2bxoxQjTXAErryXVgUOfv2YqNllqGeg==", + "dev": true, + "license": "MIT" + }, + "node_modules/tinyexec": { + "version": "1.3.1", + "resolved": "https://registry.npmjs.org/tinyexec/-/tinyexec-1.3.1.tgz", + "integrity": "sha512-GCvB3aoys96IuDFBMcTB46JOR6mdMtAToqwiW8JlWhsoh1mhHi/xn9ss/Dg7N555GiJyEt2qzoG/NHCwM6h1EA==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=18" + } + }, + "node_modules/tinyglobby": { + "version": "0.2.17", + "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.17.tgz", + "integrity": "sha512-wXR/dYpcqKmfWpEdZjiKJOwCNFndD0DMnrW/cYjVGttEkBfVgcLFHoNrlj47mjOVic9yyNu65alsgF4NQyTa2g==", + "dev": true, + "license": "MIT", + "dependencies": { + "fdir": "^6.5.0", + "picomatch": "^4.0.4" + }, + "engines": { + "node": ">=12.0.0" + }, + "funding": { + "url": "https://github.com/sponsors/SuperchupuDev" + } + }, + "node_modules/tinyrainbow": { + "version": "3.1.1", + "resolved": "https://registry.npmjs.org/tinyrainbow/-/tinyrainbow-3.1.1.tgz", + "integrity": "sha512-yau8yJdTt989Mm0Bd/236QnzEiPf2xLLTqUZRUJOo/3CB078LSwzei343DgtJVmfJKJE3TMINY1u42SQsP6mXw==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=14.0.0" + } + }, + "node_modules/tslib": { + "version": "2.8.1", + "resolved": "https://registry.npmjs.org/tslib/-/tslib-2.8.1.tgz", + "integrity": "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w==", + "dev": true, + "license": "0BSD" + }, + "node_modules/tunnel": { + "version": "0.0.6", + "resolved": "https://registry.npmjs.org/tunnel/-/tunnel-0.0.6.tgz", + "integrity": "sha512-1h/Lnq9yajKY2PEbBadPXj3VxsDDu844OnaAo52UVmIzIvwwtBPIuNvkjuzBlTWpfJyUbG3ez0KSBibQkj4ojg==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=0.6.11 <=0.7.0 || >=0.7.3" + } + }, + "node_modules/typed-rest-client": { + "version": "1.8.11", + "resolved": "https://registry.npmjs.org/typed-rest-client/-/typed-rest-client-1.8.11.tgz", + "integrity": "sha512-5UvfMpd1oelmUPRbbaVnq+rHP7ng2cE4qoQkQeAqxRL6PklkxsM0g32/HL0yfvruK6ojQ5x8EE+HF4YV6DtuCA==", + "dev": true, + "license": "MIT", + "dependencies": { + "qs": "^6.9.1", + "tunnel": "0.0.6", + "underscore": "^1.12.1" + } + }, + "node_modules/typescript": { + "version": "5.9.3", + "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.9.3.tgz", + "integrity": "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw==", + "dev": true, + "license": "Apache-2.0", + "bin": { + "tsc": "bin/tsc", + "tsserver": "bin/tsserver" + }, + "engines": { + "node": ">=14.17" + } + }, + "node_modules/underscore": { + "version": "1.13.8", + "resolved": "https://registry.npmjs.org/underscore/-/underscore-1.13.8.tgz", + "integrity": "sha512-DXtD3ZtEQzc7M8m4cXotyHR+FAS18C64asBYY5vqZexfYryNNnDc02W4hKg3rdQuqOYas1jkseX0+nZXjTXnvQ==", + "dev": true, + "license": "MIT" + }, + "node_modules/undici-types": { + "version": "6.21.0", + "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-6.21.0.tgz", + "integrity": "sha512-iwDZqg0QAGrg9Rav5H4n0M64c3mkR59cJ6wQp+7C4nI0gsmExaedaYLNO44eT4AtBBwjbTiGPMlt2Md0T9H9JQ==", + "dev": true, + "license": "MIT" + }, + "node_modules/url-join": { + "version": "4.0.1", + "resolved": "https://registry.npmjs.org/url-join/-/url-join-4.0.1.tgz", + "integrity": "sha512-jk1+QP6ZJqyOiuEI9AEWQfju/nB2Pw466kbA0LEZljHwKeMgd9WrAEgEGxjPDD2+TNbbb37rTyhEfrCXfuKXnA==", + "dev": true, + "license": "MIT" + }, + "node_modules/version-range": { + "version": "4.15.0", + "resolved": "https://registry.npmjs.org/version-range/-/version-range-4.15.0.tgz", + "integrity": "sha512-Ck0EJbAGxHwprkzFO966t4/5QkRuzh+/I1RxhLgUKKwEn+Cd8NwM60mE3AqBZg5gYODoXW0EFsQvbZjRlvdqbg==", + "dev": true, + "license": "Artistic-2.0", + "engines": { + "node": ">=4" + }, + "funding": { + "url": "https://bevry.me/fund" + } + }, + "node_modules/vite": { + "version": "8.3.0", + "resolved": "https://registry.npmjs.org/vite/-/vite-8.3.0.tgz", + "integrity": "sha512-lhZBVvEHefgE+HQZC9O7EBJgCU/nVzFNl7vkS4RE0APtWLP02/8QVIkQtzBxPquh7lq5/78NHipTj7ODQ6XuyQ==", + "dev": true, + "license": "MIT", + "dependencies": { + "lightningcss": "^1.33.0", + "picomatch": "^4.0.7", + "postcss": "^8.5.28", + "rolldown": "~1.2.6", + "tinyglobby": "^0.2.17" + }, + "bin": { + "vite": "bin/vite.js" + }, + "engines": { + "node": "^20.19.0 || >=22.12.0" + }, + "funding": { + "url": "https://github.com/vitejs/vite?sponsor=1" + }, + "optionalDependencies": { + "fsevents": "~2.3.3" + }, + "peerDependencies": { + "@types/node": "^20.19.0 || >=22.12.0", + "@vitejs/devtools": "^0.7.1", + "esbuild": "^0.27.0 || ^0.28.0", + "jiti": ">=1.21.0", + "less": "^4.0.0", + "sass": "^1.70.0", + "sass-embedded": "^1.70.0", + "stylus": ">=0.54.8", + "sugarss": "^5.0.0", + "terser": "^5.16.0", + "tsx": "^4.8.1", + "yaml": "^2.4.2" + }, + "peerDependenciesMeta": { + "@types/node": { + "optional": true + }, + "@vitejs/devtools": { + "optional": true + }, + "esbuild": { + "optional": true + }, + "jiti": { + "optional": true + }, + "less": { + "optional": true + }, + "sass": { + "optional": true + }, + "sass-embedded": { + "optional": true + }, + "stylus": { + "optional": true + }, + "sugarss": { + "optional": true + }, + "terser": { + "optional": true + }, + "tsx": { + "optional": true + }, + "yaml": { + "optional": true + } + } + }, + "node_modules/vitest": { + "version": "4.1.11", + "resolved": "https://registry.npmjs.org/vitest/-/vitest-4.1.11.tgz", + "integrity": "sha512-fhACrNXUidIbGSBr5FlbuBkO7VWC1ZyLl0DO4CU2DrQoAPxX84Ysxs+HeGQpii5lZWV1Q4gBZTTu49mF+A6Edw==", + "dev": true, + "license": "MIT", + "dependencies": { + "@vitest/expect": "4.1.11", + "@vitest/mocker": "4.1.11", + "@vitest/pretty-format": "4.1.11", + "@vitest/runner": "4.1.11", + "@vitest/snapshot": "4.1.11", + "@vitest/spy": "4.1.11", + "@vitest/utils": "4.1.11", + "es-module-lexer": "^2.0.0", + "expect-type": "^1.3.0", + "magic-string": "^0.30.21", + "obug": "^2.1.1", + "pathe": "^2.0.3", + "picomatch": "^4.0.3", + "std-env": "^4.0.0-rc.1", + "tinybench": "^2.9.0", + "tinyexec": "^1.0.2", + "tinyglobby": "^0.2.15", + "tinyrainbow": "^3.1.0", + "vite": "^6.0.0 || ^7.0.0 || ^8.0.0", + "why-is-node-running": "^2.3.0" + }, + "bin": { + "vitest": "vitest.mjs" + }, + "engines": { + "node": "^20.0.0 || ^22.0.0 || >=24.0.0" + }, + "funding": { + "url": "https://opencollective.com/vitest" + }, + "peerDependencies": { + "@edge-runtime/vm": "*", + "@opentelemetry/api": "^1.9.0", + "@types/node": "^20.0.0 || ^22.0.0 || >=24.0.0", + "@vitest/browser-playwright": "4.1.11", + "@vitest/browser-preview": "4.1.11", + "@vitest/browser-webdriverio": "4.1.11", + "@vitest/coverage-istanbul": "4.1.11", + "@vitest/coverage-v8": "4.1.11", + "@vitest/ui": "4.1.11", + "happy-dom": "*", + "jsdom": "*", + "vite": "^6.0.0 || ^7.0.0 || ^8.0.0" + }, + "peerDependenciesMeta": { + "@edge-runtime/vm": { + "optional": true + }, + "@opentelemetry/api": { + "optional": true + }, + "@types/node": { + "optional": true + }, + "@vitest/browser-playwright": { + "optional": true + }, + "@vitest/browser-preview": { + "optional": true + }, + "@vitest/browser-webdriverio": { + "optional": true + }, + "@vitest/coverage-istanbul": { + "optional": true + }, + "@vitest/coverage-v8": { + "optional": true + }, + "@vitest/ui": { + "optional": true + }, + "happy-dom": { + "optional": true + }, + "jsdom": { + "optional": true + }, + "vite": { + "optional": false + } + } + }, + "node_modules/why-is-node-running": { + "version": "2.3.0", + "resolved": "https://registry.npmjs.org/why-is-node-running/-/why-is-node-running-2.3.0.tgz", + "integrity": "sha512-hUrmaWBdVDcxvYqnyh09zunKzROWjbZTiNy8dBEjkS7ehEDQibXJ7XvlmtbwuTclUiIyN+CyXQD4Vmko8fNm8w==", + "dev": true, + "license": "MIT", + "dependencies": { + "siginfo": "^2.0.0", + "stackback": "0.0.2" + }, + "bin": { + "why-is-node-running": "cli.js" + }, + "engines": { + "node": ">=8" + } + }, + "node_modules/wsl-utils": { + "version": "0.1.0", + "resolved": "https://registry.npmjs.org/wsl-utils/-/wsl-utils-0.1.0.tgz", + "integrity": "sha512-h3Fbisa2nKGPxCpm89Hk33lBLsnaGBvctQopaBSOW/uIs6FTe1ATyAnKFJrzVs9vpGdsTe73WF3V4lIsk4Gacw==", + "dev": true, + "license": "MIT", + "dependencies": { + "is-wsl": "^3.1.0" + }, + "engines": { + "node": ">=18" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, + "node_modules/xml2js": { + "version": "0.5.0", + "resolved": "https://registry.npmjs.org/xml2js/-/xml2js-0.5.0.tgz", + "integrity": "sha512-drPFnkQJik/O+uPKpqSgr22mpuFHqKdbS835iAQrUC73L2F5WkboIRd63ai/2Yg6I1jzifPFKH2NTK+cfglkIA==", + "dev": true, + "license": "MIT", + "dependencies": { + "sax": ">=0.6.0", + "xmlbuilder": "~11.0.0" + }, + "engines": { + "node": ">=4.0.0" + } + }, + "node_modules/xmlbuilder": { + "version": "11.0.1", + "resolved": "https://registry.npmjs.org/xmlbuilder/-/xmlbuilder-11.0.1.tgz", + "integrity": "sha512-fDlsI/kFEx7gLvbecc0/ohLG50fugQp8ryHzMTuW9vSa1GJ0XYWKnhsUx7oie3G98+r56aTQIUB4kht42R3JvA==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=4.0" + } + }, + "node_modules/yallist": { + "version": "4.0.0", + "resolved": "https://registry.npmjs.org/yallist/-/yallist-4.0.0.tgz", + "integrity": "sha512-3wdGidZyq5PB084XLES5TpOSRA3wjXAlIWMhum2kRcv/41Sn2emQ0dycQW4uZXLejwKvg6EsvbdlVL+FYEct7A==", + "dev": true, + "license": "ISC" + }, + "node_modules/yauzl": { + "version": "3.4.0", + "resolved": "https://registry.npmjs.org/yauzl/-/yauzl-3.4.0.tgz", + "integrity": "sha512-jIH9yLR9wqr0wOS0TpBvo/g/2UgZH5qePVbjgRliiF0BYvOZyaBknKsF+x9Iht0O6sqgnB93rCICdOZFecJuDw==", + "dev": true, + "license": "MIT", + "dependencies": { + "pend": "~1.2.0" + }, + "engines": { + "node": ">=12" + } + }, + "node_modules/yazl": { + "version": "2.5.1", + "resolved": "https://registry.npmjs.org/yazl/-/yazl-2.5.1.tgz", + "integrity": "sha512-phENi2PLiHnHb6QBVot+dJnaAZ0xosj7p3fWl+znIjBDlnMI2PsZCJZ306BPTFOaHf5qdDEI8x5qFrSOBN5vrw==", + "dev": true, + "license": "MIT", + "dependencies": { + "buffer-crc32": "~0.2.3" + } + } + } +} diff --git a/vscode-extension/package.json b/vscode-extension/package.json new file mode 100644 index 00000000000..a8e52b14eda --- /dev/null +++ b/vscode-extension/package.json @@ -0,0 +1,87 @@ +{ + "name": "litellm-vscode", + "displayName": "LiteLLM", + "description": "Chat with every model behind your LiteLLM AI Gateway in VS Code, with live pricing and reasoning effort controls in the model picker", + "version": "0.1.0", + "publisher": "litellm", + "license": "MIT", + "repository": { + "type": "git", + "url": "https://github.com/BerriAI/litellm.git", + "directory": "vscode-extension" + }, + "homepage": "https://docs.litellm.ai", + "bugs": { + "url": "https://github.com/BerriAI/litellm/issues" + }, + "engines": { + "vscode": "^1.109.0" + }, + "categories": [ + "AI", + "Chat" + ], + "keywords": [ + "litellm", + "ai gateway", + "llm", + "chat", + "copilot" + ], + "main": "./dist/extension.js", + "activationEvents": [], + "contributes": { + "languageModelChatProviders": [ + { + "vendor": "litellm", + "displayName": "LiteLLM", + "configuration": { + "type": "object", + "properties": { + "baseUrl": { + "type": "string", + "title": "Gateway URL", + "description": "Base URL of your LiteLLM AI Gateway, for example https://litellm.example.com", + "default": "http://localhost:4000" + }, + "apiKey": { + "type": "string", + "title": "API key", + "description": "A LiteLLM virtual key. The models offered are the ones this key can access", + "secret": true + } + }, + "required": [ + "baseUrl", + "apiKey" + ] + } + } + ], + "commands": [ + { + "command": "litellm.refreshModels", + "title": "Refresh Models", + "category": "LiteLLM" + } + ] + }, + "scripts": { + "build": "esbuild src/extension.ts --bundle --outfile=dist/extension.js --external:vscode --format=cjs --platform=node --target=node22", + "typecheck": "tsc --noEmit", + "test": "vitest run", + "vscode:prepublish": "npm run typecheck && npm run build", + "package": "vsce package --no-dependencies" + }, + "dependencies": { + "openai": "^7.18.0" + }, + "devDependencies": { + "@types/node": "^22.20.3", + "@types/vscode": "^1.109.0", + "@vscode/vsce": "^4.0.0", + "esbuild": "^0.28.2", + "typescript": "^5.9.3", + "vitest": "^4.1.11" + } +} diff --git a/vscode-extension/src/extension.ts b/vscode-extension/src/extension.ts new file mode 100644 index 00000000000..fe0cafb2fbc --- /dev/null +++ b/vscode-extension/src/extension.ts @@ -0,0 +1,17 @@ +import * as vscode from "vscode"; +import { createGatewayClient } from "./gateway"; +import { LiteLLMChatProvider } from "./provider"; + +export const VENDOR = "litellm"; +export const REFRESH_COMMAND = "litellm.refreshModels"; + +export function activate(context: vscode.ExtensionContext): void { + const provider = new LiteLLMChatProvider(createGatewayClient()); + context.subscriptions.push( + provider, + vscode.lm.registerLanguageModelChatProvider(VENDOR, provider), + vscode.commands.registerCommand(REFRESH_COMMAND, () => provider.refresh()), + ); +} + +export function deactivate(): void {} diff --git a/vscode-extension/src/gateway.ts b/vscode-extension/src/gateway.ts new file mode 100644 index 00000000000..0ffe5e0f3f3 --- /dev/null +++ b/vscode-extension/src/gateway.ts @@ -0,0 +1,114 @@ +import OpenAI from "openai"; +import type { ChatCompletionChunk, ChatCompletionCreateParamsStreaming } from "openai/resources/chat/completions"; +import packageJson from "../package.json"; +import { parseModelGroups, type ConfigurationValues, type ModelGroupInfo } from "./models"; + +export interface GatewayConfig { + readonly baseUrl: string; + readonly apiKey: string; +} + +export type GatewayConfigResult = + | { readonly kind: "ok"; readonly config: GatewayConfig } + | { readonly kind: "unconfigured" } + | { readonly kind: "missing_fields"; readonly fields: readonly string[] } + | { readonly kind: "invalid_url"; readonly baseUrl: string }; + +export type ModelGroupsResult = + | { readonly kind: "ok"; readonly groups: readonly ModelGroupInfo[] } + | { readonly kind: "http_error"; readonly status: number; readonly body: string } + | { readonly kind: "invalid_response"; readonly reason: string }; + +export interface GatewayClient { + listModelGroups(config: GatewayConfig, signal: AbortSignal): Promise; + streamChatCompletion( + config: GatewayConfig, + params: ChatCompletionCreateParamsStreaming, + signal: AbortSignal, + ): Promise>; +} + +export const USER_AGENT = `litellm-vscode/${packageJson.version}`; +export const ERROR_SUMMARY_LIMIT = 200; + +const GATEWAY_PROTOCOLS: ReadonlySet = new Set(["http:", "https:"]); + +const parsesAsHttpUrl = (value: string): boolean => { + try { + return GATEWAY_PROTOCOLS.has(new URL(value).protocol); + } catch { + return false; + } +}; + +export const gatewayRoot = (baseUrl: string): string | undefined => { + const root = baseUrl.trim().replace(/\/+$/, "").replace(/\/v1$/, ""); + return parsesAsHttpUrl(root) ? root : undefined; +}; + +const nonEmptyString = (value: unknown): string | undefined => + typeof value === "string" && value.trim() !== "" ? value.trim() : undefined; + +export const gatewayConfigFrom = (configuration: ConfigurationValues | undefined): GatewayConfigResult => { + if (configuration === undefined) { + return { kind: "unconfigured" }; + } + const baseUrl = nonEmptyString(configuration.baseUrl); + const apiKey = nonEmptyString(configuration.apiKey); + if (baseUrl === undefined || apiKey === undefined) { + const fields = [...(baseUrl === undefined ? ["Gateway URL"] : []), ...(apiKey === undefined ? ["API key"] : [])]; + return { kind: "missing_fields", fields }; + } + const root = gatewayRoot(baseUrl); + return root === undefined ? { kind: "invalid_url", baseUrl } : { kind: "ok", config: { baseUrl: root, apiKey } }; +}; + +export const modelGroupInfoUrl = (root: string): string => `${root}/model_group/info`; + +export const openAiBaseUrl = (root: string): string => `${root}/v1`; + +const isRecord = (value: unknown): value is Record => typeof value === "object" && value !== null; + +const errorMessageIn = (body: string): string | undefined => { + try { + const parsed: unknown = JSON.parse(body); + if (!isRecord(parsed)) { + return undefined; + } + if (isRecord(parsed.error) && typeof parsed.error.message === "string") { + return parsed.error.message; + } + return typeof parsed.detail === "string" ? parsed.detail : undefined; + } catch { + return undefined; + } +}; + +export const summarizeErrorBody = (body: string): string => { + const message = (errorMessageIn(body) ?? body).replace(/\s+/g, " ").trim(); + return message.length > ERROR_SUMMARY_LIMIT ? `${message.slice(0, ERROR_SUMMARY_LIMIT)}...` : message; +}; + +export const createGatewayClient = (fetchImpl: typeof fetch = fetch): GatewayClient => ({ + async listModelGroups(config, signal) { + const response = await fetchImpl(modelGroupInfoUrl(config.baseUrl), { + headers: { Authorization: `Bearer ${config.apiKey}`, "User-Agent": USER_AGENT }, + signal, + }); + if (!response.ok) { + return { kind: "http_error", status: response.status, body: await response.text() }; + } + const parsed = parseModelGroups(await response.json()); + return parsed.kind === "ok" ? parsed : { kind: "invalid_response", reason: parsed.reason }; + }, + streamChatCompletion(config, params, signal) { + const client = new OpenAI({ + apiKey: config.apiKey, + baseURL: openAiBaseUrl(config.baseUrl), + defaultHeaders: { "User-Agent": USER_AGENT }, + fetch: fetchImpl, + maxRetries: 0, + }); + return client.chat.completions.create(params, { signal }); + }, +}); diff --git a/vscode-extension/src/messages.ts b/vscode-extension/src/messages.ts new file mode 100644 index 00000000000..855f85435ad --- /dev/null +++ b/vscode-extension/src/messages.ts @@ -0,0 +1,188 @@ +import type * as vscode from "vscode"; +import type { + ChatCompletionAssistantMessageParam, + ChatCompletionContentPart, + ChatCompletionCreateParamsStreaming, + ChatCompletionMessageParam, + ChatCompletionMessageToolCall, + ChatCompletionTool, + ChatCompletionToolMessageParam, +} from "openai/resources/chat/completions"; +import { estimateTokens } from "./models"; + +export interface ChatRequestInput { + readonly model: string; + readonly messages: readonly vscode.LanguageModelChatRequestMessage[]; + readonly tools: readonly vscode.LanguageModelChatTool[]; + readonly requireToolCall: boolean; + readonly reasoningEffort: string | undefined; + readonly modelOptions: { readonly [key: string]: unknown }; +} + +interface TextPart { + readonly value: string; +} + +interface ToolCallPart { + readonly callId: string; + readonly name: string; + readonly input: object; +} + +interface ToolResultPart { + readonly callId: string; + readonly content: ReadonlyArray; +} + +interface DataPart { + readonly mimeType: string; + readonly data: Uint8Array; +} + +const USER_ROLE = 1; +const ASSISTANT_ROLE = 2; +const SYSTEM_ROLE = 3; + +export const ESTIMATED_TOKENS_PER_IMAGE = 1000; + +const isRecord = (value: unknown): value is Record => typeof value === "object" && value !== null; + +const isTextPart = (part: unknown): part is TextPart => isRecord(part) && typeof part.value === "string"; + +const isToolCallPart = (part: unknown): part is ToolCallPart => + isRecord(part) && typeof part.callId === "string" && typeof part.name === "string" && isRecord(part.input); + +const isToolResultPart = (part: unknown): part is ToolResultPart => + isRecord(part) && typeof part.callId === "string" && Array.isArray(part.content); + +const isDataPart = (part: unknown): part is DataPart => + isRecord(part) && typeof part.mimeType === "string" && part.data instanceof Uint8Array; + +const isImagePart = (part: unknown): part is DataPart => isDataPart(part) && part.mimeType.startsWith("image/"); + +const dataUrl = (part: DataPart): string => `data:${part.mimeType};base64,${Buffer.from(part.data).toString("base64")}`; + +const textOf = (part: unknown): string => { + if (isTextPart(part)) { + return part.value; + } + if (isDataPart(part) && part.mimeType.startsWith("text/")) { + return Buffer.from(part.data).toString("utf8"); + } + if (isRecord(part) && "value" in part) { + return JSON.stringify(part.value); + } + return ""; +}; + +const contentParts = (parts: readonly unknown[]): readonly ChatCompletionContentPart[] => + parts.flatMap((part): readonly ChatCompletionContentPart[] => { + if (isImagePart(part)) { + return [{ type: "image_url", image_url: { url: dataUrl(part) } }]; + } + const text = textOf(part); + return text === "" ? [] : [{ type: "text", text }]; + }); + +const toolMessage = (part: ToolResultPart): ChatCompletionToolMessageParam => ({ + role: "tool", + tool_call_id: part.callId, + content: part.content.filter((item) => !isImagePart(item)).map(textOf).join(""), +}); + +const userMessages = (parts: readonly unknown[]): readonly ChatCompletionMessageParam[] => { + const toolResults = parts.filter(isToolResultPart); + const toolResultImages = toolResults.flatMap((result) => result.content.filter(isImagePart)); + const remaining = parts.filter((part) => !isToolResultPart(part)); + const userContent = contentParts([...remaining, ...toolResultImages]); + const userMessage: readonly ChatCompletionMessageParam[] = + userContent.length === 0 ? [] : [{ role: "user", content: [...userContent] }]; + return [...toolResults.map(toolMessage), ...userMessage]; +}; + +const toolCall = (part: ToolCallPart): ChatCompletionMessageToolCall => ({ + id: part.callId, + type: "function", + function: { name: part.name, arguments: JSON.stringify(part.input) }, +}); + +const assistantMessages = (parts: readonly unknown[]): readonly ChatCompletionAssistantMessageParam[] => { + const text = parts.filter(isTextPart).map((part) => part.value).join(""); + const toolCalls = parts.filter(isToolCallPart).map(toolCall); + if (text === "" && toolCalls.length === 0) { + return []; + } + return [ + { + role: "assistant", + content: text === "" ? null : text, + ...(toolCalls.length === 0 ? {} : { tool_calls: toolCalls }), + }, + ]; +}; + +const convertMessage = (message: vscode.LanguageModelChatRequestMessage): readonly ChatCompletionMessageParam[] => { + const role: number = message.role; + switch (role) { + case USER_ROLE: + return userMessages(message.content); + case ASSISTANT_ROLE: + return assistantMessages(message.content); + case SYSTEM_ROLE: + return [{ role: "system", content: message.content.map(textOf).join("") }]; + default: + return []; + } +}; + +export const toChatCompletionMessages = ( + messages: readonly vscode.LanguageModelChatRequestMessage[], +): readonly ChatCompletionMessageParam[] => messages.flatMap(convertMessage); + +const imagePartsIn = (parts: readonly unknown[]): readonly DataPart[] => [ + ...parts.filter(isImagePart), + ...parts.filter(isToolResultPart).flatMap((result) => result.content.filter(isImagePart)), +]; + +const withoutImageData = (key: string, value: unknown): unknown => (key === "image_url" ? undefined : value); + +export const estimateMessageTokens = (message: vscode.LanguageModelChatRequestMessage): number => { + const converted = convertMessage(message); + if (converted.length === 0) { + return 0; + } + const images = imagePartsIn(message.content).length; + return estimateTokens(JSON.stringify(converted, withoutImageData)) + images * ESTIMATED_TOKENS_PER_IMAGE; +}; + +const toTool = (tool: vscode.LanguageModelChatTool): ChatCompletionTool => ({ + type: "function", + function: { + name: tool.name, + description: tool.description, + ...(tool.inputSchema === undefined ? {} : { parameters: tool.inputSchema as Record }), + }, +}); + +const NUMERIC_OPTIONS = ["temperature", "top_p", "max_tokens", "presence_penalty", "frequency_penalty", "seed"] as const; + +const forwardedModelOptions = (modelOptions: { readonly [key: string]: unknown }): Record => + Object.fromEntries( + NUMERIC_OPTIONS.flatMap((key) => { + const value = modelOptions[key]; + return typeof value === "number" ? [[key, value] as const] : []; + }), + ); + +export const buildChatCompletionParams = (input: ChatRequestInput): ChatCompletionCreateParamsStreaming => ({ + model: input.model, + messages: [...toChatCompletionMessages(input.messages)], + stream: true, + stream_options: { include_usage: true }, + ...forwardedModelOptions(input.modelOptions), + ...(input.tools.length === 0 ? {} : { tools: input.tools.map(toTool) }), + ...(input.tools.length === 0 || !input.requireToolCall ? {} : { tool_choice: "required" }), + ...(input.reasoningEffort === undefined + ? {} + : { reasoning_effort: input.reasoningEffort as ChatCompletionCreateParamsStreaming["reasoning_effort"] }), +}); diff --git a/vscode-extension/src/models.ts b/vscode-extension/src/models.ts new file mode 100644 index 00000000000..d76f6557b8b --- /dev/null +++ b/vscode-extension/src/models.ts @@ -0,0 +1,170 @@ +export interface ModelGroupInfo { + readonly modelGroup: string; + readonly providers: readonly string[]; + readonly mode: string | undefined; + readonly maxInputTokens: number | undefined; + readonly maxOutputTokens: number | undefined; + readonly inputCostPerToken: number | undefined; + readonly outputCostPerToken: number | undefined; + readonly supportsVision: boolean; + readonly supportsFunctionCalling: boolean; + readonly supportedReasoningEfforts: readonly string[]; +} + +export type ConfigurationValues = { readonly [key: string]: unknown }; + +export interface ConfigurationSchemaProperty { + readonly type: "string"; + readonly title: string; + readonly enum: readonly string[]; + readonly enumItemLabels: readonly string[]; + readonly default: string; + readonly group: "navigation"; +} + +export interface ConfigurationSchema { + readonly properties: { readonly [key: string]: ConfigurationSchemaProperty }; +} + +export interface ModelDescriptor { + readonly id: string; + readonly name: string; + readonly family: string; + readonly version: string; + readonly detail: string; + readonly tooltip: string; + readonly maxInputTokens: number; + readonly maxOutputTokens: number; + readonly imageInput: boolean; + readonly toolCalling: boolean; + readonly configurationSchema: ConfigurationSchema | undefined; +} + +export type ModelGroupsParseResult = + | { readonly kind: "ok"; readonly groups: readonly ModelGroupInfo[] } + | { readonly kind: "invalid"; readonly reason: string }; + +export const REASONING_EFFORT_KEY = "reasoningEffort"; +export const GATEWAY_DEFAULT_EFFORT = "default"; +export const ASSUMED_MAX_INPUT_TOKENS = 128000; +export const ASSUMED_MAX_OUTPUT_TOKENS = 4096; +export const MARKDOWN_LINE_BREAK = " \n"; + +const isRecord = (value: unknown): value is Record => typeof value === "object" && value !== null; + +const optionalNumber = (value: unknown): number | undefined => + typeof value === "number" && Number.isFinite(value) ? value : undefined; + +const optionalString = (value: unknown): string | undefined => (typeof value === "string" ? value : undefined); + +const stringList = (value: unknown): readonly string[] => + Array.isArray(value) ? value.filter((item): item is string => typeof item === "string") : []; + +const parseGroup = (value: unknown): ModelGroupInfo | undefined => { + if (!isRecord(value) || typeof value.model_group !== "string") { + return undefined; + } + return { + modelGroup: value.model_group, + providers: stringList(value.providers), + mode: optionalString(value.mode), + maxInputTokens: optionalNumber(value.max_input_tokens), + maxOutputTokens: optionalNumber(value.max_output_tokens), + inputCostPerToken: optionalNumber(value.input_cost_per_token), + outputCostPerToken: optionalNumber(value.output_cost_per_token), + supportsVision: value.supports_vision === true, + supportsFunctionCalling: value.supports_function_calling === true, + supportedReasoningEfforts: stringList(value.supported_reasoning_efforts), + }; +}; + +export const parseModelGroups = (body: unknown): ModelGroupsParseResult => { + if (!isRecord(body) || !Array.isArray(body.data)) { + return { kind: "invalid", reason: "response has no data array" }; + } + const groups = body.data.map(parseGroup).filter((group): group is ModelGroupInfo => group !== undefined); + return { kind: "ok", groups }; +}; + +const isChatGroup = (group: ModelGroupInfo): boolean => group.mode === undefined || group.mode === "chat"; + +export const formatUsdPerMillionTokens = (costPerToken: number): string => { + const perMillion = costPerToken * 1_000_000; + const digits = perMillion === 0 || perMillion >= 0.01 ? perMillion.toFixed(2) : perMillion.toPrecision(2); + return `$${digits}`; +}; + +const priceLine = (label: string, costPerToken: number | undefined): string => + costPerToken === undefined ? `${label}: no price configured` : `${label}: ${formatUsdPerMillionTokens(costPerToken)} per 1M tokens`; + +const pricingDetail = (group: ModelGroupInfo): string => { + if (group.inputCostPerToken === undefined && group.outputCostPerToken === undefined) { + return "No pricing configured"; + } + const input = group.inputCostPerToken === undefined ? "n/a" : formatUsdPerMillionTokens(group.inputCostPerToken); + const output = group.outputCostPerToken === undefined ? "n/a" : formatUsdPerMillionTokens(group.outputCostPerToken); + return `${input} in / ${output} out per 1M tokens`; +}; + +const capitalize = (value: string): string => value.charAt(0).toUpperCase() + value.slice(1); + +const effortSchema = (efforts: readonly string[]): ConfigurationSchema | undefined => { + if (efforts.length === 0) { + return undefined; + } + return { + properties: { + [REASONING_EFFORT_KEY]: { + type: "string", + title: "Reasoning Effort", + enum: [GATEWAY_DEFAULT_EFFORT, ...efforts], + enumItemLabels: ["Gateway default", ...efforts.map(capitalize)], + default: GATEWAY_DEFAULT_EFFORT, + group: "navigation", + }, + }, + }; +}; + +const tooltipFor = (group: ModelGroupInfo): string => { + const providers = group.providers.length === 0 ? "" : ` via ${group.providers.join(", ")}`; + const context = + group.maxInputTokens === undefined || group.maxOutputTokens === undefined + ? `Context: unknown, assuming ${ASSUMED_MAX_INPUT_TOKENS} in / ${ASSUMED_MAX_OUTPUT_TOKENS} out tokens` + : `Context: ${group.maxInputTokens} in / ${group.maxOutputTokens} out tokens`; + const efforts = + group.supportedReasoningEfforts.length === 0 + ? "Reasoning effort: not configurable" + : `Reasoning effort: ${group.supportedReasoningEfforts.join(", ")}`; + return [ + `LiteLLM model group ${group.modelGroup}${providers}`, + priceLine("Input", group.inputCostPerToken), + priceLine("Output", group.outputCostPerToken), + context, + efforts, + ].join(MARKDOWN_LINE_BREAK); +}; + +const describeGroup = (group: ModelGroupInfo): ModelDescriptor => ({ + id: group.modelGroup, + name: group.modelGroup, + family: group.modelGroup, + version: "1.0", + detail: pricingDetail(group), + tooltip: tooltipFor(group), + maxInputTokens: group.maxInputTokens ?? ASSUMED_MAX_INPUT_TOKENS, + maxOutputTokens: group.maxOutputTokens ?? ASSUMED_MAX_OUTPUT_TOKENS, + imageInput: group.supportsVision, + toolCalling: group.supportsFunctionCalling, + configurationSchema: effortSchema(group.supportedReasoningEfforts), +}); + +export const describeModels = (groups: readonly ModelGroupInfo[]): readonly ModelDescriptor[] => + groups.filter(isChatGroup).map(describeGroup); + +export const reasoningEffortFrom = (configuration: ConfigurationValues | undefined): string | undefined => { + const effort = configuration?.[REASONING_EFFORT_KEY]; + return typeof effort === "string" && effort !== GATEWAY_DEFAULT_EFFORT ? effort : undefined; +}; + +export const estimateTokens = (text: string): number => Math.ceil(text.length / 4); diff --git a/vscode-extension/src/provider.ts b/vscode-extension/src/provider.ts new file mode 100644 index 00000000000..3cafd449147 --- /dev/null +++ b/vscode-extension/src/provider.ts @@ -0,0 +1,136 @@ +import * as vscode from "vscode"; +import { gatewayConfigFrom, summarizeErrorBody, type GatewayClient, type GatewayConfig, type GatewayConfigResult, type ModelGroupsResult } from "./gateway"; +import { buildChatCompletionParams, estimateMessageTokens } from "./messages"; +import { describeModels, estimateTokens, reasoningEffortFrom, type ModelDescriptor } from "./models"; +import { responseParts, type ResponsePart } from "./stream"; + +export interface LiteLLMModel extends vscode.LanguageModelChatInformation { + readonly gateway: GatewayConfig; +} + +export const TRUNCATED_MESSAGE = "The model stopped at its output token limit before finishing the response"; + +const RECONFIGURE_HINT = + 'Fix it from the gear on its row in Manage Language Models: "Update API Key" for the key, "Open in Language Models (JSON)" for the URL'; + +const configurationProblem = (result: Exclude): string => { + switch (result.kind) { + case "missing_fields": + return `LiteLLM provider is missing its ${result.fields.join(" and ")}. ${RECONFIGURE_HINT}`; + case "invalid_url": + return `LiteLLM gateway URL "${result.baseUrl}" is not an http or https URL. ${RECONFIGURE_HINT}`; + } +}; + +const discoveryFailure = (result: Exclude, baseUrl: string): string => { + switch (result.kind) { + case "http_error": + return `LiteLLM gateway at ${baseUrl} answered ${result.status} for /model_group/info: ${summarizeErrorBody(result.body)}`; + case "invalid_response": + return `LiteLLM gateway at ${baseUrl} returned an unexpected /model_group/info payload: ${result.reason}`; + } +}; + +const toModel = (descriptor: ModelDescriptor, gateway: GatewayConfig): LiteLLMModel => ({ + id: descriptor.id, + name: descriptor.name, + family: descriptor.family, + version: descriptor.version, + detail: descriptor.detail, + tooltip: descriptor.tooltip, + maxInputTokens: descriptor.maxInputTokens, + maxOutputTokens: descriptor.maxOutputTokens, + capabilities: { imageInput: descriptor.imageInput, toolCalling: descriptor.toolCalling }, + ...(descriptor.configurationSchema === undefined ? {} : { configurationSchema: descriptor.configurationSchema }), + gateway, +}); + +const toVscodePart = (part: ResponsePart): vscode.LanguageModelResponsePart => { + switch (part.kind) { + case "text": + return new vscode.LanguageModelTextPart(part.value); + case "tool_call": + return new vscode.LanguageModelToolCallPart(part.callId, part.name, part.input); + case "invalid_tool_call": + throw new Error(`Model returned invalid JSON arguments for tool ${part.name}: ${part.arguments}`); + case "truncated": + throw new Error(TRUNCATED_MESSAGE); + } +}; + +const withAbortSignal = async (token: vscode.CancellationToken, run: (signal: AbortSignal) => Promise): Promise => { + const controller = new AbortController(); + const subscription = token.onCancellationRequested(() => controller.abort()); + try { + return await run(controller.signal); + } finally { + subscription.dispose(); + } +}; + +export class LiteLLMChatProvider implements vscode.LanguageModelChatProvider, vscode.Disposable { + private readonly changeEmitter = new vscode.EventEmitter(); + readonly onDidChangeLanguageModelChatInformation = this.changeEmitter.event; + + constructor(private readonly gateway: GatewayClient) {} + + refresh(): void { + this.changeEmitter.fire(); + } + + dispose(): void { + this.changeEmitter.dispose(); + } + + async provideLanguageModelChatInformation( + options: vscode.PrepareLanguageModelChatModelOptions, + token: vscode.CancellationToken, + ): Promise { + const configured = gatewayConfigFrom(options.configuration); + if (configured.kind === "unconfigured") { + return []; + } + if (configured.kind !== "ok") { + throw new Error(configurationProblem(configured)); + } + const result = await withAbortSignal(token, (signal) => this.gateway.listModelGroups(configured.config, signal)); + if (result.kind !== "ok") { + throw new Error(discoveryFailure(result, configured.config.baseUrl)); + } + return describeModels(result.groups).map((descriptor) => toModel(descriptor, configured.config)); + } + + async provideLanguageModelChatResponse( + model: LiteLLMModel, + messages: readonly vscode.LanguageModelChatRequestMessage[], + options: vscode.ProvideLanguageModelChatResponseOptions, + progress: vscode.Progress, + token: vscode.CancellationToken, + ): Promise { + const params = buildChatCompletionParams({ + model: model.id, + messages, + tools: options.tools ?? [], + requireToolCall: options.toolMode === vscode.LanguageModelChatToolMode.Required, + reasoningEffort: reasoningEffortFrom(options.modelConfiguration), + modelOptions: options.modelOptions ?? {}, + }); + try { + await withAbortSignal(token, async (signal) => { + const chunks = await this.gateway.streamChatCompletion(model.gateway, params, signal); + for await (const part of responseParts(chunks)) { + progress.report(toVscodePart(part)); + } + }); + } catch (error) { + if (token.isCancellationRequested) { + return; + } + throw error; + } + } + + async provideTokenCount(_model: LiteLLMModel, text: string | vscode.LanguageModelChatRequestMessage): Promise { + return typeof text === "string" ? estimateTokens(text) : estimateMessageTokens(text); + } +} diff --git a/vscode-extension/src/stream.ts b/vscode-extension/src/stream.ts new file mode 100644 index 00000000000..f1677262bbf --- /dev/null +++ b/vscode-extension/src/stream.ts @@ -0,0 +1,90 @@ +import type { ChatCompletionChunk } from "openai/resources/chat/completions"; + +export type ResponsePart = + | { readonly kind: "text"; readonly value: string } + | { readonly kind: "tool_call"; readonly callId: string; readonly name: string; readonly input: object } + | { readonly kind: "invalid_tool_call"; readonly callId: string; readonly name: string; readonly arguments: string } + | { readonly kind: "truncated" }; + +interface PendingToolCall { + readonly index: number; + readonly callId: string; + readonly name: string; + readonly arguments: string; +} + +export type PendingToolCalls = readonly PendingToolCall[]; + +export interface ChunkOutcome { + readonly pending: PendingToolCalls; + readonly parts: readonly ResponsePart[]; +} + +export const NO_PENDING_TOOL_CALLS: PendingToolCalls = []; + +type ToolCallDelta = NonNullable[number]; + +const nonEmpty = (value: string | undefined): string | undefined => (value === undefined || value === "" ? undefined : value); + +const targetOf = (pending: PendingToolCalls, delta: ToolCallDelta): PendingToolCall | undefined => { + const id = nonEmpty(delta.id); + if (id !== undefined) { + return pending.find((call) => call.callId === id); + } + const sameIndex = pending.filter((call) => call.index === delta.index); + return sameIndex.at(-1) ?? (delta.index === undefined ? pending.at(-1) : undefined); +}; + +const mergeToolCallDelta = (pending: PendingToolCalls, delta: ToolCallDelta): PendingToolCalls => { + const target = targetOf(pending, delta); + const base: PendingToolCall = target ?? { index: delta.index ?? pending.length, callId: nonEmpty(delta.id) ?? "", name: "", arguments: "" }; + const merged: PendingToolCall = { + ...base, + name: nonEmpty(delta.function?.name) ?? base.name, + arguments: base.arguments + (delta.function?.arguments ?? ""), + }; + return target === undefined ? [...pending, merged] : pending.map((call) => (call === target ? merged : call)); +}; + +export const applyChunk = (pending: PendingToolCalls, chunk: ChatCompletionChunk): ChunkOutcome => { + const choice = chunk.choices[0]; + if (choice === undefined) { + return { pending, parts: [] }; + } + const text = typeof choice.delta.content === "string" && choice.delta.content !== "" ? [{ kind: "text", value: choice.delta.content } as const] : []; + const truncated = choice.finish_reason === "length" ? [{ kind: "truncated" } as const] : []; + const nextPending = (choice.delta.tool_calls ?? []).reduce(mergeToolCallDelta, pending); + return { pending: nextPending, parts: [...text, ...truncated] }; +}; + +const parseArguments = (raw: string): object | undefined => { + if (raw.trim() === "") { + return {}; + } + try { + const parsed: unknown = JSON.parse(raw); + return typeof parsed === "object" && parsed !== null ? parsed : undefined; + } catch { + return undefined; + } +}; + +const finishToolCall = (call: PendingToolCall): ResponsePart => { + const input = parseArguments(call.arguments); + return input === undefined + ? { kind: "invalid_tool_call", callId: call.callId, name: call.name, arguments: call.arguments } + : { kind: "tool_call", callId: call.callId, name: call.name, input }; +}; + +export const flushToolCalls = (pending: PendingToolCalls): readonly ResponsePart[] => + [...pending].sort((left, right) => left.index - right.index).map(finishToolCall); + +export async function* responseParts(chunks: AsyncIterable): AsyncGenerator { + let pending: PendingToolCalls = NO_PENDING_TOOL_CALLS; + for await (const chunk of chunks) { + const outcome = applyChunk(pending, chunk); + pending = outcome.pending; + yield* outcome.parts; + } + yield* flushToolCalls(pending); +} diff --git a/vscode-extension/src/vscode.proposed.d.ts b/vscode-extension/src/vscode.proposed.d.ts new file mode 100644 index 00000000000..6665aaa38ef --- /dev/null +++ b/vscode-extension/src/vscode.proposed.d.ts @@ -0,0 +1,15 @@ +import type { ConfigurationSchema, ConfigurationValues } from "./models"; + +declare module "vscode" { + interface LanguageModelChatInformation { + readonly configurationSchema?: ConfigurationSchema; + } + + interface PrepareLanguageModelChatModelOptions { + readonly configuration?: ConfigurationValues; + } + + interface ProvideLanguageModelChatResponseOptions { + readonly modelConfiguration?: ConfigurationValues; + } +} diff --git a/vscode-extension/test/gateway.test.ts b/vscode-extension/test/gateway.test.ts new file mode 100644 index 00000000000..3a70a68b377 --- /dev/null +++ b/vscode-extension/test/gateway.test.ts @@ -0,0 +1,205 @@ +import { createServer, type IncomingMessage, type Server, type ServerResponse } from "node:http"; +import type { AddressInfo } from "node:net"; +import { afterEach, describe, expect, it } from "vitest"; +import { + ERROR_SUMMARY_LIMIT, + createGatewayClient, + gatewayConfigFrom, + gatewayRoot, + modelGroupInfoUrl, + openAiBaseUrl, + summarizeErrorBody, + USER_AGENT, + type GatewayConfig, +} from "../src/gateway"; +import { buildChatCompletionParams } from "../src/messages"; + +interface RecordedRequest { + readonly method: string | undefined; + readonly url: string | undefined; + readonly authorization: string | undefined; + readonly userAgent: string | undefined; + readonly body: string; +} + +type Handler = (request: RecordedRequest, response: ServerResponse) => void; + +const readBody = (request: IncomingMessage): Promise => + new Promise((resolve) => { + const chunks: Buffer[] = []; + request.on("data", (chunk: Buffer) => chunks.push(chunk)); + request.on("end", () => resolve(Buffer.concat(chunks).toString("utf8"))); + }); + +const servers: Server[] = []; + +const startGateway = (handler: Handler): Promise<{ readonly url: string; readonly requests: readonly RecordedRequest[] }> => + new Promise((resolve) => { + const requests: RecordedRequest[] = []; + const server = createServer(async (request, response) => { + const recorded: RecordedRequest = { + method: request.method, + url: request.url, + authorization: request.headers.authorization, + userAgent: request.headers["user-agent"], + body: await readBody(request), + }; + requests.push(recorded); + handler(recorded, response); + }); + servers.push(server); + server.listen(0, "127.0.0.1", () => { + const { port } = server.address() as AddressInfo; + resolve({ url: `http://127.0.0.1:${port}`, requests }); + }); + }); + +afterEach(() => { + servers.splice(0).forEach((server) => server.close()); +}); + +const configFor = (baseUrl: string, apiKey: string): GatewayConfig => { + const result = gatewayConfigFrom({ baseUrl, apiKey }); + if (result.kind !== "ok") { + throw new Error(result.kind); + } + return result.config; +}; + +const sse = (response: ServerResponse, events: readonly object[]): void => { + response.writeHead(200, { "content-type": "text/event-stream" }); + events.forEach((event) => response.write(`data: ${JSON.stringify(event)}\n\n`)); + response.end("data: [DONE]\n\n"); +}; + +describe("gateway URLs", () => { + it("accepts the gateway root with or without a trailing slash or /v1", () => { + expect(gatewayRoot("https://litellm.example.com/")).toBe("https://litellm.example.com"); + expect(gatewayRoot("https://litellm.example.com/v1")).toBe("https://litellm.example.com"); + expect(gatewayRoot(" http://localhost:4000 ")).toBe("http://localhost:4000"); + expect(modelGroupInfoUrl("https://litellm.example.com")).toBe("https://litellm.example.com/model_group/info"); + expect(openAiBaseUrl("https://litellm.example.com")).toBe("https://litellm.example.com/v1"); + }); + + it("rejects anything that is not an http or https URL", () => { + expect(gatewayRoot("litellm.example.com")).toBeUndefined(); + expect(gatewayRoot("ftp://litellm.example.com")).toBeUndefined(); + expect(gatewayRoot("")).toBeUndefined(); + }); +}); + +describe("gatewayConfigFrom", () => { + it("distinguishes the unconfigured probe, a lost secret, and a bad URL from a usable configuration", () => { + expect(gatewayConfigFrom(undefined)).toEqual({ kind: "unconfigured" }); + expect(gatewayConfigFrom({ baseUrl: "http://localhost:4000", apiKey: undefined })).toEqual({ kind: "missing_fields", fields: ["API key"] }); + expect(gatewayConfigFrom({ baseUrl: " ", apiKey: "" })).toEqual({ kind: "missing_fields", fields: ["Gateway URL", "API key"] }); + expect(gatewayConfigFrom({ baseUrl: "localhost:4000", apiKey: "sk" })).toEqual({ kind: "invalid_url", baseUrl: "localhost:4000" }); + expect(gatewayConfigFrom({ baseUrl: " http://localhost:4000/v1/ ", apiKey: " sk-test " })).toEqual({ + kind: "ok", + config: { baseUrl: "http://localhost:4000", apiKey: "sk-test" }, + }); + }); +}); + +describe("summarizeErrorBody", () => { + it("prefers the gateway's error message and caps the length", () => { + expect(summarizeErrorBody('{"error":{"message":"invalid key","type":"auth_error","param":"sk-...abcd"}}')).toBe("invalid key"); + expect(summarizeErrorBody('{"detail":"Not Found"}')).toBe("Not Found"); + expect(summarizeErrorBody("\n 502 Bad Gateway\n")).toBe(" 502 Bad Gateway "); + const long = summarizeErrorBody("x".repeat(ERROR_SUMMARY_LIMIT + 50)); + expect(long).toBe(`${"x".repeat(ERROR_SUMMARY_LIMIT)}...`); + }); +}); + +describe("listModelGroups", () => { + it("calls /model_group/info with the virtual key and this extension's user agent", async () => { + const gateway = await startGateway((_request, response) => { + response.writeHead(200, { "content-type": "application/json" }); + response.end(JSON.stringify({ data: [{ model_group: "gpt-5.6", mode: "chat", input_cost_per_token: 4e-6 }] })); + }); + const result = await createGatewayClient().listModelGroups(configFor(`${gateway.url}/v1`, "sk-test"), new AbortController().signal); + expect(result).toEqual({ + kind: "ok", + groups: [expect.objectContaining({ modelGroup: "gpt-5.6", inputCostPerToken: 4e-6 })], + }); + expect(gateway.requests).toEqual([ + expect.objectContaining({ method: "GET", url: "/model_group/info", authorization: "Bearer sk-test", userAgent: USER_AGENT }), + ]); + }); + + it("reports the gateway's status and body when the key is rejected", async () => { + const gateway = await startGateway((_request, response) => { + response.writeHead(401, { "content-type": "application/json" }); + response.end('{"error":{"message":"invalid key"}}'); + }); + expect(await createGatewayClient().listModelGroups({ baseUrl: gateway.url, apiKey: "sk-bad" }, new AbortController().signal)).toEqual({ + kind: "http_error", + status: 401, + body: '{"error":{"message":"invalid key"}}', + }); + }); + + it("reports a payload that is not a model group listing", async () => { + const gateway = await startGateway((_request, response) => { + response.writeHead(200, { "content-type": "application/json" }); + response.end('{"object":"list","models":[]}'); + }); + expect(await createGatewayClient().listModelGroups({ baseUrl: gateway.url, apiKey: "sk" }, new AbortController().signal)).toEqual({ + kind: "invalid_response", + reason: "response has no data array", + }); + }); +}); + +describe("streamChatCompletion", () => { + it("streams /v1/chat/completions through the gateway with the chosen reasoning effort", async () => { + const gateway = await startGateway((_request, response) => + sse(response, [ + { id: "c", object: "chat.completion.chunk", created: 0, model: "gpt-5.6", choices: [{ index: 0, delta: { content: "Hi" }, finish_reason: null }] }, + { id: "c", object: "chat.completion.chunk", created: 0, model: "gpt-5.6", choices: [{ index: 0, delta: {}, finish_reason: "stop" }] }, + ]), + ); + const params = buildChatCompletionParams({ + model: "gpt-5.6", + messages: [{ role: 1, content: [{ value: "hello" }], name: undefined }], + tools: [], + requireToolCall: false, + reasoningEffort: "high", + modelOptions: {}, + }); + const chunks = await createGatewayClient().streamChatCompletion(configFor(`${gateway.url}/`, "sk-test"), params, new AbortController().signal); + const contents: string[] = []; + for await (const chunk of chunks) { + contents.push(chunk.choices[0]?.delta.content ?? ""); + } + expect(contents.join("")).toBe("Hi"); + const [request] = gateway.requests; + expect(request).toMatchObject({ method: "POST", url: "/v1/chat/completions", authorization: "Bearer sk-test", userAgent: USER_AGENT }); + expect(JSON.parse(request?.body ?? "{}")).toMatchObject({ + model: "gpt-5.6", + stream: true, + stream_options: { include_usage: true }, + reasoning_effort: "high", + messages: [{ role: "user", content: [{ type: "text", text: "hello" }] }], + }); + }); + + it("leaves retries to the gateway instead of resending a failed request", async () => { + const gateway = await startGateway((_request, response) => { + response.writeHead(502, { "content-type": "application/json" }); + response.end('{"error":{"message":"upstream unavailable"}}'); + }); + const params = buildChatCompletionParams({ + model: "gpt-5.6", + messages: [{ role: 1, content: [{ value: "hello" }], name: undefined }], + tools: [], + requireToolCall: false, + reasoningEffort: undefined, + modelOptions: {}, + }); + await expect( + createGatewayClient().streamChatCompletion({ baseUrl: gateway.url, apiKey: "sk-test" }, params, new AbortController().signal), + ).rejects.toThrow(/upstream unavailable/); + expect(gateway.requests).toHaveLength(1); + }); +}); diff --git a/vscode-extension/test/messages.test.ts b/vscode-extension/test/messages.test.ts new file mode 100644 index 00000000000..ba11e479ad2 --- /dev/null +++ b/vscode-extension/test/messages.test.ts @@ -0,0 +1,168 @@ +import { describe, expect, it } from "vitest"; +import type * as vscode from "vscode"; +import { ESTIMATED_TOKENS_PER_IMAGE, buildChatCompletionParams, estimateMessageTokens, toChatCompletionMessages, type ChatRequestInput } from "../src/messages"; + +const USER = 1 as vscode.LanguageModelChatMessageRole; +const ASSISTANT = 2 as vscode.LanguageModelChatMessageRole; +const SYSTEM = 3 as vscode.LanguageModelChatMessageRole; + +const message = (role: vscode.LanguageModelChatMessageRole, content: readonly unknown[]): vscode.LanguageModelChatRequestMessage => ({ + role, + content, + name: undefined, +}); + +const text = (value: string): unknown => ({ value }); +const image = (bytes: readonly number[], mimeType = "image/png"): unknown => ({ mimeType, data: Uint8Array.from(bytes) }); +const toolCall = (callId: string, name: string, input: object): unknown => ({ callId, name, input }); +const toolResult = (callId: string, content: readonly unknown[]): unknown => ({ callId, content }); + +const request = (overrides: Partial = {}): ChatRequestInput => ({ + model: "gpt-5.6", + messages: [message(USER, [text("hi")])], + tools: [], + requireToolCall: false, + reasoningEffort: undefined, + modelOptions: {}, + ...overrides, +}); + +describe("toChatCompletionMessages", () => { + it("maps system, user, and assistant text", () => { + expect( + toChatCompletionMessages([ + message(SYSTEM, [text("be terse")]), + message(USER, [text("hello "), text("there")]), + message(ASSISTANT, [text("hi")]), + ]), + ).toEqual([ + { role: "system", content: "be terse" }, + { role: "user", content: [{ type: "text", text: "hello " }, { type: "text", text: "there" }] }, + { role: "assistant", content: "hi" }, + ]); + }); + + it("sends user images as data URLs", () => { + expect(toChatCompletionMessages([message(USER, [text("what is this"), image([1, 2, 3])])])).toEqual([ + { + role: "user", + content: [ + { type: "text", text: "what is this" }, + { type: "image_url", image_url: { url: "data:image/png;base64,AQID" } }, + ], + }, + ]); + }); + + it("round-trips tool calls and puts tool results before the user's follow-up text", () => { + expect( + toChatCompletionMessages([ + message(ASSISTANT, [text("checking"), toolCall("call_1", "read_file", { path: "a.ts" })]), + message(USER, [toolResult("call_1", [text("export const a = 1;")]), text("thanks")]), + ]), + ).toEqual([ + { + role: "assistant", + content: "checking", + tool_calls: [{ id: "call_1", type: "function", function: { name: "read_file", arguments: '{"path":"a.ts"}' } }], + }, + { role: "tool", tool_call_id: "call_1", content: "export const a = 1;" }, + { role: "user", content: [{ type: "text", text: "thanks" }] }, + ]); + }); + + it("emits a content-less assistant turn that only called tools", () => { + expect(toChatCompletionMessages([message(ASSISTANT, [toolCall("c", "t", {})])])).toEqual([ + { role: "assistant", content: null, tool_calls: [{ id: "c", type: "function", function: { name: "t", arguments: "{}" } }] }, + ]); + }); + + it("drops an assistant turn with neither text nor tool calls", () => { + expect(toChatCompletionMessages([message(USER, [text("hi")]), message(ASSISTANT, [text("")]), message(USER, [text("again")])])).toEqual([ + { role: "user", content: [{ type: "text", text: "hi" }] }, + { role: "user", content: [{ type: "text", text: "again" }] }, + ]); + }); + + it("hoists images out of tool results into a user message and serializes prompt-tsx values", () => { + expect( + toChatCompletionMessages([ + message(USER, [toolResult("call_2", [text("screenshot:"), image([9], "image/jpeg"), { value: { node: 1 } }])]), + ]), + ).toEqual([ + { role: "tool", tool_call_id: "call_2", content: 'screenshot:{"node":1}' }, + { role: "user", content: [{ type: "image_url", image_url: { url: "data:image/jpeg;base64,CQ==" } }] }, + ]); + }); + + it("decodes text data parts and ignores unknown parts", () => { + expect(toChatCompletionMessages([message(USER, [{ mimeType: "text/plain", data: Uint8Array.from([104, 105]) }, 42])])).toEqual([ + { role: "user", content: [{ type: "text", text: "hi" }] }, + ]); + }); +}); + +describe("estimateMessageTokens", () => { + it("counts what the gateway will receive, tool results and tool calls included", () => { + const plain = estimateMessageTokens(message(USER, [text("ok")])); + const withToolResult = estimateMessageTokens(message(USER, [toolResult("call_1", [text("y".repeat(800))]), text("ok")])); + const withToolCall = estimateMessageTokens(message(ASSISTANT, [toolCall("call_1", "read_file", { path: "z".repeat(800) })])); + expect(plain).toBeGreaterThan(0); + expect(withToolResult).toBeGreaterThanOrEqual(plain + 200); + expect(withToolCall).toBeGreaterThanOrEqual(200); + }); + + it("charges each image a flat estimate rather than its base64 length", () => { + const withoutImage = estimateMessageTokens(message(USER, [text("see")])); + const withImages = estimateMessageTokens(message(USER, [text("see"), image(new Array(30000).fill(0)), image([1])])); + expect(withImages - withoutImage).toBeGreaterThanOrEqual(2 * ESTIMATED_TOKENS_PER_IMAGE); + expect(withImages - withoutImage).toBeLessThan(2 * ESTIMATED_TOKENS_PER_IMAGE + 20); + }); + + it("counts nothing for a turn the gateway will never see", () => { + expect(estimateMessageTokens(message(ASSISTANT, []))).toBe(0); + }); +}); + +describe("buildChatCompletionParams", () => { + it("streams with usage and forwards only the chosen extras", () => { + expect(buildChatCompletionParams(request())).toEqual({ + model: "gpt-5.6", + messages: [{ role: "user", content: [{ type: "text", text: "hi" }] }], + stream: true, + stream_options: { include_usage: true }, + }); + }); + + it("declares tools as functions and requires a call only when VS Code does", () => { + const tools: readonly vscode.LanguageModelChatTool[] = [ + { name: "read_file", description: "Read a file", inputSchema: { type: "object", properties: { path: { type: "string" } } } }, + { name: "noop", description: "No input" }, + ]; + const auto = buildChatCompletionParams(request({ tools })); + expect(auto.tools).toEqual([ + { + type: "function", + function: { name: "read_file", description: "Read a file", parameters: { type: "object", properties: { path: { type: "string" } } } }, + }, + { type: "function", function: { name: "noop", description: "No input" } }, + ]); + expect(auto.tool_choice).toBeUndefined(); + expect(buildChatCompletionParams(request({ tools, requireToolCall: true })).tool_choice).toBe("required"); + expect(buildChatCompletionParams(request({ requireToolCall: true })).tool_choice).toBeUndefined(); + }); + + it("sends reasoning_effort only when the user picked one", () => { + expect(buildChatCompletionParams(request({ reasoningEffort: "xhigh" })).reasoning_effort).toBe("xhigh"); + expect(buildChatCompletionParams(request()).reasoning_effort).toBeUndefined(); + }); + + it("forwards numeric sampling options and drops everything else", () => { + const params = buildChatCompletionParams( + request({ modelOptions: { temperature: 0.2, max_tokens: 500, seed: "7", foo: "bar", top_p: 0.9 } }), + ); + expect(params).toMatchObject({ temperature: 0.2, max_tokens: 500, top_p: 0.9 }); + expect(params).not.toHaveProperty("seed"); + expect(params).not.toHaveProperty("foo"); + }); +}); diff --git a/vscode-extension/test/models.test.ts b/vscode-extension/test/models.test.ts new file mode 100644 index 00000000000..584b71ef8d4 --- /dev/null +++ b/vscode-extension/test/models.test.ts @@ -0,0 +1,185 @@ +import { describe, expect, it } from "vitest"; +import { + ASSUMED_MAX_INPUT_TOKENS, + ASSUMED_MAX_OUTPUT_TOKENS, + MARKDOWN_LINE_BREAK, + describeModels, + estimateTokens, + formatUsdPerMillionTokens, + parseModelGroups, + reasoningEffortFrom, + type ModelGroupInfo, +} from "../src/models"; + +const gatewayGroup = (overrides: Partial> = {}): Record => ({ + model_group: "gpt-5.6", + providers: ["openai"], + max_input_tokens: 922000, + max_output_tokens: 128000, + input_cost_per_token: 4e-6, + output_cost_per_token: 2e-5, + mode: "chat", + supports_vision: true, + supports_function_calling: true, + supports_reasoning: true, + supported_reasoning_efforts: ["none", "low", "medium", "high", "xhigh"], + ...overrides, +}); + +const parsed = (...groups: readonly Record[]): readonly ModelGroupInfo[] => { + const result = parseModelGroups({ data: groups }); + if (result.kind !== "ok") { + throw new Error(result.reason); + } + return result.groups; +}; + +describe("parseModelGroups", () => { + it("maps the gateway's /model_group/info shape", () => { + expect(parsed(gatewayGroup())).toEqual([ + { + modelGroup: "gpt-5.6", + providers: ["openai"], + mode: "chat", + maxInputTokens: 922000, + maxOutputTokens: 128000, + inputCostPerToken: 4e-6, + outputCostPerToken: 2e-5, + supportsVision: true, + supportsFunctionCalling: true, + supportedReasoningEfforts: ["none", "low", "medium", "high", "xhigh"], + }, + ]); + }); + + it("treats null limits, prices, and efforts as unknown", () => { + const [group] = parsed( + gatewayGroup({ + max_input_tokens: null, + max_output_tokens: null, + input_cost_per_token: null, + output_cost_per_token: null, + supported_reasoning_efforts: null, + supports_vision: null, + }), + ); + expect(group).toMatchObject({ + maxInputTokens: undefined, + inputCostPerToken: undefined, + supportedReasoningEfforts: [], + supportsVision: false, + }); + }); + + it("drops entries without a model_group and rejects payloads without data", () => { + expect(parsed({ providers: ["openai"] }, gatewayGroup()).map((group) => group.modelGroup)).toEqual(["gpt-5.6"]); + expect(parseModelGroups({ detail: "Unauthorized" })).toEqual({ kind: "invalid", reason: "response has no data array" }); + }); +}); + +describe("describeModels", () => { + it("lists chat groups with USD pricing in the detail and a full tooltip", () => { + const [model] = describeModels(parsed(gatewayGroup())); + expect(model).toMatchObject({ + id: "gpt-5.6", + name: "gpt-5.6", + family: "gpt-5.6", + detail: "$4.00 in / $20.00 out per 1M tokens", + maxInputTokens: 922000, + maxOutputTokens: 128000, + imageInput: true, + toolCalling: true, + }); + expect(model?.tooltip).toBe( + [ + "LiteLLM model group gpt-5.6 via openai", + "Input: $4.00 per 1M tokens", + "Output: $20.00 per 1M tokens", + "Context: 922000 in / 128000 out tokens", + "Reasoning effort: none, low, medium, high, xhigh", + ].join(MARKDOWN_LINE_BREAK), + ); + }); + + it("offers the gateway's reasoning efforts behind a gateway default entry", () => { + const [model] = describeModels(parsed(gatewayGroup({ supported_reasoning_efforts: ["low", "high"] }))); + expect(model?.configurationSchema).toEqual({ + properties: { + reasoningEffort: { + type: "string", + title: "Reasoning Effort", + enum: ["default", "low", "high"], + enumItemLabels: ["Gateway default", "Low", "High"], + default: "default", + group: "navigation", + }, + }, + }); + }); + + it("has no configuration schema when the group lists no reasoning efforts", () => { + const [model] = describeModels(parsed(gatewayGroup({ supported_reasoning_efforts: null }))); + expect(model?.configurationSchema).toBeUndefined(); + expect(model?.tooltip).toContain("Reasoning effort: not configurable"); + }); + + it("keeps groups without a mode and skips non-chat groups", () => { + const models = describeModels( + parsed( + gatewayGroup({ model_group: "text-embedding-4", mode: "embedding" }), + gatewayGroup({ model_group: "whisper-3", mode: "audio_transcription" }), + gatewayGroup({ model_group: "gpt-image-2", mode: "image_generation" }), + gatewayGroup({ model_group: "unlabeled", mode: null }), + gatewayGroup(), + ), + ); + expect(models.map((model) => model.id)).toEqual(["unlabeled", "gpt-5.6"]); + }); + + it("falls back to assumed context limits and says so", () => { + const [model] = describeModels(parsed(gatewayGroup({ max_input_tokens: null, max_output_tokens: null }))); + expect(model).toMatchObject({ maxInputTokens: ASSUMED_MAX_INPUT_TOKENS, maxOutputTokens: ASSUMED_MAX_OUTPUT_TOKENS }); + expect(model?.tooltip).toContain(`Context: unknown, assuming ${ASSUMED_MAX_INPUT_TOKENS} in / ${ASSUMED_MAX_OUTPUT_TOKENS} out tokens`); + }); + + it("shows missing prices instead of inventing zeros", () => { + const [both, inputOnly, free] = describeModels( + parsed( + gatewayGroup({ input_cost_per_token: null, output_cost_per_token: null }), + gatewayGroup({ output_cost_per_token: null }), + gatewayGroup({ input_cost_per_token: 0, output_cost_per_token: 0 }), + ), + ); + expect(both?.detail).toBe("No pricing configured"); + expect(both?.tooltip).toContain("Input: no price configured"); + expect(inputOnly?.detail).toBe("$4.00 in / n/a out per 1M tokens"); + expect(free?.detail).toBe("$0.00 in / $0.00 out per 1M tokens"); + }); +}); + +describe("formatUsdPerMillionTokens", () => { + it("renders cents for ordinary prices and two significant digits below a cent", () => { + expect(formatUsdPerMillionTokens(4e-6)).toBe("$4.00"); + expect(formatUsdPerMillionTokens(7.5e-7)).toBe("$0.75"); + expect(formatUsdPerMillionTokens(2.5e-5)).toBe("$25.00"); + expect(formatUsdPerMillionTokens(1e-9)).toBe("$0.0010"); + expect(formatUsdPerMillionTokens(0)).toBe("$0.00"); + }); +}); + +describe("reasoningEffortFrom", () => { + it("forwards a chosen effort and leaves the gateway default unset", () => { + expect(reasoningEffortFrom({ reasoningEffort: "high" })).toBe("high"); + expect(reasoningEffortFrom({ reasoningEffort: "default" })).toBeUndefined(); + expect(reasoningEffortFrom({ reasoningEffort: 3 })).toBeUndefined(); + expect(reasoningEffortFrom(undefined)).toBeUndefined(); + }); +}); + +describe("estimateTokens", () => { + it("rounds four characters per token upward", () => { + expect(estimateTokens("")).toBe(0); + expect(estimateTokens("abcd")).toBe(1); + expect(estimateTokens("abcde")).toBe(2); + }); +}); diff --git a/vscode-extension/test/provider.test.ts b/vscode-extension/test/provider.test.ts new file mode 100644 index 00000000000..07fc4ea5b59 --- /dev/null +++ b/vscode-extension/test/provider.test.ts @@ -0,0 +1,258 @@ +import type { ChatCompletionChunk, ChatCompletionCreateParamsStreaming } from "openai/resources/chat/completions"; +import { describe, expect, it } from "vitest"; +import type * as vscode from "vscode"; +import type { GatewayClient, GatewayConfig, ModelGroupsResult } from "../src/gateway"; +import { ESTIMATED_TOKENS_PER_IMAGE } from "../src/messages"; +import { LiteLLMChatProvider, TRUNCATED_MESSAGE, type LiteLLMModel } from "../src/provider"; +import { CancellationTokenSource, LanguageModelTextPart, LanguageModelToolCallPart } from "./vscode-mock"; + +interface StreamRequest { + readonly config: GatewayConfig; + readonly params: ChatCompletionCreateParamsStreaming; +} + +interface FakeGateway extends GatewayClient { + readonly listCalls: readonly GatewayConfig[]; + readonly streamRequests: readonly StreamRequest[]; +} + +const chunk = (delta: ChatCompletionChunk.Choice.Delta, finishReason: ChatCompletionChunk.Choice["finish_reason"] = null): ChatCompletionChunk => ({ + id: "chatcmpl-1", + object: "chat.completion.chunk", + created: 0, + model: "gpt-5.6", + choices: [{ index: 0, delta, finish_reason: finishReason }], +}); + +const modelGroups: ModelGroupsResult = { + kind: "ok", + groups: [ + { + modelGroup: "gpt-5.6", + providers: ["openai"], + mode: "chat", + maxInputTokens: 922000, + maxOutputTokens: 128000, + inputCostPerToken: 4e-6, + outputCostPerToken: 2e-5, + supportsVision: true, + supportsFunctionCalling: true, + supportedReasoningEfforts: ["low", "high"], + }, + ], +}; + +const fakeGateway = ( + listResult: ModelGroupsResult = modelGroups, + stream: (signal: AbortSignal) => AsyncIterable = () => (async function* () {})(), +): FakeGateway => { + const listCalls: GatewayConfig[] = []; + const streamRequests: StreamRequest[] = []; + return { + listCalls, + streamRequests, + async listModelGroups(config) { + listCalls.push(config); + return listResult; + }, + async streamChatCompletion(config, params, signal) { + streamRequests.push({ config, params }); + return stream(signal); + }, + }; +}; + +const token = (): vscode.CancellationToken => new CancellationTokenSource().token as unknown as vscode.CancellationToken; + +const prepare = (configuration: Record | undefined): vscode.PrepareLanguageModelChatModelOptions => + ({ silent: true, configuration }) as vscode.PrepareLanguageModelChatModelOptions; + +const gateway: GatewayConfig = { baseUrl: "http://127.0.0.1:4000", apiKey: "sk-test" }; + +const model = (overrides: Partial = {}): LiteLLMModel => ({ + id: "gpt-5.6", + name: "gpt-5.6", + family: "gpt-5.6", + version: "1.0", + maxInputTokens: 922000, + maxOutputTokens: 128000, + capabilities: { imageInput: true, toolCalling: true }, + gateway, + ...overrides, +}); + +const userMessage = (parts: readonly unknown[]): vscode.LanguageModelChatRequestMessage => + ({ role: 1, content: parts, name: undefined }) as vscode.LanguageModelChatRequestMessage; + +const responseOptions = (overrides: Partial = {}): vscode.ProvideLanguageModelChatResponseOptions => + ({ toolMode: 1, ...overrides }) as vscode.ProvideLanguageModelChatResponseOptions; + +const collect = ( + provider: LiteLLMChatProvider, + cancellation: CancellationTokenSource = new CancellationTokenSource(), +): { readonly parts: readonly vscode.LanguageModelResponsePart[]; readonly run: Promise } => { + const parts: vscode.LanguageModelResponsePart[] = []; + const run = provider.provideLanguageModelChatResponse( + model(), + [userMessage([{ value: "hi" }])], + responseOptions(), + { report: (part) => parts.push(part) }, + cancellation.token as unknown as vscode.CancellationToken, + ); + return { parts, run }; +}; + +describe("provideLanguageModelChatInformation", () => { + it("returns nothing for the unconfigured probe without touching the gateway", async () => { + const client = fakeGateway(); + expect(await new LiteLLMChatProvider(client).provideLanguageModelChatInformation(prepare(undefined), token())).toEqual([]); + expect(client.listCalls).toEqual([]); + }); + + it("names the API key when the stored secret is gone instead of listing nothing", async () => { + const client = fakeGateway(); + await expect( + new LiteLLMChatProvider(client).provideLanguageModelChatInformation(prepare({ baseUrl: "http://127.0.0.1:4000" }), token()), + ).rejects.toThrow(/missing its API key/); + expect(client.listCalls).toEqual([]); + }); + + it("rejects a gateway URL that is not http or https", async () => { + await expect( + new LiteLLMChatProvider(fakeGateway()).provideLanguageModelChatInformation(prepare({ baseUrl: "litellm.example.com", apiKey: "sk" }), token()), + ).rejects.toThrow(/"litellm.example.com" is not an http or https URL/); + }); + + it("lists the gateway's chat models with pricing, effort choices, and the gateway attached", async () => { + const client = fakeGateway(); + const models = await new LiteLLMChatProvider(client).provideLanguageModelChatInformation( + prepare({ baseUrl: "http://127.0.0.1:4000/v1/", apiKey: "sk-test" }), + token(), + ); + expect(client.listCalls).toEqual([gateway]); + expect(models).toEqual([ + expect.objectContaining({ + id: "gpt-5.6", + detail: "$4.00 in / $20.00 out per 1M tokens", + maxInputTokens: 922000, + capabilities: { imageInput: true, toolCalling: true }, + configurationSchema: expect.objectContaining({ properties: expect.objectContaining({ reasoningEffort: expect.anything() }) }), + gateway, + }), + ]); + }); + + it("shows the gateway's error message, not its whole JSON body, when discovery fails", async () => { + const body = JSON.stringify({ + error: { message: "Authentication Error, Invalid proxy server token passed", type: "auth_error", param: "sk-...abcd", code: "401" }, + }); + await expect( + new LiteLLMChatProvider(fakeGateway({ kind: "http_error", status: 401, body })).provideLanguageModelChatInformation( + prepare({ baseUrl: "http://127.0.0.1:4000", apiKey: "sk-bad" }), + token(), + ), + ).rejects.toThrow("LiteLLM gateway at http://127.0.0.1:4000 answered 401 for /model_group/info: Authentication Error, Invalid proxy server token passed"); + }); +}); + +describe("provideLanguageModelChatResponse", () => { + it("streams text and tool calls with the picked reasoning effort and a required tool choice", async () => { + const client = fakeGateway(modelGroups, () => + (async function* () { + yield chunk({ content: "Reading" }); + yield chunk({ tool_calls: [{ index: 0, id: "call_1", type: "function", function: { name: "read_file", arguments: '{"path":"a"}' } }] }); + yield chunk({}, "tool_calls"); + })(), + ); + const parts: vscode.LanguageModelResponsePart[] = []; + await new LiteLLMChatProvider(client).provideLanguageModelChatResponse( + model(), + [userMessage([{ value: "read a" }])], + responseOptions({ toolMode: 2, tools: [{ name: "read_file", description: "Read" }], modelConfiguration: { reasoningEffort: "high" } }), + { report: (part) => parts.push(part) }, + token(), + ); + expect(parts).toEqual([new LanguageModelTextPart("Reading"), new LanguageModelToolCallPart("call_1", "read_file", { path: "a" })]); + expect(client.streamRequests).toEqual([ + { + config: gateway, + params: expect.objectContaining({ model: "gpt-5.6", reasoning_effort: "high", tool_choice: "required", tools: [expect.anything()] }), + }, + ]); + }); + + it("finishes quietly when the user cancels mid-stream and drops its cancellation listener", async () => { + const cancellation = new CancellationTokenSource(); + const client = fakeGateway(modelGroups, (signal) => + (async function* () { + yield chunk({ content: "partial" }); + await new Promise((resolve) => signal.addEventListener("abort", () => resolve(), { once: true })); + throw new Error("Request was aborted."); + })(), + ); + const { parts, run } = collect(new LiteLLMChatProvider(client), cancellation); + await new Promise((resolve) => setTimeout(resolve, 0)); + cancellation.cancel(); + await expect(run).resolves.toBeUndefined(); + expect(parts).toEqual([new LanguageModelTextPart("partial")]); + expect(cancellation.disposedListeners).toBe(1); + }); + + it("surfaces a gateway failure as an error and still drops its cancellation listener", async () => { + const cancellation = new CancellationTokenSource(); + const client = fakeGateway(modelGroups, () => + (async function* () { + throw new Error("502 Bad Gateway"); + })(), + ); + const { run } = collect(new LiteLLMChatProvider(client), cancellation); + await expect(run).rejects.toThrow("502 Bad Gateway"); + expect(cancellation.disposedListeners).toBe(1); + }); + + it("reports the text it got and then fails when the model hits its output limit", async () => { + const client = fakeGateway(modelGroups, () => + (async function* () { + yield chunk({ content: "half an ans" }); + yield chunk({}, "length"); + })(), + ); + const { parts, run } = collect(new LiteLLMChatProvider(client)); + await expect(run).rejects.toThrow(TRUNCATED_MESSAGE); + expect(parts).toEqual([new LanguageModelTextPart("half an ans")]); + }); + + it("fails on tool arguments that are not JSON", async () => { + const client = fakeGateway(modelGroups, () => + (async function* () { + yield chunk({ tool_calls: [{ index: 0, id: "call_1", type: "function", function: { name: "grep", arguments: "{oops" } }] }); + })(), + ); + const { run } = collect(new LiteLLMChatProvider(client)); + await expect(run).rejects.toThrow("invalid JSON arguments for tool grep"); + }); +}); + +describe("provideTokenCount", () => { + const provider = new LiteLLMChatProvider(fakeGateway()); + + it("estimates plain text at four characters per token", async () => { + expect(await provider.provideTokenCount(model(), "abcdefgh")).toBe(2); + }); + + it("counts tool results and tool calls, not only text parts", async () => { + const textOnly = await provider.provideTokenCount(model(), userMessage([{ value: "ok" }])); + const withToolResult = await provider.provideTokenCount( + model(), + userMessage([{ callId: "call_1", content: [{ value: "x".repeat(400) }] }, { value: "ok" }]), + ); + expect(withToolResult).toBeGreaterThan(textOnly + 100); + }); + + it("charges a flat estimate per image instead of counting its bytes", async () => { + const withImage = await provider.provideTokenCount(model(), userMessage([{ value: "see" }, { mimeType: "image/png", data: new Uint8Array(50000) }])); + const withoutImage = await provider.provideTokenCount(model(), userMessage([{ value: "see" }])); + expect(withImage - withoutImage).toBeGreaterThanOrEqual(ESTIMATED_TOKENS_PER_IMAGE); + expect(withImage - withoutImage).toBeLessThan(ESTIMATED_TOKENS_PER_IMAGE + 20); + }); +}); diff --git a/vscode-extension/test/stream.test.ts b/vscode-extension/test/stream.test.ts new file mode 100644 index 00000000000..eb37013fb10 --- /dev/null +++ b/vscode-extension/test/stream.test.ts @@ -0,0 +1,93 @@ +import { describe, expect, it } from "vitest"; +import type { ChatCompletionChunk } from "openai/resources/chat/completions"; +import { responseParts, type ResponsePart } from "../src/stream"; + +type ToolCallDelta = NonNullable[number]; + +const chunk = (delta: ChatCompletionChunk.Choice.Delta, finishReason: ChatCompletionChunk.Choice["finish_reason"] = null): ChatCompletionChunk => ({ + id: "chatcmpl-1", + object: "chat.completion.chunk", + created: 0, + model: "gpt-5.6", + choices: [{ index: 0, delta, finish_reason: finishReason }], +}); + +const usageChunk: ChatCompletionChunk = { + id: "chatcmpl-1", + object: "chat.completion.chunk", + created: 0, + model: "gpt-5.6", + choices: [], + usage: { prompt_tokens: 3, completion_tokens: 2, total_tokens: 5 }, +}; + +async function* stream(chunks: readonly ChatCompletionChunk[]): AsyncGenerator { + yield* chunks; +} + +const collect = async (chunks: readonly ChatCompletionChunk[]): Promise => { + const parts: ResponsePart[] = []; + for await (const part of responseParts(stream(chunks))) { + parts.push(part); + } + return parts; +}; + +describe("responseParts", () => { + it("yields text deltas as they arrive and ignores empty and usage-only chunks", async () => { + expect(await collect([chunk({ role: "assistant", content: "" }), chunk({ content: "Hel" }), chunk({ content: "lo" }), usageChunk])).toEqual([ + { kind: "text", value: "Hel" }, + { kind: "text", value: "lo" }, + ]); + }); + + it("assembles tool calls split across chunks and emits them after the text, in index order", async () => { + expect( + await collect([ + chunk({ content: "Looking" }), + chunk({ tool_calls: [{ index: 1, id: "call_b", type: "function", function: { name: "grep", arguments: "" } }] }), + chunk({ tool_calls: [{ index: 0, id: "call_a", type: "function", function: { name: "read_file", arguments: '{"pa' } }] }), + chunk({ tool_calls: [{ index: 0, function: { name: "read_file", arguments: 'th":"a"}' } }] }), + chunk({ tool_calls: [{ index: 1, function: { arguments: '{"q":"x"}' } }] }, "tool_calls"), + ]), + ).toEqual([ + { kind: "text", value: "Looking" }, + { kind: "tool_call", callId: "call_a", name: "read_file", input: { path: "a" } }, + { kind: "tool_call", callId: "call_b", name: "grep", input: { q: "x" } }, + ]); + }); + + it("starts a new call when a fresh id reuses an index and appends index-less deltas to the last call", async () => { + expect( + await collect([ + chunk({ tool_calls: [{ index: 0, id: "call_a", type: "function", function: { name: "grep", arguments: '{"q":' } }] }), + chunk({ tool_calls: [{ function: { arguments: '"a"}' } } as ToolCallDelta] }), + chunk({ tool_calls: [{ index: 0, id: "call_b", type: "function", function: { name: "grep", arguments: '{"q":"b"}' } }] }), + ]), + ).toEqual([ + { kind: "tool_call", callId: "call_a", name: "grep", input: { q: "a" } }, + { kind: "tool_call", callId: "call_b", name: "grep", input: { q: "b" } }, + ]); + }); + + it("flags a response cut off at the output token limit after the text it did produce", async () => { + expect(await collect([chunk({ content: "half" }), chunk({}, "length"), usageChunk])).toEqual([ + { kind: "text", value: "half" }, + { kind: "truncated" }, + ]); + }); + + it("treats empty arguments as an empty object and flags malformed JSON", async () => { + expect( + await collect([ + chunk({ tool_calls: [{ index: 0, id: "call_0", type: "function", function: { name: "noop", arguments: "" } }] }), + chunk({ tool_calls: [{ index: 1, id: "call_1", type: "function", function: { name: "bad", arguments: "{oops" } }] }), + chunk({ tool_calls: [{ index: 2, id: "call_2", type: "function", function: { name: "scalar", arguments: "42" } }] }), + ]), + ).toEqual([ + { kind: "tool_call", callId: "call_0", name: "noop", input: {} }, + { kind: "invalid_tool_call", callId: "call_1", name: "bad", arguments: "{oops" }, + { kind: "invalid_tool_call", callId: "call_2", name: "scalar", arguments: "42" }, + ]); + }); +}); diff --git a/vscode-extension/test/vscode-mock.ts b/vscode-extension/test/vscode-mock.ts new file mode 100644 index 00000000000..5d2e8577201 --- /dev/null +++ b/vscode-extension/test/vscode-mock.ts @@ -0,0 +1,82 @@ +export class LanguageModelTextPart { + constructor(readonly value: string) {} +} + +export class LanguageModelToolCallPart { + constructor( + readonly callId: string, + readonly name: string, + readonly input: object, + ) {} +} + +export const LanguageModelChatToolMode = { Auto: 1, Required: 2 } as const; + +type Listener = (value: T) => void; + +interface Subscription { + dispose(): void; +} + +export class EventEmitter { + private readonly listeners: Listener[] = []; + + readonly event = (listener: Listener): Subscription => { + this.listeners.push(listener); + return { + dispose: () => { + const index = this.listeners.indexOf(listener); + if (index >= 0) { + this.listeners.splice(index, 1); + } + }, + }; + }; + + fire(value: T): void { + [...this.listeners].forEach((listener) => listener(value)); + } + + dispose(): void { + this.listeners.splice(0); + } +} + +export interface MockCancellationToken { + readonly isCancellationRequested: boolean; + onCancellationRequested(listener: Listener): Subscription; +} + +export class CancellationTokenSource { + private readonly emitter = new EventEmitter(); + private cancelled = false; + private disposed = 0; + readonly token: MockCancellationToken; + + constructor() { + const source = this; + this.token = { + get isCancellationRequested(): boolean { + return source.cancelled; + }, + onCancellationRequested: (listener) => { + const subscription = source.emitter.event(listener); + return { + dispose: () => { + source.disposed += 1; + subscription.dispose(); + }, + }; + }, + }; + } + + get disposedListeners(): number { + return this.disposed; + } + + cancel(): void { + this.cancelled = true; + this.emitter.fire(); + } +} diff --git a/vscode-extension/tsconfig.json b/vscode-extension/tsconfig.json new file mode 100644 index 00000000000..6835fa9b672 --- /dev/null +++ b/vscode-extension/tsconfig.json @@ -0,0 +1,19 @@ +{ + "compilerOptions": { + "target": "ES2022", + "module": "ESNext", + "moduleResolution": "Bundler", + "lib": ["ES2022"], + "types": ["node"], + "strict": true, + "noUncheckedIndexedAccess": true, + "noImplicitOverride": true, + "isolatedModules": true, + "verbatimModuleSyntax": true, + "esModuleInterop": true, + "resolveJsonModule": true, + "skipLibCheck": true, + "noEmit": true + }, + "include": ["src", "test"] +} diff --git a/vscode-extension/vitest.config.mts b/vscode-extension/vitest.config.mts new file mode 100644 index 00000000000..cf1533f7f16 --- /dev/null +++ b/vscode-extension/vitest.config.mts @@ -0,0 +1,13 @@ +import { fileURLToPath } from "node:url"; +import { defineConfig } from "vitest/config"; + +export default defineConfig({ + resolve: { + alias: { + vscode: fileURLToPath(new URL("./test/vscode-mock.ts", import.meta.url)), + }, + }, + test: { + include: ["test/**/*.test.ts"], + }, +}); From c0b0ba20b23a83a5e382ee2b7de7739cf5dcaa10 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 12:41:40 -0700 Subject: [PATCH 108/141] fix(azure-ai): keep FLUX.2 tolerant of OpenAI-only image params --- .../image_edit/flux2_transformation.py | 17 +--------- .../image_generation/flux_transformation.py | 34 +++++++++++++++---- ...test_azure_ai_image_edit_transformation.py | 11 ++++++ .../test_azure_ai_flux2_image_generation.py | 34 +++++++++++++++---- 4 files changed, 67 insertions(+), 29 deletions(-) diff --git a/litellm/llms/azure_ai/image_edit/flux2_transformation.py b/litellm/llms/azure_ai/image_edit/flux2_transformation.py index aa8905e5601..f91a87ba0f4 100644 --- a/litellm/llms/azure_ai/image_edit/flux2_transformation.py +++ b/litellm/llms/azure_ai/image_edit/flux2_transformation.py @@ -31,22 +31,7 @@ class AzureFoundryFlux2ImageEditConfig(OpenAIImageEditConfig): """ def get_supported_openai_params(self, model: str) -> list: - """ - FLUX 2 supports a subset of OpenAI image edit params - """ - return [ - "n", - "size", - "width", - "height", - "num_images", - "seed", - "safety_tolerance", - "output_format", - "aspect_ratio", - "guidance", - "steps", - ] + return AzureFoundryFluxImageGenerationConfig().get_supported_openai_params(model) def map_openai_params( self, diff --git a/litellm/llms/azure_ai/image_generation/flux_transformation.py b/litellm/llms/azure_ai/image_generation/flux_transformation.py index 997205b1fc3..ac9ec24420b 100644 --- a/litellm/llms/azure_ai/image_generation/flux_transformation.py +++ b/litellm/llms/azure_ai/image_generation/flux_transformation.py @@ -2,9 +2,18 @@ from collections.abc import Mapping from types import MappingProxyType from typing import Final +from litellm.exceptions import BadRequestError, UnsupportedParamsError from litellm.llms.openai.image_generation import GPTImageGenerationConfig from litellm.types.llms.openai import OpenAIImageGenerationOptionalParams +FLUX2_DROPPED_OPENAI_PARAMS: Final[tuple[OpenAIImageGenerationOptionalParams, ...]] = ( + "background", + "moderation", + "output_compression", + "quality", + "user", +) + class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): """Azure Foundry BFL API configuration for FLUX image generation.""" @@ -81,10 +90,13 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): "num_images", "guidance", "steps", + *FLUX2_DROPPED_OPENAI_PARAMS, ] @staticmethod - def _map_parameter(name: str, value: object) -> tuple[tuple[str, object], ...]: + def _map_parameter(name: str, value: object, model: str) -> tuple[tuple[str, object], ...]: + if name in FLUX2_DROPPED_OPENAI_PARAMS: + return () if isinstance(value, str): if name in ("n", "num_images", "width", "height", "steps", "seed", "safety_tolerance"): return (("num_images" if name == "n" else name, int(value)),) @@ -94,11 +106,17 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): return (("num_images", value),) if name != "size": return ((name, value),) + if str(value).lower() == "auto": + return () try: width, height = (int(dimension) for dimension in str(value).lower().split("x")) except (TypeError, ValueError): - raise ValueError(f"Invalid size format '{value}'. Expected 'WxH', for example '1024x1024'.") + raise BadRequestError( + message=f"Invalid size format '{value}'. Expected 'WxH', for example '1024x1024'.", + model=model, + llm_provider="azure_ai", + ) return (("width", width), ("height", height)) def map_openai_params( @@ -118,9 +136,13 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): supported_params: Final = self.get_supported_openai_params(model) unsupported_params: Final = tuple(name for name in non_default_params if name not in supported_params) if unsupported_params and not drop_params: - raise ValueError( - f"Parameters {unsupported_params} are not supported for model {model}. " - f"Supported parameters are {supported_params}. Set drop_params=True to drop unsupported parameters." + raise UnsupportedParamsError( + message=( + f"Parameters {unsupported_params} are not supported for model {model}. " + f"Supported parameters are {supported_params}. Set drop_params=True to drop unsupported parameters." + ), + model=model, + llm_provider="azure_ai", ) mapped_params: Final[Mapping[str, object]] = MappingProxyType( @@ -128,7 +150,7 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig): mapped_name: mapped_value for name, value in non_default_params.items() if name in supported_params - for mapped_name, mapped_value in self._map_parameter(name, value) + for mapped_name, mapped_value in self._map_parameter(name, value, model) } ) return {**optional_params, **mapped_params} # mutable-ok: inherited config contract returns a dict diff --git a/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py b/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py index d74afa88a6b..39001c1795b 100644 --- a/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py +++ b/tests/test_litellm/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py @@ -176,3 +176,14 @@ def test_flux2_image_edit_preserves_controls_and_pixel_cost(dimensions: Mapping[ ) assert response._hidden_params["response_cost"] == pytest.approx(5e-08 * 2048 * 1024 * 2) + + +def test_flux2_image_edit_accepts_and_drops_openai_only_parameters(): + optional_params: Final = ImageEditRequestUtils.get_optional_params_image_edit( + model="FLUX.2-pro", + image_edit_provider_config=AzureFoundryFlux2ImageEditConfig(), + image_edit_optional_params={"n": 1, "size": "auto", "quality": "high", "user": "end-user-1"}, + drop_params=False, + ) + + assert optional_params == {"num_images": 1} diff --git a/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py index 388cca9f054..512e98b4151 100644 --- a/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py +++ b/tests/test_litellm/llms/azure_ai/image_generation/test_azure_ai_flux2_image_generation.py @@ -15,7 +15,7 @@ from litellm.llms.azure_ai.image_generation.flux_transformation import ( AzureFoundryFluxImageGenerationConfig, ) from litellm.types.utils import ImageObject, ImageResponse -from litellm.utils import _invalidate_model_cost_lowercase_map +from litellm.utils import _invalidate_model_cost_lowercase_map, get_optional_params_image_gen @pytest.fixture(autouse=True) @@ -86,15 +86,35 @@ def test_flux2_flex_maps_openai_and_provider_parameters(): } -def test_flux2_flex_rejects_invalid_size(): - with pytest.raises(ValueError, match="Expected 'WxH'"): - AzureFoundryFluxImageGenerationConfig().map_openai_params( - non_default_params={"size": "large"}, - optional_params={}, +def test_flux2_flex_rejects_invalid_size_as_bad_request(): + with pytest.raises(litellm.BadRequestError, match="Expected 'WxH'") as raised: + get_optional_params_image_gen( model="FLUX.2-flex", - drop_params=False, + custom_llm_provider="azure_ai", + provider_config=AzureFoundryFluxImageGenerationConfig(), + size="large", ) + assert raised.value.status_code == 400 + + +@pytest.mark.parametrize("model", ("FLUX.2-pro", "FLUX.2-flex")) +def test_flux2_accepts_and_drops_openai_only_image_parameters(model: str): + optional_params: Final = get_optional_params_image_gen( + model=model, + custom_llm_provider="azure_ai", + provider_config=AzureFoundryFluxImageGenerationConfig(), + n=1, + size="auto", + quality="high", + user="end-user-1", + background="transparent", + moderation="low", + output_compression=50, + ) + + assert optional_params == {"num_images": 1} + def test_flux2_flex_model_info(): model_info = litellm.get_model_info( From 4f904f69b6235497f784839be5078f1f77d5255d Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 19:44:18 +0000 Subject: [PATCH 109/141] test(proxy): restore app routes after pass-through reload tests The two pass-through reload tests register real FastAPI routes on the shared proxy app and only clean up the internal registry, so any test running after them in the same worker sees stray /v1/kept-* and /v1/deleted-* routes. test_component_allowlists counts those as uncovered and fails. Snapshot app.routes and the registry up front and restore both in a finally block. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/proxy/test_proxy_server.py | 52 +++++++++++++------ 1 file changed, 35 insertions(+), 17 deletions(-) diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index ce809eda847..19822032d1f 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -7536,25 +7536,35 @@ async def test_deleting_the_stored_pass_through_row_takes_the_route_out_of_servi deleted. The proxy's own registry of live pass-through routes is what decides whether a request is routed upstream or falls through to the auth error, so it has to lose the entry on the reload rather than at the next process restart.""" - from litellm.proxy.pass_through_endpoints.pass_through_endpoints import InitPassThroughEndpointHelpers - from litellm.proxy.proxy_server import ProxyConfig + from litellm.proxy.pass_through_endpoints.pass_through_endpoints import ( + InitPassThroughEndpointHelpers, + _registered_pass_through_routes, + ) + from litellm.proxy.proxy_server import ProxyConfig, app path: Final = f"/v1/deleted-{uuid.uuid4().hex[:8]}" db_endpoint: Final = {"id": "db-1", "path": path, "target": "https://example.com/post"} + prior_routes: Final = list(app.routes) + prior_registry: Final = dict(_registered_pass_through_routes) def live_routes() -> set[str]: return {route for route in InitPassThroughEndpointHelpers.get_all_registered_pass_through_routes() if path in route} settings: Final = patch("litellm.proxy.proxy_server.general_settings", {}) # test-quality-ok: the method reads this module global; no injection seam yaml_endpoints: Final = patch("litellm.proxy.proxy_server.config_passthrough_endpoints", None) # test-quality-ok: module global holding the YAML endpoints; this case has none - with settings, yaml_endpoints: - pc = ProxyConfig() - await pc._update_general_settings(db_general_settings={"pass_through_endpoints": [db_endpoint]}) - assert live_routes(), "the stored endpoint should be serving before the row is deleted" + try: + with settings, yaml_endpoints: + pc = ProxyConfig() + await pc._update_general_settings(db_general_settings={"pass_through_endpoints": [db_endpoint]}) + assert live_routes(), "the stored endpoint should be serving before the row is deleted" - await pc._update_general_settings(db_general_settings={}) + await pc._update_general_settings(db_general_settings={}) - assert live_routes() == set() + assert live_routes() == set() + finally: + app.routes[:] = prior_routes + _registered_pass_through_routes.clear() + _registered_pass_through_routes.update(prior_registry) @pytest.mark.asyncio @@ -7564,15 +7574,18 @@ async def test_a_stored_pass_through_row_never_disturbs_the_config_declared_rout serving untouched. The stored entry never gets a route of its own.""" from litellm.proxy.pass_through_endpoints.pass_through_endpoints import ( InitPassThroughEndpointHelpers, + _registered_pass_through_routes, initialize_pass_through_endpoints, ) - from litellm.proxy.proxy_server import ProxyConfig + from litellm.proxy.proxy_server import ProxyConfig, app marker: Final = uuid.uuid4().hex[:8] config_path: Final = f"/v1/kept-{marker}" db_path: Final = f"/v1/ignored-{marker}" config_endpoint: Final = {"id": f"cfg-{marker}", "path": config_path, "target": "https://example.com/post"} db_endpoint: Final = {"id": f"db-{marker}", "path": db_path, "target": "https://example.com/post"} + prior_routes: Final = list(app.routes) + prior_registry: Final = dict(_registered_pass_through_routes) def live_paths() -> set[str]: registered: Final = InitPassThroughEndpointHelpers.get_all_registered_pass_through_routes() @@ -7580,17 +7593,22 @@ async def test_a_stored_pass_through_row_never_disturbs_the_config_declared_rout settings: Final = patch("litellm.proxy.proxy_server.general_settings", {"pass_through_endpoints": [config_endpoint]}) # test-quality-ok: the method reads this module global; no injection seam yaml_endpoints: Final = patch("litellm.proxy.proxy_server.config_passthrough_endpoints", [config_endpoint]) # test-quality-ok: module global holding the YAML endpoints the reload merges in - with settings, yaml_endpoints: - await initialize_pass_through_endpoints(pass_through_endpoints=[config_endpoint]) - assert live_paths() == {config_path} + try: + with settings, yaml_endpoints: + await initialize_pass_through_endpoints(pass_through_endpoints=[config_endpoint]) + assert live_paths() == {config_path} - pc = ProxyConfig() - await pc._update_general_settings(db_general_settings={"pass_through_endpoints": [db_endpoint]}) - assert live_paths() == {config_path} + pc = ProxyConfig() + await pc._update_general_settings(db_general_settings={"pass_through_endpoints": [db_endpoint]}) + assert live_paths() == {config_path} - await pc._update_general_settings(db_general_settings={}) + await pc._update_general_settings(db_general_settings={}) - assert live_paths() == {config_path} + assert live_paths() == {config_path} + finally: + app.routes[:] = prior_routes + _registered_pass_through_routes.clear() + _registered_pass_through_routes.update(prior_registry) def _fill_user_api_key_cache(cache: DualCache, count: int) -> None: From 0743c97bef983d54d7881254fb7c0d5292b9b743 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 19:44:18 +0000 Subject: [PATCH 110/141] chore(deps): bump anyio to 4.14.2 in uv.lock Addresses GHSA-5p39-cfhj-2xmp and GHSA-82r6-8w77-94w6 flagged by the osv-scan job. Hand-edited the single anyio block so the lockfile stays byte-stable for CI's older uv version; dependency set is unchanged. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- uv.lock | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/uv.lock b/uv.lock index a5e60c68515..e30274c1f6a 100644 --- a/uv.lock +++ b/uv.lock @@ -315,16 +315,16 @@ vertex = [ [[package]] name = "anyio" -version = "4.13.0" +version = "4.14.2" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "exceptiongroup", marker = "python_full_version < '3.11'" }, { name = "idna" }, { name = "typing-extensions", marker = "python_full_version < '3.13'" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/19/14/2c5dd9f512b66549ae92767a9c7b330ae88e1932ca57876909410251fe13/anyio-4.13.0.tar.gz", hash = "sha256:334b70e641fd2221c1505b3890c69882fe4a2df910cba14d97019b90b24439dc", size = 231622, upload-time = "2026-03-24T12:59:09.671Z" } +sdist = { url = "https://files.pythonhosted.org/packages/61/cc/a381afa6efea9f496eff839d4a6a1aed3bfafc7b3ab4b0d1b243a12573dd/anyio-4.14.2.tar.gz", hash = "sha256:cfa139f3ed1a23ee8f88a145ddb5ac7605b8bbfd8592baacd7ce3d8bb4313c7f", size = 260176, upload-time = "2026-07-12T20:29:07.082Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/da/42/e921fccf5015463e32a3cf6ee7f980a6ed0f395ceeaa45060b61d86486c2/anyio-4.13.0-py3-none-any.whl", hash = "sha256:08b310f9e24a9594186fd75b4f73f4a4152069e3853f1ed8bfbf58369f4ad708", size = 114353, upload-time = "2026-03-24T12:59:08.246Z" }, + { url = "https://files.pythonhosted.org/packages/da/35/f2287558c17e29fafc8ef3daf819bb9834061cfa43bff8014f7df7f63bdc/anyio-4.14.2-py3-none-any.whl", hash = "sha256:9f505dda5ac9f0c8309b5e8bd445a8c2bf7246f3ce950121e45ea15bc41d1494", size = 125813, upload-time = "2026-07-12T20:29:05.763Z" }, ] [[package]] From 58beea227547cccf54dd70eb1f7ae3024b8a1a03 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 12:45:54 -0700 Subject: [PATCH 111/141] fix(passthrough): read the stream flag by truthiness in the timeout resolver --- litellm/passthrough/timeout_utils.py | 4 ++-- .../test_pass_through_endpoints.py | 14 ++++++++++++++ 2 files changed, 16 insertions(+), 2 deletions(-) diff --git a/litellm/passthrough/timeout_utils.py b/litellm/passthrough/timeout_utils.py index 829277105e3..9284f7143b0 100644 --- a/litellm/passthrough/timeout_utils.py +++ b/litellm/passthrough/timeout_utils.py @@ -10,7 +10,6 @@ DEFAULT_PASS_THROUGH_REQUEST_TIMEOUT_SECONDS: Final = 600.0 class _TimeoutFields(BaseModel): model_config = ConfigDict(frozen=True) - stream: bool = False stream_timeout: float | None = None timeout: float | None = None request_timeout: float | None = None @@ -61,10 +60,11 @@ def resolve_llm_passthrough_timeout( kwargs stream_timeout -> litellm_params stream_timeout -> router_stream_timeout, then the non-streaming chain above. """ + streaming: Final = bool((kwargs or {}).get("stream")) request: Final = _TimeoutFields.model_validate(kwargs or {}) deployment: Final = _TimeoutFields.model_validate(litellm_params or {}) stream_candidates: Final = ( - (request.stream_timeout, deployment.stream_timeout, router_stream_timeout) if request.stream else () + (request.stream_timeout, deployment.stream_timeout, router_stream_timeout) if streaming else () ) candidates: Final = ( *stream_candidates, diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index 54336800db4..abff897c4f5 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -1175,6 +1175,20 @@ def test_resolve_llm_passthrough_timeout_stream_timeout_precedence(): ) +@pytest.mark.parametrize( + "stream, expected", + [(None, 90.0), (0, 90.0), ("", 90.0), (1, 1800.0), ("yes", 1800.0)], +) +def test_resolve_llm_passthrough_timeout_reads_stream_by_truthiness(stream: object, expected: float): + assert ( + resolve_llm_passthrough_timeout( + kwargs={"stream": stream}, + litellm_params={"stream_timeout": 1800, "timeout": 90}, + ) + == expected + ) + + @pytest.mark.asyncio async def test_pass_through_request_uses_resolved_timeout(): with patch("litellm.proxy.proxy_server.proxy_logging_obj") as mock_proxy_logging: From cf22e4b171da3b923765a7e597d78378f03fb9c9 Mon Sep 17 00:00:00 2001 From: yassin Date: Fri, 18 Sep 2026 19:38:38 +0000 Subject: [PATCH 112/141] fix(proxy): make SettingsStore.clear terminate when the config file owns a key MutableMapping.clear pops items until the mapping is empty, but __delitem__ keeps config-owned keys, so clear spun forever on any store that had loaded a config file. unittest.mock.patch.dict calls clear on exit, which is why the proxy-infra and proxy-endpoints shards hung at 99 percent until the 20 minute job timeout on every run since the store started refusing config-owned writes Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/config_resolvers/settings_store.py | 4 ++++ .../config_resolvers/test_settings_store.py | 16 ++++++++++++++++ 2 files changed, 20 insertions(+) diff --git a/litellm/proxy/config_resolvers/settings_store.py b/litellm/proxy/config_resolvers/settings_store.py index 079d262319f..a3d5da4b343 100644 --- a/litellm/proxy/config_resolvers/settings_store.py +++ b/litellm/proxy/config_resolvers/settings_store.py @@ -98,6 +98,10 @@ class SettingsStore(MutableMapping[str, JsonValue]): def __len__(self) -> int: return sum(1 for _ in self) + def clear(self) -> None: + self._runtime_values = _EMPTY_VALUES + self._deleted_runtime_keys = frozenset(key for key in self._keys() if not self.owned_by_config(key)) + def _clear_runtime(self) -> None: self._runtime_values = _EMPTY_VALUES self._deleted_runtime_keys = frozenset() diff --git a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py index aa410deac43..0b472938b74 100644 --- a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py +++ b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py @@ -168,6 +168,22 @@ def test_settings_store_refuses_a_runtime_write_to_a_config_owned_key() -> None: assert store.source("max_parallel_requests") == "config" +@pytest.mark.timeout(10) +def test_settings_store_clear_terminates_and_keeps_config_owned_keys() -> None: + store: Final = SettingsStore("general_settings") + store.load_yaml({"max_parallel_requests": 3}) + store.apply_db_row("general_settings", {"max_file_size_mb": 9}) + store["ui_access_mode"] = "admin_only" + before: Final = dict(store) + + store.clear() + cleared: Final = dict(store) + store.update(before) + + assert cleared == {"max_parallel_requests": 3} + assert dict(store) == before + + def test_settings_store_reports_the_config_owned_keys_a_write_would_change() -> None: store: Final = SettingsStore("general_settings") store.load_yaml({"max_parallel_requests": 3, "ui_access_mode": "admin_only"}) From 32e6a86319ea9967c4af7a00e26411fd91046871 Mon Sep 17 00:00:00 2001 From: jesus Date: Sat, 12 Sep 2026 01:00:44 +0000 Subject: [PATCH 113/141] fix(router): report null cost for unpriced deployments instead of 0 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/proxy_server.py | 5 +- litellm/router.py | 11 ++++- litellm/utils.py | 16 +++++++ .../proxy_server/test_routes_model_info.py | 37 +++++++++++++++ tests/test_litellm/test_router.py | 47 +++++++++++++++++++ 5 files changed, 113 insertions(+), 3 deletions(-) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 3b5f4236d22..f21c907c292 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -154,7 +154,7 @@ from litellm.types.utils import ( TextCompletionResponse, TokenCountResponse, ) -from litellm.utils import load_credentials_from_list +from litellm.utils import cost_map_omits_token_price, load_credentials_from_list if TYPE_CHECKING: from aiohttp import ClientSession @@ -13618,9 +13618,10 @@ def _enrich_model_info_with_litellm_data( discovered_model_info: Final = ( llm_router.get_discovered_model_info(model_info.get("id")) if llm_router is not None else MappingProxyType({}) ) + unpriced: Final = cost_map_omits_token_price(model_info.get("id"), litellm_model_info.get("key")) for k, v in MappingProxyType({**litellm_model_info, **discovered_model_info}).items(): if k not in model_info or (model_info[k] is None and k in discovered_model_info): - model_info[k] = v + model_info[k] = None if unpriced and k in ("input_cost_per_token", "output_cost_per_token") else v model["model_info"] = model_info # don't return the api key / vertex credentials # don't return the llm credentials diff --git a/litellm/router.py b/litellm/router.py index eec393d6cef..8a543373692 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -10575,11 +10575,12 @@ class Router: 2. If not, check if litellm model name is in model info 3. If not, return None """ - from litellm.utils import _update_dictionary + from litellm.utils import _update_dictionary, cost_map_omits_token_price model_info: ModelInfo | None = None custom_model_info: dict | None = None litellm_model_name_model_info: ModelInfo | None = None + base_model_key: str | None = None try: custom_model_info = ( @@ -10606,6 +10607,7 @@ class Router: ## update litellm model info with base model info base_model_info: Final = copy.deepcopy(litellm.get_model_info(model=base_model)) if base_model_info is not None: + base_model_key = base_model_info.get("key") # Base model provides defaults, custom model info overrides custom_model_info = _update_dictionary( cast(dict, base_model_info), @@ -10633,6 +10635,13 @@ class Router: # custom_model_info already includes base_model defaults at this point, if applicable model_info = cast(ModelInfo, custom_model_info) + if model_info is None: + return None + builtin_key: Final = ( + litellm_model_name_model_info.get("key") if litellm_model_name_model_info is not None else None + ) + if cost_map_omits_token_price(model_id, builtin_key, base_model_key): + return cast(ModelInfo, {**model_info, "input_cost_per_token": None, "output_cost_per_token": None}) return model_info def _set_model_group_info(self, model_group: str, user_facing_model_group_name: str) -> ModelGroupInfo | None: diff --git a/litellm/utils.py b/litellm/utils.py index 2c9200fbad7..ec11895a8b5 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -3156,6 +3156,22 @@ def reapply_runtime_model_cost_registrations() -> None: register_model(model_cost=dict(_runtime_registered_model_cost)) # mutable-ok: snapshot, replay rewrites it +def cost_map_omits_token_price(*keys: object) -> bool: + """Whether the raw ``litellm.model_cost`` entries under ``keys`` exist but none carries a per-token price. + + ``get_model_info`` substitutes 0 for a missing price, which reads exactly like a declared + zero. Surfaces that report pricing use this to keep an unpriced deployment at ``None``. + """ + entries: Final = tuple( + entry + for entry in (litellm.model_cost.get(key) for key in keys if isinstance(key, str)) + if isinstance(entry, dict) + ) + return len(entries) > 0 and not any( + "input_cost_per_token" in entry or "output_cost_per_token" in entry for entry in entries + ) + + def register_model( model_cost: str | dict, *, diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py b/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py index a1cf838ab6b..203a237113f 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py @@ -286,6 +286,43 @@ def test_get_proxy_model_info_surfaces_supports_parallel_function_calling(local_ assert enriched["model_info"]["supports_parallel_function_calling"] is True +def test_model_info_reports_null_cost_for_unpriced_deployment_and_zero_for_declared_zero(): + """A deployment configured with no cost fields must not surface the 0 that ``get_model_info`` + defaults to, since the zero-cost budget bypass only honours a declared zero. The declared zero + and a catalog price still come through.""" + import litellm + + router = litellm.Router( + model_list=[ + { + "model_name": "vllm-unpriced", + "litellm_params": {"model": "openai/vllm-unpriced", "api_key": "x", "api_base": "http://vllm"}, + }, + { + "model_name": "vllm-free", + "litellm_params": { + "model": "openai/vllm-free", + "api_key": "x", + "api_base": "http://vllm", + "input_cost_per_token": 0, + "output_cost_per_token": 0, + }, + }, + {"model_name": "gpt-priced", "litellm_params": {"model": "gpt-4o", "api_key": "x"}}, + ] + ) + + def enriched_cost(model_name: str) -> tuple: + deployment = router.get_model_list(model_name=model_name)[0] + info = proxy_server._enrich_model_info_with_litellm_data({**deployment, "model_info": dict(deployment["model_info"])})["model_info"] + return info.get("input_cost_per_token"), info.get("output_cost_per_token") + + assert enriched_cost("vllm-unpriced") == (None, None) + assert enriched_cost("vllm-free") == (0, 0) + input_cost, output_cost = enriched_cost("gpt-priced") + assert input_cost > 0 and output_cost > 0 + + def test_v1_model_info_star_wildcard_filter_keeps_provider_expansion(monkeypatch): from litellm.proxy._types import SpecialModelNames, UserAPIKeyAuth from litellm.proxy.auth import model_checks diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 26f9803a022..ceb8fcf6fd1 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -1888,6 +1888,53 @@ def test_model_group_info_cost_none_when_db_model_info_has_no_cost(): assert result.output_cost_per_token is None +def test_model_group_info_cost_none_for_unpriced_deployment_but_zero_when_declared(): + """A deployment with no cost fields anywhere must report None, not the 0 that + get_model_info defaults to, so the reported price matches what the zero-cost + budget bypass accepts. A deployment declaring 0 keeps reporting 0.""" + router = litellm.Router( + model_list=[ + { + "model_name": "vllm-unpriced", + "litellm_params": { + "model": "openai/my-vllm-unpriced", + "api_key": "fake", + "api_base": "http://localhost:8000/v1", + }, + }, + { + "model_name": "vllm-free", + "litellm_params": { + "model": "openai/my-vllm-free", + "api_key": "fake", + "api_base": "http://localhost:8000/v1", + "input_cost_per_token": 0, + "output_cost_per_token": 0, + }, + }, + { + "model_name": "gpt-priced", + "litellm_params": {"model": "gpt-4o", "api_key": "fake"}, + }, + ] + ) + + unpriced = router.get_model_group_info(model_group="vllm-unpriced") + assert unpriced is not None + assert unpriced.input_cost_per_token is None + assert unpriced.output_cost_per_token is None + + free = router.get_model_group_info(model_group="vllm-free") + assert free is not None + assert free.input_cost_per_token == 0 + assert free.output_cost_per_token == 0 + + priced = router.get_model_group_info(model_group="gpt-priced") + assert priced is not None + assert priced.input_cost_per_token is not None and priced.input_cost_per_token > 0 + assert priced.output_cost_per_token is not None and priced.output_cost_per_token > 0 + + @pytest.mark.parametrize( "value,expected", [ From 170581e7101ba821a6bd6a2cb039718f899fc85d Mon Sep 17 00:00:00 2001 From: jesus Date: Sat, 12 Sep 2026 01:04:40 +0000 Subject: [PATCH 114/141] fix(router): annotate cast for unpriced model info Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/router.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/litellm/router.py b/litellm/router.py index 8a543373692..9ff23c597f1 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -10641,7 +10641,9 @@ class Router: litellm_model_name_model_info.get("key") if litellm_model_name_model_info is not None else None ) if cost_map_omits_token_price(model_id, builtin_key, base_model_key): - return cast(ModelInfo, {**model_info, "input_cost_per_token": None, "output_cost_per_token": None}) + return cast( # cast-ok: TypedDict spread with overridden keys loses its type + ModelInfo, {**model_info, "input_cost_per_token": None, "output_cost_per_token": None} + ) return model_info def _set_model_group_info(self, model_group: str, user_facing_model_group_name: str) -> ModelGroupInfo | None: From 1e0613d8cae45e088bf122f8adfe42abdba5ab36 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 13:00:13 -0700 Subject: [PATCH 115/141] chore(vscode): pin @types/vscode to the 1.109 engine minimum --- vscode-extension/package-lock.json | 8 ++++---- vscode-extension/package.json | 2 +- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/vscode-extension/package-lock.json b/vscode-extension/package-lock.json index 378579dc024..a4f962f2039 100644 --- a/vscode-extension/package-lock.json +++ b/vscode-extension/package-lock.json @@ -13,7 +13,7 @@ }, "devDependencies": { "@types/node": "^22.20.3", - "@types/vscode": "^1.109.0", + "@types/vscode": "1.109.0", "@vscode/vsce": "^4.0.0", "esbuild": "^0.28.2", "typescript": "^5.9.3", @@ -1301,9 +1301,9 @@ } }, "node_modules/@types/vscode": { - "version": "1.138.0", - "resolved": "https://registry.npmjs.org/@types/vscode/-/vscode-1.138.0.tgz", - "integrity": "sha512-DhlvrucJa8n6EHst7qQcXOcjLKs/VRTQy55plLO+RN74jonYsr/t9dd66zBnAyz1tpDig0okQAxuoPHpVNje2g==", + "version": "1.109.0", + "resolved": "https://registry.npmjs.org/@types/vscode/-/vscode-1.109.0.tgz", + "integrity": "sha512-0Pf95rnwEIwDbmXGC08r0B4TQhAbsHQ5UyTIgVgoieDe4cOnf92usuR5dEczb6bTKEp7ziZH4TV1TRGPPCExtw==", "dev": true, "license": "MIT" }, diff --git a/vscode-extension/package.json b/vscode-extension/package.json index a8e52b14eda..67e3bb6c727 100644 --- a/vscode-extension/package.json +++ b/vscode-extension/package.json @@ -78,7 +78,7 @@ }, "devDependencies": { "@types/node": "^22.20.3", - "@types/vscode": "^1.109.0", + "@types/vscode": "1.109.0", "@vscode/vsce": "^4.0.0", "esbuild": "^0.28.2", "typescript": "^5.9.3", From 657cf18aea50d8825b629d95f9e74467d17391d9 Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 13:01:55 -0700 Subject: [PATCH 116/141] feat(rust): port exception_type to litellm-core-utils Adds a Rust port of litellm_core_utils/exception_mapping_utils.exception_type with its provider rule tables (OpenAI-compatible, Cohere, Vertex AI), the secret redaction patterns from secret_redaction.py, and a Python repr helper for messages that quote caller values. The public failure shapes are pinned by golden JSON fixtures under tests/test_litellm/rust_bridge/fixtures/public_failures, which the Python side reads too once the bridge is wired to this module. Nothing calls the port yet. --- litellm-rust/Cargo.lock | 29 + litellm-rust/Cargo.toml | 1 + litellm-rust/crates/core-utils/Cargo.toml | 5 + .../src/exception_mapping_utils/cohere.rs | 232 ++++++ .../src/exception_mapping_utils/mod.rs | 734 ++++++++++++++++++ .../src/exception_mapping_utils/openai.rs | 586 ++++++++++++++ .../src/exception_mapping_utils/original.rs | 49 ++ .../src/exception_mapping_utils/public.rs | 262 +++++++ .../src/exception_mapping_utils/rules.rs | 403 ++++++++++ .../src/exception_mapping_utils/status.rs | 153 ++++ .../src/exception_mapping_utils/vertex_ai.rs | 574 ++++++++++++++ litellm-rust/crates/core-utils/src/lib.rs | 3 + .../crates/core-utils/src/python_repr.rs | 93 +++ .../crates/core-utils/src/secret_redaction.rs | 97 +++ .../fixtures/public_failures/api.json | 13 + .../public_failures/api_connection.json | 11 + .../status_authentication.json | 13 + .../public_failures/status_bad_gateway.json | 28 + .../public_failures/status_bad_request.json | 28 + .../status_content_policy_violation.json | 28 + .../status_context_window_exceeded.json | 28 + .../status_internal_server.json | 19 + .../public_failures/status_not_found.json | 28 + .../status_permission_denied.json | 19 + .../public_failures/status_rate_limit.json | 28 + .../status_service_unavailable.json | 28 + .../status_unsupported_params.json | 28 + .../public_failures/timeout_with_status.json | 12 + .../timeout_without_status.json | 12 + 29 files changed, 3544 insertions(+) create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs create mode 100644 litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs create mode 100644 litellm-rust/crates/core-utils/src/python_repr.rs create mode 100644 litellm-rust/crates/core-utils/src/secret_redaction.rs create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/api.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json create mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json diff --git a/litellm-rust/Cargo.lock b/litellm-rust/Cargo.lock index 3b9ff62fbaa..6c524124550 100644 --- a/litellm-rust/Cargo.lock +++ b/litellm-rust/Cargo.lock @@ -559,6 +559,21 @@ dependencies = [ "vsimd", ] +[[package]] +name = "bit-set" +version = "0.8.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "08807e080ed7f9d5433fa9b275196cfc35414f66a0c79d864dc51a0d825231a3" +dependencies = [ + "bit-vec", +] + +[[package]] +name = "bit-vec" +version = "0.8.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "5e764a1d40d510daf35e07be9eb06e75770908c27d411ee6c92109c9840eaaf7" + [[package]] name = "bitflags" version = "2.13.1" @@ -1166,6 +1181,17 @@ dependencies = [ "pin-project-lite", ] +[[package]] +name = "fancy-regex" +version = "0.19.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d301f5bf187b3c295fce6468d3875037a0bccc5f6b151c63cac2f85babf21912" +dependencies = [ + "bit-set", + "regex-automata", + "regex-syntax", +] + [[package]] name = "fastrand" version = "2.5.0" @@ -2059,11 +2085,14 @@ dependencies = [ name = "litellm-core-utils" version = "0.1.0" dependencies = [ + "fancy-regex", "litellm-types", + "rstest", "serde", "serde_json", "serde_path_to_error", "serde_with", + "strum", "thiserror 2.0.19", "url", ] diff --git a/litellm-rust/Cargo.toml b/litellm-rust/Cargo.toml index f8377138050..ffdbf64bb49 100644 --- a/litellm-rust/Cargo.toml +++ b/litellm-rust/Cargo.toml @@ -50,6 +50,7 @@ strum = { version = "0.28.0", features = ["derive"] } url = "2.5.8" time = { version = "0.3.53", features = ["parsing"] } criterion = "0.8.2" +fancy-regex = "0.19.2" veil = "0.3.0" [profile.release] diff --git a/litellm-rust/crates/core-utils/Cargo.toml b/litellm-rust/crates/core-utils/Cargo.toml index 109c3312727..59e0ee1a09d 100644 --- a/litellm-rust/crates/core-utils/Cargo.toml +++ b/litellm-rust/crates/core-utils/Cargo.toml @@ -6,10 +6,15 @@ license.workspace = true repository.workspace = true [dependencies] +fancy-regex.workspace = true litellm-types.workspace = true serde.workspace = true serde_json.workspace = true serde_path_to_error = "0.1" serde_with.workspace = true +strum.workspace = true thiserror.workspace = true url.workspace = true + +[dev-dependencies] +rstest.workspace = true diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs new file mode 100644 index 00000000000..381ebd7ad27 --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs @@ -0,0 +1,232 @@ +use super::Mapping; +use super::public::{PublicFailure, StatusClass}; +use super::rules::{Kind, ResponseChoice, Rule, apply, contains_any}; + +const fn with_response(class: StatusClass) -> Kind { + Kind::Status { + class, + response: ResponseChoice::Provider, + } +} + +fn original(mapping: &Mapping<'_>) -> String { + format!("CohereException - {}", mapping.original.message) +} + +fn status_is(mapping: &Mapping<'_>, statuses: &[u16]) -> bool { + mapping + .original + .status + .is_some_and(|status| statuses.contains(&status)) +} + +/// `_map_cohere_exception`, in its branch order. A failure no rule claims falls through to +/// the status table. +const RULES: &[Rule] = &[ + Rule { + when: |mapping| { + contains_any( + &mapping.error_str, + &["invalid api token", "No API key provided."], + ) + }, + kind: with_response(StatusClass::Authentication), + message: original, + debug: false, + }, + Rule { + when: |mapping| mapping.error_str.contains("invalid type: parameter"), + kind: with_response(StatusClass::BadRequest), + message: original, + debug: false, + }, + Rule { + when: |mapping| mapping.error_str.contains("too many tokens"), + kind: with_response(StatusClass::ContextWindowExceeded), + message: original, + debug: false, + }, + Rule { + when: |mapping| { + mapping + .error_str + .to_lowercase() + .contains("internal server error") + }, + kind: with_response(StatusClass::InternalServer), + message: |mapping| format!("CohereException - {}", mapping.error_str), + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, &[400, 498]), + kind: with_response(StatusClass::BadRequest), + message: original, + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, &[408]), + kind: Kind::Timeout(None), + message: original, + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, &[500]), + kind: with_response(StatusClass::InternalServer), + message: original, + debug: false, + }, +]; + +pub(super) fn map(mapping: &Mapping<'_>) -> Option { + apply(RULES, mapping).map(|failure| PublicFailure { + llm_provider: Some("cohere".to_string()), + ..failure + }) +} + +#[cfg(test)] +mod tests { + use super::super::testing::{context, failure, http, status, upstream}; + use super::super::{ExceptionFamily, OriginalException, PublicKind}; + use super::*; + + fn mapped(provider: &str, original: &OriginalException) -> Option { + let context = context(provider, ExceptionFamily::Cohere); + map(&Mapping::new(&context, original)) + } + + fn cohere(class: StatusClass, status_code: u16, body: &str, message: &str) -> PublicFailure { + failure( + status(class, upstream(status_code, body)), + message, + "cohere", + ) + } + + #[rstest::rstest] + #[case::invalid_token( + 500, + "invalid api token", + cohere( + StatusClass::Authentication, + 500, + "invalid api token", + "CohereException - invalid api token" + ) + )] + #[case::no_api_key( + 500, + "No API key provided.", + cohere( + StatusClass::Authentication, + 500, + "No API key provided.", + "CohereException - No API key provided." + ) + )] + #[case::invalid_parameter( + 500, + "invalid type: parameter x", + cohere( + StatusClass::BadRequest, + 500, + "invalid type: parameter x", + "CohereException - invalid type: parameter x" + ) + )] + #[case::too_many_tokens( + 500, + "too many tokens", + cohere( + StatusClass::ContextWindowExceeded, + 500, + "too many tokens", + "CohereException - too many tokens" + ) + )] + #[case::internal_server_text( + 400, + "Internal Server Error", + cohere( + StatusClass::InternalServer, + 400, + "Internal Server Error", + "CohereException - Internal Server Error" + ) + )] + #[case::bad_request( + 400, + "rejected", + cohere(StatusClass::BadRequest, 400, "rejected", "CohereException - rejected") + )] + #[case::invalid_token_status( + 498, + "rejected", + cohere(StatusClass::BadRequest, 498, "rejected", "CohereException - rejected") + )] + #[case::request_timeout(408, "rejected", failure(PublicKind::Timeout { status: None }, "CohereException - rejected", "cohere"))] + #[case::internal_server( + 500, + "rejected", + cohere( + StatusClass::InternalServer, + 500, + "rejected", + "CohereException - rejected" + ) + )] + fn each_rule_maps_and_reports_cohere( + #[case] status_code: u16, + #[case] body: &str, + #[case] expected: PublicFailure, + ) { + assert_eq!(mapped("azure_ai", &http(status_code, body)), Some(expected)); + } + + #[rstest::rstest] + #[case::unmapped_status(409)] + #[case::unauthorized(401)] + fn statuses_without_a_rule_fall_through(#[case] status_code: u16) { + assert_eq!(mapped("cohere", &http(status_code, "rejected")), None); + } + + #[test] + fn the_internal_server_rule_uses_the_redacted_text() { + let body = "internal server error Bearer abcdefghijklmnop"; + assert_eq!( + mapped("cohere", &http(400, body)), + Some(cohere( + StatusClass::InternalServer, + 400, + body, + "CohereException - internal server error REDACTED" + )) + ); + } + + #[rstest::rstest] + #[case::token_before_parameter( + "invalid api token invalid type: parameter", + StatusClass::Authentication + )] + #[case::parameter_before_tokens( + "invalid type: parameter too many tokens", + StatusClass::BadRequest + )] + #[case::tokens_before_internal( + "too many tokens Internal Server Error", + StatusClass::ContextWindowExceeded + )] + #[case::internal_before_status("Internal Server Error", StatusClass::InternalServer)] + fn the_earlier_rule_wins_when_two_apply(#[case] body: &str, #[case] class: StatusClass) { + assert_eq!( + mapped("cohere", &http(400, body)), + Some(cohere( + class, + 400, + body, + &format!("CohereException - {body}") + )) + ); + } +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs new file mode 100644 index 00000000000..8892fba1399 --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs @@ -0,0 +1,734 @@ +//! A port of Python's `exception_type` for the routes that run in Rust. +//! +//! KNOWN_GAPS: differences from the Python mapper that no Rust route can reach today. Each +//! one stops being acceptable at its trigger. +//! - The Vertex partner-model API base for "claude" models is not built into +//! `extra_information`. Trigger: a Vertex route whose models include Anthropic partner +//! models; then `api_base` gets that branch and a table row. +//! - Python reports the provider `get_llm_provider` resolves for a stripped model name when +//! that name happens to be in the model cost map. Trigger: a route whose model names +//! overlap the cost map; that needs the provider resolution port, not a classifier change. +//! - The generic `APIConnectionError` fallback appends `traceback.format_exc()` to the +//! message. Rust has no Python traceback and does not invent one; a sweep row that reaches +//! it compares the message before the traceback. + +use super::secret_redaction::{redact_string, secret_redaction_enabled}; + +mod cohere; +mod openai; +mod original; +mod public; +mod rules; +mod status; +mod vertex_ai; + +pub use original::{ExceptionFamily, LocalClass, OriginalException}; +pub use public::{HttpStub, PublicFailure, PublicKind, ResponseArg, StatusClass, UpstreamResponse}; + +const DOCS_URL: &str = "https://docs.litellm.ai/docs"; + +const TIMEOUT_MARKERS: &[&str] = &[ + "Request Timeout Error", + "Request timed out", + "Timed out generating response", + "The read operation timed out", +]; + +#[derive(Clone, Debug, Default, PartialEq, Eq)] +pub struct ExceptionContext { + pub model: String, + pub custom_llm_provider: Option, + pub family: ExceptionFamily, + pub asynchronous: bool, + pub suppress_debug_info: bool, + pub redact_messages_in_exceptions: bool, + pub vertex_project: Option, + pub vertex_location: Option, + pub model_group: Option, + pub deployment: Option, + pub user_api_key_alias: Option, + pub user_api_key_team_alias: Option, +} + +/// The attributes `exception_type` reads off the Python exception: a provider error +/// (`BaseLLMException`) carries a status, a response and a request, a plain exception +/// carries only its text. +struct Raised { + status: Option, + status_is_synthesized: bool, + message: String, + response: Option, +} + +impl Raised { + fn provider( + status: u16, + message: String, + body: String, + headers: Vec<(String, String)>, + ) -> Self { + Self { + status: Some(status), + status_is_synthesized: false, + message, + response: Some(UpstreamResponse { + status, + body, + headers, + }), + } + } + + fn plain(message: String) -> Self { + Self { + status: None, + status_is_synthesized: false, + message, + response: None, + } + } + + fn new(original: &OriginalException, asynchronous: bool) -> Self { + match original { + OriginalException::Http { + status, + body, + headers, + } => Self::provider(*status, body.clone(), body.clone(), headers.clone()), + OriginalException::Connection { message } => Self { + status_is_synthesized: true, + ..Self::provider(500, message.clone(), String::new(), Vec::new()) + }, + OriginalException::Timeout { + timeout_seconds, + elapsed_seconds, + } => Self::provider( + 408, + timeout_message(asynchronous, *timeout_seconds, *elapsed_seconds), + String::new(), + Vec::new(), + ), + OriginalException::Response { message } + | OriginalException::Local { message, .. } + | OriginalException::Public { message, .. } => Self::plain(message.clone()), + } + } +} + +/// The text `litellm.Timeout` carries when the Python HTTP handler times out: the sync +/// and async handlers word it differently. +fn timeout_message( + asynchronous: bool, + timeout_seconds: Option, + elapsed_seconds: Option, +) -> String { + let timeout = python_float(timeout_seconds); + if asynchronous { + let elapsed = + python_float(elapsed_seconds.map(|seconds| (seconds * 1000.0).round() / 1000.0)); + format!( + "litellm.Timeout: Connection timed out. Timeout passed={timeout}, time taken={elapsed} seconds" + ) + } else { + format!("litellm.Timeout: Connection timed out after {timeout} seconds.") + } +} + +fn python_float(value: Option) -> String { + match value { + None => "None".to_string(), + Some(value) if value.fract() == 0.0 => format!("{value:.1}"), + Some(value) => value.to_string(), + } +} + +/// Everything the rules read: the original as Python sees it and the text `exception_type` +/// derives from the context before any provider mapper runs. +struct Mapping<'a> { + context: &'a ExceptionContext, + original: Raised, + provider: &'a str, + error_str: String, + exception_provider: String, + extra_information: String, +} + +impl<'a> Mapping<'a> { + fn new(context: &'a ExceptionContext, original: &OriginalException) -> Self { + let original = Raised::new(original, context.asynchronous); + let error_str = if secret_redaction_enabled() { + redact_string(&original.message) + } else { + original.message.clone() + }; + Self { + context, + original, + provider: context.custom_llm_provider.as_deref().unwrap_or_default(), + error_str, + exception_provider: match &context.custom_llm_provider { + None => "None".to_string(), + Some(provider) => exception_provider(provider), + }, + extra_information: extra_information(context, api_base(context).as_deref()), + } + } + + fn failure(&self, kind: PublicKind, message: String, debug: bool) -> PublicFailure { + PublicFailure { + kind, + message, + model: self.context.model.clone(), + llm_provider: self.context.custom_llm_provider.clone(), + litellm_debug_info: debug.then(|| self.extra_information.clone()), + litellm_response_headers: None, + print_banner: false, + } + } +} + +pub fn exception_type(context: &ExceptionContext, original: &OriginalException) -> PublicFailure { + if let OriginalException::Public { class, message } = original { + return PublicFailure { + kind: PublicKind::Status { + status_class: *class, + response: None, + }, + message: message.clone(), + model: context.model.clone(), + llm_provider: context.custom_llm_provider.clone(), + litellm_debug_info: None, + litellm_response_headers: None, + print_banner: false, + }; + } + let mapping = Mapping::new(context, original); + let litellm_response_headers = mapping + .original + .response + .as_ref() + .map(|response| response.headers.clone()) + .filter(|headers| !headers.is_empty()); + PublicFailure { + litellm_response_headers, + print_banner: !context.suppress_debug_info, + ..map(&mapping) + } +} + +fn map(mapping: &Mapping<'_>) -> PublicFailure { + if rules::contains_any(&mapping.error_str, TIMEOUT_MARKERS) { + return mapping.failure( + PublicKind::Timeout { status: None }, + format!( + "APITimeoutError - Request timed out. Error_str: {}", + mapping.error_str + ), + true, + ); + } + let provider_failure = match mapping.context.family { + ExceptionFamily::OpenAiCompatible => openai::map(mapping), + ExceptionFamily::VertexAi => vertex_ai::map(mapping), + ExceptionFamily::Cohere => cohere::map(mapping), + ExceptionFamily::Other => None, + }; + provider_failure + .or_else(|| status::map(mapping)) + .unwrap_or_else(|| unmapped(mapping)) +} + +/// The `APIConnectionError` Python raises when no mapper claimed the failure: with the +/// provider prefix for a provider error, with the bare text for a plain exception. +fn unmapped(mapping: &Mapping<'_>) -> PublicFailure { + let message = match mapping.original.status { + Some(_) => format!("{} - {}", mapping.exception_provider, mapping.error_str), + None => mapping.original.message.clone(), + }; + mapping.failure(PublicKind::ApiConnection, message, false) +} + +fn exception_provider(provider: &str) -> String { + let mut characters = provider.chars(); + match characters.next() { + Some(first) => format!("{}{}Exception", first.to_uppercase(), characters.as_str()), + None => String::new(), + } +} + +fn python_capitalize(value: &str) -> String { + let mut characters = value.chars(); + match characters.next() { + Some(first) => format!( + "{}{}", + first.to_uppercase(), + characters.as_str().to_lowercase() + ), + None => String::new(), + } +} + +fn api_base(context: &ExceptionContext) -> Option { + match (&context.vertex_location, &context.vertex_project) { + (Some(location), Some(project)) => Some(format!( + "{location}-aiplatform.googleapis.com/v1/projects/{project}/locations/{location}/publishers/google/models/{}:generateContent", + context.model + )), + _ => None, + } +} + +fn extra_information(context: &ExceptionContext, api_base: Option<&str>) -> String { + let lines = [ + Some(format!("\nModel: {}", context.model)), + api_base.map(|api_base| format!("\nAPI Base: `{api_base}`")), + (!context.redact_messages_in_exceptions).then(|| "\nMessages: `None`".to_string()), + context + .model_group + .as_ref() + .map(|value| format!("\nmodel_group: `{value}`\n")), + context + .deployment + .as_ref() + .map(|value| format!("\ndeployment: `{value}`\n")), + context + .vertex_project + .as_ref() + .map(|value| format!("\nvertex_project: `{value}`\n")), + context + .vertex_location + .as_ref() + .map(|value| format!("\nvertex_location: `{value}`\n")), + ]; + let information: String = lines.into_iter().flatten().collect(); + match &context.user_api_key_alias { + Some(alias) => format!( + "\n\nKey Name: `{alias}`\nTeam: `{}`{information}", + context.user_api_key_team_alias.as_deref().unwrap_or("None") + ), + None => information, + } +} + +#[cfg(test)] +mod testing { + use super::*; + + pub(super) const DEBUG: &str = "\nModel: ocr-model\nMessages: `None`"; + + pub(super) fn context(provider: &str, family: ExceptionFamily) -> ExceptionContext { + ExceptionContext { + model: "ocr-model".into(), + custom_llm_provider: Some(provider.into()), + family, + suppress_debug_info: true, + ..ExceptionContext::default() + } + } + + pub(super) fn http(status: u16, body: &str) -> OriginalException { + OriginalException::Http { + status, + body: body.into(), + headers: vec![("retry-after".into(), "7".into())], + } + } + + pub(super) fn upstream(status: u16, body: &str) -> Option { + Some(ResponseArg::Upstream(UpstreamResponse { + status, + body: body.into(), + headers: vec![("retry-after".into(), "7".into())], + })) + } + + pub(super) fn status(class: StatusClass, response: Option) -> PublicKind { + PublicKind::Status { + status_class: class, + response, + } + } + + /// The failure a rule builds before `exception_type` adds the response headers and the + /// banner flag. + pub(super) fn failure(kind: PublicKind, message: &str, provider: &str) -> PublicFailure { + PublicFailure { + kind, + message: message.into(), + model: "ocr-model".into(), + llm_provider: Some(provider.into()), + litellm_debug_info: None, + litellm_response_headers: None, + print_banner: false, + } + } + + pub(super) fn with_debug(failure: PublicFailure) -> PublicFailure { + PublicFailure { + litellm_debug_info: Some(DEBUG.into()), + ..failure + } + } + + /// What `exception_type` returns for an `http` original the rule mapped to `failure`. + pub(super) fn with_headers(failure: PublicFailure) -> PublicFailure { + PublicFailure { + litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), + ..failure + } + } +} + +#[cfg(test)] +mod tests { + use super::testing::{DEBUG, context, failure, http, status, upstream, with_debug}; + use super::*; + + fn openai() -> ExceptionContext { + context("mistral", ExceptionFamily::OpenAiCompatible) + } + + #[test] + fn a_public_original_passes_through_without_banner_debug_or_prefix() { + let original = OriginalException::Public { + class: StatusClass::UnsupportedParams, + message: "Invalid `req_format`".into(), + }; + let context = ExceptionContext { + suppress_debug_info: false, + ..openai() + }; + assert_eq!( + exception_type(&context, &original), + failure( + status(StatusClass::UnsupportedParams, None), + "Invalid `req_format`", + "mistral" + ) + ); + } + + #[rstest::rstest] + #[case::vertex_family_status_rule(ExceptionFamily::VertexAi, "vertex_ai", PublicFailure { + litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), + ..with_debug(failure(status(StatusClass::BadRequest, upstream(409, "rejected")), "Vertex_aiException - rejected", "vertex_ai")) + })] + #[case::cohere_family(ExceptionFamily::Cohere, "cohere", PublicFailure { + litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), + ..with_debug(failure(status(StatusClass::BadRequest, upstream(409, "rejected")), "CohereException - rejected", "cohere")) + })] + #[case::other_family(ExceptionFamily::Other, "reducto", PublicFailure { + litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), + ..with_debug(failure(status(StatusClass::BadRequest, upstream(409, "rejected")), "ReductoException - rejected", "reducto")) + })] + fn families_without_a_409_rule_reach_the_status_table( + #[case] family: ExceptionFamily, + #[case] provider: &str, + #[case] expected: PublicFailure, + ) { + assert_eq!( + exception_type(&context(provider, family), &http(409, "rejected")), + expected + ); + } + + #[test] + fn the_openai_family_claims_a_409_before_the_status_table() { + assert_eq!( + exception_type(&openai(), &http(409, "rejected")), + PublicFailure { + litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), + ..with_debug(failure( + PublicKind::Api { + status: 409, + request_url: DOCS_URL + }, + "APIError: MistralException - rejected", + "mistral" + )) + } + ); + } + + #[rstest::rstest] + #[case::request_timeout_error("Request Timeout Error")] + #[case::request_timed_out("Request timed out")] + #[case::timed_out_generating("Timed out generating response")] + #[case::read_operation("The read operation timed out")] + fn timeout_markers_win_over_every_family(#[case] marker: &str) { + let body = format!("rate limit {marker}"); + for family in [ + ExceptionFamily::OpenAiCompatible, + ExceptionFamily::VertexAi, + ExceptionFamily::Cohere, + ExceptionFamily::Other, + ] { + assert_eq!( + exception_type(&context("mistral", family), &http(429, &body)), + PublicFailure { + litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), + ..with_debug(failure( + PublicKind::Timeout { status: None }, + &format!("APITimeoutError - Request timed out. Error_str: {body}"), + "mistral" + )) + } + ); + } + } + + #[rstest::rstest] + #[case::provider_error_keeps_the_prefix(http(409, "rejected"), "ReductoException - rejected")] + #[case::synthesized_status_skips_the_status_table( + OriginalException::Connection { message: "refused".into() }, + "ReductoException - refused" + )] + #[case::plain_exception_keeps_its_text( + OriginalException::Local { class: LocalClass::FileNotFound, message: "File not found: /a".into() }, + "File not found: /a" + )] + fn unmapped_failures_are_connection_errors( + #[case] original: OriginalException, + #[case] message: &str, + ) { + let context = context("reducto", ExceptionFamily::Other); + let expected = failure(PublicKind::ApiConnection, message, "reducto"); + let actual = exception_type(&context, &original); + assert_eq!( + PublicFailure { + litellm_response_headers: None, + ..actual + }, + expected + ); + } + + #[test] + fn a_missing_provider_renders_like_python_none() { + let context = ExceptionContext { + custom_llm_provider: None, + family: ExceptionFamily::Other, + ..openai() + }; + assert_eq!( + exception_type(&context, &http(401, "rejected")), + PublicFailure { + llm_provider: None, + litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), + ..with_debug(failure( + status(StatusClass::Authentication, upstream(401, "rejected")), + "None - rejected", + "unused" + )) + } + ); + } + + #[rstest::rstest] + #[case::suppressed(true, false)] + #[case::printed(false, true)] + fn the_banner_prints_unless_debug_info_is_suppressed( + #[case] suppress_debug_info: bool, + #[case] print_banner: bool, + ) { + let context = ExceptionContext { + suppress_debug_info, + ..openai() + }; + assert_eq!( + exception_type(&context, &http(400, "rejected")).print_banner, + print_banner + ); + } + + #[test] + fn empty_upstream_headers_are_not_reported() { + let original = OriginalException::Http { + status: 400, + body: "rejected".into(), + headers: Vec::new(), + }; + assert_eq!( + exception_type(&openai(), &original).litellm_response_headers, + None + ); + } + + #[test] + fn messages_are_redacted_before_markers_and_prefixes() { + let body = "rejected Bearer abcdefghijklmnop"; + assert_eq!( + exception_type( + &context("reducto", ExceptionFamily::Other), + &http(400, body) + ) + .message, + "ReductoException - rejected REDACTED" + ); + } + + const SYNC_TIMEOUT: &str = "litellm.Timeout: Connection timed out after 0.5 seconds."; + + #[rstest::rstest] + #[case::sync(false, Some(0.5), Some(0.5031), SYNC_TIMEOUT)] + #[case::async_rounds_the_elapsed_time( + true, + Some(0.5), + Some(0.5031), + "litellm.Timeout: Connection timed out. Timeout passed=0.5, time taken=0.503 seconds" + )] + #[case::whole_seconds_keep_a_decimal( + true, + Some(600.0), + Some(2.0), + "litellm.Timeout: Connection timed out. Timeout passed=600.0, time taken=2.0 seconds" + )] + #[case::unknown_values_render_as_none( + true, + None, + None, + "litellm.Timeout: Connection timed out. Timeout passed=None, time taken=None seconds" + )] + fn timeout_text_follows_the_delivery_mode( + #[case] asynchronous: bool, + #[case] timeout_seconds: Option, + #[case] elapsed_seconds: Option, + #[case] expected: &str, + ) { + assert_eq!( + timeout_message(asynchronous, timeout_seconds, elapsed_seconds), + expected + ); + } + + #[rstest::rstest] + #[case::sync(false, SYNC_TIMEOUT)] + #[case::async_( + true, + "litellm.Timeout: Connection timed out. Timeout passed=0.5, time taken=0.503 seconds" + )] + fn a_timeout_is_a_408_carrying_the_handler_text( + #[case] asynchronous: bool, + #[case] text: &str, + ) { + let context = ExceptionContext { + asynchronous, + ..openai() + }; + let original = OriginalException::Timeout { + timeout_seconds: Some(0.5), + elapsed_seconds: Some(0.5031), + }; + assert_eq!( + exception_type(&context, &original), + with_debug(failure( + PublicKind::Timeout { status: None }, + &format!("Timeout Error: MistralException - {text}"), + "mistral" + )) + ); + } + + #[test] + fn a_refused_connection_is_a_500_with_an_empty_response() { + assert_eq!( + exception_type( + &openai(), + &OriginalException::Connection { + message: "refused".into() + } + ), + with_debug(failure( + status( + StatusClass::InternalServer, + Some(ResponseArg::Upstream(UpstreamResponse { + status: 500, + body: String::new(), + headers: Vec::new(), + })) + ), + "InternalServerError: MistralException - refused", + "mistral" + )) + ); + } + + #[test] + fn debug_information_follows_the_python_layout() { + let context = ExceptionContext { + vertex_project: Some("project".into()), + vertex_location: Some("region".into()), + model_group: Some("ocr".into()), + deployment: Some("deployment".into()), + user_api_key_alias: Some("key".into()), + ..openai() + }; + assert_eq!( + extra_information(&context, api_base(&context).as_deref()), + concat!( + "\n\nKey Name: `key`\nTeam: `None`", + "\nModel: ocr-model", + "\nAPI Base: `region-aiplatform.googleapis.com/v1/projects/project/locations/region/publishers/google/models/ocr-model:generateContent`", + "\nMessages: `None`", + "\nmodel_group: `ocr`\n", + "\ndeployment: `deployment`\n", + "\nvertex_project: `project`\n", + "\nvertex_location: `region`\n", + ) + ); + } + + #[rstest::rstest] + #[case::bare(ExceptionContext::default(), "\nModel: ")] + #[case::redacted_messages(ExceptionContext { redact_messages_in_exceptions: true, model: "m".into(), ..ExceptionContext::default() }, "\nModel: m")] + #[case::messages(ExceptionContext { model: "m".into(), ..ExceptionContext::default() }, "\nModel: m\nMessages: `None`")] + #[case::team_alias( + ExceptionContext { model: "m".into(), user_api_key_alias: Some("key".into()), user_api_key_team_alias: Some("team".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + "\n\nKey Name: `key`\nTeam: `team`\nModel: m" + )] + #[case::team_alias_without_key_is_ignored( + ExceptionContext { model: "m".into(), user_api_key_team_alias: Some("team".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + "\nModel: m" + )] + #[case::project_without_location_has_no_api_base( + ExceptionContext { model: "m".into(), vertex_project: Some("p".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + "\nModel: m\nvertex_project: `p`\n" + )] + #[case::location_without_project_has_no_api_base( + ExceptionContext { model: "m".into(), vertex_location: Some("l".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + "\nModel: m\nvertex_location: `l`\n" + )] + fn each_optional_context_field_adds_its_own_line( + #[case] context: ExceptionContext, + #[case] expected: &str, + ) { + assert_eq!( + extra_information(&context, api_base(&context).as_deref()), + expected + ); + } + + #[rstest::rstest] + #[case::lowercase("mistral", "MistralException")] + #[case::keeps_the_rest("azure_ai", "Azure_aiException")] + #[case::empty("", "")] + fn exception_provider_capitalizes_only_the_first_letter( + #[case] provider: &str, + #[case] expected: &str, + ) { + assert_eq!(exception_provider(provider), expected); + } + + #[rstest::rstest] + #[case::lowers_the_rest("vERTEX_AI", "Vertex_ai")] + #[case::empty("", "")] + fn python_capitalize_lowers_the_rest(#[case] value: &str, #[case] expected: &str) { + assert_eq!(python_capitalize(value), expected); + } + + #[test] + fn debug_constant_matches_the_default_test_context() { + let context = openai(); + assert_eq!(extra_information(&context, None), DEBUG); + } +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs new file mode 100644 index 00000000000..ffbb172582b --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs @@ -0,0 +1,586 @@ +use super::public::{PublicFailure, StatusClass}; +use super::rules::{ + ApiStatus, Kind, ResponseChoice, Rule, apply, contains_any, is_context_window_exceeded, + is_rate_limit, +}; +use super::{DOCS_URL, Mapping}; + +const OPENAI_URL: &str = "https://api.openai.com/v1"; + +const ENCRYPTED_CONTENT_HELP: &str = "\n\n This error occurs when load balancing Responses API across deployments with different API keys.\n Encrypted content is tied to the organization that created it and cannot be decrypted by other organizations.\n\n Solution: Enable 'encrypted_content_affinity' to route follow-up requests to the correct deployment:\n\n router_settings:\n enable_pre_call_checks: true\n optional_pre_call_checks:\n - encrypted_content_affinity\n\n Learn more: https://docs.litellm.ai/docs/response_api#encrypted-content-affinity-multi-region-load-balancing"; + +const fn with_response(class: StatusClass) -> Kind { + Kind::Status { + class, + response: ResponseChoice::Provider, + } +} + +fn exception_provider(mapping: &Mapping<'_>) -> String { + if mapping.provider == "openai" { + "OpenAIException".to_string() + } else { + super::exception_provider(mapping.provider) + } +} + +/// The raw message with OpenAI's own names swapped for the provider's. +fn message(mapping: &Mapping<'_>) -> String { + let provider = mapping.provider; + mapping + .original + .message + .replace("OPENAI", &provider.to_uppercase()) + .replace("openai.OpenAIError", &format!("{provider}.{provider}Error")) +} + +fn prefixed(mapping: &Mapping<'_>, label: &str) -> String { + format!( + "{label}{} - {}", + exception_provider(mapping), + message(mapping) + ) +} + +fn status_is(mapping: &Mapping<'_>, statuses: &[u16]) -> bool { + mapping + .original + .status + .is_some_and(|status| statuses.contains(&status)) +} + +/// `_map_openai_exception`, in its branch order. +const RULES: &[Rule] = &[ + Rule { + when: |mapping| is_rate_limit(&mapping.error_str, mapping.original.status), + kind: with_response(StatusClass::RateLimit), + message: |mapping| prefixed(mapping, "RateLimitError: "), + debug: false, + }, + Rule { + when: |mapping| is_context_window_exceeded(&mapping.error_str), + kind: with_response(StatusClass::ContextWindowExceeded), + message: |mapping| prefixed(mapping, "ContextWindowExceededError: "), + debug: true, + }, + Rule { + when: |mapping| { + mapping.error_str.contains("invalid_request_error") + && mapping.error_str.contains("model_not_found") + }, + kind: with_response(StatusClass::NotFound), + message: |mapping| prefixed(mapping, ""), + debug: true, + }, + Rule { + when: |mapping| mapping.error_str.contains("A timeout occurred"), + kind: Kind::Timeout(None), + message: |mapping| prefixed(mapping, ""), + debug: true, + }, + Rule { + when: |mapping| { + let error_str = &mapping.error_str; + (error_str.contains("invalid_request_error") + && error_str.contains("content_policy_violation")) + || (error_str.contains("Invalid prompt") + && error_str.contains("violating our usage policy")) + || error_str + .to_lowercase() + .contains("request was rejected as a result of the safety system") + }, + kind: with_response(StatusClass::ContentPolicyViolation), + message: |mapping| prefixed(mapping, "ContentPolicyViolationError: "), + debug: true, + }, + Rule { + when: |mapping| { + contains_any( + &mapping.error_str, + &["invalid_encrypted_content", "could not be verified"], + ) + }, + kind: with_response(StatusClass::BadRequest), + message: |mapping| { + format!( + "{} - {}{ENCRYPTED_CONTENT_HELP}", + exception_provider(mapping), + message(mapping) + ) + }, + debug: true, + }, + Rule { + when: |mapping| { + mapping.error_str.contains("invalid_request_error") + && !mapping.error_str.contains("Incorrect API key provided") + }, + kind: with_response(StatusClass::BadRequest), + message: |mapping| prefixed(mapping, ""), + debug: true, + }, + Rule { + when: |mapping| { + contains_any( + &mapping.error_str, + &[ + "Web server is returning an unknown error", + "The server had an error processing your request.", + ], + ) + }, + kind: Kind::Status { + class: StatusClass::InternalServer, + response: ResponseChoice::Omitted, + }, + message: |mapping| prefixed(mapping, ""), + debug: false, + }, + Rule { + when: |mapping| mapping.error_str.contains("Request too large"), + kind: with_response(StatusClass::RateLimit), + message: |mapping| prefixed(mapping, "RateLimitError: "), + debug: true, + }, + Rule { + when: |mapping| { + mapping.error_str.contains("The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable") + }, + kind: with_response(StatusClass::Authentication), + message: |mapping| prefixed(mapping, "AuthenticationError: "), + debug: true, + }, + Rule { + when: |mapping| { + mapping + .error_str + .contains("Mistral API raised a streaming error") + }, + kind: Kind::Api { + status: ApiStatus::Fixed(500), + request_url: OPENAI_URL, + }, + message: |mapping| prefixed(mapping, ""), + debug: true, + }, + Rule { + when: |mapping| mapping.original.status.is_none(), + kind: Kind::ApiConnection, + message: |mapping| prefixed(mapping, "APIConnectionError: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[400, 422]), + kind: with_response(StatusClass::BadRequest), + message: |mapping| prefixed(mapping, ""), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[401]), + kind: with_response(StatusClass::Authentication), + message: |mapping| prefixed(mapping, "AuthenticationError: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[404]), + kind: with_response(StatusClass::NotFound), + message: |mapping| prefixed(mapping, "NotFoundError: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[408]), + kind: Kind::Timeout(None), + message: |mapping| prefixed(mapping, "Timeout Error: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[429]), + kind: with_response(StatusClass::RateLimit), + message: |mapping| prefixed(mapping, "RateLimitError: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[500]), + kind: with_response(StatusClass::InternalServer), + message: |mapping| prefixed(mapping, "InternalServerError: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[502]), + kind: with_response(StatusClass::BadGateway), + message: |mapping| prefixed(mapping, "BadGatewayError: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[503]), + kind: with_response(StatusClass::ServiceUnavailable), + message: |mapping| prefixed(mapping, "ServiceUnavailableError: "), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, &[504]), + kind: Kind::Timeout(Some(504)), + message: |mapping| prefixed(mapping, "Timeout Error: "), + debug: true, + }, + Rule { + when: |_| true, + kind: Kind::Api { + status: ApiStatus::Original, + request_url: DOCS_URL, + }, + message: |mapping| prefixed(mapping, "APIError: "), + debug: true, + }, +]; + +pub(super) fn map(mapping: &Mapping<'_>) -> Option { + apply(RULES, mapping) +} + +#[cfg(test)] +mod tests { + use super::super::testing::{context, failure, http, status, upstream, with_debug}; + use super::super::{ExceptionFamily, OriginalException, PublicKind}; + use super::*; + + fn mapped(provider: &str, original: &OriginalException) -> PublicFailure { + let context = context(provider, ExceptionFamily::OpenAiCompatible); + map(&Mapping::new(&context, original)).expect("the OpenAI table ends in a catch-all") + } + + fn kind(class: StatusClass, status_code: u16, body: &str) -> PublicKind { + status(class, upstream(status_code, body)) + } + + #[rstest::rstest] + #[case::rate_limit_phrase( + 400, + "rate limit reached", + failure( + kind(StatusClass::RateLimit, 400, "rate limit reached"), + "RateLimitError: MistralException - rate limit reached", + "mistral", + ) + )] + #[case::context_window( + 500, + "This model's maximum context length is 10", + with_debug(failure( + kind( + StatusClass::ContextWindowExceeded, + 500, + "This model's maximum context length is 10" + ), + "ContextWindowExceededError: MistralException - This model's maximum context length is 10", + "mistral", + )) + )] + #[case::model_not_found( + 400, + "invalid_request_error model_not_found", + with_debug(failure( + kind(StatusClass::NotFound, 400, "invalid_request_error model_not_found"), + "MistralException - invalid_request_error model_not_found", + "mistral", + )) + )] + #[case::timeout_occurred(400, "A timeout occurred", with_debug(failure( + PublicKind::Timeout { status: None }, + "MistralException - A timeout occurred", + "mistral", + )))] + #[case::content_policy_error_code( + 400, + "invalid_request_error content_policy_violation", + with_debug(failure( + kind( + StatusClass::ContentPolicyViolation, + 400, + "invalid_request_error content_policy_violation" + ), + "ContentPolicyViolationError: MistralException - invalid_request_error content_policy_violation", + "mistral", + )) + )] + #[case::content_policy_usage_policy( + 400, + "Invalid prompt violating our usage policy", + with_debug(failure( + kind( + StatusClass::ContentPolicyViolation, + 400, + "Invalid prompt violating our usage policy" + ), + "ContentPolicyViolationError: MistralException - Invalid prompt violating our usage policy", + "mistral", + )) + )] + #[case::content_policy_safety_system( + 400, + "Request was rejected as a result of the safety system", + with_debug(failure( + kind( + StatusClass::ContentPolicyViolation, + 400, + "Request was rejected as a result of the safety system" + ), + "ContentPolicyViolationError: MistralException - Request was rejected as a result of the safety system", + "mistral", + )) + )] + #[case::encrypted_content(400, "invalid_encrypted_content", with_debug(failure( + kind(StatusClass::BadRequest, 400, "invalid_encrypted_content"), + &format!("MistralException - invalid_encrypted_content{ENCRYPTED_CONTENT_HELP}"), + "mistral", + )))] + #[case::unverifiable_content(400, "could not be verified", with_debug(failure( + kind(StatusClass::BadRequest, 400, "could not be verified"), + &format!("MistralException - could not be verified{ENCRYPTED_CONTENT_HELP}"), + "mistral", + )))] + #[case::invalid_request( + 429, + "invalid_request_error bad field", + with_debug(failure( + kind(StatusClass::BadRequest, 429, "invalid_request_error bad field"), + "MistralException - invalid_request_error bad field", + "mistral", + )) + )] + #[case::unknown_server_error( + 400, + "Web server is returning an unknown error", + failure( + status(StatusClass::InternalServer, None), + "MistralException - Web server is returning an unknown error", + "mistral", + ) + )] + #[case::server_had_an_error( + 400, + "The server had an error processing your request.", + failure( + status(StatusClass::InternalServer, None), + "MistralException - The server had an error processing your request.", + "mistral", + ) + )] + #[case::request_too_large( + 400, + "Request too large", + with_debug(failure( + kind(StatusClass::RateLimit, 400, "Request too large"), + "RateLimitError: MistralException - Request too large", + "mistral", + )) + )] + #[case::missing_client_api_key( + 400, + "The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable", + with_debug(failure( + kind( + StatusClass::Authentication, + 400, + "The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable", + ), + "AuthenticationError: MistralException - The api_key client option must be set either by passing api_key to the client or by setting the MISTRAL_API_KEY environment variable", + "mistral", + )) + )] + #[case::mistral_streaming_error(400, "Mistral API raised a streaming error", with_debug(failure( + PublicKind::Api { status: 500, request_url: OPENAI_URL }, + "MistralException - Mistral API raised a streaming error", + "mistral", + )))] + fn each_text_rule_maps_by_the_body( + #[case] status_code: u16, + #[case] body: &str, + #[case] expected: PublicFailure, + ) { + assert_eq!(mapped("mistral", &http(status_code, body)), expected); + } + + #[rstest::rstest] + #[case::bad_request( + 400, + kind(StatusClass::BadRequest, 400, "rejected"), + "MistralException - rejected" + )] + #[case::unprocessable( + 422, + kind(StatusClass::BadRequest, 422, "rejected"), + "MistralException - rejected" + )] + #[case::authentication( + 401, + kind(StatusClass::Authentication, 401, "rejected"), + "AuthenticationError: MistralException - rejected" + )] + #[case::not_found( + 404, + kind(StatusClass::NotFound, 404, "rejected"), + "NotFoundError: MistralException - rejected" + )] + #[case::request_timeout(408, PublicKind::Timeout { status: None }, "Timeout Error: MistralException - rejected")] + #[case::rate_limited( + 429, + kind(StatusClass::RateLimit, 429, "rejected"), + "RateLimitError: MistralException - rejected" + )] + #[case::internal_server( + 500, + kind(StatusClass::InternalServer, 500, "rejected"), + "InternalServerError: MistralException - rejected" + )] + #[case::bad_gateway( + 502, + kind(StatusClass::BadGateway, 502, "rejected"), + "BadGatewayError: MistralException - rejected" + )] + #[case::service_unavailable( + 503, + kind(StatusClass::ServiceUnavailable, 503, "rejected"), + "ServiceUnavailableError: MistralException - rejected" + )] + #[case::gateway_timeout(504, PublicKind::Timeout { status: Some(504) }, "Timeout Error: MistralException - rejected")] + #[case::any_other_status(409, PublicKind::Api { status: 409, request_url: DOCS_URL }, "APIError: MistralException - rejected")] + fn each_status_rule_maps_by_the_status( + #[case] status_code: u16, + #[case] kind: PublicKind, + #[case] message: &str, + ) { + assert_eq!( + mapped("mistral", &http(status_code, "rejected")), + with_debug(failure(kind, message, "mistral")) + ); + } + + #[test] + fn a_failure_without_a_status_is_a_connection_error() { + let original = OriginalException::Response { + message: "invalid OCR response field: pages".into(), + }; + assert_eq!( + mapped("mistral", &original), + with_debug(failure( + PublicKind::ApiConnection, + "APIConnectionError: MistralException - invalid OCR response field: pages", + "mistral" + )) + ); + } + + #[rstest::rstest] + #[case::rate_limit_before_context_window( + 400, + "rate limit and This model's maximum context length is 10", + kind( + StatusClass::RateLimit, + 400, + "rate limit and This model's maximum context length is 10" + ), + "RateLimitError: MistralException - rate limit and This model's maximum context length is 10", + false + )] + #[case::context_window_before_content_policy( + 400, + "This model's maximum context length is 10 invalid_request_error content_policy_violation", + kind( + StatusClass::ContextWindowExceeded, + 400, + "This model's maximum context length is 10 invalid_request_error content_policy_violation" + ), + "ContextWindowExceededError: MistralException - This model's maximum context length is 10 invalid_request_error content_policy_violation", + true + )] + #[case::model_not_found_before_invalid_request( + 400, + "invalid_request_error model_not_found", + kind(StatusClass::NotFound, 400, "invalid_request_error model_not_found"), + "MistralException - invalid_request_error model_not_found", + true + )] + #[case::timeout_before_invalid_request( + 400, + "A timeout occurred invalid_request_error", + PublicKind::Timeout { status: None }, + "MistralException - A timeout occurred invalid_request_error", + true + )] + #[case::content_policy_before_invalid_request( + 400, + "invalid_request_error content_policy_violation", + kind( + StatusClass::ContentPolicyViolation, + 400, + "invalid_request_error content_policy_violation" + ), + "ContentPolicyViolationError: MistralException - invalid_request_error content_policy_violation", + true + )] + #[case::invalid_request_with_a_bad_key_falls_to_the_status( + 401, + "invalid_request_error Incorrect API key provided", + kind( + StatusClass::Authentication, + 401, + "invalid_request_error Incorrect API key provided" + ), + "AuthenticationError: MistralException - invalid_request_error Incorrect API key provided", + true + )] + #[case::text_rules_before_status( + 429, + "invalid_request_error bad field", + kind(StatusClass::BadRequest, 429, "invalid_request_error bad field"), + "MistralException - invalid_request_error bad field", + true + )] + #[case::echoed_429_is_not_a_rate_limit( + 400, + "token 429 in the prompt", + kind(StatusClass::BadRequest, 400, "token 429 in the prompt"), + "MistralException - token 429 in the prompt", + true + )] + fn the_earlier_rule_wins_when_two_apply( + #[case] status_code: u16, + #[case] body: &str, + #[case] kind: PublicKind, + #[case] message: &str, + #[case] debug: bool, + ) { + let expected = failure(kind, message, "mistral"); + assert_eq!( + mapped("mistral", &http(status_code, body)), + if debug { + with_debug(expected) + } else { + expected + } + ); + } + + #[rstest::rstest] + #[case::provider_names_replace_openai( + "azure_ai", + "OPENAI said openai.OpenAIError", + "Azure_aiException - AZURE_AI said azure_ai.azure_aiError" + )] + #[case::openai_keeps_its_own_name("openai", "rejected", "OpenAIException - rejected")] + fn the_message_names_the_provider( + #[case] provider: &str, + #[case] body: &str, + #[case] message: &str, + ) { + assert_eq!( + mapped(provider, &http(400, body)), + with_debug(failure( + kind(StatusClass::BadRequest, 400, body), + message, + provider + )) + ); + } +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs new file mode 100644 index 00000000000..d64868c1606 --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs @@ -0,0 +1,49 @@ +use super::public::StatusClass; + +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum LocalClass { + ValueError, + FileNotFound, + OsError, +} + +/// A route failure in the shape Python's `exception_type` receives it, before any public +/// class is chosen. +#[derive(Clone, Debug, PartialEq)] +pub enum OriginalException { + Http { + status: u16, + body: String, + headers: Vec<(String, String)>, + }, + Connection { + message: String, + }, + Timeout { + timeout_seconds: Option, + elapsed_seconds: Option, + }, + Response { + message: String, + }, + Local { + class: LocalClass, + message: String, + }, + /// A failure Python raises as a public LiteLLM exception itself, which `exception_type` + /// hands back unchanged. + Public { + class: StatusClass, + message: String, + }, +} + +/// Which of the provider-specific mappers in `exception_type` a route's provider uses. +#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)] +pub enum ExceptionFamily { + OpenAiCompatible, + VertexAi, + Cohere, + #[default] + Other, +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs new file mode 100644 index 00000000000..a7319c1287b --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs @@ -0,0 +1,262 @@ +use serde::Serialize; + +/// The public LiteLLM classes built from a status code alone: every one takes the same +/// constructor arguments. +#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, strum::EnumIter, strum::IntoStaticStr)] +#[serde(rename_all = "snake_case")] +#[strum(serialize_all = "snake_case")] +pub enum StatusClass { + BadRequest, + Authentication, + PermissionDenied, + NotFound, + RateLimit, + ContextWindowExceeded, + ContentPolicyViolation, + InternalServer, + BadGateway, + ServiceUnavailable, + UnsupportedParams, +} + +impl StatusClass { + /// The `status_code` the Python class sets on itself. + pub const fn status_code(self) -> u16 { + match self { + Self::BadRequest + | Self::ContextWindowExceeded + | Self::ContentPolicyViolation + | Self::UnsupportedParams => 400, + Self::Authentication => 401, + Self::PermissionDenied => 403, + Self::NotFound => 404, + Self::RateLimit => 429, + Self::InternalServer => 500, + Self::BadGateway => 502, + Self::ServiceUnavailable => 503, + } + } +} + +#[derive(Clone, Debug, PartialEq, Eq, Serialize)] +pub struct UpstreamResponse { + pub status: u16, + pub body: String, + pub headers: Vec<(String, String)>, +} + +#[derive(Clone, Debug, PartialEq, Eq, Serialize)] +pub struct HttpStub { + pub status: u16, + pub method: &'static str, + pub url: &'static str, + pub content: Option, +} + +#[derive(Clone, Debug, PartialEq, Eq, Serialize)] +#[serde(tag = "type", rename_all = "snake_case")] +pub enum ResponseArg { + Upstream(UpstreamResponse), + Stub(HttpStub), +} + +#[derive(Clone, Debug, PartialEq, Eq, Serialize)] +#[serde(tag = "type", rename_all = "snake_case")] +pub enum PublicKind { + Status { + status_class: StatusClass, + response: Option, + }, + Timeout { + status: Option, + }, + ApiConnection, + Api { + status: u16, + request_url: &'static str, + }, +} + +/// Constructor arguments for the public LiteLLM exception, as `exception_type` passes them. +#[derive(Clone, Debug, PartialEq, Eq, Serialize)] +pub struct PublicFailure { + pub kind: PublicKind, + pub message: String, + pub model: String, + pub llm_provider: Option, + pub litellm_debug_info: Option, + pub litellm_response_headers: Option>, + pub print_banner: bool, +} + +#[cfg(test)] +mod tests { + use std::collections::BTreeSet; + use std::path::PathBuf; + + use serde_json::Value; + use strum::IntoEnumIterator; + + use super::*; + + const REGENERATE: &str = "LITELLM_REGENERATE_PUBLIC_FAILURE_FIXTURES"; + + fn fixture_directory() -> PathBuf { + PathBuf::from(env!("CARGO_MANIFEST_DIR")) + .join("../../../tests/test_litellm/rust_bridge/fixtures/public_failures") + } + + fn upstream(status: u16) -> ResponseArg { + ResponseArg::Upstream(UpstreamResponse { + status, + body: r#"{"message": "rejected"}"#.into(), + headers: vec![("retry-after".into(), "7".into())], + }) + } + + fn status_response(class: StatusClass) -> Option { + match class { + StatusClass::Authentication => None, + StatusClass::PermissionDenied => Some(ResponseArg::Stub(HttpStub { + status: 403, + method: "POST", + url: " https://cloud.google.com/vertex-ai/", + content: None, + })), + StatusClass::InternalServer => Some(ResponseArg::Stub(HttpStub { + status: 500, + method: "completion", + url: "https://github.com/BerriAI/litellm", + content: Some("upstream text".into()), + })), + class => Some(upstream(class.status_code())), + } + } + + fn failure(kind: PublicKind, name: &str) -> PublicFailure { + let headers = matches!( + &kind, + PublicKind::Status { + response: Some(ResponseArg::Upstream(_)), + .. + } + ); + PublicFailure { + kind, + message: format!("MistralException - {name}"), + model: "ocr-model".into(), + llm_provider: Some("mistral".into()), + litellm_debug_info: Some("\nModel: ocr-model".into()), + litellm_response_headers: headers.then(|| vec![("retry-after".into(), "7".into())]), + print_banner: false, + } + } + + /// One payload per public class the constructor can build; `test_failures.py` reads the + /// same files, so a shape change on either side fails there or here. + fn fixtures() -> Vec<(String, PublicFailure)> { + let statuses = StatusClass::iter().map(|class| { + let name: &'static str = class.into(); + let name = format!("status_{name}"); + let built = failure( + PublicKind::Status { + status_class: class, + response: status_response(class), + }, + &name, + ); + (name, built) + }); + let others = [ + ( + "timeout_with_status", + PublicFailure { + print_banner: true, + ..failure( + PublicKind::Timeout { status: Some(504) }, + "timeout_with_status", + ) + }, + ), + ( + "timeout_without_status", + PublicFailure { + litellm_debug_info: None, + ..failure( + PublicKind::Timeout { status: None }, + "timeout_without_status", + ) + }, + ), + ( + "api_connection", + PublicFailure { + llm_provider: None, + ..failure(PublicKind::ApiConnection, "api_connection") + }, + ), + ( + "api", + failure( + PublicKind::Api { + status: 409, + request_url: "https://docs.litellm.ai/docs", + }, + "api", + ), + ), + ] + .map(|(name, built)| (name.to_string(), built)); + statuses.chain(others).collect() + } + + #[test] + fn serialized_payloads_match_the_golden_fixtures_python_reads() { + let directory = fixture_directory(); + let regenerate = std::env::var_os(REGENERATE).is_some(); + let expected = fixtures(); + for (name, built) in &expected { + let path = directory.join(format!("{name}.json")); + let serialized = serde_json::to_value(built).unwrap(); + if regenerate { + std::fs::create_dir_all(&directory).unwrap(); + std::fs::write( + &path, + format!("{}\n", serde_json::to_string_pretty(&serialized).unwrap()), + ) + .unwrap(); + } + let golden: Value = + serde_json::from_str(&std::fs::read_to_string(&path).unwrap()).unwrap(); + assert_eq!(serialized, golden, "{name}; set {REGENERATE}=1 to rewrite"); + } + let on_disk: BTreeSet = std::fs::read_dir(&directory) + .unwrap() + .map(|entry| entry.unwrap().file_name().to_string_lossy().into_owned()) + .collect(); + let generated: BTreeSet = expected + .iter() + .map(|(name, _)| format!("{name}.json")) + .collect(); + assert_eq!(on_disk, generated); + } + + #[rstest::rstest] + #[case(StatusClass::BadRequest, 400)] + #[case(StatusClass::Authentication, 401)] + #[case(StatusClass::PermissionDenied, 403)] + #[case(StatusClass::NotFound, 404)] + #[case(StatusClass::RateLimit, 429)] + #[case(StatusClass::ContextWindowExceeded, 400)] + #[case(StatusClass::ContentPolicyViolation, 400)] + #[case(StatusClass::InternalServer, 500)] + #[case(StatusClass::BadGateway, 502)] + #[case(StatusClass::ServiceUnavailable, 503)] + #[case(StatusClass::UnsupportedParams, 400)] + fn status_codes_are_the_ones_the_python_classes_set( + #[case] class: StatusClass, + #[case] status: u16, + ) { + assert_eq!(class.status_code(), status); + } +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs new file mode 100644 index 00000000000..5df8ad991b5 --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs @@ -0,0 +1,403 @@ +use std::sync::LazyLock; + +use fancy_regex::Regex; +use serde_json::Value; + +use super::Mapping; +use super::public::{HttpStub, PublicFailure, PublicKind, ResponseArg, StatusClass}; + +const GITHUB_URL: &str = "https://github.com/BerriAI/litellm"; + +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub(super) enum ResponseChoice { + Omitted, + Provider, + Stub { status: u16, url: &'static str }, + InternalServerStub, +} + +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub(super) enum ApiStatus { + Fixed(u16), + Original, +} + +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub(super) enum Kind { + Status { + class: StatusClass, + response: ResponseChoice, + }, + Timeout(Option), + ApiConnection, + Api { + status: ApiStatus, + request_url: &'static str, + }, +} + +/// One branch of a Python `_map_*_exception` function: when it applies, the class it +/// raises, the message it builds, and whether it passes `litellm_debug_info`. +pub(super) struct Rule { + pub(super) when: fn(&Mapping<'_>) -> bool, + pub(super) kind: Kind, + pub(super) message: fn(&Mapping<'_>) -> String, + pub(super) debug: bool, +} + +/// The first rule that applies decides the failure, as the `if`/`elif` chain does in Python. +pub(super) fn apply(rules: &[Rule], mapping: &Mapping<'_>) -> Option { + rules + .iter() + .find(|rule| (rule.when)(mapping)) + .map(|rule| rule.build(mapping)) +} + +impl Rule { + fn build(&self, mapping: &Mapping<'_>) -> PublicFailure { + let kind = match self.kind { + Kind::Status { class, response } => PublicKind::Status { + status_class: class, + response: response.resolve(mapping), + }, + Kind::Timeout(status) => PublicKind::Timeout { status }, + Kind::ApiConnection => PublicKind::ApiConnection, + Kind::Api { + status, + request_url, + } => PublicKind::Api { + status: match status { + ApiStatus::Fixed(status) => status, + ApiStatus::Original => mapping.original.status.unwrap_or(500), + }, + request_url, + }, + }; + PublicFailure { + kind, + message: (self.message)(mapping), + model: mapping.context.model.clone(), + llm_provider: mapping.context.custom_llm_provider.clone(), + litellm_debug_info: self.debug.then(|| mapping.extra_information.clone()), + litellm_response_headers: None, + print_banner: false, + } + } +} + +impl ResponseChoice { + fn resolve(self, mapping: &Mapping<'_>) -> Option { + match self { + Self::Omitted => None, + Self::Provider => mapping.original.response.clone().map(ResponseArg::Upstream), + Self::Stub { status, url } => Some(ResponseArg::Stub(HttpStub { + status, + method: "POST", + url, + content: None, + })), + Self::InternalServerStub => Some(ResponseArg::Stub(HttpStub { + status: 500, + method: "completion", + url: GITHUB_URL, + content: Some(mapping.original.message.clone()), + })), + } + } +} + +pub(super) fn contains_any(text: &str, markers: &[&str]) -> bool { + markers.iter().any(|marker| text.contains(marker)) +} + +static STANDALONE_429: LazyLock = + LazyLock::new(|| Regex::new(r"\b429\b").expect("valid regex")); +static RATE_LIMIT_PHRASE: LazyLock = + LazyLock::new(|| Regex::new(r"rate[\s_\-]*limit").expect("valid regex")); + +/// `ExceptionCheckers.is_error_str_rate_limit`. +pub(super) fn is_rate_limit(error_str: &str, status: Option) -> bool { + if STANDALONE_429.is_match(error_str).unwrap_or(false) && status == Some(429) { + return true; + } + let lower = error_str.to_lowercase(); + RATE_LIMIT_PHRASE.is_match(&lower).unwrap_or(false) + || lower.contains("service tier capacity exceeded") +} + +/// `ExceptionCheckers.is_error_str_context_window_exceeded`. +pub(super) fn is_context_window_exceeded(error_str: &str) -> bool { + let lower = error_str.to_lowercase(); + if lower.contains("string_above_max_length") { + return false; + } + if lower.contains("invalid 'user'") && lower.contains("string too long") { + return false; + } + contains_any( + &lower, + &[ + "exceed context limit", + "this model's maximum context length is", + "string too long. expected a string with maximum length", + "model's maximum context limit", + "is longer than the model's context length", + "input tokens exceed the configured limit", + "`inputs` tokens + `max_new_tokens` must be", + "exceeds the available context size", + "exceeds the maximum number of tokens allowed", + ], + ) || (lower.contains("current length is") && lower.contains("while limit is")) + || (lower.contains("maximum input length is") && lower.contains("tokens")) +} + +/// The integer `error.code` of a JSON error body, read the way Python's `int()` would. +pub(super) fn body_error_code(error_str: &str) -> Option { + let body: Value = serde_json::from_str(error_str).ok()?; + let Some(Value::Object(error)) = body.as_object()?.get("error") else { + return None; + }; + match error.get("code")? { + Value::Number(number) => number + .as_i64() + .or_else(|| number.as_f64().map(|value| value.trunc() as i64)), + Value::String(code) => code.trim().replace('_', "").parse().ok(), + Value::Bool(flag) => Some(i64::from(*flag)), + _ => None, + } +} + +#[cfg(test)] +mod tests { + use super::super::testing::{context, failure, http}; + use super::super::{ExceptionFamily, OriginalException, UpstreamResponse}; + use super::*; + + fn first_marker(mapping: &Mapping<'_>) -> bool { + mapping.error_str.contains("first") + } + + fn always(_: &Mapping<'_>) -> bool { + true + } + + fn text(mapping: &Mapping<'_>) -> String { + format!("seen {}", mapping.error_str) + } + + const ORDERED: &[Rule] = &[ + Rule { + when: first_marker, + kind: Kind::Status { + class: StatusClass::NotFound, + response: ResponseChoice::Omitted, + }, + message: text, + debug: false, + }, + Rule { + when: always, + kind: Kind::ApiConnection, + message: text, + debug: true, + }, + ]; + + fn apply_one(kind: Kind, debug: bool, original: &OriginalException) -> Option { + let context = context("mistral", ExceptionFamily::OpenAiCompatible); + let mapping = Mapping::new(&context, original); + apply( + &[Rule { + when: always, + kind, + message: text, + debug, + }], + &mapping, + ) + } + + #[rstest::rstest] + #[case::earlier_rule_wins("first and second", failure( + PublicKind::Status { status_class: StatusClass::NotFound, response: None }, + "seen first and second", + "mistral", + ))] + #[case::later_rule_when_the_earlier_does_not_apply("second", PublicFailure { + litellm_debug_info: Some("\nModel: ocr-model\nMessages: `None`".into()), + ..failure(PublicKind::ApiConnection, "seen second", "mistral") + })] + fn the_first_applicable_rule_decides(#[case] body: &str, #[case] expected: PublicFailure) { + let context = context("mistral", ExceptionFamily::OpenAiCompatible); + let original = http(400, body); + assert_eq!( + apply(ORDERED, &Mapping::new(&context, &original)), + Some(expected) + ); + } + + #[test] + fn no_applicable_rule_leaves_the_failure_to_the_caller() { + let context = context("mistral", ExceptionFamily::OpenAiCompatible); + let original = http(400, "second"); + assert_eq!( + apply(&ORDERED[..1], &Mapping::new(&context, &original)), + None + ); + } + + #[rstest::rstest] + #[case::omitted(ResponseChoice::Omitted, None)] + #[case::provider(ResponseChoice::Provider, Some(ResponseArg::Upstream(UpstreamResponse { + status: 400, + body: "body".into(), + headers: vec![("retry-after".into(), "7".into())], + })))] + #[case::stub( + ResponseChoice::Stub { status: 429, url: "https://stub.test" }, + Some(ResponseArg::Stub(HttpStub { status: 429, method: "POST", url: "https://stub.test", content: None })) + )] + #[case::internal_server_stub( + ResponseChoice::InternalServerStub, + Some(ResponseArg::Stub(HttpStub { + status: 500, + method: "completion", + url: GITHUB_URL, + content: Some("body".into()), + })) + )] + fn response_choices_resolve_against_the_original( + #[case] response: ResponseChoice, + #[case] expected: Option, + ) { + let built = apply_one( + Kind::Status { + class: StatusClass::BadRequest, + response, + }, + false, + &http(400, "body"), + ) + .unwrap(); + assert_eq!( + built.kind, + PublicKind::Status { + status_class: StatusClass::BadRequest, + response: expected, + } + ); + } + + #[rstest::rstest] + #[case::fixed(ApiStatus::Fixed(500), http(409, "body"), 500)] + #[case::original(ApiStatus::Original, http(409, "body"), 409)] + #[case::original_without_a_status( + ApiStatus::Original, + OriginalException::Response { message: "body".into() }, + 500 + )] + fn api_status_is_fixed_or_the_originals( + #[case] status: ApiStatus, + #[case] original: OriginalException, + #[case] expected: u16, + ) { + let built = apply_one( + Kind::Api { + status, + request_url: "https://api.test", + }, + false, + &original, + ) + .unwrap(); + assert_eq!( + built, + failure( + PublicKind::Api { + status: expected, + request_url: "https://api.test" + }, + "seen body", + "mistral" + ) + ); + } + + #[rstest::rstest] + #[case::with_debug(true, Some("\nModel: ocr-model\nMessages: `None`"))] + #[case::without_debug(false, None)] + fn debug_rules_carry_the_extra_information( + #[case] debug: bool, + #[case] expected: Option<&str>, + ) { + let built = apply_one(Kind::Timeout(Some(504)), debug, &http(504, "body")).unwrap(); + assert_eq!( + built, + PublicFailure { + litellm_debug_info: expected.map(str::to_string), + ..failure( + PublicKind::Timeout { status: Some(504) }, + "seen body", + "mistral" + ) + } + ); + } + + #[rstest::rstest] + #[case::standalone_429_with_429_status("got 429 back", Some(429), true)] + #[case::standalone_429_with_other_status("got 429 back", Some(400), false)] + #[case::embedded_429("token4290", Some(429), false)] + #[case::phrase_spaced("Rate Limit reached", None, true)] + #[case::phrase_underscored("rate_limit", None, true)] + #[case::phrase_hyphenated("rate-limit", None, true)] + #[case::service_tier("Service tier capacity exceeded", None, true)] + #[case::unrelated("rejected", Some(429), false)] + fn rate_limit_detection( + #[case] text: &str, + #[case] status: Option, + #[case] expected: bool, + ) { + assert_eq!(is_rate_limit(text, status), expected); + } + + #[rstest::rstest] + #[case::exceed_context_limit("Exceed context limit", true)] + #[case::maximum_context_length("This model's maximum context length is 10", true)] + #[case::string_too_long("string too long. Expected a string with maximum length 5", true)] + #[case::maximum_context_limit("the model's maximum context limit", true)] + #[case::longer_than_context("prompt is longer than the model's context length", true)] + #[case::configured_limit("input tokens exceed the configured limit", true)] + #[case::max_new_tokens("`inputs` tokens + `max_new_tokens` must be <= 10", true)] + #[case::available_context("exceeds the available context size", true)] + #[case::maximum_tokens("exceeds the maximum number of tokens allowed", true)] + #[case::current_and_limit("current length is 9 while limit is 8", true)] + #[case::current_without_limit("current length is 9", false)] + #[case::maximum_input_tokens("maximum input length is 8 tokens", true)] + #[case::maximum_input_without_tokens("maximum input length is 8", false)] + #[case::string_above_max_length_wins("string_above_max_length exceed context limit", false)] + #[case::user_field_is_not_context( + "invalid 'user': string too long. expected a string with maximum length", + false + )] + #[case::unrelated("rejected", false)] + fn context_window_detection(#[case] text: &str, #[case] expected: bool) { + assert_eq!(is_context_window_exceeded(text), expected); + } + + #[rstest::rstest] + #[case::integer(r#"{"error": {"code": 429}}"#, Some(429))] + #[case::float(r#"{"error": {"code": 429.9}}"#, Some(429))] + #[case::string(r#"{"error": {"code": " 4_29 "}}"#, Some(429))] + #[case::boolean(r#"{"error": {"code": true}}"#, Some(1))] + #[case::unparseable_string(r#"{"error": {"code": "slow"}}"#, None)] + #[case::null(r#"{"error": {"code": null}}"#, None)] + #[case::no_code(r#"{"error": {}}"#, None)] + #[case::error_not_an_object(r#"{"error": "429"}"#, None)] + #[case::no_error(r#"{"code": 429}"#, None)] + #[case::not_an_object("[429]", None)] + #[case::not_json("429", None)] + fn body_error_code_reads_the_nested_code(#[case] body: &str, #[case] expected: Option) { + assert_eq!(body_error_code(body), expected); + } +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs new file mode 100644 index 00000000000..3d817fe9ae2 --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs @@ -0,0 +1,153 @@ +use super::public::{PublicFailure, StatusClass}; +use super::rules::{ApiStatus, Kind, ResponseChoice, Rule, apply}; +use super::{DOCS_URL, Mapping}; + +const fn with_response(class: StatusClass) -> Kind { + Kind::Status { + class, + response: ResponseChoice::Provider, + } +} + +fn message(mapping: &Mapping<'_>) -> String { + format!("{} - {}", mapping.exception_provider, mapping.error_str) +} + +fn status(mapping: &Mapping<'_>) -> u16 { + mapping.original.status.unwrap_or_default() +} + +/// `_map_exception_by_status`, the fallback for a provider error no provider mapper claimed. +const RULES: &[Rule] = &[ + Rule { + when: |mapping| status(mapping) == 401, + kind: with_response(StatusClass::Authentication), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 403, + kind: with_response(StatusClass::PermissionDenied), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 404, + kind: with_response(StatusClass::NotFound), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 408, + kind: Kind::Timeout(None), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 429, + kind: with_response(StatusClass::RateLimit), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 500, + kind: with_response(StatusClass::InternalServer), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 502, + kind: with_response(StatusClass::BadGateway), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 503, + kind: with_response(StatusClass::ServiceUnavailable), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) == 504, + kind: Kind::Timeout(Some(504)), + message, + debug: true, + }, + Rule { + when: |mapping| status(mapping) < 500, + kind: with_response(StatusClass::BadRequest), + message, + debug: true, + }, + Rule { + when: |_| true, + kind: Kind::Api { + status: ApiStatus::Original, + request_url: DOCS_URL, + }, + message, + debug: true, + }, +]; + +/// Only a real provider status of 400 or more reaches the table; a status the HTTP handler +/// synthesized for a failure without a response does not. +pub(super) fn map(mapping: &Mapping<'_>) -> Option { + let status = mapping.original.status?; + if status < 400 || mapping.original.status_is_synthesized { + return None; + } + apply(RULES, mapping) +} + +#[cfg(test)] +mod tests { + use super::super::testing::{context, failure, http, upstream, with_debug}; + use super::super::{ExceptionFamily, OriginalException, PublicKind}; + use super::*; + + fn mapped(original: &OriginalException) -> Option { + let context = context("reducto", ExceptionFamily::Other); + map(&Mapping::new(&context, original)) + } + + fn classified(class: StatusClass, status_code: u16) -> PublicKind { + PublicKind::Status { + status_class: class, + response: upstream(status_code, "rejected"), + } + } + + #[rstest::rstest] + #[case::authentication(401, classified(StatusClass::Authentication, 401))] + #[case::permission_denied(403, classified(StatusClass::PermissionDenied, 403))] + #[case::not_found(404, classified(StatusClass::NotFound, 404))] + #[case::request_timeout(408, PublicKind::Timeout { status: None })] + #[case::rate_limited(429, classified(StatusClass::RateLimit, 429))] + #[case::internal_server(500, classified(StatusClass::InternalServer, 500))] + #[case::bad_gateway(502, classified(StatusClass::BadGateway, 502))] + #[case::service_unavailable(503, classified(StatusClass::ServiceUnavailable, 503))] + #[case::gateway_timeout(504, PublicKind::Timeout { status: Some(504) })] + #[case::lowest_client_error(400, classified(StatusClass::BadRequest, 400))] + #[case::other_client_error(409, classified(StatusClass::BadRequest, 409))] + #[case::highest_client_error(499, classified(StatusClass::BadRequest, 499))] + #[case::other_server_error(501, PublicKind::Api { status: 501, request_url: DOCS_URL })] + fn every_mapped_status_and_the_fallback(#[case] status_code: u16, #[case] kind: PublicKind) { + assert_eq!( + mapped(&http(status_code, "rejected")), + Some(with_debug(failure( + kind, + "ReductoException - rejected", + "reducto" + ))) + ); + } + + #[rstest::rstest] + #[case::below_client_errors(http(399, "rejected"))] + #[case::synthesized(OriginalException::Connection { message: "refused".into() })] + #[case::no_status(OriginalException::Response { message: "bad body".into() })] + fn failures_the_table_does_not_claim(#[case] original: OriginalException) { + assert_eq!(mapped(&original), None); + } +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs new file mode 100644 index 00000000000..0fa8c19c4f6 --- /dev/null +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs @@ -0,0 +1,574 @@ +use super::public::{PublicFailure, PublicKind, ResponseArg, StatusClass, UpstreamResponse}; +use super::rules::{ + Kind, ResponseChoice, Rule, apply, body_error_code, contains_any, is_context_window_exceeded, +}; +use super::{Mapping, python_capitalize}; + +const VERTEX_URL: &str = "https://cloud.google.com/vertex-ai/"; +const VERTEX_URL_WITH_SPACE: &str = " https://cloud.google.com/vertex-ai/"; + +const QUOTA_MARKERS: &[&str] = &[ + "429 Quota exceeded", + "Quota exceeded for", + "Resource exhausted", + "IndexError: list index out of range", + "429 Unable to submit request because the service is temporarily out of capacity.", +]; + +const fn stubbed(class: StatusClass, status: u16, url: &'static str) -> Kind { + Kind::Status { + class, + response: ResponseChoice::Stub { status, url }, + } +} + +const fn bare(class: StatusClass) -> Kind { + Kind::Status { + class, + response: ResponseChoice::Omitted, + } +} + +/// `{Provider}Exception{label} - {error_str}` with Python's `str.capitalize()`. +fn capitalized(mapping: &Mapping<'_>, label: &str) -> String { + format!( + "{}Exception{label} - {}", + python_capitalize(mapping.provider), + mapping.error_str + ) +} + +/// `litellm.{Class}: {provider}Exception - {error_str}` with the provider as given. +fn litellm_prefixed(mapping: &Mapping<'_>, class: &str) -> String { + format!( + "litellm.{class}: {}Exception - {}", + mapping.provider, mapping.error_str + ) +} + +fn status_is(mapping: &Mapping<'_>, status: u16) -> bool { + mapping.original.status == Some(status) +} + +/// `_map_vertex_exception`, in its branch order. A failure no rule claims falls through +/// to the status table. +const RULES: &[Rule] = &[ + Rule { + when: |mapping| { + contains_any( + &mapping.error_str, + &[ + "Vertex AI API has not been used in project", + "Unable to find your project", + ], + ) + }, + kind: stubbed(StatusClass::BadRequest, 400, VERTEX_URL_WITH_SPACE), + message: |mapping| litellm_prefixed(mapping, "BadRequestError"), + debug: true, + }, + Rule { + when: |mapping| { + mapping + .error_str + .contains("400 Request payload size exceeds") + }, + kind: bare(StatusClass::ContextWindowExceeded), + message: |mapping| capitalized(mapping, ""), + debug: false, + }, + Rule { + when: |mapping| is_context_window_exceeded(&mapping.error_str), + kind: bare(StatusClass::ContextWindowExceeded), + message: |mapping| format!("ContextWindowExceededError: {}", capitalized(mapping, "")), + debug: true, + }, + Rule { + when: |mapping| { + contains_any( + &mapping.error_str, + &["None Unknown Error.", "Content has no parts."], + ) + }, + kind: Kind::Status { + class: StatusClass::InternalServer, + response: ResponseChoice::InternalServerStub, + }, + message: |mapping| litellm_prefixed(mapping, "InternalServerError"), + debug: true, + }, + Rule { + when: |mapping| mapping.error_str.contains("API key not valid."), + kind: bare(StatusClass::Authentication), + message: |mapping| capitalized(mapping, ""), + debug: true, + }, + Rule { + when: |mapping| mapping.error_str.contains("403"), + kind: stubbed(StatusClass::BadRequest, 403, VERTEX_URL_WITH_SPACE), + message: |mapping| capitalized(mapping, " BadRequestError"), + debug: true, + }, + Rule { + when: |mapping| { + contains_any( + &mapping.error_str, + &[ + "The response was blocked.", + "Output blocked by content filtering policy", + ], + ) + }, + kind: stubbed( + StatusClass::ContentPolicyViolation, + 400, + VERTEX_URL_WITH_SPACE, + ), + message: |mapping| capitalized(mapping, " ContentPolicyViolationError"), + debug: true, + }, + Rule { + when: |mapping| { + contains_any(&mapping.error_str, QUOTA_MARKERS) + || (mapping + .original + .status + .is_some_and(|status| (500..600).contains(&status)) + && body_error_code(&mapping.error_str) == Some(429)) + }, + kind: stubbed(StatusClass::RateLimit, 429, VERTEX_URL_WITH_SPACE), + message: |mapping| litellm_prefixed(mapping, "RateLimitError"), + debug: true, + }, + Rule { + when: |mapping| { + contains_any( + &mapping.error_str, + &["500 Internal Server Error", "The model is overloaded."], + ) + }, + kind: bare(StatusClass::InternalServer), + message: |mapping| litellm_prefixed(mapping, "InternalServerError"), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, 400), + kind: stubbed(StatusClass::BadRequest, 400, VERTEX_URL), + message: |mapping| capitalized(mapping, " BadRequestError"), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, 401), + kind: bare(StatusClass::Authentication), + message: |mapping| capitalized(mapping, ""), + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, 403), + kind: stubbed(StatusClass::PermissionDenied, 403, VERTEX_URL), + message: |mapping| capitalized(mapping, ""), + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, 404), + kind: bare(StatusClass::NotFound), + message: |mapping| capitalized(mapping, ""), + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, 408), + kind: Kind::Timeout(None), + message: |mapping| capitalized(mapping, ""), + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, 429), + kind: stubbed(StatusClass::RateLimit, 429, VERTEX_URL_WITH_SPACE), + message: |mapping| format!("litellm.RateLimitError: {}", capitalized(mapping, "")), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, 500), + kind: Kind::Status { + class: StatusClass::InternalServer, + response: ResponseChoice::InternalServerStub, + }, + message: |mapping| capitalized(mapping, " InternalServerError"), + debug: true, + }, + Rule { + when: |mapping| status_is(mapping, 502), + kind: Kind::ApiConnection, + message: |mapping| capitalized(mapping, ""), + debug: false, + }, + Rule { + when: |mapping| status_is(mapping, 503), + kind: bare(StatusClass::ServiceUnavailable), + message: |mapping| capitalized(mapping, ""), + debug: false, + }, +]; + +pub(super) fn map(mapping: &Mapping<'_>) -> Option { + apply(RULES, mapping).map(|failure| keep_upstream_response(mapping, failure)) +} + +/// Deliberate divergence from `_map_vertex_exception`, which replaces the provider response +/// with a stub and so drops the upstream body and `retry-after`. The response keeps the +/// status the public class carries. +fn keep_upstream_response(mapping: &Mapping<'_>, failure: PublicFailure) -> PublicFailure { + let (PublicKind::Status { status_class, .. }, Some(upstream), false) = ( + &failure.kind, + &mapping.original.response, + mapping.original.status_is_synthesized, + ) else { + return failure; + }; + PublicFailure { + kind: PublicKind::Status { + status_class: *status_class, + response: Some(ResponseArg::Upstream(UpstreamResponse { + status: status_class.status_code(), + ..upstream.clone() + })), + }, + ..failure + } +} + +#[cfg(test)] +mod tests { + use super::super::testing::{context, failure, http, status, upstream, with_debug}; + use super::super::{ExceptionFamily, HttpStub, OriginalException}; + use super::*; + + fn mapped(original: &OriginalException) -> Option { + let context = context("vertex_ai", ExceptionFamily::VertexAi); + map(&Mapping::new(&context, original)) + } + + fn kept(class: StatusClass, body: &str) -> PublicKind { + status(class, upstream(class.status_code(), body)) + } + + #[rstest::rstest] + #[case::api_not_enabled( + 400, + "Vertex AI API has not been used in project x", + with_debug(failure( + kept( + StatusClass::BadRequest, + "Vertex AI API has not been used in project x" + ), + "litellm.BadRequestError: vertex_aiException - Vertex AI API has not been used in project x", + "vertex_ai", + )) + )] + #[case::project_not_found( + 400, + "Unable to find your project", + with_debug(failure( + kept(StatusClass::BadRequest, "Unable to find your project"), + "litellm.BadRequestError: vertex_aiException - Unable to find your project", + "vertex_ai", + )) + )] + #[case::payload_too_large( + 400, + "400 Request payload size exceeds the limit", + failure( + kept( + StatusClass::ContextWindowExceeded, + "400 Request payload size exceeds the limit" + ), + "Vertex_aiException - 400 Request payload size exceeds the limit", + "vertex_ai", + ) + )] + #[case::context_window( + 500, + "This model's maximum context length is 10", + with_debug(failure( + kept( + StatusClass::ContextWindowExceeded, + "This model's maximum context length is 10" + ), + "ContextWindowExceededError: Vertex_aiException - This model's maximum context length is 10", + "vertex_ai", + )) + )] + #[case::unknown_error( + 400, + "None Unknown Error.", + with_debug(failure( + kept(StatusClass::InternalServer, "None Unknown Error."), + "litellm.InternalServerError: vertex_aiException - None Unknown Error.", + "vertex_ai", + )) + )] + #[case::no_parts( + 400, + "Content has no parts.", + with_debug(failure( + kept(StatusClass::InternalServer, "Content has no parts."), + "litellm.InternalServerError: vertex_aiException - Content has no parts.", + "vertex_ai", + )) + )] + #[case::api_key_not_valid( + 400, + "API key not valid.", + with_debug(failure( + kept(StatusClass::Authentication, "API key not valid."), + "Vertex_aiException - API key not valid.", + "vertex_ai", + )) + )] + #[case::forbidden_text( + 400, + "got a 403", + with_debug(failure( + kept(StatusClass::BadRequest, "got a 403"), + "Vertex_aiException BadRequestError - got a 403", + "vertex_ai", + )) + )] + #[case::response_blocked( + 400, + "The response was blocked.", + with_debug(failure( + kept(StatusClass::ContentPolicyViolation, "The response was blocked."), + "Vertex_aiException ContentPolicyViolationError - The response was blocked.", + "vertex_ai", + )) + )] + #[case::output_blocked( + 400, + "Output blocked by content filtering policy", + with_debug(failure( + kept( + StatusClass::ContentPolicyViolation, + "Output blocked by content filtering policy" + ), + "Vertex_aiException ContentPolicyViolationError - Output blocked by content filtering policy", + "vertex_ai", + )) + )] + #[case::quota_marker( + 400, + "Quota exceeded for aiplatform", + with_debug(failure( + kept(StatusClass::RateLimit, "Quota exceeded for aiplatform"), + "litellm.RateLimitError: vertex_aiException - Quota exceeded for aiplatform", + "vertex_ai", + )) + )] + #[case::wrapped_429( + 503, + r#"{"error": {"code": "429"}}"#, + with_debug(failure( + kept(StatusClass::RateLimit, r#"{"error": {"code": "429"}}"#), + r#"litellm.RateLimitError: vertex_aiException - {"error": {"code": "429"}}"#, + "vertex_ai", + )) + )] + #[case::overloaded( + 400, + "The model is overloaded.", + with_debug(failure( + kept(StatusClass::InternalServer, "The model is overloaded."), + "litellm.InternalServerError: vertex_aiException - The model is overloaded.", + "vertex_ai", + )) + )] + #[case::internal_server_text( + 400, + "500 Internal Server Error", + with_debug(failure( + kept(StatusClass::InternalServer, "500 Internal Server Error"), + "litellm.InternalServerError: vertex_aiException - 500 Internal Server Error", + "vertex_ai", + )) + )] + fn each_text_rule_maps_by_the_body( + #[case] status_code: u16, + #[case] body: &str, + #[case] expected: PublicFailure, + ) { + assert_eq!(mapped(&http(status_code, body)), Some(expected)); + } + + #[rstest::rstest] + #[case::bad_request( + 400, + with_debug(failure( + kept(StatusClass::BadRequest, "rejected"), + "Vertex_aiException BadRequestError - rejected", + "vertex_ai" + )) + )] + #[case::authentication( + 401, + failure( + kept(StatusClass::Authentication, "rejected"), + "Vertex_aiException - rejected", + "vertex_ai" + ) + )] + #[case::permission_denied( + 403, + failure( + kept(StatusClass::PermissionDenied, "rejected"), + "Vertex_aiException - rejected", + "vertex_ai" + ) + )] + #[case::not_found( + 404, + failure( + kept(StatusClass::NotFound, "rejected"), + "Vertex_aiException - rejected", + "vertex_ai" + ) + )] + #[case::request_timeout(408, failure(PublicKind::Timeout { status: None }, "Vertex_aiException - rejected", "vertex_ai"))] + #[case::rate_limited( + 429, + with_debug(failure( + kept(StatusClass::RateLimit, "rejected"), + "litellm.RateLimitError: Vertex_aiException - rejected", + "vertex_ai" + )) + )] + #[case::internal_server( + 500, + with_debug(failure( + kept(StatusClass::InternalServer, "rejected"), + "Vertex_aiException InternalServerError - rejected", + "vertex_ai" + )) + )] + #[case::bad_gateway( + 502, + failure( + PublicKind::ApiConnection, + "Vertex_aiException - rejected", + "vertex_ai" + ) + )] + #[case::service_unavailable( + 503, + failure( + kept(StatusClass::ServiceUnavailable, "rejected"), + "Vertex_aiException - rejected", + "vertex_ai" + ) + )] + fn each_status_rule_maps_by_the_status( + #[case] status_code: u16, + #[case] expected: PublicFailure, + ) { + assert_eq!(mapped(&http(status_code, "rejected")), Some(expected)); + } + + #[rstest::rstest] + #[case::unmapped_status(409)] + #[case::gateway_timeout(504)] + fn statuses_without_a_rule_fall_through(#[case] status_code: u16) { + assert_eq!(mapped(&http(status_code, "rejected")), None); + } + + #[rstest::rstest] + #[case::stub_without_an_upstream_response( + OriginalException::Response { message: "got a 403".into() }, + status(StatusClass::BadRequest, Some(ResponseArg::Stub(HttpStub { status: 403, method: "POST", url: VERTEX_URL_WITH_SPACE, content: None }))) + )] + #[case::stub_for_a_synthesized_status( + OriginalException::Connection { message: "got a 403".into() }, + status(StatusClass::BadRequest, Some(ResponseArg::Stub(HttpStub { status: 403, method: "POST", url: VERTEX_URL_WITH_SPACE, content: None }))) + )] + fn the_rule_response_stays_when_there_is_no_real_upstream_response( + #[case] original: OriginalException, + #[case] kind: PublicKind, + ) { + assert_eq!(mapped(&original).map(|failure| failure.kind), Some(kind)); + } + + #[test] + fn a_synthesized_500_keeps_the_internal_server_stub() { + let original = OriginalException::Connection { + message: "refused".into(), + }; + assert_eq!( + mapped(&original), + Some(with_debug(failure( + status( + StatusClass::InternalServer, + Some(ResponseArg::Stub(HttpStub { + status: 500, + method: "completion", + url: "https://github.com/BerriAI/litellm", + content: Some("refused".into()), + })) + ), + "Vertex_aiException InternalServerError - refused", + "vertex_ai" + ))) + ); + } + + #[rstest::rstest] + #[case::project_before_payload_size( + "Unable to find your project 400 Request payload size exceeds", + StatusClass::BadRequest, + "litellm.BadRequestError: vertex_aiException - Unable to find your project 400 Request payload size exceeds", + true + )] + #[case::payload_size_before_context_window( + "400 Request payload size exceeds; This model's maximum context length is 10", + StatusClass::ContextWindowExceeded, + "Vertex_aiException - 400 Request payload size exceeds; This model's maximum context length is 10", + false + )] + #[case::api_key_before_forbidden( + "API key not valid. 403", + StatusClass::Authentication, + "Vertex_aiException - API key not valid. 403", + true + )] + #[case::forbidden_before_blocked( + "403 The response was blocked.", + StatusClass::BadRequest, + "Vertex_aiException BadRequestError - 403 The response was blocked.", + true + )] + #[case::blocked_before_quota( + "The response was blocked. Resource exhausted", + StatusClass::ContentPolicyViolation, + "Vertex_aiException ContentPolicyViolationError - The response was blocked. Resource exhausted", + true + )] + #[case::quota_before_overloaded( + "Resource exhausted The model is overloaded.", + StatusClass::RateLimit, + "litellm.RateLimitError: vertex_aiException - Resource exhausted The model is overloaded.", + true + )] + fn the_earlier_rule_wins_when_two_apply( + #[case] body: &str, + #[case] class: StatusClass, + #[case] message: &str, + #[case] debug: bool, + ) { + let expected = failure(kept(class, body), message, "vertex_ai"); + assert_eq!( + mapped(&http(401, body)), + Some(if debug { + with_debug(expected) + } else { + expected + }) + ); + } +} diff --git a/litellm-rust/crates/core-utils/src/lib.rs b/litellm-rust/crates/core-utils/src/lib.rs index a8895cccf2c..0c26aa50cc3 100644 --- a/litellm-rust/crates/core-utils/src/lib.rs +++ b/litellm-rust/crates/core-utils/src/lib.rs @@ -1,7 +1,10 @@ pub mod call_arguments; pub mod core_helpers; +pub mod exception_mapping_utils; pub mod get_llm_provider_logic; pub mod params; pub mod prompt_templates; +pub mod python_repr; +pub mod secret_redaction; pub mod serde_compat; pub mod url_utils; diff --git a/litellm-rust/crates/core-utils/src/python_repr.rs b/litellm-rust/crates/core-utils/src/python_repr.rs new file mode 100644 index 00000000000..7ccfb826377 --- /dev/null +++ b/litellm-rust/crates/core-utils/src/python_repr.rs @@ -0,0 +1,93 @@ +/// `repr()` of a Python `str`: single quotes unless the text holds a single quote and no +/// double quote, with backslashes, the chosen quote and control characters escaped. +pub fn python_str_repr(value: &str) -> String { + let quote = if value.contains('\'') && !value.contains('"') { + '"' + } else { + '\'' + }; + let escaped: String = value + .chars() + .map(|character| match character { + '\\' => "\\\\".to_string(), + '\t' => "\\t".to_string(), + '\n' => "\\n".to_string(), + '\r' => "\\r".to_string(), + character if character == quote => format!("\\{character}"), + character + if (character as u32) < 0x20 || (0x7f..0xa0).contains(&(character as u32)) => + { + format!("\\x{:02x}", character as u32) + } + character => character.to_string(), + }) + .collect(); + format!("{quote}{escaped}{quote}") +} + +/// `repr()` of the Python value a JSON value decodes to. +pub fn python_value_repr(value: &serde_json::Value) -> String { + use serde_json::Value; + match value { + Value::Null => "None".to_string(), + Value::Bool(true) => "True".to_string(), + Value::Bool(false) => "False".to_string(), + Value::Number(number) => number.to_string(), + Value::String(text) => python_str_repr(text), + Value::Array(items) => format!( + "[{}]", + items + .iter() + .map(python_value_repr) + .collect::>() + .join(", ") + ), + Value::Object(fields) => format!( + "{{{}}}", + fields + .iter() + .map(|(key, value)| format!( + "{}: {}", + python_str_repr(key), + python_value_repr(value) + )) + .collect::>() + .join(", ") + ), + } +} + +#[cfg(test)] +mod tests { + use serde_json::json; + + use super::{python_str_repr, python_value_repr}; + + #[rstest::rstest] + #[case::null(json!(null), "None")] + #[case::true_(json!(true), "True")] + #[case::false_(json!(false), "False")] + #[case::integer(json!(5), "5")] + #[case::float(json!(1.5), "1.5")] + #[case::string(json!("it's"), "\"it's\"")] + #[case::list(json!(["a", 1]), "['a', 1]")] + #[case::dict(json!({"format": "native"}), "{'format': 'native'}")] + #[case::empty_list(json!([]), "[]")] + fn value_repr_matches_python(#[case] value: serde_json::Value, #[case] expected: &str) { + assert_eq!(python_value_repr(&value), expected); + } + + #[rstest::rstest] + #[case::plain("native", "'native'")] + #[case::single_quote("it's", "\"it's\"")] + #[case::both_quotes("it's \"x\"", "'it\\'s \"x\"'")] + #[case::double_quote("say \"x\"", "'say \"x\"'")] + #[case::backslash("a\\b", "'a\\\\b'")] + #[case::whitespace("a\tb\nc\rd", "'a\\tb\\nc\\rd'")] + #[case::control("a\u{1}b\u{7f}c\u{85}", "'a\\x01b\\x7fc\\x85'")] + #[case::unicode("café", "'café'")] + #[case::empty("", "''")] + fn matches_python_repr(#[case] value: &str, #[case] expected: &str) { + assert_eq!(python_str_repr(value), expected); + } +} diff --git a/litellm-rust/crates/core-utils/src/secret_redaction.rs b/litellm-rust/crates/core-utils/src/secret_redaction.rs new file mode 100644 index 00000000000..32922d7430c --- /dev/null +++ b/litellm-rust/crates/core-utils/src/secret_redaction.rs @@ -0,0 +1,97 @@ +use std::sync::LazyLock; + +use fancy_regex::Regex; + +pub const REDACTED: &str = "REDACTED"; + +const DEFAULT_MINIMUM_CUSTOM_KEY_LENGTH: usize = 16; + +fn minimum_custom_key_length() -> usize { + std::env::var("MINIMUM_CUSTOM_KEY_LENGTH") + .ok() + .and_then(|value| value.trim().parse().ok()) + .unwrap_or(DEFAULT_MINIMUM_CUSTOM_KEY_LENGTH) +} + +fn secret_patterns(minimum_custom_key_length: usize) -> String { + let sk_suffix_length = minimum_custom_key_length.saturating_sub("sk-".len()); + [ + r"-----BEGIN[A-Z \-]*PRIVATE KEY-----[\s\S]*?-----END[A-Z \-]*PRIVATE KEY-----", + r"\bya29\.[A-Za-z0-9_.~+/-]+", + r#"(?:client_secret|azure_password|azure_username)\s+[^\s,'"})\]{}>]+"#, + r"(?:AKIA|ASIA)[0-9A-Z]{16}", + r"Bearer\s+[A-Za-z0-9\-._~+/]{10,}=*", + r"Basic\s+[A-Za-z0-9+/]{10,}={0,2}", + &format!(r"sk-[A-Za-z0-9\-_]{{{sk_suffix_length},}}"), + r#"(?<=[?&])(?:api[_-]?key|\w*(?:token|password|passwd|client_secret|secret_key|_secret))=[^\s&'"]+"#, + r#"(?:api[_-]?key)['"]?\s*[:=]\s*['"]?[^\s,'"})\]{}>]{8,}"#, + r#"(?:x-api-key|api-key)['"]?\s*[:=]\s*['"]?[^\s,'"})\]{}>]+"#, + r"x-ak-[A-Za-z0-9\-_]{20,}", + r"AIza[0-9A-Za-z\-_]{35}", + r#"(?<=[?&])key=[^\s&'"]{8,}"#, + r#"(?:^|(?<=\W))\w*(?:password|passwd|client_secret|secret_key|_secret)['"]?\s*[:=]\s*['"]?[^\s,'"})\]{}>]+"#, + r#"(?<=://)[^\s'":]{0,4096}:[^\s'"]{1,4096}(?=@)"#, + r"dapi[0-9a-f]{32}", + r#"litellm\.[A-Za-z0-9_]*_key['"]?\s*[:=]\s*['"]?[^\s,'"})\]{}>]+"#, + r#"private_key['"]?\s*[:=]\s*['"]?(?:-----BEGIN[A-Z \-]*PRIVATE KEY-----[\s\S]*?-----END[A-Z \-]*PRIVATE KEY-----|[^\s,'"})\]{}>]+)"#, + concat!( + r"(?:master_key|xai_key|database_url|db_url|connection_string|", + r"aws_secret_access_key|aws_session_token|aws_access_key_id|", + r"signing_key|encryption_key|", + r"auth_token|access_token|refresh_token|", + r"slack_webhook_url|webhook_url|", + r"database_connection_string|", + r"huggingface_token|jwt_secret)", + r#"['"]?\s*[:=]\s*['"]?[^\s,'"})\]{}>]+"#, + ), + r"\beyJ[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]+\.[A-Za-z0-9_-]*", + r"(?<=[?&])sig=[A-Za-z0-9%+/=]+", + r#"\{[^{}]*"type"\s*:\s*"service_account"[^{}]*(?:\{[^{}]*\}[^{}]*)*\}"#, + ] + .join("|") +} + +static SECRET_RE: LazyLock = LazyLock::new(|| { + Regex::new(&format!( + "(?i){}", + secret_patterns(minimum_custom_key_length()) + )) + .expect("secret redaction patterns compile") +}); + +pub fn redact_string(value: &str) -> String { + SECRET_RE.replace_all(value, REDACTED).into_owned() +} + +pub fn secret_redaction_enabled() -> bool { + !std::env::var("LITELLM_DISABLE_REDACT_SECRETS") + .is_ok_and(|value| value.eq_ignore_ascii_case("true")) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[rstest::rstest] + #[case::bearer("auth failed: Bearer abcdefghijklmnop", "auth failed: REDACTED")] + #[case::sk_key("key sk-abcdefghijklmnopqrstuvwxyz rejected", "key REDACTED rejected")] + #[case::short_sk_key_is_kept("sk-abc", "sk-abc")] + #[case::query_param("GET /v1?api_key=secret123&x=1", "GET /v1?REDACTED&x=1")] + #[case::dict_repr("{'api_key': 'abcdefghij'}", "{'REDACTED'}")] + #[case::url_credentials("postgres://user:pass@host/db", "postgres://REDACTED@host/db")] + #[case::case_insensitive("BEARER ABCDEFGHIJKLMNOP", "REDACTED")] + #[case::aws_key("AKIAABCDEFGHIJKLMNOP", "REDACTED")] + #[case::sas_signature("https://x.blob/a?sv=1&sig=abc%2B=", "https://x.blob/a?sv=1&REDACTED")] + #[case::password_needs_word_boundary("db_password=hunter2", "REDACTED")] + #[case::plain_text_is_kept(r#"{"message": "rejected"}"#, r#"{"message": "rejected"}"#)] + fn redacts_the_same_spans_as_the_python_patterns(#[case] input: &str, #[case] expected: &str) { + assert_eq!(redact_string(input), expected); + } + + #[test] + fn sk_threshold_follows_the_minimum_custom_key_length() { + let patterns = Regex::new(&format!("(?i){}", secret_patterns(8))).unwrap(); + assert_eq!(patterns.replace_all("sk-abcde", REDACTED), REDACTED); + assert_eq!(patterns.replace_all("sk-abcd", REDACTED), "sk-abcd"); + } +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/api.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/api.json new file mode 100644 index 00000000000..ae629114589 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/api.json @@ -0,0 +1,13 @@ +{ + "kind": { + "type": "api", + "status": 409, + "request_url": "https://docs.litellm.ai/docs" + }, + "message": "MistralException - api", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": null, + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json new file mode 100644 index 00000000000..ab400ed9e01 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json @@ -0,0 +1,11 @@ +{ + "kind": { + "type": "api_connection" + }, + "message": "MistralException - api_connection", + "model": "ocr-model", + "llm_provider": null, + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": null, + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json new file mode 100644 index 00000000000..057392575a1 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json @@ -0,0 +1,13 @@ +{ + "kind": { + "type": "status", + "status_class": "authentication", + "response": null + }, + "message": "MistralException - status_authentication", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": null, + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json new file mode 100644 index 00000000000..abb1425f686 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "bad_gateway", + "response": { + "type": "upstream", + "status": 502, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_bad_gateway", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json new file mode 100644 index 00000000000..171a994cd35 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "bad_request", + "response": { + "type": "upstream", + "status": 400, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_bad_request", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json new file mode 100644 index 00000000000..ae9c1e145de --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "content_policy_violation", + "response": { + "type": "upstream", + "status": 400, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_content_policy_violation", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json new file mode 100644 index 00000000000..61e1a56a622 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "context_window_exceeded", + "response": { + "type": "upstream", + "status": 400, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_context_window_exceeded", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json new file mode 100644 index 00000000000..b3c5c51a785 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json @@ -0,0 +1,19 @@ +{ + "kind": { + "type": "status", + "status_class": "internal_server", + "response": { + "type": "stub", + "status": 500, + "method": "completion", + "url": "https://github.com/BerriAI/litellm", + "content": "upstream text" + } + }, + "message": "MistralException - status_internal_server", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": null, + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json new file mode 100644 index 00000000000..26ffd872961 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "not_found", + "response": { + "type": "upstream", + "status": 404, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_not_found", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json new file mode 100644 index 00000000000..42772f98830 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json @@ -0,0 +1,19 @@ +{ + "kind": { + "type": "status", + "status_class": "permission_denied", + "response": { + "type": "stub", + "status": 403, + "method": "POST", + "url": " https://cloud.google.com/vertex-ai/", + "content": null + } + }, + "message": "MistralException - status_permission_denied", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": null, + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json new file mode 100644 index 00000000000..c9b88822ebd --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "rate_limit", + "response": { + "type": "upstream", + "status": 429, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_rate_limit", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json new file mode 100644 index 00000000000..5af65202ef1 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "service_unavailable", + "response": { + "type": "upstream", + "status": 503, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_service_unavailable", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json new file mode 100644 index 00000000000..a1b318ce03c --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json @@ -0,0 +1,28 @@ +{ + "kind": { + "type": "status", + "status_class": "unsupported_params", + "response": { + "type": "upstream", + "status": 400, + "body": "{\"message\": \"rejected\"}", + "headers": [ + [ + "retry-after", + "7" + ] + ] + } + }, + "message": "MistralException - status_unsupported_params", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": [ + [ + "retry-after", + "7" + ] + ], + "print_banner": false +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json new file mode 100644 index 00000000000..8b215bdb7e6 --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json @@ -0,0 +1,12 @@ +{ + "kind": { + "type": "timeout", + "status": 504 + }, + "message": "MistralException - timeout_with_status", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": "\nModel: ocr-model", + "litellm_response_headers": null, + "print_banner": true +} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json new file mode 100644 index 00000000000..4a79169776e --- /dev/null +++ b/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json @@ -0,0 +1,12 @@ +{ + "kind": { + "type": "timeout", + "status": null + }, + "message": "MistralException - timeout_without_status", + "model": "ocr-model", + "llm_provider": "mistral", + "litellm_debug_info": null, + "litellm_response_headers": null, + "print_banner": false +} From 74e9fb23233c8d3bfeed73485bc0270da79370d3 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 13:07:33 -0700 Subject: [PATCH 117/141] fix(passthrough): validate only the winning timeout value in the resolver --- litellm/passthrough/timeout_utils.py | 36 +++++++++---------- .../test_pass_through_endpoints.py | 20 +++++++++++ 2 files changed, 38 insertions(+), 18 deletions(-) diff --git a/litellm/passthrough/timeout_utils.py b/litellm/passthrough/timeout_utils.py index 9284f7143b0..fc67aa8c553 100644 --- a/litellm/passthrough/timeout_utils.py +++ b/litellm/passthrough/timeout_utils.py @@ -1,18 +1,14 @@ import sys from collections.abc import Mapping +from types import MappingProxyType from typing import Final -from pydantic import BaseModel, ConfigDict +from pydantic import TypeAdapter DEFAULT_PASS_THROUGH_REQUEST_TIMEOUT_SECONDS: Final = 600.0 - -class _TimeoutFields(BaseModel): - model_config = ConfigDict(frozen=True) - - stream_timeout: float | None = None - timeout: float | None = None - request_timeout: float | None = None +_SECONDS: Final = TypeAdapter(float) +_NO_PARAMS: Final[Mapping[str, object]] = MappingProxyType({}) def resolve_pass_through_request_timeout( @@ -59,20 +55,24 @@ def resolve_llm_passthrough_timeout( any generic timeout, matching ``Router._get_stream_timeout`` on the completion route: kwargs stream_timeout -> litellm_params stream_timeout -> router_stream_timeout, then the non-streaming chain above. + + Only the first set value is validated as seconds, so a value in a lower-precedence + field never fails the call. """ - streaming: Final = bool((kwargs or {}).get("stream")) - request: Final = _TimeoutFields.model_validate(kwargs or {}) - deployment: Final = _TimeoutFields.model_validate(litellm_params or {}) + request: Final = kwargs if kwargs is not None else _NO_PARAMS + deployment: Final = litellm_params if litellm_params is not None else _NO_PARAMS stream_candidates: Final = ( - (request.stream_timeout, deployment.stream_timeout, router_stream_timeout) if streaming else () + (request.get("stream_timeout"), deployment.get("stream_timeout"), router_stream_timeout) + if request.get("stream") + else () ) candidates: Final = ( *stream_candidates, - request.timeout, - request.request_timeout, - deployment.timeout, - deployment.request_timeout, + request.get("timeout"), + request.get("request_timeout"), + deployment.get("timeout"), + deployment.get("request_timeout"), router_timeout, ) - resolved: Final = next((float(val) for val in candidates if val is not None), None) - return resolved if resolved is not None else resolve_pass_through_request_timeout() + winner: Final = next((val for val in candidates if val is not None), None) + return resolve_pass_through_request_timeout() if winner is None else _SECONDS.validate_python(winner) diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index abff897c4f5..13cd41278b9 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -12,6 +12,7 @@ from unittest.mock import AsyncMock, MagicMock, patch import httpx import pytest from fastapi import Request, Response, UploadFile +from pydantic import ValidationError from starlette.datastructures import FormData, Headers, QueryParams from starlette.datastructures import UploadFile as StarletteUploadFile @@ -1189,6 +1190,25 @@ def test_resolve_llm_passthrough_timeout_reads_stream_by_truthiness(stream: obje ) +@pytest.mark.parametrize( + "kwargs, litellm_params, expected", + [ + ({"stream": True, "stream_timeout": 1800, "timeout": httpx.Timeout(30.0)}, {}, 1800.0), + ({"stream": False}, {"stream_timeout": httpx.Timeout(30.0), "timeout": 90}, 90.0), + ({"timeout": 45}, {"request_timeout": httpx.Timeout(30.0)}, 45.0), + ], +) +def test_resolve_llm_passthrough_timeout_validates_only_the_winning_value( + kwargs: dict[str, object], litellm_params: dict[str, object], expected: float +): + assert resolve_llm_passthrough_timeout(kwargs=kwargs, litellm_params=litellm_params) == expected + + +def test_resolve_llm_passthrough_timeout_rejects_a_non_numeric_winner(): + with pytest.raises(ValidationError): + resolve_llm_passthrough_timeout(kwargs={"timeout": httpx.Timeout(30.0)}) + + @pytest.mark.asyncio async def test_pass_through_request_uses_resolved_timeout(): with patch("litellm.proxy.proxy_server.proxy_logging_obj") as mock_proxy_logging: From d37320f8cf242f310bad3e759c9a7dbbdc27b699 Mon Sep 17 00:00:00 2001 From: ryan-crabbe-berri Date: Fri, 18 Sep 2026 13:08:55 -0700 Subject: [PATCH 118/141] fix(proxy): report null cost for unpriced deployments on the /model/info id lookup GET /model/info?litellm_model_id= went through _get_proxy_model_info, a copy of _enrich_model_info_with_litellm_data that missed the unpriced check, so the same deployment read as null in the list and as 0 on the id lookup. The id lookup now delegates to the shared helper, so the two paths cannot drift again --- litellm/proxy/proxy_server.py | 40 +-------------- .../proxy_server/test_routes_model_info.py | 50 +++++++++++++++++++ 2 files changed, 51 insertions(+), 39 deletions(-) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index f21c907c292..f3b2dbd0b03 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -15057,45 +15057,7 @@ def _translate_model_name_for_response(model: dict) -> dict: def _get_proxy_model_info(model: dict) -> dict: - # provided model_info in config.yaml - model_info: Final = model.get("model_info", {}) - - # read litellm model_prices_and_context_window.json to get the following: - # input_cost_per_token, output_cost_per_token, max_tokens - litellm_model_info = get_litellm_model_info(model=model) - - # 2nd pass on the model, try seeing if we can find model in litellm model_cost map - if litellm_model_info == {}: - # use litellm_param model_name to get model_info - litellm_params = model.get("litellm_params", {}) - litellm_model = litellm_params.get("model", None) - try: - litellm_model_info = litellm.get_model_info(model=litellm_model) - except Exception: - litellm_model_info = {} - # 3rd pass on the model, try seeing if we can find model but without the "/" in model cost map - if litellm_model_info == {}: - # use litellm_param model_name to get model_info - litellm_params = model.get("litellm_params", {}) - litellm_model = litellm_params.get("model", None) - split_model: Final = litellm_model.split("/") - if len(split_model) > 0: - litellm_model = split_model[-1] - try: - litellm_model_info = litellm.get_model_info(model=litellm_model, custom_llm_provider=split_model[0]) - except Exception: - litellm_model_info = {} - discovered_model_info: Final = ( - llm_router.get_discovered_model_info(model_info.get("id")) if llm_router is not None else MappingProxyType({}) - ) - for k, v in MappingProxyType({**litellm_model_info, **discovered_model_info}).items(): - if k not in model_info or (model_info[k] is None and k in discovered_model_info): - model_info[k] = v - model["model_info"] = model_info - # don't return the llm credentials - model = remove_sensitive_info_from_deployment(deployment_dict=model, excluded_keys={"litellm_credential_name"}) - - return _translate_model_name_for_response(model) + return _translate_model_name_for_response(_enrich_model_info_with_litellm_data(model=model, llm_router=llm_router)) def _model_info_json_response(data: Sequence[Mapping[str, object]] | Mapping[str, object]) -> Response: diff --git a/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py b/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py index 203a237113f..75a8657356a 100644 --- a/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py +++ b/tests/test_litellm/proxy/proxy_server/test_routes_model_info.py @@ -323,6 +323,56 @@ def test_model_info_reports_null_cost_for_unpriced_deployment_and_zero_for_decla assert input_cost > 0 and output_cost > 0 +def test_model_info_id_lookup_reports_the_same_cost_as_the_list( + client: TestClient, + auth_as: Callable[[], AbstractContextManager[object]], + monkeypatch: pytest.MonkeyPatch, +) -> None: + """``GET /model/info?litellm_model_id=`` must agree with the ``GET /model/info`` list, so an + unpriced deployment cannot read as null in the list and as free on the id lookup.""" + monkeypatch.setattr(litellm, "model_cost", copy.deepcopy(litellm.model_cost)) + declared: Final = {"input_cost_per_token": 0.001, "output_cost_per_token": 0.002} + free: Final = {"input_cost_per_token": 0, "output_cost_per_token": 0} + router: Final = litellm.Router( + model_list=[ + { + "model_name": name, + "litellm_params": {"model": f"openai/{name}", "api_key": "x", "api_base": "http://vllm", **costs}, + "model_info": {"id": f"{name}-id"}, + } + for name, costs in (("vllm-unpriced", {}), ("vllm-free", free), ("vllm-priced", declared)) + ] + ) + monkeypatch.setattr(proxy_server, "llm_router", router) + monkeypatch.setattr(proxy_server, "llm_model_list", router.get_model_list()) + monkeypatch.setattr(proxy_server, "user_model", None) + + def costs_of(response: httpx.Response) -> dict[str, tuple[object, object]]: + assert response.status_code == 200, response.text + return { + row["model_info"]["id"]: ( + row["model_info"].get("input_cost_per_token"), + row["model_info"].get("output_cost_per_token"), + ) + for row in response.json()["data"] + } + + with auth_as(): + listed: Final = costs_of(client.get("/model/info")) + by_id: Final = { + model_id: costs_of(client.get("/model/info", params={"litellm_model_id": model_id}))[model_id] + for model_id in listed + } + + assert listed == { + "vllm-unpriced-id": (None, None), + "vllm-free-id": (0, 0), + "vllm-priced-id": (declared["input_cost_per_token"], declared["output_cost_per_token"]), + } + assert by_id == listed + _invalidate_model_cost_lowercase_map() + + def test_v1_model_info_star_wildcard_filter_keeps_provider_expansion(monkeypatch): from litellm.proxy._types import SpecialModelNames, UserAPIKeyAuth from litellm.proxy.auth import model_checks From 04193745a65cc18a36820be8ef1d901d6cbe19db Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 20:10:20 +0000 Subject: [PATCH 119/141] fix(docs-test): restore line-anchored table regex in router settings check A merge on this branch dropped the ^ anchor and MULTILINE flag from the doc_key_pattern, so the unanchored match started swallowing key names into captured fields whenever a row's description cell itself contains a pipe (Literal unions and similar). The check then reported 27 long-documented keys as undocumented. Restores the exact pattern used on main, verified against a live litellm-docs checkout: all 58 Router init params resolve as documented. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/documentation_tests/test_router_settings.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/tests/documentation_tests/test_router_settings.py b/tests/documentation_tests/test_router_settings.py index 75032f80dfa..7e3d0c07459 100644 --- a/tests/documentation_tests/test_router_settings.py +++ b/tests/documentation_tests/test_router_settings.py @@ -51,9 +51,7 @@ try: if general_settings_section: # Extract the table rows, which contain the documented keys table_content = general_settings_section.group(1) - doc_key_pattern = re.compile( - r"\|\s*([^\|]+?)\s*\|" - ) # Capture the key from each row of the table + doc_key_pattern = re.compile(r"^\|\s*([^\|]+?)\s*\|", re.MULTILINE) documented_keys.update(doc_key_pattern.findall(table_content)) except Exception as e: raise Exception( From d16d17d2d28f56cb0775a8c42ebc95b8ccf4f032 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 13:15:17 -0700 Subject: [PATCH 120/141] ci(osv-scan): scan the VS Code extension lockfile --- .github/workflows/osv-scan.yml | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/.github/workflows/osv-scan.yml b/.github/workflows/osv-scan.yml index 31104002dab..0aedeaec12a 100644 --- a/.github/workflows/osv-scan.yml +++ b/.github/workflows/osv-scan.yml @@ -41,4 +41,5 @@ jobs: "$RUNNER_TEMP/osv-scanner" scan source \ --config osv-scanner.toml \ -L uv.lock \ - -L ui/litellm-dashboard/package-lock.json + -L ui/litellm-dashboard/package-lock.json \ + -L vscode-extension/package-lock.json From 0f59e61f24dc72c6f4a9af15f1618d1b13591cde Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 20:17:11 +0000 Subject: [PATCH 121/141] fix(rust): align exception mapping tests with Python Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../src/exception_mapping_utils/mod.rs | 25 +++++++++---------- .../src/exception_mapping_utils/rules.rs | 4 +-- 2 files changed, 14 insertions(+), 15 deletions(-) diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs index 8892fba1399..acc7820c770 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs @@ -282,7 +282,6 @@ fn extra_information(context: &ExceptionContext, api_base: Option<&str>) -> Stri let lines = [ Some(format!("\nModel: {}", context.model)), api_base.map(|api_base| format!("\nAPI Base: `{api_base}`")), - (!context.redact_messages_in_exceptions).then(|| "\nMessages: `None`".to_string()), context .model_group .as_ref() @@ -314,7 +313,7 @@ fn extra_information(context: &ExceptionContext, api_base: Option<&str>) -> Stri mod testing { use super::*; - pub(super) const DEBUG: &str = "\nModel: ocr-model\nMessages: `None`"; + pub(super) const DEBUG: &str = "\nModel: ocr-model"; pub(super) fn context(provider: &str, family: ExceptionFamily) -> ExceptionContext { ExceptionContext { @@ -369,14 +368,6 @@ mod testing { ..failure } } - - /// What `exception_type` returns for an `http` original the rule mapped to `failure`. - pub(super) fn with_headers(failure: PublicFailure) -> PublicFailure { - PublicFailure { - litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), - ..failure - } - } } #[cfg(test)] @@ -492,7 +483,17 @@ mod tests { #[case] message: &str, ) { let context = context("reducto", ExceptionFamily::Other); - let expected = failure(PublicKind::ApiConnection, message, "reducto"); + let expected = match original { + OriginalException::Http { .. } => with_debug(failure( + status(StatusClass::BadRequest, upstream(409, "rejected")), + message, + "reducto", + )), + OriginalException::Connection { .. } | OriginalException::Local { .. } => { + failure(PublicKind::ApiConnection, message, "reducto") + } + _ => unreachable!(), + }; let actual = exception_type(&context, &original); assert_eq!( PublicFailure { @@ -669,7 +670,6 @@ mod tests { "\n\nKey Name: `key`\nTeam: `None`", "\nModel: ocr-model", "\nAPI Base: `region-aiplatform.googleapis.com/v1/projects/project/locations/region/publishers/google/models/ocr-model:generateContent`", - "\nMessages: `None`", "\nmodel_group: `ocr`\n", "\ndeployment: `deployment`\n", "\nvertex_project: `project`\n", @@ -681,7 +681,6 @@ mod tests { #[rstest::rstest] #[case::bare(ExceptionContext::default(), "\nModel: ")] #[case::redacted_messages(ExceptionContext { redact_messages_in_exceptions: true, model: "m".into(), ..ExceptionContext::default() }, "\nModel: m")] - #[case::messages(ExceptionContext { model: "m".into(), ..ExceptionContext::default() }, "\nModel: m\nMessages: `None`")] #[case::team_alias( ExceptionContext { model: "m".into(), user_api_key_alias: Some("key".into()), user_api_key_team_alias: Some("team".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, "\n\nKey Name: `key`\nTeam: `team`\nModel: m" diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs index 5df8ad991b5..34cf2c390c3 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs @@ -224,7 +224,7 @@ mod tests { "mistral", ))] #[case::later_rule_when_the_earlier_does_not_apply("second", PublicFailure { - litellm_debug_info: Some("\nModel: ocr-model\nMessages: `None`".into()), + litellm_debug_info: Some("\nModel: ocr-model".into()), ..failure(PublicKind::ApiConnection, "seen second", "mistral") })] fn the_first_applicable_rule_decides(#[case] body: &str, #[case] expected: PublicFailure) { @@ -324,7 +324,7 @@ mod tests { } #[rstest::rstest] - #[case::with_debug(true, Some("\nModel: ocr-model\nMessages: `None`"))] + #[case::with_debug(true, Some("\nModel: ocr-model"))] #[case::without_debug(false, None)] fn debug_rules_carry_the_extra_information( #[case] debug: bool, From 6464c1fd6c9a58efcf75f4b73650d541aa4c76bd Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 19:17:08 +0000 Subject: [PATCH 122/141] fix(proxy): make SettingsStore.clear() terminate when the config file owns a key Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../proxy/config_resolvers/settings_store.py | 4 ++ .../config_resolvers/test_settings_store.py | 49 +++++++++++++++++++ 2 files changed, 53 insertions(+) diff --git a/litellm/proxy/config_resolvers/settings_store.py b/litellm/proxy/config_resolvers/settings_store.py index 079d262319f..49f5a833b8f 100644 --- a/litellm/proxy/config_resolvers/settings_store.py +++ b/litellm/proxy/config_resolvers/settings_store.py @@ -87,6 +87,10 @@ class SettingsStore(MutableMapping[str, JsonValue]): ) self._deleted_runtime_keys = self._deleted_runtime_keys | frozenset((key,)) + def clear(self) -> None: + self._deleted_runtime_keys = frozenset(key for key in self._keys() if not self.owned_by_config(key)) + self._runtime_values = _EMPTY_VALUES + def __iter__(self) -> Iterator[str]: return iter( key diff --git a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py index aa410deac43..aec78933c89 100644 --- a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py +++ b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py @@ -1,6 +1,7 @@ from __future__ import annotations from typing import Final +from unittest.mock import patch import pytest @@ -168,6 +169,54 @@ def test_settings_store_refuses_a_runtime_write_to_a_config_owned_key() -> None: assert store.source("max_parallel_requests") == "config" +@pytest.mark.timeout(10) +def test_settings_store_clear_removes_every_key_the_config_file_does_not_own() -> None: + store: Final = SettingsStore("general_settings") + store.load_yaml({"master_key": "sk-config"}) + store.apply_db_row("general_settings", {"max_parallel_requests": 3, "alerting": ["slack"]}) + store["allow_requests_on_db_unavailable"] = True + del store["alerting"] + + store.clear() + + assert dict(store) == {"master_key": "sk-config"} + assert "alerting" not in store + with pytest.raises(KeyError): + store["max_parallel_requests"] + + +@pytest.mark.timeout(10) +def test_settings_store_clear_then_refill_matches_a_plain_dict() -> None: + expected: Final[dict[str, JsonValue]] = {"max_parallel_requests": 3, "alerting": ["slack"]} + store: Final = SettingsStore("general_settings") + store.update(expected) + + expected.clear() + store.clear() + expected.update({"alerting": ["email"], "max_parallel_requests": 11}) + store.update({"alerting": ["email"], "max_parallel_requests": 11}) + + assert dict(store) == expected + assert tuple(store) == tuple(expected) + assert len(store) == len(expected) + + +@pytest.mark.timeout(10) +@pytest.mark.parametrize("clear", (False, True)) +def test_settings_store_survives_a_patch_dict_round_trip_when_the_config_file_owns_a_key(clear: bool) -> None: + store: Final = SettingsStore("general_settings") + store.load_yaml({"master_key": "sk-config"}) + store.apply_db_row("general_settings", {"max_parallel_requests": 3}) + before: Final = dict(store) + + with patch.dict(store, {"allow_requests_on_db_unavailable": True}, clear=clear): + assert store["allow_requests_on_db_unavailable"] is True + assert store["master_key"] == "sk-config" + assert ("max_parallel_requests" in store) is not clear + + assert dict(store) == before + + def test_settings_store_reports_the_config_owned_keys_a_write_would_change() -> None: store: Final = SettingsStore("general_settings") store.load_yaml({"max_parallel_requests": 3, "ui_access_mode": "admin_only"}) From d6b6ab31b75992150f800ec5a2bc71cf78453e8c Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 19:27:23 +0000 Subject: [PATCH 123/141] test(proxy): compare the cleared settings store against an unmutated refill mapping Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../proxy/config_resolvers/test_settings_store.py | 14 ++++++-------- 1 file changed, 6 insertions(+), 8 deletions(-) diff --git a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py index aec78933c89..d59fb3e5b14 100644 --- a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py +++ b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py @@ -187,18 +187,16 @@ def test_settings_store_clear_removes_every_key_the_config_file_does_not_own() - @pytest.mark.timeout(10) def test_settings_store_clear_then_refill_matches_a_plain_dict() -> None: - expected: Final[dict[str, JsonValue]] = {"max_parallel_requests": 3, "alerting": ["slack"]} + refilled: Final[dict[str, JsonValue]] = {"alerting": ["email"], "max_parallel_requests": 11} store: Final = SettingsStore("general_settings") - store.update(expected) + store.update({"max_parallel_requests": 3, "alerting": ["slack"]}) - expected.clear() store.clear() - expected.update({"alerting": ["email"], "max_parallel_requests": 11}) - store.update({"alerting": ["email"], "max_parallel_requests": 11}) + store.update(refilled) - assert dict(store) == expected - assert tuple(store) == tuple(expected) - assert len(store) == len(expected) + assert dict(store) == refilled + assert tuple(store) == tuple(refilled) + assert len(store) == len(refilled) @pytest.mark.timeout(10) From 08d08191592b822396ad655b55d5b57a47449244 Mon Sep 17 00:00:00 2001 From: yucheng Date: Fri, 18 Sep 2026 19:39:15 +0000 Subject: [PATCH 124/141] fix(proxy): keep the resolved values of config-owned keys across SettingsStore.clear() Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/config_resolvers/settings_store.py | 4 +++- .../proxy/config_resolvers/test_settings_store.py | 10 ++++++---- 2 files changed, 9 insertions(+), 5 deletions(-) diff --git a/litellm/proxy/config_resolvers/settings_store.py b/litellm/proxy/config_resolvers/settings_store.py index 49f5a833b8f..d4ca0e87d2b 100644 --- a/litellm/proxy/config_resolvers/settings_store.py +++ b/litellm/proxy/config_resolvers/settings_store.py @@ -89,7 +89,9 @@ class SettingsStore(MutableMapping[str, JsonValue]): def clear(self) -> None: self._deleted_runtime_keys = frozenset(key for key in self._keys() if not self.owned_by_config(key)) - self._runtime_values = _EMPTY_VALUES + self._runtime_values = MappingProxyType( + {key: value for key, value in self._runtime_values.items() if self.owned_by_config(key)} + ) def __iter__(self) -> Iterator[str]: return iter( diff --git a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py index d59fb3e5b14..88ec382b013 100644 --- a/tests/test_litellm/proxy/config_resolvers/test_settings_store.py +++ b/tests/test_litellm/proxy/config_resolvers/test_settings_store.py @@ -172,14 +172,15 @@ def test_settings_store_refuses_a_runtime_write_to_a_config_owned_key() -> None: @pytest.mark.timeout(10) def test_settings_store_clear_removes_every_key_the_config_file_does_not_own() -> None: store: Final = SettingsStore("general_settings") - store.load_yaml({"master_key": "sk-config"}) + store.load_yaml({"master_key": "os.environ/MASTER_KEY"}) store.apply_db_row("general_settings", {"max_parallel_requests": 3, "alerting": ["slack"]}) + store.apply_runtime_values({"master_key": "sk-resolved", "alerting": ["slack"]}) store["allow_requests_on_db_unavailable"] = True del store["alerting"] store.clear() - assert dict(store) == {"master_key": "sk-config"} + assert dict(store) == {"master_key": "sk-resolved"} assert "alerting" not in store with pytest.raises(KeyError): store["max_parallel_requests"] @@ -203,13 +204,14 @@ def test_settings_store_clear_then_refill_matches_a_plain_dict() -> None: @pytest.mark.parametrize("clear", (False, True)) def test_settings_store_survives_a_patch_dict_round_trip_when_the_config_file_owns_a_key(clear: bool) -> None: store: Final = SettingsStore("general_settings") - store.load_yaml({"master_key": "sk-config"}) + store.load_yaml({"master_key": "os.environ/MASTER_KEY"}) store.apply_db_row("general_settings", {"max_parallel_requests": 3}) + store.apply_runtime_values({"master_key": "sk-resolved", "max_parallel_requests": 3}) before: Final = dict(store) with patch.dict(store, {"allow_requests_on_db_unavailable": True}, clear=clear): assert store["allow_requests_on_db_unavailable"] is True - assert store["master_key"] == "sk-config" + assert store["master_key"] == "sk-resolved" assert ("max_parallel_requests" in store) is not clear assert dict(store) == before From 2703fab3c810f1c82b65c035acf4920ad3ac85e3 Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 20:21:21 +0000 Subject: [PATCH 125/141] fix(deps): update anyio for osv scan Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- uv.lock | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/uv.lock b/uv.lock index a5e60c68515..29027c2a75f 100644 --- a/uv.lock +++ b/uv.lock @@ -315,16 +315,16 @@ vertex = [ [[package]] name = "anyio" -version = "4.13.0" +version = "4.15.1" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "exceptiongroup", marker = "python_full_version < '3.11'" }, { name = "idna" }, - { name = "typing-extensions", marker = "python_full_version < '3.13'" }, + { name = "typing-extensions" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/19/14/2c5dd9f512b66549ae92767a9c7b330ae88e1932ca57876909410251fe13/anyio-4.13.0.tar.gz", hash = "sha256:334b70e641fd2221c1505b3890c69882fe4a2df910cba14d97019b90b24439dc", size = 231622, upload-time = "2026-03-24T12:59:09.671Z" } +sdist = { url = "https://files.pythonhosted.org/packages/a9/d2/f4d173e22df740bc37b1db102b386ba719b66e95b0f0d751f556b387e6d2/anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94", size = 276966, upload-time = "2026-09-05T10:42:39.44Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/da/42/e921fccf5015463e32a3cf6ee7f980a6ed0f395ceeaa45060b61d86486c2/anyio-4.13.0-py3-none-any.whl", hash = "sha256:08b310f9e24a9594186fd75b4f73f4a4152069e3853f1ed8bfbf58369f4ad708", size = 114353, upload-time = "2026-03-24T12:59:08.246Z" }, + { url = "https://files.pythonhosted.org/packages/12/b8/4bd346e22b28902df4d651910f5242c28d84e4a5c2435ca5c3f797ed7e2e/anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101", size = 132079, upload-time = "2026-09-05T10:42:37.923Z" }, ] [[package]] From e036e256ef00be1afbed0cfb0a9ce87f67d18fa5 Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 20:23:53 +0000 Subject: [PATCH 126/141] revert: drop redundant anyio lockfile bump Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- uv.lock | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/uv.lock b/uv.lock index 29027c2a75f..a5e60c68515 100644 --- a/uv.lock +++ b/uv.lock @@ -315,16 +315,16 @@ vertex = [ [[package]] name = "anyio" -version = "4.15.1" +version = "4.13.0" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "exceptiongroup", marker = "python_full_version < '3.11'" }, { name = "idna" }, - { name = "typing-extensions" }, + { name = "typing-extensions", marker = "python_full_version < '3.13'" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/a9/d2/f4d173e22df740bc37b1db102b386ba719b66e95b0f0d751f556b387e6d2/anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94", size = 276966, upload-time = "2026-09-05T10:42:39.44Z" } +sdist = { url = "https://files.pythonhosted.org/packages/19/14/2c5dd9f512b66549ae92767a9c7b330ae88e1932ca57876909410251fe13/anyio-4.13.0.tar.gz", hash = "sha256:334b70e641fd2221c1505b3890c69882fe4a2df910cba14d97019b90b24439dc", size = 231622, upload-time = "2026-03-24T12:59:09.671Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/12/b8/4bd346e22b28902df4d651910f5242c28d84e4a5c2435ca5c3f797ed7e2e/anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101", size = 132079, upload-time = "2026-09-05T10:42:37.923Z" }, + { url = "https://files.pythonhosted.org/packages/da/42/e921fccf5015463e32a3cf6ee7f980a6ed0f395ceeaa45060b61d86486c2/anyio-4.13.0-py3-none-any.whl", hash = "sha256:08b310f9e24a9594186fd75b4f73f4a4152069e3853f1ed8bfbf58369f4ad708", size = 114353, upload-time = "2026-03-24T12:59:08.246Z" }, ] [[package]] From eeca8b682b4c681d8fa0489959a63dd9ba294550 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 13:28:28 -0700 Subject: [PATCH 127/141] fix(vscode): raise the VS Code minimum to 1.115 for per-model configuration --- vscode-extension/README.md | 2 +- vscode-extension/package-lock.json | 10 +++++----- vscode-extension/package.json | 4 ++-- 3 files changed, 8 insertions(+), 8 deletions(-) diff --git a/vscode-extension/README.md b/vscode-extension/README.md index 47c1a3afd27..cea8af6b578 100644 --- a/vscode-extension/README.md +++ b/vscode-extension/README.md @@ -17,7 +17,7 @@ Add the same provider again with another name to reach a second gateway or a sec ## Requirements -VS Code 1.109 or newer and a LiteLLM AI Gateway the key can reach. The key needs access to at least one model group whose mode is `chat` +VS Code 1.115 or newer and a LiteLLM AI Gateway the key can reach. The key needs access to at least one model group whose mode is `chat` ## Development diff --git a/vscode-extension/package-lock.json b/vscode-extension/package-lock.json index a4f962f2039..453dd8e1ff8 100644 --- a/vscode-extension/package-lock.json +++ b/vscode-extension/package-lock.json @@ -13,14 +13,14 @@ }, "devDependencies": { "@types/node": "^22.20.3", - "@types/vscode": "1.109.0", + "@types/vscode": "1.115.0", "@vscode/vsce": "^4.0.0", "esbuild": "^0.28.2", "typescript": "^5.9.3", "vitest": "^4.1.11" }, "engines": { - "vscode": "^1.109.0" + "vscode": "^1.115.0" } }, "node_modules/@azure/abort-controller": { @@ -1301,9 +1301,9 @@ } }, "node_modules/@types/vscode": { - "version": "1.109.0", - "resolved": "https://registry.npmjs.org/@types/vscode/-/vscode-1.109.0.tgz", - "integrity": "sha512-0Pf95rnwEIwDbmXGC08r0B4TQhAbsHQ5UyTIgVgoieDe4cOnf92usuR5dEczb6bTKEp7ziZH4TV1TRGPPCExtw==", + "version": "1.115.0", + "resolved": "https://registry.npmjs.org/@types/vscode/-/vscode-1.115.0.tgz", + "integrity": "sha512-/M8cdznOlqtMqduHKKlIF00v4eum4ZWKgn8YoPRKcN6PDdvoWeeqDaQSnw63ipDbq1Uzz78Wndk/d0uSPwORfA==", "dev": true, "license": "MIT" }, diff --git a/vscode-extension/package.json b/vscode-extension/package.json index 67e3bb6c727..431dad316a1 100644 --- a/vscode-extension/package.json +++ b/vscode-extension/package.json @@ -15,7 +15,7 @@ "url": "https://github.com/BerriAI/litellm/issues" }, "engines": { - "vscode": "^1.109.0" + "vscode": "^1.115.0" }, "categories": [ "AI", @@ -78,7 +78,7 @@ }, "devDependencies": { "@types/node": "^22.20.3", - "@types/vscode": "1.109.0", + "@types/vscode": "1.115.0", "@vscode/vsce": "^4.0.0", "esbuild": "^0.28.2", "typescript": "^5.9.3", From e75ad61fceaf08b49830dc4c737db2ea7a28f49c Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 13:32:43 -0700 Subject: [PATCH 128/141] fix(router): rank litellm_settings.request_timeout on the passthrough route like the completion route --- litellm/router.py | 8 +++++++- tests/test_litellm/test_router.py | 10 ++++++++++ 2 files changed, 17 insertions(+), 1 deletion(-) diff --git a/litellm/router.py b/litellm/router.py index a5523d7af79..e3e9454ddce 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -3877,11 +3877,17 @@ class Router: ) _router_timeout: Final = ( - float(self._explicit_timeout) if isinstance(self._explicit_timeout, (int, float)) else None + self.request_timeout + if self.request_timeout is not None + else float(self._explicit_timeout) + if isinstance(self._explicit_timeout, (int, float)) + else None ) _router_stream_timeout: Final = ( self.stream_timeout if self.stream_timeout is not None + else self.request_timeout + if self.request_timeout is not None else self.default_litellm_params.get("stream_timeout") ) kwargs["timeout"] = resolve_llm_passthrough_timeout( diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index e92e5de23dc..148ca0b0ffa 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -8230,6 +8230,16 @@ class TestRouterRequestTimeoutPropagation: == 60 ) + def test_passthrough_prefers_request_timeout_over_router_timeout(self, explicit_request_timeout): + router = self._make_router(timeout=330) + deployment: Final = router.model_list[0] + assert _passthrough_timeout(router, deployment, stream=False) == 300.0 + assert _passthrough_timeout(router, deployment, stream=True) == 300.0 + + def test_passthrough_stream_timeout_still_wins_over_request_timeout(self, explicit_request_timeout): + router = self._make_router(timeout=330, stream_timeout=45) + assert _passthrough_timeout(router, router.model_list[0], stream=True) == 45.0 + # --------------------------------------------------------------------------- # Deferred-stream eager-fetch tests From de4b520153718864dddfa53a37dd6b56825c958b Mon Sep 17 00:00:00 2001 From: ryan Date: Wed, 16 Sep 2026 01:12:56 +0000 Subject: [PATCH 129/141] fix(proxy): track team member spend when the member has no budget add_new_member only wrote a LiteLLM_TeamMembership row when a budget id resolved, and the spend writer used update_many so a missing row failed silently. Members without a budget therefore never accrued per-member spend. The membership row is now always upserted (budget_id NULL when no budget applies) and the spend write is an upsert so members added before this fix start accruing on their next request. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/db/db_spend_update_writer.py | 16 +++- litellm/proxy/management_helpers/utils.py | 23 +++--- litellm/repositories/prisma_protocols.py | 2 + .../proxy/db/test_db_spend_update_writer.py | 37 ++++++--- .../test_team_endpoints.py | 10 +-- .../test_management_helpers_utils.py | 75 ++++++++++++++----- 6 files changed, 114 insertions(+), 49 deletions(-) diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index c13b852484e..913675705e3 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -1693,11 +1693,19 @@ class DBSpendUpdateWriter: team_id = key.split("::")[1] user_id = key.split("::")[3] - batcher.litellm_teammembership.update_many( # 'update_many' prevents error from being raised if no row exists - where={"team_id": team_id, "user_id": user_id}, + batcher.litellm_teammembership.upsert( + where={"user_id_team_id": {"user_id": user_id, "team_id": team_id}}, data={ - "spend": {"increment": response_cost}, - "total_spend": {"increment": response_cost}, + "create": { + "team_id": team_id, + "user_id": user_id, + "spend": response_cost, + "total_spend": response_cost, + }, + "update": { + "spend": {"increment": response_cost}, + "total_spend": {"increment": response_cost}, + }, }, ) # Transaction succeeded, break out of retry loop diff --git a/litellm/proxy/management_helpers/utils.py b/litellm/proxy/management_helpers/utils.py index f3bd4b0f6dd..d940eba86d3 100644 --- a/litellm/proxy/management_helpers/utils.py +++ b/litellm/proxy/management_helpers/utils.py @@ -86,7 +86,9 @@ class _PrismaUserTable(Protocol): class _PrismaTeamMembershipTable(Protocol): """Team membership table actions the management helpers issue.""" - async def create(self, *, data: Mapping[str, object], include: Mapping[str, bool]) -> _PrismaRecord: ... + async def upsert( + self, *, where: Mapping[str, object], data: Mapping[str, Mapping[str, object]], include: Mapping[str, bool] + ) -> _PrismaRecord: ... class MemberWriteTx(Protocol): @@ -348,7 +350,7 @@ async def _resolve_member_budget_id( default member budget is cloned (with ``budget_duration`` overriding its reset window while keeping its other limits). A lone ``budget_duration`` with no team default creates a window-only budget. With nothing set the - member gets no budget. + member gets no budget, though ``add_new_member`` still writes its membership row. """ has_explicit_limit: Final = max_budget_in_team is not None or allowed_models is not None @@ -415,9 +417,9 @@ async def add_new_member( Add a new member to a team - add team id to user table - - add team member w/ budget to team member table + - add team member to team member table, linked to a budget when one resolves - Returns created/existing user + team membership w/ budget id + Returns created/existing user + team membership (``budget_id`` is ``None`` when no budget applies) Callers already inside a transaction pass it as ``tx`` so every write here runs on that connection instead of borrowing more from the pool while the caller's locks are held. @@ -471,14 +473,13 @@ async def add_new_member( tx=tx, ) - if _budget_id and returned_user is not None and returned_user.user_id is not None: + if returned_user is not None and returned_user.user_id is not None: membership_table: Final[_PrismaTeamMembershipTable] = _team_membership_table(prisma_client, tx) - _returned_team_membership: Final = await membership_table.create( - data={ - "team_id": team_id, - "user_id": returned_user.user_id, - "budget_id": _budget_id, - }, + membership_key: Final[Mapping[str, object]] = {"user_id": returned_user.user_id, "team_id": team_id} + budget_link: Final[Mapping[str, object]] = {"budget_id": _budget_id} if _budget_id is not None else {} + _returned_team_membership: Final = await membership_table.upsert( + where={"user_id_team_id": membership_key}, + data={"create": {**membership_key, **budget_link}, "update": {**budget_link}}, include={"litellm_budget_table": True}, ) diff --git a/litellm/repositories/prisma_protocols.py b/litellm/repositories/prisma_protocols.py index 93b8c5c7cd7..0919ae9f808 100644 --- a/litellm/repositories/prisma_protocols.py +++ b/litellm/repositories/prisma_protocols.py @@ -123,6 +123,8 @@ class BatchTable(Protocol): def update_many(self, *, where: Mapping[str, object], data: Mapping[str, object]) -> None: ... + def upsert(self, *, where: Mapping[str, object], data: Mapping[str, Mapping[str, object]]) -> None: ... + class PrismaBatch(Protocol): @property diff --git a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py index b3f5a60877d..bed903d7a1f 100644 --- a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py +++ b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py @@ -918,7 +918,11 @@ async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total """ Verify that _commit_spend_updates_to_db increments BOTH spend (cycle-scoped) and total_spend (non-resetting) on LiteLLM_TeamMembership in a single - update_many call, using the same response_cost. + upsert call, using the same response_cost. + + Regression (LIT-5502): members added without a budget had no membership row, and + the previous update_many matched zero rows, so their spend was silently dropped. + The upsert has to create the row seeded with this call's cost in that case. """ db_writer = DBSpendUpdateWriter() @@ -930,7 +934,7 @@ async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total mock_batcher.litellm_teamtable = MagicMock() mock_batcher.litellm_teamtable.update_many = MagicMock() mock_batcher.litellm_teammembership = MagicMock() - mock_batcher.litellm_teammembership.update_many = MagicMock() + mock_batcher.litellm_teammembership.upsert = MagicMock() mock_batcher.litellm_organizationtable = MagicMock() mock_batcher.litellm_organizationtable.update_many = MagicMock() mock_batcher.litellm_tagtable = MagicMock() @@ -979,12 +983,21 @@ async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total db_spend_update_transactions=db_spend_update_transactions, ) - mock_batcher.litellm_teammembership.update_many.assert_called_once() - call_kwargs = mock_batcher.litellm_teammembership.update_many.call_args[1] - assert call_kwargs["where"] == {"team_id": team_id, "user_id": user_id} + mock_batcher.litellm_teammembership.upsert.assert_called_once() + mock_batcher.litellm_teammembership.update_many.assert_not_called() + call_kwargs = mock_batcher.litellm_teammembership.upsert.call_args.kwargs + assert call_kwargs["where"] == {"user_id_team_id": {"user_id": user_id, "team_id": team_id}} assert call_kwargs["data"] == { - "spend": {"increment": response_cost}, - "total_spend": {"increment": response_cost}, + "create": { + "team_id": team_id, + "user_id": user_id, + "spend": response_cost, + "total_spend": response_cost, + }, + "update": { + "spend": {"increment": response_cost}, + "total_spend": {"increment": response_cost}, + }, } @@ -2219,9 +2232,13 @@ async def test_commit_daily_tag_spend_no_requeue_on_success(): "team_id::team_b::user_id::user_x": 0.3, }, "litellm_teammembership", - "update_many", - "team_id", - ["team_a", "team_b", "team_c"], + "upsert", + "user_id_team_id", + [ + {"user_id": "user_x", "team_id": "team_a"}, + {"user_id": "user_x", "team_id": "team_b"}, + {"user_id": "user_x", "team_id": "team_c"}, + ], id="team_member", ), pytest.param( diff --git a/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py index 4f5c5066367..f7157da3100 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py @@ -2086,7 +2086,7 @@ async def test_add_team_members_runs_member_writes_on_the_lock_holding_transacti tx.litellm_usertable.upsert = AsyncMock(return_value=added_user) tx.litellm_usertable.update_many = AsyncMock() tx.litellm_budgettable.create = AsyncMock(return_value=created_budget) - tx.litellm_teammembership.create = AsyncMock(return_value=membership) + tx.litellm_teammembership.upsert = AsyncMock(return_value=membership) tx_cm = MagicMock() tx_cm.__aenter__ = AsyncMock(return_value=tx) @@ -5772,7 +5772,7 @@ async def test_new_team_max_budget_within_user_limit(): "budget_id": None, } mock_prisma.db.litellm_teammembership = MagicMock() - mock_prisma.db.litellm_teammembership.create = AsyncMock( + mock_prisma.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_membership ) @@ -5915,7 +5915,7 @@ async def test_new_team_org_scoped_budget_bypasses_user_limit(): "budget_id": None, } mock_prisma.db.litellm_teammembership = MagicMock() - mock_prisma.db.litellm_teammembership.create = AsyncMock( + mock_prisma.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_membership ) @@ -6063,7 +6063,7 @@ async def test_new_team_org_scoped_models_bypasses_user_limit(): "budget_id": None, } mock_prisma.db.litellm_teammembership = MagicMock() - mock_prisma.db.litellm_teammembership.create = AsyncMock( + mock_prisma.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_membership ) @@ -9525,7 +9525,7 @@ async def test_new_team_soft_budget_validation( "budget_id": None, } mock_prisma.db.litellm_teammembership = MagicMock() - mock_prisma.db.litellm_teammembership.create = AsyncMock( + mock_prisma.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_membership ) diff --git a/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py b/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py index a6b1fc32eda..0dcf7e1b8ad 100644 --- a/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py +++ b/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py @@ -234,7 +234,7 @@ async def test_add_new_member_clones_default_team_budget_id(): "budget_id": test_cloned_budget_id, "litellm_budget_table": None, } - mock_prisma_client.db.litellm_teammembership.create = AsyncMock( + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_team_membership_response ) @@ -257,7 +257,7 @@ async def test_add_new_member_clones_default_team_budget_id(): assert result_team_membership.budget_id != test_default_budget_id mock_prisma_client.db.litellm_usertable.upsert.assert_called_once() - mock_prisma_client.db.litellm_teammembership.create.assert_called_once() + mock_prisma_client.db.litellm_teammembership.upsert.assert_called_once() # The clone must have happened: find_unique on the default, create for the clone. mock_prisma_client.db.litellm_budgettable.find_unique.assert_called_once_with( @@ -274,9 +274,9 @@ async def test_add_new_member_clones_default_team_budget_id(): assert cloned_create_data["created_by"] == user_api_key_dict.user_id team_membership_call_args = ( - mock_prisma_client.db.litellm_teammembership.create.call_args + mock_prisma_client.db.litellm_teammembership.upsert.call_args ) - create_data = team_membership_call_args.kwargs["data"] + create_data = team_membership_call_args.kwargs["data"]["create"] assert create_data["budget_id"] == test_cloned_budget_id @@ -332,7 +332,7 @@ async def test_add_new_member_budget_duration_only_clones_default_max_budget(): "budget_id": "cloned-dc", "litellm_budget_table": None, } - mock_prisma_client.db.litellm_teammembership.create = AsyncMock( + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_team_membership_response ) @@ -362,7 +362,8 @@ async def test_add_new_member_no_budget_when_no_default_and_no_max_budget(): Test that add_new_member links no budget to the team membership when neither max_budget_in_team nor default_team_budget_id is provided. - When the team has no default member budget, new members get nothing. + When the team has no default member budget, no budget row is created, but the + membership row still is, otherwise the member's spend has nowhere to accrue. """ from litellm.proxy._types import LitellmUserRoles @@ -393,7 +394,19 @@ async def test_add_new_member_no_budget_when_no_default_and_no_max_budget(): # Even though we mock these, they must NOT be called on the no-budget path. mock_prisma_client.db.litellm_budgettable.find_unique = AsyncMock() mock_prisma_client.db.litellm_budgettable.create = AsyncMock() - mock_prisma_client.db.litellm_teammembership.create = AsyncMock() + + mock_team_membership_response = MagicMock() + mock_team_membership_response.model_dump.return_value = { + "team_id": test_team_id, + "user_id": test_user_id, + "budget_id": None, + "spend": 0.0, + "total_spend": 0.0, + "litellm_budget_table": None, + } + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock( + return_value=mock_team_membership_response + ) result_user, result_team_membership = await add_new_member( new_member=new_member, @@ -408,11 +421,20 @@ async def test_add_new_member_no_budget_when_no_default_and_no_max_budget(): assert result_user is not None assert result_user.user_id == test_user_id - # No budget id, so no team membership row is created. - assert result_team_membership is None mock_prisma_client.db.litellm_budgettable.find_unique.assert_not_called() mock_prisma_client.db.litellm_budgettable.create.assert_not_called() - mock_prisma_client.db.litellm_teammembership.create.assert_not_called() + + # Regression (LIT-5502): the membership row is what per-member spend increments land on, + # so it has to exist even when the member has no budget. Skipping it silently dropped spend. + assert result_team_membership is not None + assert result_team_membership.budget_id is None + mock_prisma_client.db.litellm_teammembership.upsert.assert_awaited_once() + upsert_kwargs = mock_prisma_client.db.litellm_teammembership.upsert.call_args.kwargs + assert upsert_kwargs["where"] == { + "user_id_team_id": {"user_id": test_user_id, "team_id": test_team_id} + } + assert upsert_kwargs["data"]["create"] == {"user_id": test_user_id, "team_id": test_team_id} + assert "budget_id" not in upsert_kwargs["data"]["update"] @pytest.mark.asyncio @@ -473,7 +495,7 @@ async def test_add_new_member_creates_new_budget_when_max_budget_provided(): "budget_id": test_new_budget_id, "litellm_budget_table": None, } - mock_prisma_client.db.litellm_teammembership.create = AsyncMock( + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_team_membership_response ) @@ -502,10 +524,10 @@ async def test_add_new_member_creates_new_budget_when_max_budget_provided(): # Verify the team membership was created with the correct budget_id team_membership_call_args = ( - mock_prisma_client.db.litellm_teammembership.create.call_args + mock_prisma_client.db.litellm_teammembership.upsert.call_args ) assert team_membership_call_args is not None - create_data = team_membership_call_args.kwargs["data"] + create_data = team_membership_call_args.kwargs["data"]["create"] assert create_data["budget_id"] == test_new_budget_id @@ -546,7 +568,7 @@ async def test_add_new_member_persists_budget_duration(): "budget_id": "budget-dur", "litellm_budget_table": None, } - mock_prisma_client.db.litellm_teammembership.create = AsyncMock( + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_team_membership_response ) @@ -610,7 +632,7 @@ async def test_add_new_member_persists_budget_duration_without_max_budget(): "budget_id": "budget-dur2", "litellm_budget_table": None, } - mock_prisma_client.db.litellm_teammembership.create = AsyncMock( + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_team_membership_response ) @@ -700,7 +722,7 @@ async def test_add_new_member_with_user_email_clones_default_budget(): "budget_id": test_cloned_budget_id, "litellm_budget_table": None, } - mock_prisma_client.db.litellm_teammembership.create = AsyncMock( + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock( return_value=mock_team_membership_response ) @@ -1031,8 +1053,15 @@ async def test_add_new_member_appends_team_only_if_absent_for_existing_user(): } mock_prisma_client.db.litellm_usertable.upsert = AsyncMock(return_value=mock_user_after) mock_prisma_client.db.litellm_usertable.update_many = AsyncMock() - # no team default budget and no explicit budget -> no team membership row mock_prisma_client.db.litellm_budgettable.find_unique = AsyncMock(return_value=None) + mock_membership = MagicMock() + mock_membership.model_dump.return_value = { + "team_id": "team-1", + "user_id": "existing-user", + "budget_id": None, + "litellm_budget_table": None, + } + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock(return_value=mock_membership) result_user, _ = await add_new_member( new_member=new_member, @@ -1099,6 +1128,14 @@ async def test_add_new_member_creates_missing_user_atomically_via_upsert(): mock_prisma_client.db.litellm_usertable.update_many = AsyncMock() mock_prisma_client.db.litellm_usertable.create = AsyncMock() mock_prisma_client.db.litellm_budgettable.find_unique = AsyncMock(return_value=None) + mock_membership = MagicMock() + mock_membership.model_dump.return_value = { + "team_id": "team-1", + "user_id": "brand-new-user", + "budget_id": None, + "litellm_budget_table": None, + } + mock_prisma_client.db.litellm_teammembership.upsert = AsyncMock(return_value=mock_membership) result_user, _ = await add_new_member( new_member=new_member, @@ -1147,7 +1184,7 @@ def _member_write_tx() -> MagicMock: tx.litellm_usertable.find_many = AsyncMock(return_value=[]) tx.litellm_budgettable.find_unique = AsyncMock(return_value=None) tx.litellm_budgettable.create = AsyncMock(return_value=created_budget) - tx.litellm_teammembership.create = AsyncMock(return_value=membership) + tx.litellm_teammembership.upsert = AsyncMock(return_value=membership) return tx @@ -1192,7 +1229,7 @@ async def test_add_new_member_runs_every_write_on_the_caller_transaction(new_mem assert result_membership.budget_id == "budget-pool" assert tx.litellm_budgettable.create.await_count == 1 - assert tx.litellm_teammembership.create.await_count == 1 + assert tx.litellm_teammembership.upsert.await_count == 1 assert tx.litellm_usertable.upsert.await_count + tx.litellm_usertable.create.await_count == 1 prisma_client.db.assert_not_called() From d9b48ac9414adbfbcf5c5f15521a7cb4293d278d Mon Sep 17 00:00:00 2001 From: ryan Date: Wed, 16 Sep 2026 01:35:53 +0000 Subject: [PATCH 130/141] test(proxy): mock team membership upsert in team admin member add test Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/proxy_unit_tests/test_proxy_server.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/tests/proxy_unit_tests/test_proxy_server.py b/tests/proxy_unit_tests/test_proxy_server.py index 1fcdaa67143..659097abbb9 100644 --- a/tests/proxy_unit_tests/test_proxy_server.py +++ b/tests/proxy_unit_tests/test_proxy_server.py @@ -1372,6 +1372,7 @@ async def test_create_team_member_add_team_admin( from fastapi import Request from litellm.proxy._types import ( + LiteLLM_TeamMembership, LiteLLM_TeamTableCachedObj, LiteLLM_UserTable, Member, @@ -1454,6 +1455,10 @@ async def test_create_team_member_add_team_admin( team_mock_client.update = AsyncMock( return_value=LiteLLM_TeamTableCachedObj(team_id="1234") ) + membership_mock_client = AsyncMock() + membership_mock_client.upsert = AsyncMock( + return_value=LiteLLM_TeamMembership(user_id="1234", team_id=_team_id) + ) tx_cm = _member_add_tx_cm(team_mock_client) @@ -1463,6 +1468,11 @@ async def test_create_team_member_add_team_admin( "litellm_teamtable", team_mock_client, ), + patch.object( # test-quality-ok: legacy test swaps the prisma table on the module-level client + litellm.proxy.proxy_server.prisma_client.db, + "litellm_teammembership", + membership_mock_client, + ), patch.object( litellm.proxy.proxy_server.prisma_client, "tx", From 6dce85c7285e7e7a6ea1b0b450ff4041815470ba Mon Sep 17 00:00:00 2001 From: ryan Date: Wed, 16 Sep 2026 02:10:29 +0000 Subject: [PATCH 131/141] fix(proxy): skip recreating membership rows for members removed before a spend flush Take the team advisory lock in the spend flush transaction and read the roster through it, so a delayed flush after /team/member_delete cannot recreate the deleted LiteLLM_TeamMembership row. TEAM_ADVISORY_LOCK_SQL moves to team_repository so the spend writer can import it without a circular import Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/db/db_spend_update_writer.py | 72 +++++++--- .../access_group_team_sync.py | 8 +- litellm/repositories/prisma_protocols.py | 8 +- litellm/repositories/team_repository.py | 12 +- .../proxy/db/test_db_spend_update_writer.py | 134 ++++++++++++------ 5 files changed, 160 insertions(+), 74 deletions(-) diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index 913675705e3..5c86c52c539 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -75,7 +75,8 @@ from litellm.proxy.spend_tracking.savings import ( marks_gateway_injection, ) from litellm.proxy.spend_tracking.spend_log_error_logger import spend_log_error -from litellm.repositories.prisma_protocols import BatchTable +from litellm.repositories.prisma_protocols import BatchTable, RawQueryTransaction +from litellm.repositories.team_repository import TEAM_ADVISORY_LOCK_SQL, TeamRepository from litellm.types.utils import CallTypes if TYPE_CHECKING: @@ -133,7 +134,7 @@ class _SpendBatchManager(Protocol): async def __aexit__(self, exc_type: object, exc_value: object, traceback: object) -> bool | None: ... -class _SpendTransaction(Protocol): +class _SpendTransaction(RawQueryTransaction, Protocol): def batch_(self) -> _SpendBatchManager: ... @@ -161,6 +162,50 @@ def _spend_update_tx(prisma_client: PrismaClient) -> _SpendTransactionManager: return tx +async def _lock_and_read_rosters( + prisma_client: PrismaClient, transaction: _SpendTransaction, team_ids: Sequence[str] +) -> frozenset[tuple[str, str]]: + """Take each team's advisory lock on ``transaction`` and return its rostered (user_id, team_id) pairs. + + A spend flush can land after ``/team/member_delete`` removed the member. Holding the same + lock that endpoint takes, until this transaction commits, means a member read here is on + the team for the whole flush, so only they may have a missing membership row created. + """ + repository: Final = TeamRepository(prisma_client) + + async def locked_roster(team_id: str) -> tuple[tuple[str, str], ...]: + await transaction.query_raw(TEAM_ADVISORY_LOCK_SQL, team_id) + roster: Final = await repository.get_members_with_roles_locked(transaction, team_id) + return tuple((member.user_id, team_id) for member in roster or () if member.user_id is not None) + + rosters: Final = tuple([await locked_roster(team_id) for team_id in sorted(frozenset(team_ids))]) + return frozenset(pair for roster in rosters for pair in roster) + + +def _queue_team_member_spend( + memberships: BatchTable, user_id: str, team_id: str, response_cost: float, rostered: bool +) -> None: + increments: Final = { + "spend": {"increment": response_cost}, + "total_spend": {"increment": response_cost}, + } + if not rostered: + memberships.update_many(where={"team_id": team_id, "user_id": user_id}, data=increments) + return + memberships.upsert( + where={"user_id_team_id": {"user_id": user_id, "team_id": team_id}}, + data={ + "create": { + "team_id": team_id, + "user_id": user_id, + "spend": response_cost, + "total_spend": response_cost, + }, + "update": increments, + }, + ) + + def get_llm_router(): """The proxy's router, or None outside a running proxy. @@ -1685,6 +1730,9 @@ class DBSpendUpdateWriter: start_time = time.time() try: async with _spend_update_tx(prisma_client) as transaction: + rostered_members = await _lock_and_read_rosters( + prisma_client, transaction, tuple(team_id for _, team_id in team_memberships_to_invalidate) + ) async with transaction.batch_() as batcher: # Sort by composite key for consistent lock ordering across pods to prevent deadlocks. # Key format "team_id::::user_id::" makes the string sort equivalent to sorting by (team_id, user_id). @@ -1693,20 +1741,12 @@ class DBSpendUpdateWriter: team_id = key.split("::")[1] user_id = key.split("::")[3] - batcher.litellm_teammembership.upsert( - where={"user_id_team_id": {"user_id": user_id, "team_id": team_id}}, - data={ - "create": { - "team_id": team_id, - "user_id": user_id, - "spend": response_cost, - "total_spend": response_cost, - }, - "update": { - "spend": {"increment": response_cost}, - "total_spend": {"increment": response_cost}, - }, - }, + _queue_team_member_spend( + batcher.litellm_teammembership, + user_id, + team_id, + response_cost, + (user_id, team_id) in rostered_members, ) # Transaction succeeded, break out of retry loop break diff --git a/litellm/proxy/management_helpers/access_group_team_sync.py b/litellm/proxy/management_helpers/access_group_team_sync.py index 664e36c9f10..cfe207e66ea 100644 --- a/litellm/proxy/management_helpers/access_group_team_sync.py +++ b/litellm/proxy/management_helpers/access_group_team_sync.py @@ -21,13 +21,7 @@ from typing import Final, Protocol from pydantic import BaseModel, TypeAdapter from litellm.proxy.auth.auth_checks import _delete_cache_access_object - -# hashtext collisions only cost two unrelated teams a little serialization, and the -# lock is never taken by the access-group endpoints as a SELECT ... FOR UPDATE row lock, -# so it cannot join their access-group-then-team lock order to form a cycle. team_endpoints -# reuses this exact statement to serialize /team/member_add and /team/delete against each -# other and against this mirror, rather than defining a second, divergent lock on the same key. -TEAM_ADVISORY_LOCK_SQL: Final = "SELECT pg_advisory_xact_lock(hashtext($1)) IS NULL AS locked" +from litellm.repositories.team_repository import TEAM_ADVISORY_LOCK_SQL _READ_TEAM_SQL: Final = 'SELECT access_group_ids FROM "LiteLLM_TeamTable" WHERE team_id = $1' diff --git a/litellm/repositories/prisma_protocols.py b/litellm/repositories/prisma_protocols.py index 0919ae9f808..c742f3f5f3f 100644 --- a/litellm/repositories/prisma_protocols.py +++ b/litellm/repositories/prisma_protocols.py @@ -7,7 +7,7 @@ private ones per file. """ from collections.abc import Mapping, Sequence -from typing import Protocol, TypeVar +from typing import LiteralString, Protocol, TypeVar RowT_co = TypeVar("RowT_co", covariant=True) @@ -108,6 +108,12 @@ class PrismaRecord(Protocol): def dict(self) -> Mapping[str, object]: ... +class RawQueryTransaction(Protocol): + """A prisma transaction handle that can run raw SQL, e.g. an advisory lock or a locked read.""" + + async def query_raw(self, query: LiteralString, *args: str) -> Sequence[Mapping[str, object]]: ... + + class ReadOnlyTable(Protocol): async def find_many(self, *, where: Mapping[str, object]) -> Sequence[PrismaRecord]: ... diff --git a/litellm/repositories/team_repository.py b/litellm/repositories/team_repository.py index 5ff07d76b5d..1b9cba48e41 100644 --- a/litellm/repositories/team_repository.py +++ b/litellm/repositories/team_repository.py @@ -15,10 +15,9 @@ from litellm.repositories.base_repository import ( DbRecord, record_to_dict, ) -from litellm.repositories.prisma_protocols import TableActions +from litellm.repositories.prisma_protocols import RawQueryTransaction, TableActions if TYPE_CHECKING: - from prisma import Prisma from prisma import models as prisma_models @@ -40,6 +39,13 @@ def _team_arrays(team: LiteLLM_TeamTable) -> _TeamArrays: return team +# hashtext collisions only cost two unrelated teams a little serialization, and the +# lock is never taken by the access-group endpoints as a SELECT ... FOR UPDATE row lock, +# so it cannot join their access-group-then-team lock order to form a cycle. team_endpoints, +# the access-group mirror and the team member spend flush all reuse this exact statement to +# serialize against each other, rather than defining a second, divergent lock on the same key. +TEAM_ADVISORY_LOCK_SQL: Final = "SELECT pg_advisory_xact_lock(hashtext($1)) IS NULL AS locked" + _MEMBERS_WITH_ROLES_ADAPTER: Final = TypeAdapter(list[Member]) _JSON_ENCODED_TEAM_FIELDS: Final = ( "metadata", @@ -78,7 +84,7 @@ class TeamRepository(BaseRepository[LiteLLM_TeamTable]): return LiteLLM_TeamTable.model_validate(data) - async def get_members_with_roles_locked(self, tx: "Prisma", team_id: str) -> list[Member] | None: + async def get_members_with_roles_locked(self, tx: RawQueryTransaction, team_id: str) -> list[Member] | None: """Return the team's members_with_roles. The caller must already hold ``TEAM_ADVISORY_LOCK_SQL`` for this team_id on ``tx`` before calling this. diff --git a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py index bed903d7a1f..b12dc68b7b1 100644 --- a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py +++ b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py @@ -20,6 +20,7 @@ from litellm.proxy.db.db_spend_update_writer import DBSpendUpdateWriter from litellm.proxy.db.db_transaction_queue.window_spend_update_queue import ( build_window_spend_transaction, ) +from litellm.repositories.team_repository import TEAM_ADVISORY_LOCK_SQL @pytest.mark.asyncio @@ -913,38 +914,22 @@ async def test_commit_spend_updates_to_db_increments_agent_spend(): assert call_kwargs["data"] == {"spend": {"increment": response_cost}} -@pytest.mark.asyncio -async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total_spend(): - """ - Verify that _commit_spend_updates_to_db increments BOTH spend (cycle-scoped) - and total_spend (non-resetting) on LiteLLM_TeamMembership in a single - upsert call, using the same response_cost. - - Regression (LIT-5502): members added without a budget had no membership row, and - the previous update_many matched zero rows, so their spend was silently dropped. - The upsert has to create the row seeded with this call's cost in that case. - """ - db_writer = DBSpendUpdateWriter() - +def _team_member_flush_fixtures(team_id: str, rostered_user_ids: list[str]) -> tuple[MagicMock, AsyncMock, MagicMock]: + """A batcher, transaction and prisma client whose locked roster read for `team_id` lists `rostered_user_ids`.""" mock_batcher = MagicMock() - mock_batcher.litellm_verificationtoken = MagicMock() - mock_batcher.litellm_verificationtoken.update_many = MagicMock() - mock_batcher.litellm_usertable = MagicMock() - mock_batcher.litellm_usertable.update_many = MagicMock() - mock_batcher.litellm_teamtable = MagicMock() - mock_batcher.litellm_teamtable.update_many = MagicMock() mock_batcher.litellm_teammembership = MagicMock() mock_batcher.litellm_teammembership.upsert = MagicMock() - mock_batcher.litellm_organizationtable = MagicMock() - mock_batcher.litellm_organizationtable.update_many = MagicMock() - mock_batcher.litellm_tagtable = MagicMock() - mock_batcher.litellm_tagtable.update_many = MagicMock() - mock_batcher.litellm_agentstable = MagicMock() - mock_batcher.litellm_agentstable.update_many = MagicMock() + mock_batcher.litellm_teammembership.update_many = MagicMock() + + roster_row = {"members_with_roles": json.dumps([{"user_id": uid, "role": "user"} for uid in rostered_user_ids])} + + async def query_raw(query: str, *args: str) -> list[dict[str, object]]: + return [] if query == TEAM_ADVISORY_LOCK_SQL else [roster_row] mock_transaction = AsyncMock() mock_transaction.__aenter__ = AsyncMock(return_value=mock_transaction) mock_transaction.__aexit__ = AsyncMock(return_value=False) + mock_transaction.query_raw = AsyncMock(side_effect=query_raw) mock_transaction.batch_ = MagicMock( return_value=AsyncMock( __aenter__=AsyncMock(return_value=mock_batcher), @@ -955,16 +940,11 @@ async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total mock_prisma_client = MagicMock() mock_prisma_client.db = MagicMock() mock_prisma_client.db.tx = MagicMock(return_value=mock_transaction) + return mock_batcher, mock_transaction, mock_prisma_client - mock_proxy_logging = MagicMock() - # Skip team-membership cache invalidation — out of scope for this test. - mock_proxy_logging.call_details.get = MagicMock(return_value=None) - team_id = "team-abc" - user_id = "user-xyz" - response_cost = 0.75 - entity_id = f"team_id::{team_id}::user_id::{user_id}" - db_spend_update_transactions = { +def _team_member_only_transactions(entity_id: str, response_cost: float) -> dict[str, dict[str, float]]: + return { "user_list_transactions": {}, "end_user_list_transactions": {}, "key_list_transactions": {}, @@ -975,14 +955,38 @@ async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total "agent_list_transactions": {}, } - with patch("litellm.proxy.utils._raise_failed_update_spend_exception"): - await db_writer._commit_spend_updates_to_db( - prisma_client=mock_prisma_client, - n_retry_times=0, - proxy_logging_obj=mock_proxy_logging, - db_spend_update_transactions=db_spend_update_transactions, - ) +@pytest.mark.asyncio +async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total_spend(): + """ + Verify that _commit_spend_updates_to_db increments BOTH spend (cycle-scoped) + and total_spend (non-resetting) on LiteLLM_TeamMembership in a single + upsert call, using the same response_cost. + + Regression (LIT-5502): members added without a budget had no membership row, and + the previous update_many matched zero rows, so their spend was silently dropped. + For a member still on the team roster the upsert has to create the row seeded + with this call's cost in that case. + """ + db_writer = DBSpendUpdateWriter() + team_id = "team-abc" + user_id = "user-xyz" + response_cost = 0.75 + mock_batcher, mock_transaction, mock_prisma_client = _team_member_flush_fixtures(team_id, [user_id]) + + mock_proxy_logging = MagicMock() + mock_proxy_logging.call_details.get = MagicMock(return_value=None) + + await db_writer._commit_spend_updates_to_db( + prisma_client=mock_prisma_client, + n_retry_times=0, + proxy_logging_obj=mock_proxy_logging, + db_spend_update_transactions=_team_member_only_transactions( + f"team_id::{team_id}::user_id::{user_id}", response_cost + ), + ) + + assert mock_transaction.query_raw.await_args_list[0] == call(TEAM_ADVISORY_LOCK_SQL, team_id) mock_batcher.litellm_teammembership.upsert.assert_called_once() mock_batcher.litellm_teammembership.update_many.assert_not_called() call_kwargs = mock_batcher.litellm_teammembership.upsert.call_args.kwargs @@ -1001,6 +1005,43 @@ async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total } +@pytest.mark.asyncio +async def test_commit_spend_updates_to_db_does_not_recreate_membership_of_removed_team_member(): + """ + A spend flush that lands after /team/member_delete must not resurrect the deleted + membership row: a user missing from the team roster, read under the team's advisory + lock inside the flush transaction, only gets an increment on whatever row still + exists, never a create. + """ + db_writer = DBSpendUpdateWriter() + team_id = "team-abc" + removed_user_id = "user-removed" + response_cost = 0.75 + mock_batcher, mock_transaction, mock_prisma_client = _team_member_flush_fixtures(team_id, ["user-still-here"]) + + mock_proxy_logging = MagicMock() + mock_proxy_logging.call_details.get = MagicMock(return_value=None) + + await db_writer._commit_spend_updates_to_db( + prisma_client=mock_prisma_client, + n_retry_times=0, + proxy_logging_obj=mock_proxy_logging, + db_spend_update_transactions=_team_member_only_transactions( + f"team_id::{team_id}::user_id::{removed_user_id}", response_cost + ), + ) + + assert mock_transaction.query_raw.await_args_list[0] == call(TEAM_ADVISORY_LOCK_SQL, team_id) + mock_batcher.litellm_teammembership.upsert.assert_not_called() + mock_batcher.litellm_teammembership.update_many.assert_called_once_with( + where={"team_id": team_id, "user_id": removed_user_id}, + data={ + "spend": {"increment": response_cost}, + "total_spend": {"increment": response_cost}, + }, + ) + + @pytest.mark.asyncio async def test_org_spend_increments_organization_membership_row_for_the_calling_user(): """A request made with a user_id inside an org must increment that user's @@ -2232,13 +2273,9 @@ async def test_commit_daily_tag_spend_no_requeue_on_success(): "team_id::team_b::user_id::user_x": 0.3, }, "litellm_teammembership", - "upsert", - "user_id_team_id", - [ - {"user_id": "user_x", "team_id": "team_a"}, - {"user_id": "user_x", "team_id": "team_b"}, - {"user_id": "user_x", "team_id": "team_c"}, - ], + "update_many", + "team_id", + ["team_a", "team_b", "team_c"], id="team_member", ), pytest.param( @@ -2312,6 +2349,8 @@ async def test_commit_spend_updates_iterates_in_sorted_order( ) ) + mock_transaction.query_raw = AsyncMock(return_value=[]) + mock_prisma_client = MagicMock() mock_prisma_client.db = MagicMock() mock_prisma_client.db.tx = MagicMock(return_value=mock_transaction) @@ -3071,6 +3110,7 @@ def _good_tx(mock_batcher): tx = AsyncMock() tx.__aenter__ = AsyncMock(return_value=tx) tx.__aexit__ = AsyncMock(return_value=False) + tx.query_raw = AsyncMock(return_value=[]) tx.batch_ = MagicMock( return_value=AsyncMock( __aenter__=AsyncMock(return_value=mock_batcher), From 7c068a4cf7772b45246568346e36cef7e4f27ec2 Mon Sep 17 00:00:00 2001 From: ryan Date: Wed, 16 Sep 2026 02:16:11 +0000 Subject: [PATCH 132/141] fix(repositories): import LiteralString from typing_extensions for python 3.10 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/repositories/prisma_protocols.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/litellm/repositories/prisma_protocols.py b/litellm/repositories/prisma_protocols.py index c742f3f5f3f..c7421c9284a 100644 --- a/litellm/repositories/prisma_protocols.py +++ b/litellm/repositories/prisma_protocols.py @@ -7,7 +7,9 @@ private ones per file. """ from collections.abc import Mapping, Sequence -from typing import LiteralString, Protocol, TypeVar +from typing import Protocol, TypeVar + +from typing_extensions import LiteralString RowT_co = TypeVar("RowT_co", covariant=True) From efef6ab68491299a550739286990cf922330dd89 Mon Sep 17 00:00:00 2001 From: ryan Date: Wed, 16 Sep 2026 02:53:47 +0000 Subject: [PATCH 133/141] fix(proxy): write team member spend as one roster checked upsert statement Replaces the per team advisory lock and Pydantic roster parse in the spend flush with a single INSERT ... ON CONFLICT statement that checks the stored roster in SQL, so malformed roster JSON cannot fail the whole flush and large batches no longer issue two queries per team inside the fixed transaction deadline Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/db/db_spend_update_writer.py | 95 +++++-------- .../access_group_team_sync.py | 8 +- litellm/repositories/prisma_protocols.py | 10 -- litellm/repositories/team_repository.py | 12 +- .../proxy/db/test_db_spend_update_writer.py | 126 ++++++------------ 5 files changed, 87 insertions(+), 164 deletions(-) diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index 5c86c52c539..ecf107ef4be 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -18,7 +18,7 @@ from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, cast, overload from urllib.parse import quote, unquote -from typing_extensions import ReadOnly, TypedDict +from typing_extensions import LiteralString, ReadOnly, TypedDict import litellm from litellm._logging import verbose_proxy_logger @@ -75,8 +75,7 @@ from litellm.proxy.spend_tracking.savings import ( marks_gateway_injection, ) from litellm.proxy.spend_tracking.spend_log_error_logger import spend_log_error -from litellm.repositories.prisma_protocols import BatchTable, RawQueryTransaction -from litellm.repositories.team_repository import TEAM_ADVISORY_LOCK_SQL, TeamRepository +from litellm.repositories.prisma_protocols import BatchTable from litellm.types.utils import CallTypes if TYPE_CHECKING: @@ -134,9 +133,11 @@ class _SpendBatchManager(Protocol): async def __aexit__(self, exc_type: object, exc_value: object, traceback: object) -> bool | None: ... -class _SpendTransaction(RawQueryTransaction, Protocol): +class _SpendTransaction(Protocol): def batch_(self) -> _SpendBatchManager: ... + async def execute_raw(self, query: LiteralString, *args: object) -> int: ... + class _SpendTransactionManager(Protocol): async def __aenter__(self) -> _SpendTransaction: ... @@ -162,47 +163,36 @@ def _spend_update_tx(prisma_client: PrismaClient) -> _SpendTransactionManager: return tx -async def _lock_and_read_rosters( - prisma_client: PrismaClient, transaction: _SpendTransaction, team_ids: Sequence[str] -) -> frozenset[tuple[str, str]]: - """Take each team's advisory lock on ``transaction`` and return its rostered (user_id, team_id) pairs. - - A spend flush can land after ``/team/member_delete`` removed the member. Holding the same - lock that endpoint takes, until this transaction commits, means a member read here is on - the team for the whole flush, so only they may have a missing membership row created. - """ - repository: Final = TeamRepository(prisma_client) - - async def locked_roster(team_id: str) -> tuple[tuple[str, str], ...]: - await transaction.query_raw(TEAM_ADVISORY_LOCK_SQL, team_id) - roster: Final = await repository.get_members_with_roles_locked(transaction, team_id) - return tuple((member.user_id, team_id) for member in roster or () if member.user_id is not None) - - rosters: Final = tuple([await locked_roster(team_id) for team_id in sorted(frozenset(team_ids))]) - return frozenset(pair for roster in rosters for pair in roster) +# One statement adds every member's cost to their membership row. A missing row is created only +# while the user is still on the team's roster; FOR SHARE on the team row makes that check and +# the insert atomic against /team/member_delete and /team/delete, which update or delete that row +# before removing memberships, so a spend flush landing after a removal never recreates the member. +_TEAM_MEMBER_SPEND_SQL: Final = """ +INSERT INTO "LiteLLM_TeamMembership" (user_id, team_id, spend, total_spend) +SELECT p.user_id, p.team_id, p.cost, p.cost +FROM unnest($1::text[], $2::text[], $3::float8[]) AS p(user_id, team_id, cost) +WHERE EXISTS ( + SELECT 1 FROM "LiteLLM_TeamTable" t + WHERE t.team_id = p.team_id + AND t.members_with_roles @> jsonb_build_array(jsonb_build_object('user_id', p.user_id)) + FOR SHARE +) + OR EXISTS (SELECT 1 FROM "LiteLLM_TeamMembership" m WHERE m.user_id = p.user_id AND m.team_id = p.team_id) +ON CONFLICT (user_id, team_id) DO UPDATE +SET spend = "LiteLLM_TeamMembership".spend + EXCLUDED.spend, + total_spend = "LiteLLM_TeamMembership".total_spend + EXCLUDED.total_spend +""" -def _queue_team_member_spend( - memberships: BatchTable, user_id: str, team_id: str, response_cost: float, rostered: bool -) -> None: - increments: Final = { - "spend": {"increment": response_cost}, - "total_spend": {"increment": response_cost}, - } - if not rostered: - memberships.update_many(where={"team_id": team_id, "user_id": user_id}, data=increments) - return - memberships.upsert( - where={"user_id_team_id": {"user_id": user_id, "team_id": team_id}}, - data={ - "create": { - "team_id": team_id, - "user_id": user_id, - "spend": response_cost, - "total_spend": response_cost, - }, - "update": increments, - }, +async def _write_team_member_spend(transaction: _SpendTransaction, spend_by_member_key: Mapping[str, float]) -> None: + # key is "team_id::::user_id::"; the string sort orders rows by (team_id, user_id), + # keeping lock order consistent across pods to prevent deadlocks + keys: Final = tuple(sorted(spend_by_member_key)) + _ = await transaction.execute_raw( + _TEAM_MEMBER_SPEND_SQL, + [key.split("::")[3] for key in keys], + [key.split("::")[1] for key in keys], + [spend_by_member_key[key] for key in keys], ) @@ -1730,24 +1720,7 @@ class DBSpendUpdateWriter: start_time = time.time() try: async with _spend_update_tx(prisma_client) as transaction: - rostered_members = await _lock_and_read_rosters( - prisma_client, transaction, tuple(team_id for _, team_id in team_memberships_to_invalidate) - ) - async with transaction.batch_() as batcher: - # Sort by composite key for consistent lock ordering across pods to prevent deadlocks. - # Key format "team_id::::user_id::" makes the string sort equivalent to sorting by (team_id, user_id). - for key, response_cost in sorted(team_member_list_transactions.items()): - # key is "team_id::::user_id::" - team_id = key.split("::")[1] - user_id = key.split("::")[3] - - _queue_team_member_spend( - batcher.litellm_teammembership, - user_id, - team_id, - response_cost, - (user_id, team_id) in rostered_members, - ) + await _write_team_member_spend(transaction, team_member_list_transactions) # Transaction succeeded, break out of retry loop break except Exception as e: diff --git a/litellm/proxy/management_helpers/access_group_team_sync.py b/litellm/proxy/management_helpers/access_group_team_sync.py index cfe207e66ea..664e36c9f10 100644 --- a/litellm/proxy/management_helpers/access_group_team_sync.py +++ b/litellm/proxy/management_helpers/access_group_team_sync.py @@ -21,7 +21,13 @@ from typing import Final, Protocol from pydantic import BaseModel, TypeAdapter from litellm.proxy.auth.auth_checks import _delete_cache_access_object -from litellm.repositories.team_repository import TEAM_ADVISORY_LOCK_SQL + +# hashtext collisions only cost two unrelated teams a little serialization, and the +# lock is never taken by the access-group endpoints as a SELECT ... FOR UPDATE row lock, +# so it cannot join their access-group-then-team lock order to form a cycle. team_endpoints +# reuses this exact statement to serialize /team/member_add and /team/delete against each +# other and against this mirror, rather than defining a second, divergent lock on the same key. +TEAM_ADVISORY_LOCK_SQL: Final = "SELECT pg_advisory_xact_lock(hashtext($1)) IS NULL AS locked" _READ_TEAM_SQL: Final = 'SELECT access_group_ids FROM "LiteLLM_TeamTable" WHERE team_id = $1' diff --git a/litellm/repositories/prisma_protocols.py b/litellm/repositories/prisma_protocols.py index c7421c9284a..93b8c5c7cd7 100644 --- a/litellm/repositories/prisma_protocols.py +++ b/litellm/repositories/prisma_protocols.py @@ -9,8 +9,6 @@ private ones per file. from collections.abc import Mapping, Sequence from typing import Protocol, TypeVar -from typing_extensions import LiteralString - RowT_co = TypeVar("RowT_co", covariant=True) @@ -110,12 +108,6 @@ class PrismaRecord(Protocol): def dict(self) -> Mapping[str, object]: ... -class RawQueryTransaction(Protocol): - """A prisma transaction handle that can run raw SQL, e.g. an advisory lock or a locked read.""" - - async def query_raw(self, query: LiteralString, *args: str) -> Sequence[Mapping[str, object]]: ... - - class ReadOnlyTable(Protocol): async def find_many(self, *, where: Mapping[str, object]) -> Sequence[PrismaRecord]: ... @@ -131,8 +123,6 @@ class BatchTable(Protocol): def update_many(self, *, where: Mapping[str, object], data: Mapping[str, object]) -> None: ... - def upsert(self, *, where: Mapping[str, object], data: Mapping[str, Mapping[str, object]]) -> None: ... - class PrismaBatch(Protocol): @property diff --git a/litellm/repositories/team_repository.py b/litellm/repositories/team_repository.py index 1b9cba48e41..5ff07d76b5d 100644 --- a/litellm/repositories/team_repository.py +++ b/litellm/repositories/team_repository.py @@ -15,9 +15,10 @@ from litellm.repositories.base_repository import ( DbRecord, record_to_dict, ) -from litellm.repositories.prisma_protocols import RawQueryTransaction, TableActions +from litellm.repositories.prisma_protocols import TableActions if TYPE_CHECKING: + from prisma import Prisma from prisma import models as prisma_models @@ -39,13 +40,6 @@ def _team_arrays(team: LiteLLM_TeamTable) -> _TeamArrays: return team -# hashtext collisions only cost two unrelated teams a little serialization, and the -# lock is never taken by the access-group endpoints as a SELECT ... FOR UPDATE row lock, -# so it cannot join their access-group-then-team lock order to form a cycle. team_endpoints, -# the access-group mirror and the team member spend flush all reuse this exact statement to -# serialize against each other, rather than defining a second, divergent lock on the same key. -TEAM_ADVISORY_LOCK_SQL: Final = "SELECT pg_advisory_xact_lock(hashtext($1)) IS NULL AS locked" - _MEMBERS_WITH_ROLES_ADAPTER: Final = TypeAdapter(list[Member]) _JSON_ENCODED_TEAM_FIELDS: Final = ( "metadata", @@ -84,7 +78,7 @@ class TeamRepository(BaseRepository[LiteLLM_TeamTable]): return LiteLLM_TeamTable.model_validate(data) - async def get_members_with_roles_locked(self, tx: RawQueryTransaction, team_id: str) -> list[Member] | None: + async def get_members_with_roles_locked(self, tx: "Prisma", team_id: str) -> list[Member] | None: """Return the team's members_with_roles. The caller must already hold ``TEAM_ADVISORY_LOCK_SQL`` for this team_id on ``tx`` before calling this. diff --git a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py index b12dc68b7b1..05769eea343 100644 --- a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py +++ b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py @@ -16,11 +16,10 @@ from redis.exceptions import DataError import litellm from litellm.proxy._types import Litellm_EntityType -from litellm.proxy.db.db_spend_update_writer import DBSpendUpdateWriter +from litellm.proxy.db.db_spend_update_writer import _TEAM_MEMBER_SPEND_SQL, DBSpendUpdateWriter from litellm.proxy.db.db_transaction_queue.window_spend_update_queue import ( build_window_spend_transaction, ) -from litellm.repositories.team_repository import TEAM_ADVISORY_LOCK_SQL @pytest.mark.asyncio @@ -914,42 +913,26 @@ async def test_commit_spend_updates_to_db_increments_agent_spend(): assert call_kwargs["data"] == {"spend": {"increment": response_cost}} -def _team_member_flush_fixtures(team_id: str, rostered_user_ids: list[str]) -> tuple[MagicMock, AsyncMock, MagicMock]: - """A batcher, transaction and prisma client whose locked roster read for `team_id` lists `rostered_user_ids`.""" - mock_batcher = MagicMock() - mock_batcher.litellm_teammembership = MagicMock() - mock_batcher.litellm_teammembership.upsert = MagicMock() - mock_batcher.litellm_teammembership.update_many = MagicMock() - - roster_row = {"members_with_roles": json.dumps([{"user_id": uid, "role": "user"} for uid in rostered_user_ids])} - - async def query_raw(query: str, *args: str) -> list[dict[str, object]]: - return [] if query == TEAM_ADVISORY_LOCK_SQL else [roster_row] - +def _team_member_flush_fixtures() -> tuple[AsyncMock, MagicMock]: + """A transaction and prisma client that record the raw statement the member spend flush runs.""" mock_transaction = AsyncMock() mock_transaction.__aenter__ = AsyncMock(return_value=mock_transaction) mock_transaction.__aexit__ = AsyncMock(return_value=False) - mock_transaction.query_raw = AsyncMock(side_effect=query_raw) - mock_transaction.batch_ = MagicMock( - return_value=AsyncMock( - __aenter__=AsyncMock(return_value=mock_batcher), - __aexit__=AsyncMock(return_value=False), - ) - ) + mock_transaction.execute_raw = AsyncMock(return_value=1) mock_prisma_client = MagicMock() mock_prisma_client.db = MagicMock() mock_prisma_client.db.tx = MagicMock(return_value=mock_transaction) - return mock_batcher, mock_transaction, mock_prisma_client + return mock_transaction, mock_prisma_client -def _team_member_only_transactions(entity_id: str, response_cost: float) -> dict[str, dict[str, float]]: +def _team_member_only_transactions(spend_by_member_key: dict[str, float]) -> dict[str, dict[str, float]]: return { "user_list_transactions": {}, "end_user_list_transactions": {}, "key_list_transactions": {}, "team_list_transactions": {}, - "team_member_list_transactions": {entity_id: response_cost}, + "team_member_list_transactions": spend_by_member_key, "org_list_transactions": {}, "tag_list_transactions": {}, "agent_list_transactions": {}, @@ -957,22 +940,20 @@ def _team_member_only_transactions(entity_id: str, response_cost: float) -> dict @pytest.mark.asyncio -async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total_spend(): +async def test_commit_spend_updates_to_db_writes_team_member_spend_in_one_roster_checked_upsert(): """ - Verify that _commit_spend_updates_to_db increments BOTH spend (cycle-scoped) - and total_spend (non-resetting) on LiteLLM_TeamMembership in a single - upsert call, using the same response_cost. + Regression (LIT-5502): members added without a budget had no membership row, and the + previous update_many matched zero rows, so their spend was silently dropped. - Regression (LIT-5502): members added without a budget had no membership row, and - the previous update_many matched zero rows, so their spend was silently dropped. - For a member still on the team roster the upsert has to create the row seeded - with this call's cost in that case. + The flush now runs one INSERT ... ON CONFLICT statement for the whole batch that adds + the cost to both spend and total_spend and creates the missing row for a user still on + the team roster, so no per-team read can fail or time out ahead of the writes. """ db_writer = DBSpendUpdateWriter() team_id = "team-abc" user_id = "user-xyz" response_cost = 0.75 - mock_batcher, mock_transaction, mock_prisma_client = _team_member_flush_fixtures(team_id, [user_id]) + mock_transaction, mock_prisma_client = _team_member_flush_fixtures() mock_proxy_logging = MagicMock() mock_proxy_logging.call_details.get = MagicMock(return_value=None) @@ -982,42 +963,31 @@ async def test_commit_spend_updates_to_db_increments_team_member_spend_and_total n_retry_times=0, proxy_logging_obj=mock_proxy_logging, db_spend_update_transactions=_team_member_only_transactions( - f"team_id::{team_id}::user_id::{user_id}", response_cost + {f"team_id::{team_id}::user_id::{user_id}": response_cost} ), ) - assert mock_transaction.query_raw.await_args_list[0] == call(TEAM_ADVISORY_LOCK_SQL, team_id) - mock_batcher.litellm_teammembership.upsert.assert_called_once() - mock_batcher.litellm_teammembership.update_many.assert_not_called() - call_kwargs = mock_batcher.litellm_teammembership.upsert.call_args.kwargs - assert call_kwargs["where"] == {"user_id_team_id": {"user_id": user_id, "team_id": team_id}} - assert call_kwargs["data"] == { - "create": { - "team_id": team_id, - "user_id": user_id, - "spend": response_cost, - "total_spend": response_cost, - }, - "update": { - "spend": {"increment": response_cost}, - "total_spend": {"increment": response_cost}, - }, - } + mock_transaction.execute_raw.assert_awaited_once() + statement, user_ids, team_ids, costs = mock_transaction.execute_raw.await_args.args + assert statement is _TEAM_MEMBER_SPEND_SQL + assert (user_ids, team_ids, costs) == ([user_id], [team_id], [response_cost]) + assert 'INSERT INTO "LiteLLM_TeamMembership"' in statement + assert "members_with_roles @> jsonb_build_array(jsonb_build_object('user_id', p.user_id))" in statement + assert "FOR SHARE" in statement + assert "ON CONFLICT (user_id, team_id) DO UPDATE" in statement + assert 'spend = "LiteLLM_TeamMembership".spend + EXCLUDED.spend' in statement + assert 'total_spend = "LiteLLM_TeamMembership".total_spend + EXCLUDED.total_spend' in statement @pytest.mark.asyncio -async def test_commit_spend_updates_to_db_does_not_recreate_membership_of_removed_team_member(): +async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_user(): """ - A spend flush that lands after /team/member_delete must not resurrect the deleted - membership row: a user missing from the team roster, read under the team's advisory - lock inside the flush transaction, only gets an increment on whatever row still - exists, never a create. + The single member spend statement locks rows in the order of its input arrays, so the + batch is handed over sorted by (team_id, user_id), with each cost kept next to its + member, so concurrent pods lock in the same order and cannot deadlock. """ db_writer = DBSpendUpdateWriter() - team_id = "team-abc" - removed_user_id = "user-removed" - response_cost = 0.75 - mock_batcher, mock_transaction, mock_prisma_client = _team_member_flush_fixtures(team_id, ["user-still-here"]) + mock_transaction, mock_prisma_client = _team_member_flush_fixtures() mock_proxy_logging = MagicMock() mock_proxy_logging.call_details.get = MagicMock(return_value=None) @@ -1027,19 +997,22 @@ async def test_commit_spend_updates_to_db_does_not_recreate_membership_of_remove n_retry_times=0, proxy_logging_obj=mock_proxy_logging, db_spend_update_transactions=_team_member_only_transactions( - f"team_id::{team_id}::user_id::{removed_user_id}", response_cost + { + "team_id::team_c::user_id::user_x": 0.1, + "team_id::team_a::user_id::user_y": 0.2, + "team_id::team_a::user_id::user_x": 0.3, + "team_id::team_b::user_id::user_x": 0.4, + } ), ) - assert mock_transaction.query_raw.await_args_list[0] == call(TEAM_ADVISORY_LOCK_SQL, team_id) - mock_batcher.litellm_teammembership.upsert.assert_not_called() - mock_batcher.litellm_teammembership.update_many.assert_called_once_with( - where={"team_id": team_id, "user_id": removed_user_id}, - data={ - "spend": {"increment": response_cost}, - "total_spend": {"increment": response_cost}, - }, - ) + _statement, user_ids, team_ids, costs = mock_transaction.execute_raw.await_args.args + assert list(zip(team_ids, user_ids, costs)) == [ + ("team_a", "user_x", 0.3), + ("team_a", "user_y", 0.2), + ("team_b", "user_x", 0.4), + ("team_c", "user_x", 0.1), + ] @pytest.mark.asyncio @@ -2265,19 +2238,6 @@ async def test_commit_daily_tag_spend_no_requeue_on_success(): ["team_a", "team_b", "team_c"], id="team", ), - pytest.param( - "team_member_list_transactions", - { - "team_id::team_c::user_id::user_x": 0.1, - "team_id::team_a::user_id::user_x": 0.2, - "team_id::team_b::user_id::user_x": 0.3, - }, - "litellm_teammembership", - "update_many", - "team_id", - ["team_a", "team_b", "team_c"], - id="team_member", - ), pytest.param( "org_list_transactions", {"org_c": 0.1, "org_a": 0.2, "org_b": 0.3}, From 60077e90aabfcb52c9babeb19c28efb3ad0c1fdd Mon Sep 17 00:00:00 2001 From: ryan Date: Wed, 16 Sep 2026 03:16:54 +0000 Subject: [PATCH 134/141] fix(proxy): take the team advisory lock before the member spend flush Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/db/db_spend_update_writer.py | 18 +++++++--- .../proxy/db/test_db_spend_update_writer.py | 34 +++++++++++++------ 2 files changed, 36 insertions(+), 16 deletions(-) diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index ecf107ef4be..f617302d8ad 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -163,10 +163,17 @@ def _spend_update_tx(prisma_client: PrismaClient) -> _SpendTransactionManager: return tx +# The per-team advisory lock the team endpoints hold while changing a roster (TEAM_ADVISORY_LOCK_SQL), +# taken in sorted order so the roster check below cannot interleave with their writes. A row lock would +# deadlock with the access-group endpoints, which lock a team row after an access-group lock. +_TEAM_ADVISORY_LOCKS_SQL: Final = """ +SELECT pg_advisory_xact_lock(hashtext(teams.team_id)) +FROM (SELECT DISTINCT team_id FROM unnest($1::text[]) AS team_id ORDER BY team_id) AS teams +""" + # One statement adds every member's cost to their membership row. A missing row is created only -# while the user is still on the team's roster; FOR SHARE on the team row makes that check and -# the insert atomic against /team/member_delete and /team/delete, which update or delete that row -# before removing memberships, so a spend flush landing after a removal never recreates the member. +# while the user is still on the team's roster, so a spend flush landing after a removal never +# recreates the member. _TEAM_MEMBER_SPEND_SQL: Final = """ INSERT INTO "LiteLLM_TeamMembership" (user_id, team_id, spend, total_spend) SELECT p.user_id, p.team_id, p.cost, p.cost @@ -175,7 +182,6 @@ WHERE EXISTS ( SELECT 1 FROM "LiteLLM_TeamTable" t WHERE t.team_id = p.team_id AND t.members_with_roles @> jsonb_build_array(jsonb_build_object('user_id', p.user_id)) - FOR SHARE ) OR EXISTS (SELECT 1 FROM "LiteLLM_TeamMembership" m WHERE m.user_id = p.user_id AND m.team_id = p.team_id) ON CONFLICT (user_id, team_id) DO UPDATE @@ -188,10 +194,12 @@ async def _write_team_member_spend(transaction: _SpendTransaction, spend_by_memb # key is "team_id::::user_id::"; the string sort orders rows by (team_id, user_id), # keeping lock order consistent across pods to prevent deadlocks keys: Final = tuple(sorted(spend_by_member_key)) + team_ids: Final = [key.split("::")[1] for key in keys] + _ = await transaction.execute_raw(_TEAM_ADVISORY_LOCKS_SQL, team_ids) _ = await transaction.execute_raw( _TEAM_MEMBER_SPEND_SQL, [key.split("::")[3] for key in keys], - [key.split("::")[1] for key in keys], + team_ids, [spend_by_member_key[key] for key in keys], ) diff --git a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py index 05769eea343..9f3bc5acd92 100644 --- a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py +++ b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py @@ -16,7 +16,11 @@ from redis.exceptions import DataError import litellm from litellm.proxy._types import Litellm_EntityType -from litellm.proxy.db.db_spend_update_writer import _TEAM_MEMBER_SPEND_SQL, DBSpendUpdateWriter +from litellm.proxy.db.db_spend_update_writer import ( + _TEAM_ADVISORY_LOCKS_SQL, + _TEAM_MEMBER_SPEND_SQL, + DBSpendUpdateWriter, +) from litellm.proxy.db.db_transaction_queue.window_spend_update_queue import ( build_window_spend_transaction, ) @@ -945,9 +949,10 @@ async def test_commit_spend_updates_to_db_writes_team_member_spend_in_one_roster Regression (LIT-5502): members added without a budget had no membership row, and the previous update_many matched zero rows, so their spend was silently dropped. - The flush now runs one INSERT ... ON CONFLICT statement for the whole batch that adds - the cost to both spend and total_spend and creates the missing row for a user still on - the team roster, so no per-team read can fail or time out ahead of the writes. + The flush now takes the same per-team advisory lock the team endpoints hold, then runs + one INSERT ... ON CONFLICT statement for the whole batch that adds the cost to both spend + and total_spend and creates the missing row for a user still on the team roster, so no + per-team read can fail or time out ahead of the writes. """ db_writer = DBSpendUpdateWriter() team_id = "team-abc" @@ -967,13 +972,17 @@ async def test_commit_spend_updates_to_db_writes_team_member_spend_in_one_roster ), ) - mock_transaction.execute_raw.assert_awaited_once() - statement, user_ids, team_ids, costs = mock_transaction.execute_raw.await_args.args + lock_call, spend_call = mock_transaction.execute_raw.await_args_list + lock_statement, locked_team_ids = lock_call.args + assert lock_statement is _TEAM_ADVISORY_LOCKS_SQL + assert locked_team_ids == [team_id] + assert "pg_advisory_xact_lock(hashtext(teams.team_id))" in lock_statement + assert "ORDER BY team_id" in lock_statement + statement, user_ids, team_ids, costs = spend_call.args assert statement is _TEAM_MEMBER_SPEND_SQL assert (user_ids, team_ids, costs) == ([user_id], [team_id], [response_cost]) assert 'INSERT INTO "LiteLLM_TeamMembership"' in statement assert "members_with_roles @> jsonb_build_array(jsonb_build_object('user_id', p.user_id))" in statement - assert "FOR SHARE" in statement assert "ON CONFLICT (user_id, team_id) DO UPDATE" in statement assert 'spend = "LiteLLM_TeamMembership".spend + EXCLUDED.spend' in statement assert 'total_spend = "LiteLLM_TeamMembership".total_spend + EXCLUDED.total_spend' in statement @@ -982,9 +991,10 @@ async def test_commit_spend_updates_to_db_writes_team_member_spend_in_one_roster @pytest.mark.asyncio async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_user(): """ - The single member spend statement locks rows in the order of its input arrays, so the - batch is handed over sorted by (team_id, user_id), with each cost kept next to its - member, so concurrent pods lock in the same order and cannot deadlock. + The member spend statement touches rows in the order of its input arrays, so the batch + is handed over sorted by (team_id, user_id), with each cost kept next to its member, and + the advisory lock statement receives the same team ids, so concurrent pods lock in the + same order and cannot deadlock. """ db_writer = DBSpendUpdateWriter() mock_transaction, mock_prisma_client = _team_member_flush_fixtures() @@ -1006,7 +1016,9 @@ async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_u ), ) - _statement, user_ids, team_ids, costs = mock_transaction.execute_raw.await_args.args + lock_call, spend_call = mock_transaction.execute_raw.await_args_list + _statement, user_ids, team_ids, costs = spend_call.args + assert lock_call.args == (_TEAM_ADVISORY_LOCKS_SQL, ["team_a", "team_a", "team_b", "team_c"]) assert list(zip(team_ids, user_ids, costs)) == [ ("team_a", "user_x", 0.3), ("team_a", "user_y", 0.2), From 499c334fce491fa67fc1db92e5e1c160cd00b33d Mon Sep 17 00:00:00 2001 From: ryan Date: Wed, 16 Sep 2026 03:35:34 +0000 Subject: [PATCH 135/141] fix(proxy): lock each team in sorted order before the member spend upsert Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/db/db_spend_update_writer.py | 12 +++++------ .../proxy/db/test_db_spend_update_writer.py | 21 +++++++++++-------- 2 files changed, 17 insertions(+), 16 deletions(-) diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index f617302d8ad..09ce6b6f451 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -164,12 +164,9 @@ def _spend_update_tx(prisma_client: PrismaClient) -> _SpendTransactionManager: # The per-team advisory lock the team endpoints hold while changing a roster (TEAM_ADVISORY_LOCK_SQL), -# taken in sorted order so the roster check below cannot interleave with their writes. A row lock would -# deadlock with the access-group endpoints, which lock a team row after an access-group lock. -_TEAM_ADVISORY_LOCKS_SQL: Final = """ -SELECT pg_advisory_xact_lock(hashtext(teams.team_id)) -FROM (SELECT DISTINCT team_id FROM unnest($1::text[]) AS team_id ORDER BY team_id) AS teams -""" +# so the roster check below cannot interleave with their writes. A row lock would deadlock with the +# access-group endpoints, which lock a team row after an access-group lock. +_TEAM_ADVISORY_LOCK_SQL: Final = "SELECT pg_advisory_xact_lock(hashtext($1)) IS NULL AS locked" # One statement adds every member's cost to their membership row. A missing row is created only # while the user is still on the team's roster, so a spend flush landing after a removal never @@ -195,7 +192,8 @@ async def _write_team_member_spend(transaction: _SpendTransaction, spend_by_memb # keeping lock order consistent across pods to prevent deadlocks keys: Final = tuple(sorted(spend_by_member_key)) team_ids: Final = [key.split("::")[1] for key in keys] - _ = await transaction.execute_raw(_TEAM_ADVISORY_LOCKS_SQL, team_ids) + for team_id in dict.fromkeys(team_ids): + _ = await transaction.execute_raw(_TEAM_ADVISORY_LOCK_SQL, team_id) _ = await transaction.execute_raw( _TEAM_MEMBER_SPEND_SQL, [key.split("::")[3] for key in keys], diff --git a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py index 9f3bc5acd92..2ac9fc8eccd 100644 --- a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py +++ b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py @@ -17,7 +17,7 @@ from redis.exceptions import DataError import litellm from litellm.proxy._types import Litellm_EntityType from litellm.proxy.db.db_spend_update_writer import ( - _TEAM_ADVISORY_LOCKS_SQL, + _TEAM_ADVISORY_LOCK_SQL, _TEAM_MEMBER_SPEND_SQL, DBSpendUpdateWriter, ) @@ -973,11 +973,10 @@ async def test_commit_spend_updates_to_db_writes_team_member_spend_in_one_roster ) lock_call, spend_call = mock_transaction.execute_raw.await_args_list - lock_statement, locked_team_ids = lock_call.args - assert lock_statement is _TEAM_ADVISORY_LOCKS_SQL - assert locked_team_ids == [team_id] - assert "pg_advisory_xact_lock(hashtext(teams.team_id))" in lock_statement - assert "ORDER BY team_id" in lock_statement + lock_statement, locked_team_id = lock_call.args + assert lock_statement is _TEAM_ADVISORY_LOCK_SQL + assert locked_team_id == team_id + assert "pg_advisory_xact_lock(hashtext($1))" in lock_statement statement, user_ids, team_ids, costs = spend_call.args assert statement is _TEAM_MEMBER_SPEND_SQL assert (user_ids, team_ids, costs) == ([user_id], [team_id], [response_cost]) @@ -993,7 +992,7 @@ async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_u """ The member spend statement touches rows in the order of its input arrays, so the batch is handed over sorted by (team_id, user_id), with each cost kept next to its member, and - the advisory lock statement receives the same team ids, so concurrent pods lock in the + each distinct team is locked once, in that same order, so concurrent pods lock in the same order and cannot deadlock. """ db_writer = DBSpendUpdateWriter() @@ -1016,9 +1015,13 @@ async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_u ), ) - lock_call, spend_call = mock_transaction.execute_raw.await_args_list + *lock_calls, spend_call = mock_transaction.execute_raw.await_args_list _statement, user_ids, team_ids, costs = spend_call.args - assert lock_call.args == (_TEAM_ADVISORY_LOCKS_SQL, ["team_a", "team_a", "team_b", "team_c"]) + assert [lock_call.args for lock_call in lock_calls] == [ + (_TEAM_ADVISORY_LOCK_SQL, "team_a"), + (_TEAM_ADVISORY_LOCK_SQL, "team_b"), + (_TEAM_ADVISORY_LOCK_SQL, "team_c"), + ] assert list(zip(team_ids, user_ids, costs)) == [ ("team_a", "user_x", 0.3), ("team_a", "user_y", 0.2), From 2d61fa66b165d8e5e9b4295bba2587346709aab9 Mon Sep 17 00:00:00 2001 From: ryan Date: Thu, 17 Sep 2026 00:37:03 +0000 Subject: [PATCH 136/141] fix(proxy): lock teams in sorted team id order and keep existing member budgets on re-add Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/db/db_spend_update_writer.py | 12 ++++----- litellm/proxy/management_helpers/utils.py | 2 +- .../proxy/db/test_db_spend_update_writer.py | 27 ++++++++++--------- .../test_management_helpers_utils.py | 3 +++ 4 files changed, 24 insertions(+), 20 deletions(-) diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index 09ce6b6f451..77d16e90706 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -188,17 +188,17 @@ SET spend = "LiteLLM_TeamMembership".spend + EXCLUDED.spend, async def _write_team_member_spend(transaction: _SpendTransaction, spend_by_member_key: Mapping[str, float]) -> None: - # key is "team_id::::user_id::"; the string sort orders rows by (team_id, user_id), - # keeping lock order consistent across pods to prevent deadlocks - keys: Final = tuple(sorted(spend_by_member_key)) - team_ids: Final = [key.split("::")[1] for key in keys] + # key is "team_id::::user_id::"; rows are sorted by (team_id, user_id) so the teams are + # locked in the same `sorted(team_ids)` order the team endpoints use, preventing deadlocks + rows: Final = sorted((key.split("::")[1], key.split("::")[3], cost) for key, cost in spend_by_member_key.items()) + team_ids: Final = [team_id for team_id, _user_id, _cost in rows] for team_id in dict.fromkeys(team_ids): _ = await transaction.execute_raw(_TEAM_ADVISORY_LOCK_SQL, team_id) _ = await transaction.execute_raw( _TEAM_MEMBER_SPEND_SQL, - [key.split("::")[3] for key in keys], + [user_id for _team_id, user_id, _cost in rows], team_ids, - [spend_by_member_key[key] for key in keys], + [cost for _team_id, _user_id, cost in rows], ) diff --git a/litellm/proxy/management_helpers/utils.py b/litellm/proxy/management_helpers/utils.py index d940eba86d3..d84d32431ec 100644 --- a/litellm/proxy/management_helpers/utils.py +++ b/litellm/proxy/management_helpers/utils.py @@ -479,7 +479,7 @@ async def add_new_member( budget_link: Final[Mapping[str, object]] = {"budget_id": _budget_id} if _budget_id is not None else {} _returned_team_membership: Final = await membership_table.upsert( where={"user_id_team_id": membership_key}, - data={"create": {**membership_key, **budget_link}, "update": {**budget_link}}, + data={"create": {**membership_key, **budget_link}, "update": {}}, include={"litellm_budget_table": True}, ) diff --git a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py index 2ac9fc8eccd..70199ee0c73 100644 --- a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py +++ b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py @@ -992,8 +992,9 @@ async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_u """ The member spend statement touches rows in the order of its input arrays, so the batch is handed over sorted by (team_id, user_id), with each cost kept next to its member, and - each distinct team is locked once, in that same order, so concurrent pods lock in the - same order and cannot deadlock. + each distinct team is locked once, in `sorted(team_ids)` order, the order /team/delete + locks in, so a concurrent flush and delete cannot deadlock. `eng` and `eng2` pin that: + sorting the composite keys instead would lock `eng2` first because `2` < `:`. """ db_writer = DBSpendUpdateWriter() mock_transaction, mock_prisma_client = _team_member_flush_fixtures() @@ -1007,10 +1008,10 @@ async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_u proxy_logging_obj=mock_proxy_logging, db_spend_update_transactions=_team_member_only_transactions( { - "team_id::team_c::user_id::user_x": 0.1, - "team_id::team_a::user_id::user_y": 0.2, - "team_id::team_a::user_id::user_x": 0.3, - "team_id::team_b::user_id::user_x": 0.4, + "team_id::eng2::user_id::user_x": 0.1, + "team_id::eng::user_id::user_y": 0.2, + "team_id::eng::user_id::user_x": 0.3, + "team_id::eng-b::user_id::user_x": 0.4, } ), ) @@ -1018,15 +1019,15 @@ async def test_commit_spend_updates_to_db_orders_team_member_rows_by_team_then_u *lock_calls, spend_call = mock_transaction.execute_raw.await_args_list _statement, user_ids, team_ids, costs = spend_call.args assert [lock_call.args for lock_call in lock_calls] == [ - (_TEAM_ADVISORY_LOCK_SQL, "team_a"), - (_TEAM_ADVISORY_LOCK_SQL, "team_b"), - (_TEAM_ADVISORY_LOCK_SQL, "team_c"), + (_TEAM_ADVISORY_LOCK_SQL, "eng"), + (_TEAM_ADVISORY_LOCK_SQL, "eng-b"), + (_TEAM_ADVISORY_LOCK_SQL, "eng2"), ] assert list(zip(team_ids, user_ids, costs)) == [ - ("team_a", "user_x", 0.3), - ("team_a", "user_y", 0.2), - ("team_b", "user_x", 0.4), - ("team_c", "user_x", 0.1), + ("eng", "user_x", 0.3), + ("eng", "user_y", 0.2), + ("eng-b", "user_x", 0.4), + ("eng2", "user_x", 0.1), ] diff --git a/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py b/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py index 0dcf7e1b8ad..09d5f684f3d 100644 --- a/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py +++ b/tests/test_litellm/proxy/management_helpers/test_management_helpers_utils.py @@ -446,6 +446,8 @@ async def test_add_new_member_creates_new_budget_when_max_budget_provided(): 1. When max_budget_in_team is provided 2. A new budget is created in the litellm_budgettable 3. The new budget_id is used for the team membership + 4. The upsert's update branch stays empty, so a bulk /team/member_add that names a member + already on the team does not replace the budget_id (and the spend) their existing row carries """ from litellm.proxy._types import LitellmUserRoles @@ -529,6 +531,7 @@ async def test_add_new_member_creates_new_budget_when_max_budget_provided(): assert team_membership_call_args is not None create_data = team_membership_call_args.kwargs["data"]["create"] assert create_data["budget_id"] == test_new_budget_id + assert team_membership_call_args.kwargs["data"]["update"] == {} @pytest.mark.asyncio From b6f4ad190e00a21935479091a423e16e94530778 Mon Sep 17 00:00:00 2001 From: ryan Date: Thu, 17 Sep 2026 00:49:48 +0000 Subject: [PATCH 137/141] refactor(proxy): freeze the member spend arrays and budget link to stay within the type discipline budget Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/db/db_spend_update_writer.py | 9 ++++----- litellm/proxy/management_helpers/utils.py | 4 +++- .../test_litellm/proxy/db/test_db_spend_update_writer.py | 2 +- 3 files changed, 8 insertions(+), 7 deletions(-) diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index 77d16e90706..8bcfe28488e 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -188,17 +188,16 @@ SET spend = "LiteLLM_TeamMembership".spend + EXCLUDED.spend, async def _write_team_member_spend(transaction: _SpendTransaction, spend_by_member_key: Mapping[str, float]) -> None: - # key is "team_id::::user_id::"; rows are sorted by (team_id, user_id) so the teams are - # locked in the same `sorted(team_ids)` order the team endpoints use, preventing deadlocks + # key is "team_id::::user_id::"; locks are taken in sorted team_id order like the team endpoints rows: Final = sorted((key.split("::")[1], key.split("::")[3], cost) for key, cost in spend_by_member_key.items()) - team_ids: Final = [team_id for team_id, _user_id, _cost in rows] + team_ids: Final = tuple(team_id for team_id, _user_id, _cost in rows) for team_id in dict.fromkeys(team_ids): _ = await transaction.execute_raw(_TEAM_ADVISORY_LOCK_SQL, team_id) _ = await transaction.execute_raw( _TEAM_MEMBER_SPEND_SQL, - [user_id for _team_id, user_id, _cost in rows], + tuple(user_id for _team_id, user_id, _cost in rows), team_ids, - [cost for _team_id, _user_id, cost in rows], + tuple(cost for _team_id, _user_id, cost in rows), ) diff --git a/litellm/proxy/management_helpers/utils.py b/litellm/proxy/management_helpers/utils.py index d84d32431ec..c10e5f9b23d 100644 --- a/litellm/proxy/management_helpers/utils.py +++ b/litellm/proxy/management_helpers/utils.py @@ -476,7 +476,9 @@ async def add_new_member( if returned_user is not None and returned_user.user_id is not None: membership_table: Final[_PrismaTeamMembershipTable] = _team_membership_table(prisma_client, tx) membership_key: Final[Mapping[str, object]] = {"user_id": returned_user.user_id, "team_id": team_id} - budget_link: Final[Mapping[str, object]] = {"budget_id": _budget_id} if _budget_id is not None else {} + budget_link: Final[Mapping[str, str]] = ( + MappingProxyType({"budget_id": _budget_id}) if _budget_id is not None else MappingProxyType({}) + ) _returned_team_membership: Final = await membership_table.upsert( where={"user_id_team_id": membership_key}, data={"create": {**membership_key, **budget_link}, "update": {}}, diff --git a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py index 70199ee0c73..e38a65f2b12 100644 --- a/tests/test_litellm/proxy/db/test_db_spend_update_writer.py +++ b/tests/test_litellm/proxy/db/test_db_spend_update_writer.py @@ -979,7 +979,7 @@ async def test_commit_spend_updates_to_db_writes_team_member_spend_in_one_roster assert "pg_advisory_xact_lock(hashtext($1))" in lock_statement statement, user_ids, team_ids, costs = spend_call.args assert statement is _TEAM_MEMBER_SPEND_SQL - assert (user_ids, team_ids, costs) == ([user_id], [team_id], [response_cost]) + assert (list(user_ids), list(team_ids), list(costs)) == ([user_id], [team_id], [response_cost]) assert 'INSERT INTO "LiteLLM_TeamMembership"' in statement assert "members_with_roles @> jsonb_build_array(jsonb_build_object('user_id', p.user_id))" in statement assert "ON CONFLICT (user_id, team_id) DO UPDATE" in statement From 61d4c5b9b5b427929f2d97d216f3aafb611f0abb Mon Sep 17 00:00:00 2001 From: ryan Date: Thu, 17 Sep 2026 01:12:54 +0000 Subject: [PATCH 138/141] fix(proxy): skip members already on the team before resolving a per-member budget A mixed /team/member_add list that names an existing member used to run add_new_member for them, which created or cloned a budget that the empty upsert update branch never linked to their membership row. Filter the requested members against the freshly locked roster first so budgets and membership rows are only written for members who are actually new Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../management_endpoints/team_endpoints.py | 34 +++------ .../test_team_endpoints.py | 71 +++++++++++++++++++ 2 files changed, 79 insertions(+), 26 deletions(-) diff --git a/litellm/proxy/management_endpoints/team_endpoints.py b/litellm/proxy/management_endpoints/team_endpoints.py index 4957b3bd925..28c12173ea7 100644 --- a/litellm/proxy/management_endpoints/team_endpoints.py +++ b/litellm/proxy/management_endpoints/team_endpoints.py @@ -2904,10 +2904,15 @@ async def _process_team_members( if member_allowed_models is None and team_default_member_models: member_allowed_models = team_default_member_models - if isinstance(data.member, Member): + requested_members: Final[Sequence[Member]] = ( + (data.member,) if isinstance(data.member, Member) else tuple(data.member) + ) + for m in requested_members: + if _member_already_in_team(m, complete_team_data): + continue try: updated_user, updated_tm = await add_new_member( - new_member=data.member, + new_member=m, max_budget_in_team=data.max_budget_in_team, prisma_client=prisma_client, user_api_key_dict=user_api_key_dict, @@ -2921,34 +2926,11 @@ async def _process_team_members( except Exception as e: raise HTTPException( status_code=500, - detail={"error": f"Unable to add user - {data.member}, to team - {data.team_id}, for reason - {e}"}, + detail={"error": f"Unable to add user - {m}, to team - {data.team_id}, for reason - {e}"}, ) updated_users.append(updated_user) if updated_tm is not None: updated_team_memberships.append(updated_tm) - elif isinstance(data.member, list): - for m in data.member: - try: - updated_user, updated_tm = await add_new_member( - new_member=m, - max_budget_in_team=data.max_budget_in_team, - prisma_client=prisma_client, - user_api_key_dict=user_api_key_dict, - litellm_proxy_admin_name=litellm_proxy_admin_name, - team_id=data.team_id, - default_team_budget_id=default_team_budget_id, - allowed_models=member_allowed_models, - budget_duration=data.budget_duration, - tx=tx, - ) - except Exception as e: - raise HTTPException( - status_code=500, - detail={"error": f"Unable to add user - {m}, to team - {data.team_id}, for reason - {e}"}, - ) - updated_users.append(updated_user) - if updated_tm is not None: - updated_team_memberships.append(updated_tm) return updated_users, updated_team_memberships diff --git a/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py index f7157da3100..690b5ae80b6 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py @@ -1794,6 +1794,7 @@ async def test_process_team_members_single_member(): mock_team = MagicMock(spec=LiteLLM_TeamTable) mock_team.metadata = {"team_member_budget_id": "budget-123"} mock_team.default_team_member_models = None + mock_team.members_with_roles = [] # Mock user and membership objects mock_user = MagicMock(spec=LiteLLM_UserTable) @@ -1854,6 +1855,7 @@ async def test_process_team_members_multiple_members(): mock_team = MagicMock(spec=LiteLLM_TeamTable) mock_team.metadata = None mock_team.default_team_member_models = None + mock_team.members_with_roles = [] # Create multiple members as dictionaries (they will be converted to Member objects) members = [ @@ -2114,6 +2116,75 @@ async def test_add_team_members_runs_member_writes_on_the_lock_holding_transacti assert [tm.budget_id for tm in updated_team_memberships] == ["budget-pool"] +@pytest.mark.asyncio +async def test_add_team_members_skips_budget_and_membership_writes_for_members_already_on_the_roster(): + """ + Regression pin for orphaned budgets on a mixed /team/member_add list. + + A list naming one member already on the team and one new member must only create a + budget and membership row for the new member. Running add_new_member for the existing + member would create a per-member budget that nothing links to, since their membership + row (and the budget it already carries) is left untouched. + """ + from litellm.proxy.management_endpoints.team_endpoints import ( + _add_team_members_to_team, + ) + + added_user = MagicMock() + added_user.user_id = "bob" + added_user.model_dump.return_value = {"user_id": "bob", "teams": ["team-mixed"]} + created_budget = MagicMock() + created_budget.budget_id = "budget-bob" + membership = MagicMock() + membership.model_dump.return_value = { + "team_id": "team-mixed", + "user_id": "bob", + "budget_id": "budget-bob", + "litellm_budget_table": None, + } + + tx = MagicMock() + tx.query_raw = AsyncMock( + return_value=[{"members_with_roles": [{"user_id": "alice", "user_email": None, "role": "user"}]}] + ) + tx.litellm_teamtable.update = AsyncMock( + return_value=LiteLLM_TeamTable(team_id="team-mixed", members_with_roles=[]) + ) + tx.litellm_usertable.upsert = AsyncMock(return_value=added_user) + tx.litellm_usertable.update_many = AsyncMock() + tx.litellm_budgettable.create = AsyncMock(return_value=created_budget) + tx.litellm_teammembership.upsert = AsyncMock(return_value=membership) + + tx_cm = MagicMock() + tx_cm.__aenter__ = AsyncMock(return_value=tx) + tx_cm.__aexit__ = AsyncMock(return_value=None) + + prisma_client = MagicMock() + prisma_client.tx = MagicMock(return_value=tx_cm) + + _, updated_users, updated_team_memberships = await _add_team_members_to_team( + data=TeamMemberAddRequest( + team_id="team-mixed", + member=[Member(user_id="alice", role="user"), Member(user_id="bob", role="user")], + max_budget_in_team=50.0, + ), + complete_team_data=LiteLLM_TeamTable(team_id="team-mixed", members_with_roles=[]), + prisma_client=cast(object, prisma_client), + user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN), + litellm_proxy_admin_name="admin", + ) + + tx.litellm_budgettable.create.assert_awaited_once() + tx.litellm_teammembership.upsert.assert_awaited_once() + assert tx.litellm_teammembership.upsert.call_args.kwargs["where"] == { + "user_id_team_id": {"user_id": "bob", "team_id": "team-mixed"} + } + assert [user.user_id for user in updated_users] == ["bob"] + assert [tm.user_id for tm in updated_team_memberships] == ["bob"] + written_ids = [m["user_id"] for m in json.loads(tx.litellm_teamtable.update.call_args.kwargs["data"]["members_with_roles"])] + assert written_ids == ["alice", "bob"] + + @pytest.mark.asyncio async def test_add_team_members_writes_nothing_when_the_team_is_deleted_mid_request(): """ From c21e86a4436abdd48e0ad83cd3284adc35a50ea7 Mon Sep 17 00:00:00 2001 From: ryan-crabbe-berri Date: Fri, 18 Sep 2026 12:27:39 -0700 Subject: [PATCH 139/141] feat(ui): reset a team member's spend from the Members tab Every member now carries a membership row, so a member who spent with no budget is over budget the moment a member budget is added later. The only fix was POST /team/{team_id}/member/{user_id}/reset_spend, which had no UI. The Members tab gets a Reset spend action on rows that have current cycle spend. It confirms in a dialog, posts reset_to 0 through the typed client, and refreshes the team without remounting the page so the tab stays open. A team admin does not see it on their own row because the backend rejects that reset --- .../hooks/teams/useResetTeamMemberSpend.ts | 17 +++ .../TableIconActionButton.tsx | 1 + .../common_components/MemberTable.tsx | 18 +++ .../src/components/team/TeamInfo.tsx | 10 ++ .../components/team/TeamMemberTab.test.tsx | 114 +++++++++++++++++- .../src/components/team/TeamMemberTab.tsx | 94 ++++++++++++--- 6 files changed, 234 insertions(+), 20 deletions(-) create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/hooks/teams/useResetTeamMemberSpend.ts diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/teams/useResetTeamMemberSpend.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/teams/useResetTeamMemberSpend.ts new file mode 100644 index 00000000000..1fdf8fbd98d --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/teams/useResetTeamMemberSpend.ts @@ -0,0 +1,17 @@ +import { useMutation } from "@tanstack/react-query"; +import { fetchClient } from "@/lib/http/api"; + +export interface ResetTeamMemberSpendParams { + teamId: string; + userId: string; +} + +export const resetTeamMemberSpend = async ({ teamId, userId }: ResetTeamMemberSpendParams): Promise => { + await fetchClient.POST("/team/{team_id}/member/{user_id}/reset_spend", { + params: { path: { team_id: teamId, user_id: userId } }, + body: { reset_to: 0 }, + }); +}; + +export const useResetTeamMemberSpend = () => + useMutation({ mutationFn: resetTeamMemberSpend }); diff --git a/ui/litellm-dashboard/src/components/common_components/IconActionButton/TableIconActionButtons/TableIconActionButton.tsx b/ui/litellm-dashboard/src/components/common_components/IconActionButton/TableIconActionButtons/TableIconActionButton.tsx index 396355e1c0b..9da9c2dc702 100644 --- a/ui/litellm-dashboard/src/components/common_components/IconActionButton/TableIconActionButtons/TableIconActionButton.tsx +++ b/ui/litellm-dashboard/src/components/common_components/IconActionButton/TableIconActionButtons/TableIconActionButton.tsx @@ -30,6 +30,7 @@ export const TableIconActionButtonMap: Record boolean; + onResetSpend?: (member: Member) => void; + showResetSpendForMember?: (member: Member) => boolean; emptyText?: string; } @@ -73,6 +75,8 @@ interface MemberColumnDeps { roleTooltip?: string; extraColumns: MemberTableColumn[]; showDeleteForMember?: (member: Member) => boolean; + onResetSpend?: (member: Member) => void; + showResetSpendForMember?: (member: Member) => boolean; } const extraColumnDef = (column: MemberTableColumn): ColumnDef => { @@ -105,6 +109,8 @@ const buildColumns = ({ roleTooltip, extraColumns, showDeleteForMember, + onResetSpend, + showResetSpendForMember, }: MemberColumnDeps): ColumnDef[] => [ { id: "user_alias", @@ -173,6 +179,14 @@ const buildColumns = ({ dataTestId="edit-member" onClick={() => onEdit(row.original)} /> + {onResetSpend && (showResetSpendForMember?.(row.original) ?? true) && ( + onResetSpend(row.original)} + /> + )} {(!showDeleteForMember || showDeleteForMember(row.original)) && ( = ({ } }; + const refreshTeamData = async () => { + if (!accessToken) return; + try { + setTeamData(await teamInfoCall(accessToken, teamId)); + } catch { + toast.fromError("Failed to load team information"); + } + }; + useEffect(() => { fetchTeamInfo(); }, [teamId, accessToken]); @@ -1351,6 +1360,7 @@ const TeamInfoView: React.FC = ({ teamData={teamData} canEditTeam={canEditTeam} handleMemberDelete={handleMemberDelete} + onMemberSpendReset={refreshTeamData} setSelectedEditMember={setSelectedEditMember} setIsEditMemberModalVisible={setIsEditMemberModalVisible} setIsAddMemberModalVisible={setIsAddMemberModalVisible} diff --git a/ui/litellm-dashboard/src/components/team/TeamMemberTab.test.tsx b/ui/litellm-dashboard/src/components/team/TeamMemberTab.test.tsx index 760074d5dc9..52cba1e6330 100644 --- a/ui/litellm-dashboard/src/components/team/TeamMemberTab.test.tsx +++ b/ui/litellm-dashboard/src/components/team/TeamMemberTab.test.tsx @@ -1,4 +1,4 @@ -import { fireEvent, screen, within } from "@testing-library/react"; +import { fireEvent, screen, waitFor, within } from "@testing-library/react"; import userEvent from "@testing-library/user-event"; import { beforeEach, describe, expect, it, vi } from "vitest"; import { renderWithProviders } from "../../../tests/test-utils"; @@ -13,6 +13,9 @@ vi.mock("@/app/(dashboard)/hooks/useAuthorized", () => ({ default: vi.fn(), })); +const { POST } = vi.hoisted(() => ({ POST: vi.fn() })); +vi.mock("@/lib/http/api", () => ({ fetchClient: { POST } })); + vi.mock("@/utils/roles", () => ({ isUserTeamAdminForSingleTeam: vi.fn(() => false), isProxyAdminRole: vi.fn(() => false), @@ -26,6 +29,7 @@ const mockHandleMemberDelete = vi.fn(); const mockSetSelectedEditMember = vi.fn(); const mockSetIsEditMemberModalVisible = vi.fn(); const mockSetIsAddMemberModalVisible = vi.fn(); +const mockOnMemberSpendReset = vi.fn(); const budgetResetIso = new Date(2026, 6, 15, 12, 0, 0).toISOString(); @@ -121,6 +125,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -136,6 +141,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -154,6 +160,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -172,6 +179,7 @@ describe("TeamMembersComponent", () => { const props = { canEditTeam: false, handleMemberDelete: mockHandleMemberDelete, + onMemberSpendReset: mockOnMemberSpendReset, setSelectedEditMember: mockSetSelectedEditMember, setIsEditMemberModalVisible: mockSetIsEditMemberModalVisible, setIsAddMemberModalVisible: mockSetIsAddMemberModalVisible, @@ -195,6 +203,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={true} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -221,6 +230,7 @@ describe("TeamMembersComponent", () => { })} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -247,6 +257,7 @@ describe("TeamMembersComponent", () => { })} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -262,6 +273,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -280,6 +292,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -295,6 +308,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -311,6 +325,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -330,6 +345,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={true} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -364,6 +380,7 @@ describe("TeamMembersComponent", () => { teamData={teamData} canEditTeam={true} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -417,6 +434,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={true} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -447,6 +465,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={true} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -466,6 +485,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={true} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -482,6 +502,7 @@ describe("TeamMembersComponent", () => { teamData={createMockTeamData()} canEditTeam={false} handleMemberDelete={mockHandleMemberDelete} + onMemberSpendReset={mockOnMemberSpendReset} setSelectedEditMember={mockSetSelectedEditMember} setIsEditMemberModalVisible={mockSetIsEditMemberModalVisible} setIsAddMemberModalVisible={mockSetIsAddMemberModalVisible} @@ -491,4 +512,95 @@ describe("TeamMembersComponent", () => { expect(screen.queryByTestId("edit-member")).not.toBeInTheDocument(); expect(screen.queryByTestId("delete-member")).not.toBeInTheDocument(); }); + + describe("reset spend", () => { + const renderEditableTab = () => + renderWithProviders( + , + ); + + it("resets the member's current cycle spend to $0 after confirming, then refreshes the team", async () => { + const user = userEvent.setup(); + POST.mockResolvedValue({ data: {} }); + renderEditableTab(); + + const memberRow = screen.getByRole("row", { name: /user1@test\.com/ }); + await user.click(within(memberRow).getByTestId("reset-member-spend")); + + const dialog = await screen.findByRole("dialog", { name: "Reset Team Member Spend" }); + expect(dialog).toHaveTextContent("user1@test.com"); + expect(dialog).toHaveTextContent("$100.5000"); + expect(POST).not.toHaveBeenCalled(); + + await user.click(within(dialog).getByRole("button", { name: "Reset" })); + + await waitFor(() => expect(mockOnMemberSpendReset).toHaveBeenCalledTimes(1)); + expect(POST).toHaveBeenCalledExactlyOnceWith("/team/{team_id}/member/{user_id}/reset_spend", { + params: { path: { team_id: "team-123", user_id: "user1@test.com" } }, + body: { reset_to: 0 }, + }); + expect(screen.queryByRole("dialog")).not.toBeInTheDocument(); + }); + + it("keeps the dialog open and does not refresh the team when the reset fails", async () => { + const user = userEvent.setup(); + POST.mockRejectedValue(new Error("Cannot reset your own spend. Ask a proxy admin.")); + renderEditableTab(); + + await user.click(screen.getByTestId("reset-member-spend")); + const dialog = await screen.findByRole("dialog", { name: "Reset Team Member Spend" }); + await user.click(within(dialog).getByRole("button", { name: "Reset" })); + + await waitFor(() => expect(POST).toHaveBeenCalledTimes(1)); + expect(mockOnMemberSpendReset).not.toHaveBeenCalled(); + expect(screen.getByRole("dialog", { name: "Reset Team Member Spend" })).toBeInTheDocument(); + }); + + it("does not call the API when the dialog is cancelled", async () => { + const user = userEvent.setup(); + renderEditableTab(); + + await user.click(screen.getByTestId("reset-member-spend")); + const dialog = await screen.findByRole("dialog", { name: "Reset Team Member Spend" }); + await user.click(within(dialog).getByRole("button", { name: "Cancel" })); + + await waitFor(() => expect(screen.queryByRole("dialog")).not.toBeInTheDocument()); + expect(POST).not.toHaveBeenCalled(); + }); + + it("only offers the reset on members that have current cycle spend", () => { + renderEditableTab(); + + expect( + within(screen.getByRole("row", { name: /user1@test\.com/ })).getByTestId("reset-member-spend"), + ).toBeVisible(); + expect( + within(screen.getByRole("row", { name: /user2@test\.com/ })).queryByTestId("reset-member-spend"), + ).not.toBeInTheDocument(); + }); + + it("hides the reset on the caller's own row for a team admin, since the backend rejects it", () => { + vi.mocked(useAuthorized).mockReturnValue({ userId: "user1@test.com", userRole: "Internal User" } as never); + vi.mocked(isProxyAdminRole).mockReturnValue(false); + renderEditableTab(); + + expect(screen.queryByTestId("reset-member-spend")).not.toBeInTheDocument(); + }); + + it("shows the reset on the caller's own row for a proxy admin", () => { + vi.mocked(useAuthorized).mockReturnValue({ userId: "user1@test.com", userRole: "Admin" } as never); + vi.mocked(isProxyAdminRole).mockReturnValue(true); + renderEditableTab(); + + expect(screen.getByTestId("reset-member-spend")).toBeVisible(); + }); + }); }); diff --git a/ui/litellm-dashboard/src/components/team/TeamMemberTab.tsx b/ui/litellm-dashboard/src/components/team/TeamMemberTab.tsx index 16f3d12d71c..a869c1ad624 100644 --- a/ui/litellm-dashboard/src/components/team/TeamMemberTab.tsx +++ b/ui/litellm-dashboard/src/components/team/TeamMemberTab.tsx @@ -1,13 +1,18 @@ +import { useResetTeamMemberSpend } from "@/app/(dashboard)/hooks/teams/useResetTeamMemberSpend"; import { useUISettings } from "@/app/(dashboard)/hooks/uiSettings/useUISettings"; import useAuthorized from "@/app/(dashboard)/hooks/useAuthorized"; +import { Button } from "@/components/ui/button"; +import { Dialog, DialogContent, DialogFooter, DialogHeader, DialogTitle } from "@/components/ui/dialog"; import { SimpleTooltip } from "@/components/ui/tooltip"; import MemberTable from "@/components/common_components/MemberTable"; import { Member } from "@/components/networking"; +import { parseErrorMessage } from "@/components/shared/errorUtils"; import { DateCell, MoneyCell } from "@/components/shared/table_cells"; +import { toast } from "@/lib/toast"; import { formatNumberWithCommas } from "@/utils/dataUtils"; import { isProxyAdminRole, isUserTeamAdminForSingleTeam } from "@/utils/roles"; import { CircleHelp } from "lucide-react"; -import type { ComponentProps } from "react"; +import { useState, type ComponentProps } from "react"; import { TeamData, TeamMembership } from "./TeamInfo"; export const seedMemberBudgetFields = ( @@ -31,6 +36,7 @@ interface TeamMemberTabProps { setSelectedEditMember: (member: Member) => void; setIsEditMemberModalVisible: (visible: boolean) => void; setIsAddMemberModalVisible: (visible: boolean) => void; + onMemberSpendReset: () => void; } export default function TeamMemberTab({ @@ -40,7 +46,11 @@ export default function TeamMemberTab({ setSelectedEditMember, setIsEditMemberModalVisible, setIsAddMemberModalVisible, + onMemberSpendReset, }: TeamMemberTabProps) { + const [memberToResetSpend, setMemberToResetSpend] = useState(null); + const { mutate: resetMemberSpend, isPending: isResettingSpend } = useResetTeamMemberSpend(); + const formatNumber = (value: number | null): string => { if (value === null || value === undefined) return "0"; @@ -199,24 +209,70 @@ export default function TeamMemberTab({ }, ]; + const handleResetSpend = () => { + if (!memberToResetSpend?.user_id) return; + resetMemberSpend( + { teamId: teamData.team_id, userId: memberToResetSpend.user_id }, + { + onSuccess: () => { + toast.success("Team member spend reset to $0"); + setMemberToResetSpend(null); + onMemberSpendReset(); + }, + onError: (error) => toast.fromError(parseErrorMessage(error)), + }, + ); + }; + return ( - { - const membership = teamData.team_memberships.find((tm) => tm.user_id === record.user_id); - setSelectedEditMember(seedMemberBudgetFields(record, membership?.litellm_budget_table)); - setIsEditMemberModalVisible(true); - }} - onDelete={handleMemberDelete} - onAddMember={() => setIsAddMemberModalVisible(true)} - roleColumnTitle="Team Role" - roleTooltip="This role applies only to this team and is independent from the user's proxy-level role." - extraColumns={extraColumns} - showDeleteForMember={() => - isProxyAdmin || (canEditTeam && !isUserTeamAdmin) || (isUserTeamAdmin && !disableTeamAdminDeleteTeamUser) - } - /> + <> + { + const membership = teamData.team_memberships.find((tm) => tm.user_id === record.user_id); + setSelectedEditMember(seedMemberBudgetFields(record, membership?.litellm_budget_table)); + setIsEditMemberModalVisible(true); + }} + onDelete={handleMemberDelete} + onAddMember={() => setIsAddMemberModalVisible(true)} + roleColumnTitle="Team Role" + roleTooltip="This role applies only to this team and is independent from the user's proxy-level role." + extraColumns={extraColumns} + showDeleteForMember={() => + isProxyAdmin || (canEditTeam && !isUserTeamAdmin) || (isUserTeamAdmin && !disableTeamAdminDeleteTeamUser) + } + onResetSpend={setMemberToResetSpend} + showResetSpendForMember={(record) => + getUserCurrentCycleSpend(record.user_id) > 0 && (isProxyAdmin || record.user_id !== userId) + } + /> + !open && setMemberToResetSpend(null)}> + + + Reset Team Member Spend + +

+ Reset current cycle spend for{" "} + {memberToResetSpend?.user_email || memberToResetSpend?.user_id} in this team to{" "} + $0? +

+

+ Current cycle spend:{" "} + ${formatNumberWithCommas(getUserCurrentCycleSpend(memberToResetSpend?.user_id ?? null), 4)} + . This is the value checked against the member's budget. Total spend and logs are preserved. +

+ + + + + + + ); } From 836bf7d8974bc3fc3cdd71a3fe6d9565201c2f2e Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 13:59:36 -0700 Subject: [PATCH 140/141] refactor(rust): make the exception mapper a pure rule table Rebuild exception_type around text rules per provider family and one shared status table. The mapper takes the context, an injected redactor and the original failure, and returns a PublicError with the message, the real upstream response and the debug text. Every divergence from the Python mapper and every known gap is listed in the module header Match Python on a standalone 429 with an unknown status and on Cohere's rules for failures without a status. Drop python_repr and the unread public_failures fixtures --- litellm-rust/Cargo.lock | 1 - litellm-rust/crates/core-utils/Cargo.toml | 1 - .../src/exception_mapping_utils/cohere.rs | 275 ++----- .../src/exception_mapping_utils/mod.rs | 757 +++++++----------- .../src/exception_mapping_utils/openai.rs | 600 +++----------- .../src/exception_mapping_utils/original.rs | 145 +++- .../src/exception_mapping_utils/public.rs | 257 +----- .../src/exception_mapping_utils/rules.rs | 283 +------ .../src/exception_mapping_utils/status.rs | 182 +---- .../src/exception_mapping_utils/vertex_ai.rs | 573 ++----------- litellm-rust/crates/core-utils/src/lib.rs | 1 - .../crates/core-utils/src/python_repr.rs | 93 --- .../crates/core-utils/src/secret_redaction.rs | 50 +- .../fixtures/public_failures/api.json | 13 - .../public_failures/api_connection.json | 11 - .../status_authentication.json | 13 - .../public_failures/status_bad_gateway.json | 28 - .../public_failures/status_bad_request.json | 28 - .../status_content_policy_violation.json | 28 - .../status_context_window_exceeded.json | 28 - .../status_internal_server.json | 19 - .../public_failures/status_not_found.json | 28 - .../status_permission_denied.json | 19 - .../public_failures/status_rate_limit.json | 28 - .../status_service_unavailable.json | 28 - .../status_unsupported_params.json | 28 - .../public_failures/timeout_with_status.json | 12 - .../timeout_without_status.json | 12 - 28 files changed, 807 insertions(+), 2734 deletions(-) delete mode 100644 litellm-rust/crates/core-utils/src/python_repr.rs delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/api.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json delete mode 100644 tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json diff --git a/litellm-rust/Cargo.lock b/litellm-rust/Cargo.lock index 6c524124550..c359ca19986 100644 --- a/litellm-rust/Cargo.lock +++ b/litellm-rust/Cargo.lock @@ -2092,7 +2092,6 @@ dependencies = [ "serde_json", "serde_path_to_error", "serde_with", - "strum", "thiserror 2.0.19", "url", ] diff --git a/litellm-rust/crates/core-utils/Cargo.toml b/litellm-rust/crates/core-utils/Cargo.toml index 59e0ee1a09d..eb353bc060c 100644 --- a/litellm-rust/crates/core-utils/Cargo.toml +++ b/litellm-rust/crates/core-utils/Cargo.toml @@ -12,7 +12,6 @@ serde.workspace = true serde_json.workspace = true serde_path_to_error = "0.1" serde_with.workspace = true -strum.workspace = true thiserror.workspace = true url.workspace = true diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs index 381ebd7ad27..c2a391ee223 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/cohere.rs @@ -1,232 +1,115 @@ -use super::Mapping; -use super::public::{PublicFailure, StatusClass}; -use super::rules::{Kind, ResponseChoice, Rule, apply, contains_any}; +use super::public::PublicError; +use super::rules::{Rule, contains_any}; -const fn with_response(class: StatusClass) -> Kind { - Kind::Status { - class, - response: ResponseChoice::Provider, - } -} - -fn original(mapping: &Mapping<'_>) -> String { - format!("CohereException - {}", mapping.original.message) -} - -fn status_is(mapping: &Mapping<'_>, statuses: &[u16]) -> bool { - mapping - .original - .status - .is_some_and(|status| statuses.contains(&status)) -} - -/// `_map_cohere_exception`, in its branch order. A failure no rule claims falls through to -/// the status table. -const RULES: &[Rule] = &[ - Rule { - when: |mapping| { +/// The text branches of `_map_cohere_exception`, in its order. +pub(super) const RULES: &[Rule] = &[ + Rule::new( + |mapping| { contains_any( &mapping.error_str, &["invalid api token", "No API key provided."], ) }, - kind: with_response(StatusClass::Authentication), - message: original, - debug: false, - }, - Rule { - when: |mapping| mapping.error_str.contains("invalid type: parameter"), - kind: with_response(StatusClass::BadRequest), - message: original, - debug: false, - }, - Rule { - when: |mapping| mapping.error_str.contains("too many tokens"), - kind: with_response(StatusClass::ContextWindowExceeded), - message: original, - debug: false, - }, - Rule { - when: |mapping| { + PublicError::Authentication, + ), + Rule::new( + |mapping| mapping.error_str.contains("invalid type: parameter"), + PublicError::BadRequest, + ), + Rule::new( + |mapping| mapping.error_str.contains("too many tokens"), + PublicError::ContextWindowExceeded, + ), + Rule::new( + |mapping| { mapping .error_str .to_lowercase() .contains("internal server error") }, - kind: with_response(StatusClass::InternalServer), - message: |mapping| format!("CohereException - {}", mapping.error_str), - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, &[400, 498]), - kind: with_response(StatusClass::BadRequest), - message: original, - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, &[408]), - kind: Kind::Timeout(None), - message: original, - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, &[500]), - kind: with_response(StatusClass::InternalServer), - message: original, - debug: false, - }, + PublicError::InternalServer, + ), + Rule::new( + |mapping| mapping.status.is_none() && mapping.error_str.contains("invalid type:"), + PublicError::BadRequest, + ), + Rule::new( + |mapping| mapping.status.is_none() && mapping.error_str.contains("Unexpected server error"), + PublicError::InternalServer, + ), ]; -pub(super) fn map(mapping: &Mapping<'_>) -> Option { - apply(RULES, mapping).map(|failure| PublicFailure { - llm_provider: Some("cohere".to_string()), - ..failure - }) -} - #[cfg(test)] mod tests { - use super::super::testing::{context, failure, http, status, upstream}; - use super::super::{ExceptionFamily, OriginalException, PublicKind}; + use super::super::rules::first_match; + use super::super::testing::mapping; use super::*; - fn mapped(provider: &str, original: &OriginalException) -> Option { - let context = context(provider, ExceptionFamily::Cohere); - map(&Mapping::new(&context, original)) + fn classified(text: &str) -> Option { + classified_with(Some(400), text) } - fn cohere(class: StatusClass, status_code: u16, body: &str, message: &str) -> PublicFailure { - failure( - status(class, upstream(status_code, body)), - message, - "cohere", - ) + fn classified_with(status: Option, text: &str) -> Option { + first_match(RULES, &mapping(status, text)).map(|rule| rule.error) } #[rstest::rstest] - #[case::invalid_token( - 500, - "invalid api token", - cohere( - StatusClass::Authentication, - 500, - "invalid api token", - "CohereException - invalid api token" - ) - )] - #[case::no_api_key( - 500, - "No API key provided.", - cohere( - StatusClass::Authentication, - 500, - "No API key provided.", - "CohereException - No API key provided." - ) - )] - #[case::invalid_parameter( - 500, - "invalid type: parameter x", - cohere( - StatusClass::BadRequest, - 500, - "invalid type: parameter x", - "CohereException - invalid type: parameter x" - ) - )] - #[case::too_many_tokens( - 500, - "too many tokens", - cohere( - StatusClass::ContextWindowExceeded, - 500, - "too many tokens", - "CohereException - too many tokens" - ) - )] - #[case::internal_server_text( - 400, - "Internal Server Error", - cohere( - StatusClass::InternalServer, - 400, - "Internal Server Error", - "CohereException - Internal Server Error" - ) - )] - #[case::bad_request( - 400, - "rejected", - cohere(StatusClass::BadRequest, 400, "rejected", "CohereException - rejected") - )] - #[case::invalid_token_status( - 498, - "rejected", - cohere(StatusClass::BadRequest, 498, "rejected", "CohereException - rejected") - )] - #[case::request_timeout(408, "rejected", failure(PublicKind::Timeout { status: None }, "CohereException - rejected", "cohere"))] - #[case::internal_server( - 500, - "rejected", - cohere( - StatusClass::InternalServer, - 500, - "rejected", - "CohereException - rejected" - ) - )] - fn each_rule_maps_and_reports_cohere( - #[case] status_code: u16, - #[case] body: &str, - #[case] expected: PublicFailure, - ) { - assert_eq!(mapped("azure_ai", &http(status_code, body)), Some(expected)); - } - - #[rstest::rstest] - #[case::unmapped_status(409)] - #[case::unauthorized(401)] - fn statuses_without_a_rule_fall_through(#[case] status_code: u16) { - assert_eq!(mapped("cohere", &http(status_code, "rejected")), None); - } - - #[test] - fn the_internal_server_rule_uses_the_redacted_text() { - let body = "internal server error Bearer abcdefghijklmnop"; - assert_eq!( - mapped("cohere", &http(400, body)), - Some(cohere( - StatusClass::InternalServer, - 400, - body, - "CohereException - internal server error REDACTED" - )) - ); + #[case::invalid_token("invalid api token", PublicError::Authentication)] + #[case::no_api_key("No API key provided.", PublicError::Authentication)] + #[case::invalid_parameter("invalid type: parameter x", PublicError::BadRequest)] + #[case::too_many_tokens("too many tokens", PublicError::ContextWindowExceeded)] + #[case::internal_server_text("Internal Server Error", PublicError::InternalServer)] + #[case::internal_server_any_case("INTERNAL server ERROR", PublicError::InternalServer)] + fn each_text_rule_claims_its_marker(#[case] text: &str, #[case] expected: PublicError) { + assert_eq!(classified(text), Some(expected)); } #[rstest::rstest] #[case::token_before_parameter( "invalid api token invalid type: parameter", - StatusClass::Authentication + PublicError::Authentication )] #[case::parameter_before_tokens( "invalid type: parameter too many tokens", - StatusClass::BadRequest + PublicError::BadRequest )] #[case::tokens_before_internal( "too many tokens Internal Server Error", - StatusClass::ContextWindowExceeded + PublicError::ContextWindowExceeded )] - #[case::internal_before_status("Internal Server Error", StatusClass::InternalServer)] - fn the_earlier_rule_wins_when_two_apply(#[case] body: &str, #[case] class: StatusClass) { - assert_eq!( - mapped("cohere", &http(400, body)), - Some(cohere( - class, - 400, - body, - &format!("CohereException - {body}") - )) - ); + fn the_earlier_rule_wins_when_two_apply(#[case] text: &str, #[case] expected: PublicError) { + assert_eq!(classified(text), Some(expected)); + } + + #[rstest::rstest] + #[case::invalid_type(None, "invalid type: x", Some(PublicError::BadRequest))] + #[case::unexpected_server_error( + None, + "Unexpected server error", + Some(PublicError::InternalServer) + )] + #[case::invalid_type_before_unexpected( + None, + "invalid type: x Unexpected server error", + Some(PublicError::BadRequest) + )] + #[case::internal_before_invalid_type( + None, + "internal server error invalid type: x", + Some(PublicError::InternalServer) + )] + #[case::invalid_type_with_a_status(Some(500), "invalid type: x", None)] + #[case::unexpected_with_a_status(Some(400), "Unexpected server error", None)] + fn the_trailing_rules_only_claim_failures_without_a_status( + #[case] status: Option, + #[case] text: &str, + #[case] expected: Option, + ) { + assert_eq!(classified_with(status, text), expected); + } + + #[test] + fn text_without_a_marker_is_left_to_the_status_table() { + assert_eq!(classified("rejected"), None); } } diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs index acc7820c770..162d325e4f4 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/mod.rs @@ -1,18 +1,47 @@ -//! A port of Python's `exception_type` for the routes that run in Rust. +//! A port of Python's `exception_type` for the routes that run in Rust. Rust decides the +//! public class, the message and the debug text; Python only builds the class. +//! +//! DIVERGENCES: where the Python mapper is inconsistent, the port follows one rule instead. +//! - The message is always `{Provider}Exception - {redacted text}`. Python's per-branch +//! labels (`RateLimitError: `, `litellm.RateLimitError: `, `Vertex_aiException BadRequestError`) +//! are dropped because every public class already prefixes `litellm.{Class}: `. +//! - The upstream response is always the real one. Python swaps in made-up `httpx.Response` +//! stubs on some Vertex branches, losing the body and `retry-after`. +//! - The debug text is always attached; Python passes it on some branches only. +//! - No family rule turns a status into a class; the shared status table owns that. So a +//! Vertex 502 is a `BadGatewayError` and an OpenAI-family 403 is a `PermissionDeniedError`. +//! Three rules read the status only to gate a text match, as Python does: the standalone +//! `429`, Vertex's wrapped 429 behind a 5xx, and Cohere's rules for failures with no status. +//! - A timeout text marker on an HTTP failure keeps the upstream response. Python's `Timeout` +//! carries none. +//! - Every family matches and reports the redacted text. Python's OpenAI mapper builds the +//! message from the unredacted text. +//! - A refused connection is an `APIConnectionError`, not the 500 Python's HTTP handler +//! synthesizes. +//! - Dropped Python rules: Vertex's bare `403` substring (it matches `4031 tokens`), Vertex's +//! `IndexError` quota marker (a Python client crash), the OpenAI SDK's missing-`api_key` +//! text and its `OPENAI` renaming, Cohere's `llm_provider="cohere"` override, and Cohere's +//! `CohereConnectionError` check (a Python SDK class name). //! //! KNOWN_GAPS: differences from the Python mapper that no Rust route can reach today. Each //! one stops being acceptable at its trigger. -//! - The Vertex partner-model API base for "claude" models is not built into -//! `extra_information`. Trigger: a Vertex route whose models include Anthropic partner -//! models; then `api_base` gets that branch and a table row. +//! - The Vertex partner-model API base for "claude" models is not built into the debug text. +//! Trigger: a Vertex route whose models include Anthropic partner models. +//! - The debug text's `API Base` line is only the non-streaming Vertex URL. Python prefers an +//! explicit or provider-resolved `api_base`, uses `:streamGenerateContent` when streaming, +//! and has Gemini and OpenAI defaults. Trigger: the first route wired to this mapper, since +//! every route knows its `api_base`. +//! - The debug text has no `Messages:` line, which Python adds when +//! `redact_messages_in_exceptions` is off. Trigger: a wired route that carries messages. //! - Python reports the provider `get_llm_provider` resolves for a stripped model name when //! that name happens to be in the model cost map. Trigger: a route whose model names //! overlap the cost map; that needs the provider resolution port, not a classifier change. -//! - The generic `APIConnectionError` fallback appends `traceback.format_exc()` to the -//! message. Rust has no Python traceback and does not invent one; a sweep row that reaches -//! it compares the message before the traceback. +//! - `litellm_proxy` errors are not unwrapped into the proxied exception. Trigger: a Rust +//! route that calls a LiteLLM proxy. +//! - Only the OpenAI-compatible, Vertex AI and Cohere mappers are ported; every other +//! provider goes straight to the status table. Trigger: a Rust route for such a provider. -use super::secret_redaction::{redact_string, secret_redaction_enabled}; +use super::secret_redaction::SecretRedactor; mod cohere; mod openai; @@ -22,10 +51,10 @@ mod rules; mod status; mod vertex_ai; -pub use original::{ExceptionFamily, LocalClass, OriginalException}; -pub use public::{HttpStub, PublicFailure, PublicKind, ResponseArg, StatusClass, UpstreamResponse}; +pub use original::{ExceptionFamily, OriginalException}; +pub use public::{MappedFailure, PublicError, UpstreamResponse}; -const DOCS_URL: &str = "https://docs.litellm.ai/docs"; +use rules::{Rule, contains_any, first_match}; const TIMEOUT_MARKERS: &[&str] = &[ "Request Timeout Error", @@ -37,11 +66,8 @@ const TIMEOUT_MARKERS: &[&str] = &[ #[derive(Clone, Debug, Default, PartialEq, Eq)] pub struct ExceptionContext { pub model: String, - pub custom_llm_provider: Option, - pub family: ExceptionFamily, + pub custom_llm_provider: String, pub asynchronous: bool, - pub suppress_debug_info: bool, - pub redact_messages_in_exceptions: bool, pub vertex_project: Option, pub vertex_location: Option, pub model_group: Option, @@ -50,73 +76,93 @@ pub struct ExceptionContext { pub user_api_key_team_alias: Option, } -/// The attributes `exception_type` reads off the Python exception: a provider error -/// (`BaseLLMException`) carries a status, a response and a request, a plain exception -/// carries only its text. -struct Raised { +/// What the rules read: the status of a provider response, if any, and the redacted text. +struct Mapping { status: Option, - status_is_synthesized: bool, - message: String, - response: Option, + error_str: String, } -impl Raised { - fn provider( - status: u16, - message: String, - body: String, - headers: Vec<(String, String)>, - ) -> Self { - Self { - status: Some(status), - status_is_synthesized: false, - message, - response: Some(UpstreamResponse { - status, - body, - headers, +pub fn exception_type( + context: &ExceptionContext, + redactor: Option<&SecretRedactor>, + original: &OriginalException, +) -> MappedFailure { + let (status, text, upstream) = match original { + OriginalException::Http { + status, + body, + headers, + } => ( + Some(*status), + body.clone(), + Some(UpstreamResponse { + status: *status, + body: body.clone(), + headers: headers.clone(), }), + ), + OriginalException::Connection { message } | OriginalException::Plain { message } => { + (None, message.clone(), None) } - } - - fn plain(message: String) -> Self { - Self { - status: None, - status_is_synthesized: false, - message, - response: None, - } - } - - fn new(original: &OriginalException, asynchronous: bool) -> Self { - match original { - OriginalException::Http { - status, - body, - headers, - } => Self::provider(*status, body.clone(), body.clone(), headers.clone()), - OriginalException::Connection { message } => Self { - status_is_synthesized: true, - ..Self::provider(500, message.clone(), String::new(), Vec::new()) - }, - OriginalException::Timeout { - timeout_seconds, - elapsed_seconds, - } => Self::provider( - 408, - timeout_message(asynchronous, *timeout_seconds, *elapsed_seconds), - String::new(), - Vec::new(), - ), - OriginalException::Response { message } - | OriginalException::Local { message, .. } - | OriginalException::Public { message, .. } => Self::plain(message.clone()), - } + OriginalException::Timeout { + timeout_seconds, + elapsed_seconds, + } => ( + None, + timeout_message(context.asynchronous, *timeout_seconds, *elapsed_seconds), + None, + ), + }; + let mapping = Mapping { + status, + error_str: match redactor { + Some(redactor) => redactor.redact(&text), + None => text, + }, + }; + let family = ExceptionFamily::for_provider(&context.custom_llm_provider); + let (error, hint) = classify(family, original, &mapping); + MappedFailure { + error, + message: format!( + "{} - {}{hint}", + exception_provider(&context.custom_llm_provider), + mapping.error_str + ), + upstream, + debug_info: extra_information(context, api_base(context).as_deref()), } } -/// The text `litellm.Timeout` carries when the Python HTTP handler times out: the sync -/// and async handlers word it differently. +fn classify( + family: ExceptionFamily, + original: &OriginalException, + mapping: &Mapping, +) -> (PublicError, &'static str) { + const TIMEOUT: PublicError = PublicError::Timeout { status: 408 }; + if matches!(original, OriginalException::Timeout { .. }) + || contains_any(&mapping.error_str, TIMEOUT_MARKERS) + { + return (TIMEOUT, ""); + } + if let Some(rule) = first_match(family_rules(family), mapping) { + return (rule.error, rule.hint); + } + let by_status = mapping.status.and_then(status::classify); + (by_status.unwrap_or(PublicError::ApiConnection), "") +} + +fn family_rules(family: ExceptionFamily) -> &'static [Rule] { + match family { + ExceptionFamily::OpenAiCompatible => openai::RULES, + ExceptionFamily::VertexAi => vertex_ai::RULES, + ExceptionFamily::Cohere => cohere::RULES, + ExceptionFamily::Other => &[], + } +} + +/// The text the Python HTTP handler's timeout carries: the sync and async handlers word it +/// differently. fn timeout_message( asynchronous: bool, timeout_seconds: Option, @@ -126,11 +172,9 @@ fn timeout_message( if asynchronous { let elapsed = python_float(elapsed_seconds.map(|seconds| (seconds * 1000.0).round() / 1000.0)); - format!( - "litellm.Timeout: Connection timed out. Timeout passed={timeout}, time taken={elapsed} seconds" - ) + format!("Connection timed out. Timeout passed={timeout}, time taken={elapsed} seconds") } else { - format!("litellm.Timeout: Connection timed out after {timeout} seconds.") + format!("Connection timed out after {timeout} seconds.") } } @@ -142,113 +186,10 @@ fn python_float(value: Option) -> String { } } -/// Everything the rules read: the original as Python sees it and the text `exception_type` -/// derives from the context before any provider mapper runs. -struct Mapping<'a> { - context: &'a ExceptionContext, - original: Raised, - provider: &'a str, - error_str: String, - exception_provider: String, - extra_information: String, -} - -impl<'a> Mapping<'a> { - fn new(context: &'a ExceptionContext, original: &OriginalException) -> Self { - let original = Raised::new(original, context.asynchronous); - let error_str = if secret_redaction_enabled() { - redact_string(&original.message) - } else { - original.message.clone() - }; - Self { - context, - original, - provider: context.custom_llm_provider.as_deref().unwrap_or_default(), - error_str, - exception_provider: match &context.custom_llm_provider { - None => "None".to_string(), - Some(provider) => exception_provider(provider), - }, - extra_information: extra_information(context, api_base(context).as_deref()), - } - } - - fn failure(&self, kind: PublicKind, message: String, debug: bool) -> PublicFailure { - PublicFailure { - kind, - message, - model: self.context.model.clone(), - llm_provider: self.context.custom_llm_provider.clone(), - litellm_debug_info: debug.then(|| self.extra_information.clone()), - litellm_response_headers: None, - print_banner: false, - } - } -} - -pub fn exception_type(context: &ExceptionContext, original: &OriginalException) -> PublicFailure { - if let OriginalException::Public { class, message } = original { - return PublicFailure { - kind: PublicKind::Status { - status_class: *class, - response: None, - }, - message: message.clone(), - model: context.model.clone(), - llm_provider: context.custom_llm_provider.clone(), - litellm_debug_info: None, - litellm_response_headers: None, - print_banner: false, - }; - } - let mapping = Mapping::new(context, original); - let litellm_response_headers = mapping - .original - .response - .as_ref() - .map(|response| response.headers.clone()) - .filter(|headers| !headers.is_empty()); - PublicFailure { - litellm_response_headers, - print_banner: !context.suppress_debug_info, - ..map(&mapping) - } -} - -fn map(mapping: &Mapping<'_>) -> PublicFailure { - if rules::contains_any(&mapping.error_str, TIMEOUT_MARKERS) { - return mapping.failure( - PublicKind::Timeout { status: None }, - format!( - "APITimeoutError - Request timed out. Error_str: {}", - mapping.error_str - ), - true, - ); - } - let provider_failure = match mapping.context.family { - ExceptionFamily::OpenAiCompatible => openai::map(mapping), - ExceptionFamily::VertexAi => vertex_ai::map(mapping), - ExceptionFamily::Cohere => cohere::map(mapping), - ExceptionFamily::Other => None, - }; - provider_failure - .or_else(|| status::map(mapping)) - .unwrap_or_else(|| unmapped(mapping)) -} - -/// The `APIConnectionError` Python raises when no mapper claimed the failure: with the -/// provider prefix for a provider error, with the bare text for a plain exception. -fn unmapped(mapping: &Mapping<'_>) -> PublicFailure { - let message = match mapping.original.status { - Some(_) => format!("{} - {}", mapping.exception_provider, mapping.error_str), - None => mapping.original.message.clone(), - }; - mapping.failure(PublicKind::ApiConnection, message, false) -} - fn exception_provider(provider: &str) -> String { + if provider == "openai" { + return "OpenAIException".to_string(); + } let mut characters = provider.chars(); match characters.next() { Some(first) => format!("{}{}Exception", first.to_uppercase(), characters.as_str()), @@ -256,18 +197,6 @@ fn exception_provider(provider: &str) -> String { } } -fn python_capitalize(value: &str) -> String { - let mut characters = value.chars(); - match characters.next() { - Some(first) => format!( - "{}{}", - first.to_uppercase(), - characters.as_str().to_lowercase() - ), - None => String::new(), - } -} - fn api_base(context: &ExceptionContext) -> Option { match (&context.vertex_location, &context.vertex_project) { (Some(location), Some(project)) => Some(format!( @@ -311,132 +240,99 @@ fn extra_information(context: &ExceptionContext, api_base: Option<&str>) -> Stri #[cfg(test)] mod testing { - use super::*; + use super::Mapping; - pub(super) const DEBUG: &str = "\nModel: ocr-model"; - - pub(super) fn context(provider: &str, family: ExceptionFamily) -> ExceptionContext { - ExceptionContext { - model: "ocr-model".into(), - custom_llm_provider: Some(provider.into()), - family, - suppress_debug_info: true, - ..ExceptionContext::default() - } - } - - pub(super) fn http(status: u16, body: &str) -> OriginalException { - OriginalException::Http { + pub(super) fn mapping(status: Option, text: &str) -> Mapping { + Mapping { status, - body: body.into(), - headers: vec![("retry-after".into(), "7".into())], - } - } - - pub(super) fn upstream(status: u16, body: &str) -> Option { - Some(ResponseArg::Upstream(UpstreamResponse { - status, - body: body.into(), - headers: vec![("retry-after".into(), "7".into())], - })) - } - - pub(super) fn status(class: StatusClass, response: Option) -> PublicKind { - PublicKind::Status { - status_class: class, - response, - } - } - - /// The failure a rule builds before `exception_type` adds the response headers and the - /// banner flag. - pub(super) fn failure(kind: PublicKind, message: &str, provider: &str) -> PublicFailure { - PublicFailure { - kind, - message: message.into(), - model: "ocr-model".into(), - llm_provider: Some(provider.into()), - litellm_debug_info: None, - litellm_response_headers: None, - print_banner: false, - } - } - - pub(super) fn with_debug(failure: PublicFailure) -> PublicFailure { - PublicFailure { - litellm_debug_info: Some(DEBUG.into()), - ..failure + error_str: text.into(), } } } #[cfg(test)] mod tests { - use super::testing::{DEBUG, context, failure, http, status, upstream, with_debug}; use super::*; - fn openai() -> ExceptionContext { - context("mistral", ExceptionFamily::OpenAiCompatible) + const DEBUG: &str = "\nModel: ocr-model"; + + fn context(provider: &str) -> ExceptionContext { + ExceptionContext { + model: "ocr-model".into(), + custom_llm_provider: provider.into(), + ..ExceptionContext::default() + } + } + + fn redactor() -> SecretRedactor { + SecretRedactor::new(16) + } + + fn headers() -> Vec<(String, String)> { + vec![("retry-after".into(), "7".into())] + } + + fn http(status: u16, body: &str) -> OriginalException { + OriginalException::Http { + status, + body: body.into(), + headers: headers(), + } + } + + fn upstream(status: u16, body: &str) -> Option { + Some(UpstreamResponse { + status, + body: body.into(), + headers: headers(), + }) + } + + fn mapped(provider: &str, original: &OriginalException) -> MappedFailure { + exception_type(&context(provider), Some(&redactor()), original) + } + + #[rstest::rstest] + #[case::openai_family("mistral", "rate limit reached", PublicError::RateLimit)] + #[case::vertex_family("vertex_ai", "Resource exhausted", PublicError::RateLimit)] + #[case::cohere_family("cohere", "too many tokens", PublicError::ContextWindowExceeded)] + fn a_family_text_rule_beats_the_status_and_keeps_the_real_response( + #[case] provider: &str, + #[case] body: &str, + #[case] expected: PublicError, + ) { + let failure = mapped(provider, &http(401, body)); + assert_eq!(failure.error, expected); + assert_eq!(failure.upstream, upstream(401, body)); } #[test] - fn a_public_original_passes_through_without_banner_debug_or_prefix() { - let original = OriginalException::Public { - class: StatusClass::UnsupportedParams, - message: "Invalid `req_format`".into(), - }; - let context = ExceptionContext { - suppress_debug_info: false, - ..openai() - }; + fn the_other_family_has_no_text_rules() { assert_eq!( - exception_type(&context, &original), - failure( - status(StatusClass::UnsupportedParams, None), - "Invalid `req_format`", - "mistral" - ) + mapped("reducto", &http(401, "rate limit reached")).error, + PublicError::Authentication ); } #[rstest::rstest] - #[case::vertex_family_status_rule(ExceptionFamily::VertexAi, "vertex_ai", PublicFailure { - litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), - ..with_debug(failure(status(StatusClass::BadRequest, upstream(409, "rejected")), "Vertex_aiException - rejected", "vertex_ai")) - })] - #[case::cohere_family(ExceptionFamily::Cohere, "cohere", PublicFailure { - litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), - ..with_debug(failure(status(StatusClass::BadRequest, upstream(409, "rejected")), "CohereException - rejected", "cohere")) - })] - #[case::other_family(ExceptionFamily::Other, "reducto", PublicFailure { - litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), - ..with_debug(failure(status(StatusClass::BadRequest, upstream(409, "rejected")), "ReductoException - rejected", "reducto")) - })] - fn families_without_a_409_rule_reach_the_status_table( - #[case] family: ExceptionFamily, + #[case::openai_403_is_permission_denied("mistral", 403, PublicError::PermissionDenied)] + #[case::openai_409_is_bad_request("mistral", 409, PublicError::BadRequest)] + #[case::vertex_502_is_bad_gateway("vertex_ai", 502, PublicError::BadGateway)] + #[case::vertex_504_is_a_timeout("vertex_ai", 504, PublicError::Timeout { status: 504 })] + #[case::cohere_498_is_bad_request("cohere", 498, PublicError::BadRequest)] + #[case::other_503("reducto", 503, PublicError::ServiceUnavailable)] + fn without_a_text_rule_every_family_uses_the_status_table( #[case] provider: &str, - #[case] expected: PublicFailure, + #[case] status: u16, + #[case] expected: PublicError, ) { assert_eq!( - exception_type(&context(provider, family), &http(409, "rejected")), - expected - ); - } - - #[test] - fn the_openai_family_claims_a_409_before_the_status_table() { - assert_eq!( - exception_type(&openai(), &http(409, "rejected")), - PublicFailure { - litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), - ..with_debug(failure( - PublicKind::Api { - status: 409, - request_url: DOCS_URL - }, - "APIError: MistralException - rejected", - "mistral" - )) + mapped(provider, &http(status, "rejected")), + MappedFailure { + error: expected, + message: format!("{} - rejected", exception_provider(provider)), + upstream: upstream(status, "rejected"), + debug_info: DEBUG.into(), } ); } @@ -447,148 +343,126 @@ mod tests { #[case::timed_out_generating("Timed out generating response")] #[case::read_operation("The read operation timed out")] fn timeout_markers_win_over_every_family(#[case] marker: &str) { - let body = format!("rate limit {marker}"); - for family in [ - ExceptionFamily::OpenAiCompatible, - ExceptionFamily::VertexAi, - ExceptionFamily::Cohere, - ExceptionFamily::Other, - ] { + let body = format!("rate limit invalid api token {marker}"); + for provider in ["mistral", "vertex_ai", "cohere", "reducto"] { assert_eq!( - exception_type(&context("mistral", family), &http(429, &body)), - PublicFailure { - litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), - ..with_debug(failure( - PublicKind::Timeout { status: None }, - &format!("APITimeoutError - Request timed out. Error_str: {body}"), - "mistral" - )) - } + mapped(provider, &http(429, &body)).error, + PublicError::Timeout { status: 408 }, + "{provider}" ); } } + #[test] + fn a_handler_timeout_is_a_408_without_a_response() { + let original = OriginalException::Timeout { + timeout_seconds: Some(0.5), + elapsed_seconds: Some(0.5031), + }; + assert_eq!( + mapped("mistral", &original), + MappedFailure { + error: PublicError::Timeout { status: 408 }, + message: "MistralException - Connection timed out after 0.5 seconds.".into(), + upstream: None, + debug_info: DEBUG.into(), + } + ); + } + #[rstest::rstest] - #[case::provider_error_keeps_the_prefix(http(409, "rejected"), "ReductoException - rejected")] - #[case::synthesized_status_skips_the_status_table( - OriginalException::Connection { message: "refused".into() }, - "ReductoException - refused" - )] - #[case::plain_exception_keeps_its_text( - OriginalException::Local { class: LocalClass::FileNotFound, message: "File not found: /a".into() }, - "File not found: /a" - )] - fn unmapped_failures_are_connection_errors( + #[case::refused_connection(OriginalException::Connection { message: "refused".into() })] + #[case::unparseable_response(OriginalException::Plain { message: "refused".into() })] + #[case::informational_status(OriginalException::Http { status: 399, body: "refused".into(), headers: Vec::new() })] + fn a_failure_no_rule_or_status_claims_is_a_connection_error( #[case] original: OriginalException, - #[case] message: &str, ) { - let context = context("reducto", ExceptionFamily::Other); - let expected = match original { - OriginalException::Http { .. } => with_debug(failure( - status(StatusClass::BadRequest, upstream(409, "rejected")), - message, - "reducto", - )), - OriginalException::Connection { .. } | OriginalException::Local { .. } => { - failure(PublicKind::ApiConnection, message, "reducto") - } - _ => unreachable!(), - }; - let actual = exception_type(&context, &original); - assert_eq!( - PublicFailure { - litellm_response_headers: None, - ..actual - }, - expected - ); + let failure = mapped("reducto", &original); + assert_eq!(failure.error, PublicError::ApiConnection); + assert_eq!(failure.message, "ReductoException - refused"); } #[test] - fn a_missing_provider_renders_like_python_none() { - let context = ExceptionContext { - custom_llm_provider: None, - family: ExceptionFamily::Other, - ..openai() + fn a_timeout_marker_on_a_response_keeps_the_response() { + let failure = mapped("reducto", &http(429, "Request timed out")); + assert_eq!(failure.error, PublicError::Timeout { status: 408 }); + assert_eq!(failure.upstream, upstream(429, "Request timed out")); + } + + #[test] + fn family_text_rules_also_classify_failures_without_a_response() { + let original = OriginalException::Plain { + message: "Request too large".into(), }; - assert_eq!( - exception_type(&context, &http(401, "rejected")), - PublicFailure { - llm_provider: None, - litellm_response_headers: Some(vec![("retry-after".into(), "7".into())]), - ..with_debug(failure( - status(StatusClass::Authentication, upstream(401, "rejected")), - "None - rejected", - "unused" - )) - } - ); + assert_eq!(mapped("mistral", &original).error, PublicError::RateLimit); } #[rstest::rstest] - #[case::suppressed(true, false)] - #[case::printed(false, true)] - fn the_banner_prints_unless_debug_info_is_suppressed( - #[case] suppress_debug_info: bool, - #[case] print_banner: bool, - ) { - let context = ExceptionContext { - suppress_debug_info, - ..openai() - }; + #[case::openai_family("mistral", "MistralException - rejected REDACTED")] + #[case::vertex_family("vertex_ai", "Vertex_aiException - rejected REDACTED")] + #[case::other_family("reducto", "ReductoException - rejected REDACTED")] + fn every_family_reports_the_redacted_text(#[case] provider: &str, #[case] message: &str) { + let failure = mapped(provider, &http(400, "rejected Bearer abcdefghijklmnop")); + assert_eq!(failure.message, message); + } + + #[test] + fn redaction_runs_before_the_rules_see_the_text() { + let body = "db_password=rate_limit"; assert_eq!( - exception_type(&context, &http(400, "rejected")).print_banner, - print_banner + mapped("mistral", &http(400, body)).error, + PublicError::BadRequest + ); + assert_eq!( + exception_type(&context("mistral"), None, &http(400, body)).error, + PublicError::RateLimit ); } #[test] - fn empty_upstream_headers_are_not_reported() { - let original = OriginalException::Http { - status: 400, - body: "rejected".into(), - headers: Vec::new(), - }; - assert_eq!( - exception_type(&openai(), &original).litellm_response_headers, - None - ); - } - - #[test] - fn messages_are_redacted_before_markers_and_prefixes() { + fn without_a_redactor_the_text_is_kept() { let body = "rejected Bearer abcdefghijklmnop"; assert_eq!( - exception_type( - &context("reducto", ExceptionFamily::Other), - &http(400, body) - ) - .message, - "ReductoException - rejected REDACTED" + exception_type(&context("reducto"), None, &http(400, body)).message, + format!("ReductoException - {body}") ); } - const SYNC_TIMEOUT: &str = "litellm.Timeout: Connection timed out after 0.5 seconds."; + #[test] + fn a_rule_hint_follows_the_message() { + let failure = mapped("mistral", &http(400, "invalid_encrypted_content")); + assert_eq!(failure.error, PublicError::BadRequest); + assert!( + failure + .message + .starts_with("MistralException - invalid_encrypted_content\n\n This error occurs") + ); + } #[rstest::rstest] - #[case::sync(false, Some(0.5), Some(0.5031), SYNC_TIMEOUT)] + #[case::sync( + false, + Some(0.5), + Some(0.5031), + "Connection timed out after 0.5 seconds." + )] #[case::async_rounds_the_elapsed_time( true, Some(0.5), Some(0.5031), - "litellm.Timeout: Connection timed out. Timeout passed=0.5, time taken=0.503 seconds" + "Connection timed out. Timeout passed=0.5, time taken=0.503 seconds" )] #[case::whole_seconds_keep_a_decimal( true, Some(600.0), Some(2.0), - "litellm.Timeout: Connection timed out. Timeout passed=600.0, time taken=2.0 seconds" + "Connection timed out. Timeout passed=600.0, time taken=2.0 seconds" )] #[case::unknown_values_render_as_none( true, None, None, - "litellm.Timeout: Connection timed out. Timeout passed=None, time taken=None seconds" + "Connection timed out. Timeout passed=None, time taken=None seconds" )] fn timeout_text_follows_the_delivery_mode( #[case] asynchronous: bool, @@ -602,58 +476,6 @@ mod tests { ); } - #[rstest::rstest] - #[case::sync(false, SYNC_TIMEOUT)] - #[case::async_( - true, - "litellm.Timeout: Connection timed out. Timeout passed=0.5, time taken=0.503 seconds" - )] - fn a_timeout_is_a_408_carrying_the_handler_text( - #[case] asynchronous: bool, - #[case] text: &str, - ) { - let context = ExceptionContext { - asynchronous, - ..openai() - }; - let original = OriginalException::Timeout { - timeout_seconds: Some(0.5), - elapsed_seconds: Some(0.5031), - }; - assert_eq!( - exception_type(&context, &original), - with_debug(failure( - PublicKind::Timeout { status: None }, - &format!("Timeout Error: MistralException - {text}"), - "mistral" - )) - ); - } - - #[test] - fn a_refused_connection_is_a_500_with_an_empty_response() { - assert_eq!( - exception_type( - &openai(), - &OriginalException::Connection { - message: "refused".into() - } - ), - with_debug(failure( - status( - StatusClass::InternalServer, - Some(ResponseArg::Upstream(UpstreamResponse { - status: 500, - body: String::new(), - headers: Vec::new(), - })) - ), - "InternalServerError: MistralException - refused", - "mistral" - )) - ); - } - #[test] fn debug_information_follows_the_python_layout() { let context = ExceptionContext { @@ -662,10 +484,10 @@ mod tests { model_group: Some("ocr".into()), deployment: Some("deployment".into()), user_api_key_alias: Some("key".into()), - ..openai() + ..context("vertex_ai") }; assert_eq!( - extra_information(&context, api_base(&context).as_deref()), + exception_type(&context, None, &http(400, "rejected")).debug_info, concat!( "\n\nKey Name: `key`\nTeam: `None`", "\nModel: ocr-model", @@ -680,21 +502,20 @@ mod tests { #[rstest::rstest] #[case::bare(ExceptionContext::default(), "\nModel: ")] - #[case::redacted_messages(ExceptionContext { redact_messages_in_exceptions: true, model: "m".into(), ..ExceptionContext::default() }, "\nModel: m")] #[case::team_alias( - ExceptionContext { model: "m".into(), user_api_key_alias: Some("key".into()), user_api_key_team_alias: Some("team".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + ExceptionContext { model: "m".into(), user_api_key_alias: Some("key".into()), user_api_key_team_alias: Some("team".into()), ..ExceptionContext::default() }, "\n\nKey Name: `key`\nTeam: `team`\nModel: m" )] #[case::team_alias_without_key_is_ignored( - ExceptionContext { model: "m".into(), user_api_key_team_alias: Some("team".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + ExceptionContext { model: "m".into(), user_api_key_team_alias: Some("team".into()), ..ExceptionContext::default() }, "\nModel: m" )] #[case::project_without_location_has_no_api_base( - ExceptionContext { model: "m".into(), vertex_project: Some("p".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + ExceptionContext { model: "m".into(), vertex_project: Some("p".into()), ..ExceptionContext::default() }, "\nModel: m\nvertex_project: `p`\n" )] #[case::location_without_project_has_no_api_base( - ExceptionContext { model: "m".into(), vertex_location: Some("l".into()), redact_messages_in_exceptions: true, ..ExceptionContext::default() }, + ExceptionContext { model: "m".into(), vertex_location: Some("l".into()), ..ExceptionContext::default() }, "\nModel: m\nvertex_location: `l`\n" )] fn each_optional_context_field_adds_its_own_line( @@ -708,6 +529,7 @@ mod tests { } #[rstest::rstest] + #[case::openai_keeps_its_brand("openai", "OpenAIException")] #[case::lowercase("mistral", "MistralException")] #[case::keeps_the_rest("azure_ai", "Azure_aiException")] #[case::empty("", "")] @@ -717,17 +539,4 @@ mod tests { ) { assert_eq!(exception_provider(provider), expected); } - - #[rstest::rstest] - #[case::lowers_the_rest("vERTEX_AI", "Vertex_ai")] - #[case::empty("", "")] - fn python_capitalize_lowers_the_rest(#[case] value: &str, #[case] expected: &str) { - assert_eq!(python_capitalize(value), expected); - } - - #[test] - fn debug_constant_matches_the_default_test_context() { - let context = openai(); - assert_eq!(extra_information(&context, None), DEBUG); - } } diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs index ffbb172582b..b4f4496dbf0 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs @@ -1,85 +1,31 @@ -use super::public::{PublicFailure, StatusClass}; -use super::rules::{ - ApiStatus, Kind, ResponseChoice, Rule, apply, contains_any, is_context_window_exceeded, - is_rate_limit, -}; -use super::{DOCS_URL, Mapping}; - -const OPENAI_URL: &str = "https://api.openai.com/v1"; +use super::public::PublicError; +use super::rules::{Rule, contains_any, is_context_window_exceeded, is_rate_limit}; const ENCRYPTED_CONTENT_HELP: &str = "\n\n This error occurs when load balancing Responses API across deployments with different API keys.\n Encrypted content is tied to the organization that created it and cannot be decrypted by other organizations.\n\n Solution: Enable 'encrypted_content_affinity' to route follow-up requests to the correct deployment:\n\n router_settings:\n enable_pre_call_checks: true\n optional_pre_call_checks:\n - encrypted_content_affinity\n\n Learn more: https://docs.litellm.ai/docs/response_api#encrypted-content-affinity-multi-region-load-balancing"; -const fn with_response(class: StatusClass) -> Kind { - Kind::Status { - class, - response: ResponseChoice::Provider, - } -} - -fn exception_provider(mapping: &Mapping<'_>) -> String { - if mapping.provider == "openai" { - "OpenAIException".to_string() - } else { - super::exception_provider(mapping.provider) - } -} - -/// The raw message with OpenAI's own names swapped for the provider's. -fn message(mapping: &Mapping<'_>) -> String { - let provider = mapping.provider; - mapping - .original - .message - .replace("OPENAI", &provider.to_uppercase()) - .replace("openai.OpenAIError", &format!("{provider}.{provider}Error")) -} - -fn prefixed(mapping: &Mapping<'_>, label: &str) -> String { - format!( - "{label}{} - {}", - exception_provider(mapping), - message(mapping) - ) -} - -fn status_is(mapping: &Mapping<'_>, statuses: &[u16]) -> bool { - mapping - .original - .status - .is_some_and(|status| statuses.contains(&status)) -} - -/// `_map_openai_exception`, in its branch order. -const RULES: &[Rule] = &[ - Rule { - when: |mapping| is_rate_limit(&mapping.error_str, mapping.original.status), - kind: with_response(StatusClass::RateLimit), - message: |mapping| prefixed(mapping, "RateLimitError: "), - debug: false, - }, - Rule { - when: |mapping| is_context_window_exceeded(&mapping.error_str), - kind: with_response(StatusClass::ContextWindowExceeded), - message: |mapping| prefixed(mapping, "ContextWindowExceededError: "), - debug: true, - }, - Rule { - when: |mapping| { +/// The text branches of `_map_openai_exception`, in its order. +pub(super) const RULES: &[Rule] = &[ + Rule::new( + |mapping| is_rate_limit(&mapping.error_str, mapping.status), + PublicError::RateLimit, + ), + Rule::new( + |mapping| is_context_window_exceeded(&mapping.error_str), + PublicError::ContextWindowExceeded, + ), + Rule::new( + |mapping| { mapping.error_str.contains("invalid_request_error") && mapping.error_str.contains("model_not_found") }, - kind: with_response(StatusClass::NotFound), - message: |mapping| prefixed(mapping, ""), - debug: true, - }, - Rule { - when: |mapping| mapping.error_str.contains("A timeout occurred"), - kind: Kind::Timeout(None), - message: |mapping| prefixed(mapping, ""), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::NotFound, + ), + Rule::new( + |mapping| mapping.error_str.contains("A timeout occurred"), + PublicError::Timeout { status: 408 }, + ), + Rule::new( + |mapping| { let error_str = &mapping.error_str; (error_str.contains("invalid_request_error") && error_str.contains("content_policy_violation")) @@ -89,38 +35,29 @@ const RULES: &[Rule] = &[ .to_lowercase() .contains("request was rejected as a result of the safety system") }, - kind: with_response(StatusClass::ContentPolicyViolation), - message: |mapping| prefixed(mapping, "ContentPolicyViolationError: "), - debug: true, - }, + PublicError::ContentPolicyViolation, + ), Rule { - when: |mapping| { - contains_any( - &mapping.error_str, - &["invalid_encrypted_content", "could not be verified"], - ) - }, - kind: with_response(StatusClass::BadRequest), - message: |mapping| { - format!( - "{} - {}{ENCRYPTED_CONTENT_HELP}", - exception_provider(mapping), - message(mapping) - ) - }, - debug: true, + hint: ENCRYPTED_CONTENT_HELP, + ..Rule::new( + |mapping| { + contains_any( + &mapping.error_str, + &["invalid_encrypted_content", "could not be verified"], + ) + }, + PublicError::BadRequest, + ) }, - Rule { - when: |mapping| { + Rule::new( + |mapping| { mapping.error_str.contains("invalid_request_error") && !mapping.error_str.contains("Incorrect API key provided") }, - kind: with_response(StatusClass::BadRequest), - message: |mapping| prefixed(mapping, ""), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::BadRequest, + ), + Rule::new( + |mapping| { contains_any( &mapping.error_str, &[ @@ -129,458 +66,127 @@ const RULES: &[Rule] = &[ ], ) }, - kind: Kind::Status { - class: StatusClass::InternalServer, - response: ResponseChoice::Omitted, - }, - message: |mapping| prefixed(mapping, ""), - debug: false, - }, - Rule { - when: |mapping| mapping.error_str.contains("Request too large"), - kind: with_response(StatusClass::RateLimit), - message: |mapping| prefixed(mapping, "RateLimitError: "), - debug: true, - }, - Rule { - when: |mapping| { - mapping.error_str.contains("The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable") - }, - kind: with_response(StatusClass::Authentication), - message: |mapping| prefixed(mapping, "AuthenticationError: "), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::InternalServer, + ), + Rule::new( + |mapping| mapping.error_str.contains("Request too large"), + PublicError::RateLimit, + ), + Rule::new( + |mapping| { mapping .error_str .contains("Mistral API raised a streaming error") }, - kind: Kind::Api { - status: ApiStatus::Fixed(500), - request_url: OPENAI_URL, - }, - message: |mapping| prefixed(mapping, ""), - debug: true, - }, - Rule { - when: |mapping| mapping.original.status.is_none(), - kind: Kind::ApiConnection, - message: |mapping| prefixed(mapping, "APIConnectionError: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[400, 422]), - kind: with_response(StatusClass::BadRequest), - message: |mapping| prefixed(mapping, ""), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[401]), - kind: with_response(StatusClass::Authentication), - message: |mapping| prefixed(mapping, "AuthenticationError: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[404]), - kind: with_response(StatusClass::NotFound), - message: |mapping| prefixed(mapping, "NotFoundError: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[408]), - kind: Kind::Timeout(None), - message: |mapping| prefixed(mapping, "Timeout Error: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[429]), - kind: with_response(StatusClass::RateLimit), - message: |mapping| prefixed(mapping, "RateLimitError: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[500]), - kind: with_response(StatusClass::InternalServer), - message: |mapping| prefixed(mapping, "InternalServerError: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[502]), - kind: with_response(StatusClass::BadGateway), - message: |mapping| prefixed(mapping, "BadGatewayError: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[503]), - kind: with_response(StatusClass::ServiceUnavailable), - message: |mapping| prefixed(mapping, "ServiceUnavailableError: "), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, &[504]), - kind: Kind::Timeout(Some(504)), - message: |mapping| prefixed(mapping, "Timeout Error: "), - debug: true, - }, - Rule { - when: |_| true, - kind: Kind::Api { - status: ApiStatus::Original, - request_url: DOCS_URL, - }, - message: |mapping| prefixed(mapping, "APIError: "), - debug: true, - }, + PublicError::Api { status: 500 }, + ), ]; -pub(super) fn map(mapping: &Mapping<'_>) -> Option { - apply(RULES, mapping) -} - #[cfg(test)] mod tests { - use super::super::testing::{context, failure, http, status, upstream, with_debug}; - use super::super::{ExceptionFamily, OriginalException, PublicKind}; + use super::super::rules::first_match; + use super::super::testing::mapping; use super::*; - fn mapped(provider: &str, original: &OriginalException) -> PublicFailure { - let context = context(provider, ExceptionFamily::OpenAiCompatible); - map(&Mapping::new(&context, original)).expect("the OpenAI table ends in a catch-all") - } - - fn kind(class: StatusClass, status_code: u16, body: &str) -> PublicKind { - status(class, upstream(status_code, body)) + fn classified(status: Option, text: &str) -> Option { + first_match(RULES, &mapping(status, text)).map(|rule| rule.error) } #[rstest::rstest] - #[case::rate_limit_phrase( - 400, - "rate limit reached", - failure( - kind(StatusClass::RateLimit, 400, "rate limit reached"), - "RateLimitError: MistralException - rate limit reached", - "mistral", - ) - )] + #[case::rate_limit_phrase("rate limit reached", PublicError::RateLimit)] #[case::context_window( - 500, "This model's maximum context length is 10", - with_debug(failure( - kind( - StatusClass::ContextWindowExceeded, - 500, - "This model's maximum context length is 10" - ), - "ContextWindowExceededError: MistralException - This model's maximum context length is 10", - "mistral", - )) + PublicError::ContextWindowExceeded )] - #[case::model_not_found( - 400, - "invalid_request_error model_not_found", - with_debug(failure( - kind(StatusClass::NotFound, 400, "invalid_request_error model_not_found"), - "MistralException - invalid_request_error model_not_found", - "mistral", - )) - )] - #[case::timeout_occurred(400, "A timeout occurred", with_debug(failure( - PublicKind::Timeout { status: None }, - "MistralException - A timeout occurred", - "mistral", - )))] + #[case::model_not_found("invalid_request_error model_not_found", PublicError::NotFound)] + #[case::timeout_occurred("A timeout occurred", PublicError::Timeout { status: 408 })] #[case::content_policy_error_code( - 400, "invalid_request_error content_policy_violation", - with_debug(failure( - kind( - StatusClass::ContentPolicyViolation, - 400, - "invalid_request_error content_policy_violation" - ), - "ContentPolicyViolationError: MistralException - invalid_request_error content_policy_violation", - "mistral", - )) + PublicError::ContentPolicyViolation )] #[case::content_policy_usage_policy( - 400, "Invalid prompt violating our usage policy", - with_debug(failure( - kind( - StatusClass::ContentPolicyViolation, - 400, - "Invalid prompt violating our usage policy" - ), - "ContentPolicyViolationError: MistralException - Invalid prompt violating our usage policy", - "mistral", - )) + PublicError::ContentPolicyViolation )] #[case::content_policy_safety_system( - 400, "Request was rejected as a result of the safety system", - with_debug(failure( - kind( - StatusClass::ContentPolicyViolation, - 400, - "Request was rejected as a result of the safety system" - ), - "ContentPolicyViolationError: MistralException - Request was rejected as a result of the safety system", - "mistral", - )) - )] - #[case::encrypted_content(400, "invalid_encrypted_content", with_debug(failure( - kind(StatusClass::BadRequest, 400, "invalid_encrypted_content"), - &format!("MistralException - invalid_encrypted_content{ENCRYPTED_CONTENT_HELP}"), - "mistral", - )))] - #[case::unverifiable_content(400, "could not be verified", with_debug(failure( - kind(StatusClass::BadRequest, 400, "could not be verified"), - &format!("MistralException - could not be verified{ENCRYPTED_CONTENT_HELP}"), - "mistral", - )))] - #[case::invalid_request( - 429, - "invalid_request_error bad field", - with_debug(failure( - kind(StatusClass::BadRequest, 429, "invalid_request_error bad field"), - "MistralException - invalid_request_error bad field", - "mistral", - )) + PublicError::ContentPolicyViolation )] + #[case::encrypted_content("invalid_encrypted_content", PublicError::BadRequest)] + #[case::unverifiable_content("could not be verified", PublicError::BadRequest)] + #[case::invalid_request("invalid_request_error bad field", PublicError::BadRequest)] #[case::unknown_server_error( - 400, "Web server is returning an unknown error", - failure( - status(StatusClass::InternalServer, None), - "MistralException - Web server is returning an unknown error", - "mistral", - ) + PublicError::InternalServer )] #[case::server_had_an_error( - 400, "The server had an error processing your request.", - failure( - status(StatusClass::InternalServer, None), - "MistralException - The server had an error processing your request.", - "mistral", - ) + PublicError::InternalServer )] - #[case::request_too_large( - 400, - "Request too large", - with_debug(failure( - kind(StatusClass::RateLimit, 400, "Request too large"), - "RateLimitError: MistralException - Request too large", - "mistral", - )) + #[case::request_too_large("Request too large", PublicError::RateLimit)] + #[case::mistral_streaming_error( + "Mistral API raised a streaming error", + PublicError::Api { status: 500 } )] - #[case::missing_client_api_key( - 400, - "The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable", - with_debug(failure( - kind( - StatusClass::Authentication, - 400, - "The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable", - ), - "AuthenticationError: MistralException - The api_key client option must be set either by passing api_key to the client or by setting the MISTRAL_API_KEY environment variable", - "mistral", - )) - )] - #[case::mistral_streaming_error(400, "Mistral API raised a streaming error", with_debug(failure( - PublicKind::Api { status: 500, request_url: OPENAI_URL }, - "MistralException - Mistral API raised a streaming error", - "mistral", - )))] - fn each_text_rule_maps_by_the_body( - #[case] status_code: u16, - #[case] body: &str, - #[case] expected: PublicFailure, - ) { - assert_eq!(mapped("mistral", &http(status_code, body)), expected); - } - - #[rstest::rstest] - #[case::bad_request( - 400, - kind(StatusClass::BadRequest, 400, "rejected"), - "MistralException - rejected" - )] - #[case::unprocessable( - 422, - kind(StatusClass::BadRequest, 422, "rejected"), - "MistralException - rejected" - )] - #[case::authentication( - 401, - kind(StatusClass::Authentication, 401, "rejected"), - "AuthenticationError: MistralException - rejected" - )] - #[case::not_found( - 404, - kind(StatusClass::NotFound, 404, "rejected"), - "NotFoundError: MistralException - rejected" - )] - #[case::request_timeout(408, PublicKind::Timeout { status: None }, "Timeout Error: MistralException - rejected")] - #[case::rate_limited( - 429, - kind(StatusClass::RateLimit, 429, "rejected"), - "RateLimitError: MistralException - rejected" - )] - #[case::internal_server( - 500, - kind(StatusClass::InternalServer, 500, "rejected"), - "InternalServerError: MistralException - rejected" - )] - #[case::bad_gateway( - 502, - kind(StatusClass::BadGateway, 502, "rejected"), - "BadGatewayError: MistralException - rejected" - )] - #[case::service_unavailable( - 503, - kind(StatusClass::ServiceUnavailable, 503, "rejected"), - "ServiceUnavailableError: MistralException - rejected" - )] - #[case::gateway_timeout(504, PublicKind::Timeout { status: Some(504) }, "Timeout Error: MistralException - rejected")] - #[case::any_other_status(409, PublicKind::Api { status: 409, request_url: DOCS_URL }, "APIError: MistralException - rejected")] - fn each_status_rule_maps_by_the_status( - #[case] status_code: u16, - #[case] kind: PublicKind, - #[case] message: &str, - ) { - assert_eq!( - mapped("mistral", &http(status_code, "rejected")), - with_debug(failure(kind, message, "mistral")) - ); - } - - #[test] - fn a_failure_without_a_status_is_a_connection_error() { - let original = OriginalException::Response { - message: "invalid OCR response field: pages".into(), - }; - assert_eq!( - mapped("mistral", &original), - with_debug(failure( - PublicKind::ApiConnection, - "APIConnectionError: MistralException - invalid OCR response field: pages", - "mistral" - )) - ); + fn each_text_rule_claims_its_marker(#[case] text: &str, #[case] expected: PublicError) { + assert_eq!(classified(Some(400), text), Some(expected)); } #[rstest::rstest] #[case::rate_limit_before_context_window( - 400, "rate limit and This model's maximum context length is 10", - kind( - StatusClass::RateLimit, - 400, - "rate limit and This model's maximum context length is 10" - ), - "RateLimitError: MistralException - rate limit and This model's maximum context length is 10", - false + PublicError::RateLimit )] #[case::context_window_before_content_policy( - 400, "This model's maximum context length is 10 invalid_request_error content_policy_violation", - kind( - StatusClass::ContextWindowExceeded, - 400, - "This model's maximum context length is 10 invalid_request_error content_policy_violation" - ), - "ContextWindowExceededError: MistralException - This model's maximum context length is 10 invalid_request_error content_policy_violation", - true + PublicError::ContextWindowExceeded )] #[case::model_not_found_before_invalid_request( - 400, "invalid_request_error model_not_found", - kind(StatusClass::NotFound, 400, "invalid_request_error model_not_found"), - "MistralException - invalid_request_error model_not_found", - true + PublicError::NotFound )] #[case::timeout_before_invalid_request( - 400, "A timeout occurred invalid_request_error", - PublicKind::Timeout { status: None }, - "MistralException - A timeout occurred invalid_request_error", - true + PublicError::Timeout { status: 408 } )] #[case::content_policy_before_invalid_request( - 400, "invalid_request_error content_policy_violation", - kind( - StatusClass::ContentPolicyViolation, - 400, - "invalid_request_error content_policy_violation" - ), - "ContentPolicyViolationError: MistralException - invalid_request_error content_policy_violation", - true + PublicError::ContentPolicyViolation )] - #[case::invalid_request_with_a_bad_key_falls_to_the_status( - 401, - "invalid_request_error Incorrect API key provided", - kind( - StatusClass::Authentication, - 401, - "invalid_request_error Incorrect API key provided" - ), - "AuthenticationError: MistralException - invalid_request_error Incorrect API key provided", - true + #[case::encrypted_content_before_invalid_request( + "invalid_request_error invalid_encrypted_content", + PublicError::BadRequest )] - #[case::text_rules_before_status( - 429, - "invalid_request_error bad field", - kind(StatusClass::BadRequest, 429, "invalid_request_error bad field"), - "MistralException - invalid_request_error bad field", - true - )] - #[case::echoed_429_is_not_a_rate_limit( - 400, - "token 429 in the prompt", - kind(StatusClass::BadRequest, 400, "token 429 in the prompt"), - "MistralException - token 429 in the prompt", - true - )] - fn the_earlier_rule_wins_when_two_apply( - #[case] status_code: u16, - #[case] body: &str, - #[case] kind: PublicKind, - #[case] message: &str, - #[case] debug: bool, + fn the_earlier_rule_wins_when_two_apply(#[case] text: &str, #[case] expected: PublicError) { + assert_eq!(classified(Some(400), text), Some(expected)); + } + + #[rstest::rstest] + #[case::encrypted_content("invalid_encrypted_content", ENCRYPTED_CONTENT_HELP)] + #[case::plain_invalid_request("invalid_request_error bad field", "")] + fn only_encrypted_content_failures_carry_the_affinity_help( + #[case] text: &str, + #[case] hint: &str, ) { - let expected = failure(kind, message, "mistral"); assert_eq!( - mapped("mistral", &http(status_code, body)), - if debug { - with_debug(expected) - } else { - expected - } + first_match(RULES, &mapping(Some(400), text)).map(|rule| rule.hint), + Some(hint) ); } #[rstest::rstest] - #[case::provider_names_replace_openai( - "azure_ai", - "OPENAI said openai.OpenAIError", - "Azure_aiException - AZURE_AI said azure_ai.azure_aiError" - )] - #[case::openai_keeps_its_own_name("openai", "rejected", "OpenAIException - rejected")] - fn the_message_names_the_provider( - #[case] provider: &str, - #[case] body: &str, - #[case] message: &str, - ) { + #[case::bad_key_is_left_to_the_status("invalid_request_error Incorrect API key provided")] + #[case::echoed_429_is_not_a_rate_limit("token 429 in the prompt")] + #[case::unmarked("rejected")] + fn text_without_a_marker_is_left_to_the_status_table(#[case] text: &str) { + assert_eq!(classified(Some(400), text), None); + } + + #[test] + fn a_standalone_429_counts_only_with_a_429_status() { assert_eq!( - mapped(provider, &http(400, body)), - with_debug(failure( - kind(StatusClass::BadRequest, 400, body), - message, - provider - )) + classified(Some(429), "got 429 back"), + Some(PublicError::RateLimit) ); } } diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs index d64868c1606..82392cbd7ee 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs @@ -1,14 +1,4 @@ -use super::public::StatusClass; - -#[derive(Clone, Copy, Debug, PartialEq, Eq)] -pub enum LocalClass { - ValueError, - FileNotFound, - OsError, -} - -/// A route failure in the shape Python's `exception_type` receives it, before any public -/// class is chosen. +/// A failure a Rust route produced, before any public class is chosen. #[derive(Clone, Debug, PartialEq)] pub enum OriginalException { Http { @@ -23,27 +13,132 @@ pub enum OriginalException { timeout_seconds: Option, elapsed_seconds: Option, }, - Response { - message: String, - }, - Local { - class: LocalClass, - message: String, - }, - /// A failure Python raises as a public LiteLLM exception itself, which `exception_type` - /// hands back unchanged. - Public { - class: StatusClass, + /// A failure with no HTTP response behind it, such as an unparseable body or a local + /// file error. + Plain { message: String, }, } -/// Which of the provider-specific mappers in `exception_type` a route's provider uses. -#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)] +/// Which provider-specific text rules apply before the shared status table. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] pub enum ExceptionFamily { OpenAiCompatible, VertexAi, Cohere, - #[default] Other, } + +/// `openai_compatible_providers` in `litellm/constants.py`. +const OPENAI_COMPATIBLE_PROVIDERS: &[&str] = &[ + "anyscale", + "groq", + "nvidia_nim", + "cerebras", + "baseten", + "sambanova", + "ai21_chat", + "ai21", + "volcengine", + "codestral", + "deepseek", + "tencent", + "deepinfra", + "perplexity", + "xinference", + "xai", + "zai", + "together_ai", + "fireworks_ai", + "empower", + "friendliai", + "azure_ai", + "github", + "litellm_proxy", + "hosted_vllm", + "llamafile", + "lm_studio", + "galadriel", + "github_copilot", + "chatgpt", + "novita", + "meta_llama", + "publicai", + "synthetic", + "tensormesh", + "apertis", + "nano-gpt", + "poe", + "chutes", + "parasail", + "libertai", + "featherless_ai", + "nscale", + "nebius", + "dashscope", + "qwencloud", + "qwen_ai_platform", + "modelscope", + "moonshot", + "v0", + "helicone", + "morph", + "lambda_ai", + "inception", + "hyperbolic", + "vercel_ai_gateway", + "aiml", + "wandb", + "cometapi", + "clarifai", + "docker_model_runner", + "ragflow", + "pinstripes", + "darkbloom", + "meta", + "cognition", + "scx-ai", +]; + +impl ExceptionFamily { + /// The provider dispatch at the top of Python's `exception_type`, in its order. + pub fn for_provider(provider: &str) -> Self { + match provider { + "openai" | "text-completion-openai" | "custom_openai" | "mistral" | "runwayml" => { + Self::OpenAiCompatible + } + provider if OPENAI_COMPATIBLE_PROVIDERS.contains(&provider) => Self::OpenAiCompatible, + "vertex_ai" | "vertex_ai_beta" | "gemini" => Self::VertexAi, + "cohere" | "cohere_chat" => Self::Cohere, + _ => Self::Other, + } + } +} + +#[cfg(test)] +mod tests { + use super::*; + + #[rstest::rstest] + #[case::openai("openai", ExceptionFamily::OpenAiCompatible)] + #[case::text_completion_openai("text-completion-openai", ExceptionFamily::OpenAiCompatible)] + #[case::custom_openai("custom_openai", ExceptionFamily::OpenAiCompatible)] + #[case::mistral("mistral", ExceptionFamily::OpenAiCompatible)] + #[case::runwayml("runwayml", ExceptionFamily::OpenAiCompatible)] + #[case::listed_compatible("azure_ai", ExceptionFamily::OpenAiCompatible)] + #[case::compatible_list_wins_over_its_own_mapper( + "together_ai", + ExceptionFamily::OpenAiCompatible + )] + #[case::vertex_ai("vertex_ai", ExceptionFamily::VertexAi)] + #[case::vertex_ai_beta("vertex_ai_beta", ExceptionFamily::VertexAi)] + #[case::gemini("gemini", ExceptionFamily::VertexAi)] + #[case::cohere("cohere", ExceptionFamily::Cohere)] + #[case::cohere_chat("cohere_chat", ExceptionFamily::Cohere)] + #[case::unported_mapper("anthropic", ExceptionFamily::Other)] + #[case::unknown("reducto", ExceptionFamily::Other)] + #[case::empty("", ExceptionFamily::Other)] + fn provider_selects_the_family(#[case] provider: &str, #[case] family: ExceptionFamily) { + assert_eq!(ExceptionFamily::for_provider(provider), family); + } +} diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs index a7319c1287b..c567185aa6d 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/public.rs @@ -1,262 +1,77 @@ -use serde::Serialize; - -/// The public LiteLLM classes built from a status code alone: every one takes the same -/// constructor arguments. -#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, strum::EnumIter, strum::IntoStaticStr)] -#[serde(rename_all = "snake_case")] -#[strum(serialize_all = "snake_case")] -pub enum StatusClass { +/// The public LiteLLM exception classes a Rust route failure can become. Python builds the +/// class; Rust decides which one. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum PublicError { BadRequest, + ContextWindowExceeded, + ContentPolicyViolation, Authentication, PermissionDenied, NotFound, + Timeout { status: u16 }, RateLimit, - ContextWindowExceeded, - ContentPolicyViolation, InternalServer, BadGateway, ServiceUnavailable, - UnsupportedParams, + ApiConnection, + Api { status: u16 }, } -impl StatusClass { - /// The `status_code` the Python class sets on itself. +impl PublicError { + /// The `status_code` the Python class carries. pub const fn status_code(self) -> u16 { match self { - Self::BadRequest - | Self::ContextWindowExceeded - | Self::ContentPolicyViolation - | Self::UnsupportedParams => 400, + Self::BadRequest | Self::ContextWindowExceeded | Self::ContentPolicyViolation => 400, Self::Authentication => 401, Self::PermissionDenied => 403, Self::NotFound => 404, Self::RateLimit => 429, - Self::InternalServer => 500, + Self::InternalServer | Self::ApiConnection => 500, Self::BadGateway => 502, Self::ServiceUnavailable => 503, + Self::Timeout { status } | Self::Api { status } => status, } } } -#[derive(Clone, Debug, PartialEq, Eq, Serialize)] +#[derive(Clone, Debug, PartialEq, Eq)] pub struct UpstreamResponse { pub status: u16, pub body: String, pub headers: Vec<(String, String)>, } -#[derive(Clone, Debug, PartialEq, Eq, Serialize)] -pub struct HttpStub { - pub status: u16, - pub method: &'static str, - pub url: &'static str, - pub content: Option, -} - -#[derive(Clone, Debug, PartialEq, Eq, Serialize)] -#[serde(tag = "type", rename_all = "snake_case")] -pub enum ResponseArg { - Upstream(UpstreamResponse), - Stub(HttpStub), -} - -#[derive(Clone, Debug, PartialEq, Eq, Serialize)] -#[serde(tag = "type", rename_all = "snake_case")] -pub enum PublicKind { - Status { - status_class: StatusClass, - response: Option, - }, - Timeout { - status: Option, - }, - ApiConnection, - Api { - status: u16, - request_url: &'static str, - }, -} - -/// Constructor arguments for the public LiteLLM exception, as `exception_type` passes them. -#[derive(Clone, Debug, PartialEq, Eq, Serialize)] -pub struct PublicFailure { - pub kind: PublicKind, +#[derive(Clone, Debug, PartialEq, Eq)] +pub struct MappedFailure { + pub error: PublicError, pub message: String, - pub model: String, - pub llm_provider: Option, - pub litellm_debug_info: Option, - pub litellm_response_headers: Option>, - pub print_banner: bool, + pub upstream: Option, + pub debug_info: String, } #[cfg(test)] mod tests { - use std::collections::BTreeSet; - use std::path::PathBuf; - - use serde_json::Value; - use strum::IntoEnumIterator; - use super::*; - const REGENERATE: &str = "LITELLM_REGENERATE_PUBLIC_FAILURE_FIXTURES"; - - fn fixture_directory() -> PathBuf { - PathBuf::from(env!("CARGO_MANIFEST_DIR")) - .join("../../../tests/test_litellm/rust_bridge/fixtures/public_failures") - } - - fn upstream(status: u16) -> ResponseArg { - ResponseArg::Upstream(UpstreamResponse { - status, - body: r#"{"message": "rejected"}"#.into(), - headers: vec![("retry-after".into(), "7".into())], - }) - } - - fn status_response(class: StatusClass) -> Option { - match class { - StatusClass::Authentication => None, - StatusClass::PermissionDenied => Some(ResponseArg::Stub(HttpStub { - status: 403, - method: "POST", - url: " https://cloud.google.com/vertex-ai/", - content: None, - })), - StatusClass::InternalServer => Some(ResponseArg::Stub(HttpStub { - status: 500, - method: "completion", - url: "https://github.com/BerriAI/litellm", - content: Some("upstream text".into()), - })), - class => Some(upstream(class.status_code())), - } - } - - fn failure(kind: PublicKind, name: &str) -> PublicFailure { - let headers = matches!( - &kind, - PublicKind::Status { - response: Some(ResponseArg::Upstream(_)), - .. - } - ); - PublicFailure { - kind, - message: format!("MistralException - {name}"), - model: "ocr-model".into(), - llm_provider: Some("mistral".into()), - litellm_debug_info: Some("\nModel: ocr-model".into()), - litellm_response_headers: headers.then(|| vec![("retry-after".into(), "7".into())]), - print_banner: false, - } - } - - /// One payload per public class the constructor can build; `test_failures.py` reads the - /// same files, so a shape change on either side fails there or here. - fn fixtures() -> Vec<(String, PublicFailure)> { - let statuses = StatusClass::iter().map(|class| { - let name: &'static str = class.into(); - let name = format!("status_{name}"); - let built = failure( - PublicKind::Status { - status_class: class, - response: status_response(class), - }, - &name, - ); - (name, built) - }); - let others = [ - ( - "timeout_with_status", - PublicFailure { - print_banner: true, - ..failure( - PublicKind::Timeout { status: Some(504) }, - "timeout_with_status", - ) - }, - ), - ( - "timeout_without_status", - PublicFailure { - litellm_debug_info: None, - ..failure( - PublicKind::Timeout { status: None }, - "timeout_without_status", - ) - }, - ), - ( - "api_connection", - PublicFailure { - llm_provider: None, - ..failure(PublicKind::ApiConnection, "api_connection") - }, - ), - ( - "api", - failure( - PublicKind::Api { - status: 409, - request_url: "https://docs.litellm.ai/docs", - }, - "api", - ), - ), - ] - .map(|(name, built)| (name.to_string(), built)); - statuses.chain(others).collect() - } - - #[test] - fn serialized_payloads_match_the_golden_fixtures_python_reads() { - let directory = fixture_directory(); - let regenerate = std::env::var_os(REGENERATE).is_some(); - let expected = fixtures(); - for (name, built) in &expected { - let path = directory.join(format!("{name}.json")); - let serialized = serde_json::to_value(built).unwrap(); - if regenerate { - std::fs::create_dir_all(&directory).unwrap(); - std::fs::write( - &path, - format!("{}\n", serde_json::to_string_pretty(&serialized).unwrap()), - ) - .unwrap(); - } - let golden: Value = - serde_json::from_str(&std::fs::read_to_string(&path).unwrap()).unwrap(); - assert_eq!(serialized, golden, "{name}; set {REGENERATE}=1 to rewrite"); - } - let on_disk: BTreeSet = std::fs::read_dir(&directory) - .unwrap() - .map(|entry| entry.unwrap().file_name().to_string_lossy().into_owned()) - .collect(); - let generated: BTreeSet = expected - .iter() - .map(|(name, _)| format!("{name}.json")) - .collect(); - assert_eq!(on_disk, generated); - } - #[rstest::rstest] - #[case(StatusClass::BadRequest, 400)] - #[case(StatusClass::Authentication, 401)] - #[case(StatusClass::PermissionDenied, 403)] - #[case(StatusClass::NotFound, 404)] - #[case(StatusClass::RateLimit, 429)] - #[case(StatusClass::ContextWindowExceeded, 400)] - #[case(StatusClass::ContentPolicyViolation, 400)] - #[case(StatusClass::InternalServer, 500)] - #[case(StatusClass::BadGateway, 502)] - #[case(StatusClass::ServiceUnavailable, 503)] - #[case(StatusClass::UnsupportedParams, 400)] + #[case::bad_request(PublicError::BadRequest, 400)] + #[case::context_window(PublicError::ContextWindowExceeded, 400)] + #[case::content_policy(PublicError::ContentPolicyViolation, 400)] + #[case::authentication(PublicError::Authentication, 401)] + #[case::permission_denied(PublicError::PermissionDenied, 403)] + #[case::not_found(PublicError::NotFound, 404)] + #[case::request_timeout(PublicError::Timeout { status: 408 }, 408)] + #[case::gateway_timeout(PublicError::Timeout { status: 504 }, 504)] + #[case::rate_limit(PublicError::RateLimit, 429)] + #[case::internal_server(PublicError::InternalServer, 500)] + #[case::api_connection(PublicError::ApiConnection, 500)] + #[case::bad_gateway(PublicError::BadGateway, 502)] + #[case::service_unavailable(PublicError::ServiceUnavailable, 503)] + #[case::api(PublicError::Api { status: 501 }, 501)] fn status_codes_are_the_ones_the_python_classes_set( - #[case] class: StatusClass, + #[case] error: PublicError, #[case] status: u16, ) { - assert_eq!(class.status_code(), status); + assert_eq!(error.status_code(), status); } } diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs index 34cf2c390c3..0346a8bc718 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/rules.rs @@ -4,106 +4,29 @@ use fancy_regex::Regex; use serde_json::Value; use super::Mapping; -use super::public::{HttpStub, PublicFailure, PublicKind, ResponseArg, StatusClass}; +use super::public::PublicError; -const GITHUB_URL: &str = "https://github.com/BerriAI/litellm"; - -#[derive(Clone, Copy, Debug, PartialEq, Eq)] -pub(super) enum ResponseChoice { - Omitted, - Provider, - Stub { status: u16, url: &'static str }, - InternalServerStub, -} - -#[derive(Clone, Copy, Debug, PartialEq, Eq)] -pub(super) enum ApiStatus { - Fixed(u16), - Original, -} - -#[derive(Clone, Copy, Debug, PartialEq, Eq)] -pub(super) enum Kind { - Status { - class: StatusClass, - response: ResponseChoice, - }, - Timeout(Option), - ApiConnection, - Api { - status: ApiStatus, - request_url: &'static str, - }, -} - -/// One branch of a Python `_map_*_exception` function: when it applies, the class it -/// raises, the message it builds, and whether it passes `litellm_debug_info`. +/// One text branch of a Python `_map_*_exception` function: when it applies, the class it +/// raises, and any help text appended to the message. pub(super) struct Rule { - pub(super) when: fn(&Mapping<'_>) -> bool, - pub(super) kind: Kind, - pub(super) message: fn(&Mapping<'_>) -> String, - pub(super) debug: bool, -} - -/// The first rule that applies decides the failure, as the `if`/`elif` chain does in Python. -pub(super) fn apply(rules: &[Rule], mapping: &Mapping<'_>) -> Option { - rules - .iter() - .find(|rule| (rule.when)(mapping)) - .map(|rule| rule.build(mapping)) + pub(super) when: fn(&Mapping) -> bool, + pub(super) error: PublicError, + pub(super) hint: &'static str, } impl Rule { - fn build(&self, mapping: &Mapping<'_>) -> PublicFailure { - let kind = match self.kind { - Kind::Status { class, response } => PublicKind::Status { - status_class: class, - response: response.resolve(mapping), - }, - Kind::Timeout(status) => PublicKind::Timeout { status }, - Kind::ApiConnection => PublicKind::ApiConnection, - Kind::Api { - status, - request_url, - } => PublicKind::Api { - status: match status { - ApiStatus::Fixed(status) => status, - ApiStatus::Original => mapping.original.status.unwrap_or(500), - }, - request_url, - }, - }; - PublicFailure { - kind, - message: (self.message)(mapping), - model: mapping.context.model.clone(), - llm_provider: mapping.context.custom_llm_provider.clone(), - litellm_debug_info: self.debug.then(|| mapping.extra_information.clone()), - litellm_response_headers: None, - print_banner: false, + pub(super) const fn new(when: fn(&Mapping) -> bool, error: PublicError) -> Self { + Self { + when, + error, + hint: "", } } } -impl ResponseChoice { - fn resolve(self, mapping: &Mapping<'_>) -> Option { - match self { - Self::Omitted => None, - Self::Provider => mapping.original.response.clone().map(ResponseArg::Upstream), - Self::Stub { status, url } => Some(ResponseArg::Stub(HttpStub { - status, - method: "POST", - url, - content: None, - })), - Self::InternalServerStub => Some(ResponseArg::Stub(HttpStub { - status: 500, - method: "completion", - url: GITHUB_URL, - content: Some(mapping.original.message.clone()), - })), - } - } +/// The first rule that applies decides the class, as the `if`/`elif` chain does in Python. +pub(super) fn first_match<'r>(rules: &'r [Rule], mapping: &Mapping) -> Option<&'r Rule> { + rules.iter().find(|rule| (rule.when)(mapping)) } pub(super) fn contains_any(text: &str, markers: &[&str]) -> bool { @@ -117,7 +40,7 @@ static RATE_LIMIT_PHRASE: LazyLock = /// `ExceptionCheckers.is_error_str_rate_limit`. pub(super) fn is_rate_limit(error_str: &str, status: Option) -> bool { - if STANDALONE_429.is_match(error_str).unwrap_or(false) && status == Some(429) { + if STANDALONE_429.is_match(error_str).unwrap_or(false) && matches!(status, None | Some(429)) { return true; } let lower = error_str.to_lowercase(); @@ -169,184 +92,34 @@ pub(super) fn body_error_code(error_str: &str) -> Option { #[cfg(test)] mod tests { - use super::super::testing::{context, failure, http}; - use super::super::{ExceptionFamily, OriginalException, UpstreamResponse}; + use super::super::testing::mapping; use super::*; - fn first_marker(mapping: &Mapping<'_>) -> bool { - mapping.error_str.contains("first") - } - - fn always(_: &Mapping<'_>) -> bool { - true - } - - fn text(mapping: &Mapping<'_>) -> String { - format!("seen {}", mapping.error_str) - } - const ORDERED: &[Rule] = &[ - Rule { - when: first_marker, - kind: Kind::Status { - class: StatusClass::NotFound, - response: ResponseChoice::Omitted, - }, - message: text, - debug: false, - }, - Rule { - when: always, - kind: Kind::ApiConnection, - message: text, - debug: true, - }, + Rule::new( + |mapping| mapping.error_str.contains("first"), + PublicError::NotFound, + ), + Rule::new(|_| true, PublicError::ApiConnection), ]; - fn apply_one(kind: Kind, debug: bool, original: &OriginalException) -> Option { - let context = context("mistral", ExceptionFamily::OpenAiCompatible); - let mapping = Mapping::new(&context, original); - apply( - &[Rule { - when: always, - kind, - message: text, - debug, - }], - &mapping, - ) - } - #[rstest::rstest] - #[case::earlier_rule_wins("first and second", failure( - PublicKind::Status { status_class: StatusClass::NotFound, response: None }, - "seen first and second", - "mistral", - ))] - #[case::later_rule_when_the_earlier_does_not_apply("second", PublicFailure { - litellm_debug_info: Some("\nModel: ocr-model".into()), - ..failure(PublicKind::ApiConnection, "seen second", "mistral") - })] - fn the_first_applicable_rule_decides(#[case] body: &str, #[case] expected: PublicFailure) { - let context = context("mistral", ExceptionFamily::OpenAiCompatible); - let original = http(400, body); - assert_eq!( - apply(ORDERED, &Mapping::new(&context, &original)), - Some(expected) - ); + #[case::earlier_rule_wins("first and second", PublicError::NotFound)] + #[case::later_rule_when_the_earlier_does_not_apply("second", PublicError::ApiConnection)] + fn the_first_applicable_rule_decides(#[case] text: &str, #[case] expected: PublicError) { + let rule = first_match(ORDERED, &mapping(Some(400), text)); + assert_eq!(rule.map(|rule| rule.error), Some(expected)); } #[test] fn no_applicable_rule_leaves_the_failure_to_the_caller() { - let context = context("mistral", ExceptionFamily::OpenAiCompatible); - let original = http(400, "second"); - assert_eq!( - apply(&ORDERED[..1], &Mapping::new(&context, &original)), - None - ); - } - - #[rstest::rstest] - #[case::omitted(ResponseChoice::Omitted, None)] - #[case::provider(ResponseChoice::Provider, Some(ResponseArg::Upstream(UpstreamResponse { - status: 400, - body: "body".into(), - headers: vec![("retry-after".into(), "7".into())], - })))] - #[case::stub( - ResponseChoice::Stub { status: 429, url: "https://stub.test" }, - Some(ResponseArg::Stub(HttpStub { status: 429, method: "POST", url: "https://stub.test", content: None })) - )] - #[case::internal_server_stub( - ResponseChoice::InternalServerStub, - Some(ResponseArg::Stub(HttpStub { - status: 500, - method: "completion", - url: GITHUB_URL, - content: Some("body".into()), - })) - )] - fn response_choices_resolve_against_the_original( - #[case] response: ResponseChoice, - #[case] expected: Option, - ) { - let built = apply_one( - Kind::Status { - class: StatusClass::BadRequest, - response, - }, - false, - &http(400, "body"), - ) - .unwrap(); - assert_eq!( - built.kind, - PublicKind::Status { - status_class: StatusClass::BadRequest, - response: expected, - } - ); - } - - #[rstest::rstest] - #[case::fixed(ApiStatus::Fixed(500), http(409, "body"), 500)] - #[case::original(ApiStatus::Original, http(409, "body"), 409)] - #[case::original_without_a_status( - ApiStatus::Original, - OriginalException::Response { message: "body".into() }, - 500 - )] - fn api_status_is_fixed_or_the_originals( - #[case] status: ApiStatus, - #[case] original: OriginalException, - #[case] expected: u16, - ) { - let built = apply_one( - Kind::Api { - status, - request_url: "https://api.test", - }, - false, - &original, - ) - .unwrap(); - assert_eq!( - built, - failure( - PublicKind::Api { - status: expected, - request_url: "https://api.test" - }, - "seen body", - "mistral" - ) - ); - } - - #[rstest::rstest] - #[case::with_debug(true, Some("\nModel: ocr-model"))] - #[case::without_debug(false, None)] - fn debug_rules_carry_the_extra_information( - #[case] debug: bool, - #[case] expected: Option<&str>, - ) { - let built = apply_one(Kind::Timeout(Some(504)), debug, &http(504, "body")).unwrap(); - assert_eq!( - built, - PublicFailure { - litellm_debug_info: expected.map(str::to_string), - ..failure( - PublicKind::Timeout { status: Some(504) }, - "seen body", - "mistral" - ) - } - ); + assert!(first_match(&ORDERED[..1], &mapping(Some(400), "second")).is_none()); } #[rstest::rstest] #[case::standalone_429_with_429_status("got 429 back", Some(429), true)] #[case::standalone_429_with_other_status("got 429 back", Some(400), false)] + #[case::standalone_429_with_unknown_status("got 429 back", None, true)] #[case::embedded_429("token4290", Some(429), false)] #[case::phrase_spaced("Rate Limit reached", None, true)] #[case::phrase_underscored("rate_limit", None, true)] diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs index 3d817fe9ae2..cb8924d6c2a 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/status.rs @@ -1,153 +1,49 @@ -use super::public::{PublicFailure, StatusClass}; -use super::rules::{ApiStatus, Kind, ResponseChoice, Rule, apply}; -use super::{DOCS_URL, Mapping}; +use super::public::PublicError; -const fn with_response(class: StatusClass) -> Kind { - Kind::Status { - class, - response: ResponseChoice::Provider, - } -} - -fn message(mapping: &Mapping<'_>) -> String { - format!("{} - {}", mapping.exception_provider, mapping.error_str) -} - -fn status(mapping: &Mapping<'_>) -> u16 { - mapping.original.status.unwrap_or_default() -} - -/// `_map_exception_by_status`, the fallback for a provider error no provider mapper claimed. -const RULES: &[Rule] = &[ - Rule { - when: |mapping| status(mapping) == 401, - kind: with_response(StatusClass::Authentication), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 403, - kind: with_response(StatusClass::PermissionDenied), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 404, - kind: with_response(StatusClass::NotFound), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 408, - kind: Kind::Timeout(None), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 429, - kind: with_response(StatusClass::RateLimit), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 500, - kind: with_response(StatusClass::InternalServer), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 502, - kind: with_response(StatusClass::BadGateway), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 503, - kind: with_response(StatusClass::ServiceUnavailable), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) == 504, - kind: Kind::Timeout(Some(504)), - message, - debug: true, - }, - Rule { - when: |mapping| status(mapping) < 500, - kind: with_response(StatusClass::BadRequest), - message, - debug: true, - }, - Rule { - when: |_| true, - kind: Kind::Api { - status: ApiStatus::Original, - request_url: DOCS_URL, - }, - message, - debug: true, - }, -]; - -/// Only a real provider status of 400 or more reaches the table; a status the HTTP handler -/// synthesized for a failure without a response does not. -pub(super) fn map(mapping: &Mapping<'_>) -> Option { - let status = mapping.original.status?; - if status < 400 || mapping.original.status_is_synthesized { - return None; - } - apply(RULES, mapping) +/// `_map_exception_by_status`, the one place a provider status picks a class. Statuses +/// below 400 are not failures the table claims. +pub(super) fn classify(status: u16) -> Option { + let error = match status { + ..400 => return None, + 401 => PublicError::Authentication, + 403 => PublicError::PermissionDenied, + 404 => PublicError::NotFound, + 408 | 504 => PublicError::Timeout { status }, + 429 => PublicError::RateLimit, + 500 => PublicError::InternalServer, + 502 => PublicError::BadGateway, + 503 => PublicError::ServiceUnavailable, + 400..500 => PublicError::BadRequest, + _ => PublicError::Api { status }, + }; + Some(error) } #[cfg(test)] mod tests { - use super::super::testing::{context, failure, http, upstream, with_debug}; - use super::super::{ExceptionFamily, OriginalException, PublicKind}; use super::*; - fn mapped(original: &OriginalException) -> Option { - let context = context("reducto", ExceptionFamily::Other); - map(&Mapping::new(&context, original)) - } - - fn classified(class: StatusClass, status_code: u16) -> PublicKind { - PublicKind::Status { - status_class: class, - response: upstream(status_code, "rejected"), - } - } - #[rstest::rstest] - #[case::authentication(401, classified(StatusClass::Authentication, 401))] - #[case::permission_denied(403, classified(StatusClass::PermissionDenied, 403))] - #[case::not_found(404, classified(StatusClass::NotFound, 404))] - #[case::request_timeout(408, PublicKind::Timeout { status: None })] - #[case::rate_limited(429, classified(StatusClass::RateLimit, 429))] - #[case::internal_server(500, classified(StatusClass::InternalServer, 500))] - #[case::bad_gateway(502, classified(StatusClass::BadGateway, 502))] - #[case::service_unavailable(503, classified(StatusClass::ServiceUnavailable, 503))] - #[case::gateway_timeout(504, PublicKind::Timeout { status: Some(504) })] - #[case::lowest_client_error(400, classified(StatusClass::BadRequest, 400))] - #[case::other_client_error(409, classified(StatusClass::BadRequest, 409))] - #[case::highest_client_error(499, classified(StatusClass::BadRequest, 499))] - #[case::other_server_error(501, PublicKind::Api { status: 501, request_url: DOCS_URL })] - fn every_mapped_status_and_the_fallback(#[case] status_code: u16, #[case] kind: PublicKind) { - assert_eq!( - mapped(&http(status_code, "rejected")), - Some(with_debug(failure( - kind, - "ReductoException - rejected", - "reducto" - ))) - ); - } - - #[rstest::rstest] - #[case::below_client_errors(http(399, "rejected"))] - #[case::synthesized(OriginalException::Connection { message: "refused".into() })] - #[case::no_status(OriginalException::Response { message: "bad body".into() })] - fn failures_the_table_does_not_claim(#[case] original: OriginalException) { - assert_eq!(mapped(&original), None); + #[case::below_client_errors(399, None)] + #[case::lowest_client_error(400, Some(PublicError::BadRequest))] + #[case::authentication(401, Some(PublicError::Authentication))] + #[case::permission_denied(403, Some(PublicError::PermissionDenied))] + #[case::not_found(404, Some(PublicError::NotFound))] + #[case::request_timeout(408, Some(PublicError::Timeout { status: 408 }))] + #[case::other_client_error(409, Some(PublicError::BadRequest))] + #[case::unprocessable(422, Some(PublicError::BadRequest))] + #[case::rate_limited(429, Some(PublicError::RateLimit))] + #[case::highest_client_error(499, Some(PublicError::BadRequest))] + #[case::internal_server(500, Some(PublicError::InternalServer))] + #[case::other_server_error(501, Some(PublicError::Api { status: 501 }))] + #[case::bad_gateway(502, Some(PublicError::BadGateway))] + #[case::service_unavailable(503, Some(PublicError::ServiceUnavailable))] + #[case::gateway_timeout(504, Some(PublicError::Timeout { status: 504 }))] + #[case::highest_server_error(599, Some(PublicError::Api { status: 599 }))] + fn every_mapped_status_and_the_fallback( + #[case] status: u16, + #[case] expected: Option, + ) { + assert_eq!(classify(status), expected); } } diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs index 0fa8c19c4f6..dab1adb2329 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/vertex_ai.rs @@ -1,60 +1,17 @@ -use super::public::{PublicFailure, PublicKind, ResponseArg, StatusClass, UpstreamResponse}; -use super::rules::{ - Kind, ResponseChoice, Rule, apply, body_error_code, contains_any, is_context_window_exceeded, -}; -use super::{Mapping, python_capitalize}; - -const VERTEX_URL: &str = "https://cloud.google.com/vertex-ai/"; -const VERTEX_URL_WITH_SPACE: &str = " https://cloud.google.com/vertex-ai/"; +use super::public::PublicError; +use super::rules::{Rule, body_error_code, contains_any, is_context_window_exceeded}; const QUOTA_MARKERS: &[&str] = &[ "429 Quota exceeded", "Quota exceeded for", "Resource exhausted", - "IndexError: list index out of range", "429 Unable to submit request because the service is temporarily out of capacity.", ]; -const fn stubbed(class: StatusClass, status: u16, url: &'static str) -> Kind { - Kind::Status { - class, - response: ResponseChoice::Stub { status, url }, - } -} - -const fn bare(class: StatusClass) -> Kind { - Kind::Status { - class, - response: ResponseChoice::Omitted, - } -} - -/// `{Provider}Exception{label} - {error_str}` with Python's `str.capitalize()`. -fn capitalized(mapping: &Mapping<'_>, label: &str) -> String { - format!( - "{}Exception{label} - {}", - python_capitalize(mapping.provider), - mapping.error_str - ) -} - -/// `litellm.{Class}: {provider}Exception - {error_str}` with the provider as given. -fn litellm_prefixed(mapping: &Mapping<'_>, class: &str) -> String { - format!( - "litellm.{class}: {}Exception - {}", - mapping.provider, mapping.error_str - ) -} - -fn status_is(mapping: &Mapping<'_>, status: u16) -> bool { - mapping.original.status == Some(status) -} - -/// `_map_vertex_exception`, in its branch order. A failure no rule claims falls through -/// to the status table. -const RULES: &[Rule] = &[ - Rule { - when: |mapping| { +/// The text branches of `_map_vertex_exception`, in its order. +pub(super) const RULES: &[Rule] = &[ + Rule::new( + |mapping| { contains_any( &mapping.error_str, &[ @@ -63,54 +20,32 @@ const RULES: &[Rule] = &[ ], ) }, - kind: stubbed(StatusClass::BadRequest, 400, VERTEX_URL_WITH_SPACE), - message: |mapping| litellm_prefixed(mapping, "BadRequestError"), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::BadRequest, + ), + Rule::new( + |mapping| { mapping .error_str .contains("400 Request payload size exceeds") + || is_context_window_exceeded(&mapping.error_str) }, - kind: bare(StatusClass::ContextWindowExceeded), - message: |mapping| capitalized(mapping, ""), - debug: false, - }, - Rule { - when: |mapping| is_context_window_exceeded(&mapping.error_str), - kind: bare(StatusClass::ContextWindowExceeded), - message: |mapping| format!("ContextWindowExceededError: {}", capitalized(mapping, "")), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::ContextWindowExceeded, + ), + Rule::new( + |mapping| { contains_any( &mapping.error_str, &["None Unknown Error.", "Content has no parts."], ) }, - kind: Kind::Status { - class: StatusClass::InternalServer, - response: ResponseChoice::InternalServerStub, - }, - message: |mapping| litellm_prefixed(mapping, "InternalServerError"), - debug: true, - }, - Rule { - when: |mapping| mapping.error_str.contains("API key not valid."), - kind: bare(StatusClass::Authentication), - message: |mapping| capitalized(mapping, ""), - debug: true, - }, - Rule { - when: |mapping| mapping.error_str.contains("403"), - kind: stubbed(StatusClass::BadRequest, 403, VERTEX_URL_WITH_SPACE), - message: |mapping| capitalized(mapping, " BadRequestError"), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::InternalServer, + ), + Rule::new( + |mapping| mapping.error_str.contains("API key not valid."), + PublicError::Authentication, + ), + Rule::new( + |mapping| { contains_any( &mapping.error_str, &[ @@ -119,456 +54,124 @@ const RULES: &[Rule] = &[ ], ) }, - kind: stubbed( - StatusClass::ContentPolicyViolation, - 400, - VERTEX_URL_WITH_SPACE, - ), - message: |mapping| capitalized(mapping, " ContentPolicyViolationError"), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::ContentPolicyViolation, + ), + Rule::new( + |mapping| { contains_any(&mapping.error_str, QUOTA_MARKERS) || (mapping - .original .status .is_some_and(|status| (500..600).contains(&status)) && body_error_code(&mapping.error_str) == Some(429)) }, - kind: stubbed(StatusClass::RateLimit, 429, VERTEX_URL_WITH_SPACE), - message: |mapping| litellm_prefixed(mapping, "RateLimitError"), - debug: true, - }, - Rule { - when: |mapping| { + PublicError::RateLimit, + ), + Rule::new( + |mapping| { contains_any( &mapping.error_str, &["500 Internal Server Error", "The model is overloaded."], ) }, - kind: bare(StatusClass::InternalServer), - message: |mapping| litellm_prefixed(mapping, "InternalServerError"), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, 400), - kind: stubbed(StatusClass::BadRequest, 400, VERTEX_URL), - message: |mapping| capitalized(mapping, " BadRequestError"), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, 401), - kind: bare(StatusClass::Authentication), - message: |mapping| capitalized(mapping, ""), - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, 403), - kind: stubbed(StatusClass::PermissionDenied, 403, VERTEX_URL), - message: |mapping| capitalized(mapping, ""), - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, 404), - kind: bare(StatusClass::NotFound), - message: |mapping| capitalized(mapping, ""), - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, 408), - kind: Kind::Timeout(None), - message: |mapping| capitalized(mapping, ""), - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, 429), - kind: stubbed(StatusClass::RateLimit, 429, VERTEX_URL_WITH_SPACE), - message: |mapping| format!("litellm.RateLimitError: {}", capitalized(mapping, "")), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, 500), - kind: Kind::Status { - class: StatusClass::InternalServer, - response: ResponseChoice::InternalServerStub, - }, - message: |mapping| capitalized(mapping, " InternalServerError"), - debug: true, - }, - Rule { - when: |mapping| status_is(mapping, 502), - kind: Kind::ApiConnection, - message: |mapping| capitalized(mapping, ""), - debug: false, - }, - Rule { - when: |mapping| status_is(mapping, 503), - kind: bare(StatusClass::ServiceUnavailable), - message: |mapping| capitalized(mapping, ""), - debug: false, - }, + PublicError::InternalServer, + ), ]; -pub(super) fn map(mapping: &Mapping<'_>) -> Option { - apply(RULES, mapping).map(|failure| keep_upstream_response(mapping, failure)) -} - -/// Deliberate divergence from `_map_vertex_exception`, which replaces the provider response -/// with a stub and so drops the upstream body and `retry-after`. The response keeps the -/// status the public class carries. -fn keep_upstream_response(mapping: &Mapping<'_>, failure: PublicFailure) -> PublicFailure { - let (PublicKind::Status { status_class, .. }, Some(upstream), false) = ( - &failure.kind, - &mapping.original.response, - mapping.original.status_is_synthesized, - ) else { - return failure; - }; - PublicFailure { - kind: PublicKind::Status { - status_class: *status_class, - response: Some(ResponseArg::Upstream(UpstreamResponse { - status: status_class.status_code(), - ..upstream.clone() - })), - }, - ..failure - } -} - #[cfg(test)] mod tests { - use super::super::testing::{context, failure, http, status, upstream, with_debug}; - use super::super::{ExceptionFamily, HttpStub, OriginalException}; + use super::super::rules::first_match; + use super::super::testing::mapping; use super::*; - fn mapped(original: &OriginalException) -> Option { - let context = context("vertex_ai", ExceptionFamily::VertexAi); - map(&Mapping::new(&context, original)) - } - - fn kept(class: StatusClass, body: &str) -> PublicKind { - status(class, upstream(class.status_code(), body)) + fn classified(status: Option, text: &str) -> Option { + first_match(RULES, &mapping(status, text)).map(|rule| rule.error) } #[rstest::rstest] #[case::api_not_enabled( - 400, "Vertex AI API has not been used in project x", - with_debug(failure( - kept( - StatusClass::BadRequest, - "Vertex AI API has not been used in project x" - ), - "litellm.BadRequestError: vertex_aiException - Vertex AI API has not been used in project x", - "vertex_ai", - )) - )] - #[case::project_not_found( - 400, - "Unable to find your project", - with_debug(failure( - kept(StatusClass::BadRequest, "Unable to find your project"), - "litellm.BadRequestError: vertex_aiException - Unable to find your project", - "vertex_ai", - )) + PublicError::BadRequest )] + #[case::project_not_found("Unable to find your project", PublicError::BadRequest)] #[case::payload_too_large( - 400, "400 Request payload size exceeds the limit", - failure( - kept( - StatusClass::ContextWindowExceeded, - "400 Request payload size exceeds the limit" - ), - "Vertex_aiException - 400 Request payload size exceeds the limit", - "vertex_ai", - ) + PublicError::ContextWindowExceeded )] #[case::context_window( - 500, "This model's maximum context length is 10", - with_debug(failure( - kept( - StatusClass::ContextWindowExceeded, - "This model's maximum context length is 10" - ), - "ContextWindowExceededError: Vertex_aiException - This model's maximum context length is 10", - "vertex_ai", - )) - )] - #[case::unknown_error( - 400, - "None Unknown Error.", - with_debug(failure( - kept(StatusClass::InternalServer, "None Unknown Error."), - "litellm.InternalServerError: vertex_aiException - None Unknown Error.", - "vertex_ai", - )) - )] - #[case::no_parts( - 400, - "Content has no parts.", - with_debug(failure( - kept(StatusClass::InternalServer, "Content has no parts."), - "litellm.InternalServerError: vertex_aiException - Content has no parts.", - "vertex_ai", - )) - )] - #[case::api_key_not_valid( - 400, - "API key not valid.", - with_debug(failure( - kept(StatusClass::Authentication, "API key not valid."), - "Vertex_aiException - API key not valid.", - "vertex_ai", - )) - )] - #[case::forbidden_text( - 400, - "got a 403", - with_debug(failure( - kept(StatusClass::BadRequest, "got a 403"), - "Vertex_aiException BadRequestError - got a 403", - "vertex_ai", - )) - )] - #[case::response_blocked( - 400, - "The response was blocked.", - with_debug(failure( - kept(StatusClass::ContentPolicyViolation, "The response was blocked."), - "Vertex_aiException ContentPolicyViolationError - The response was blocked.", - "vertex_ai", - )) + PublicError::ContextWindowExceeded )] + #[case::unknown_error("None Unknown Error.", PublicError::InternalServer)] + #[case::no_parts("Content has no parts.", PublicError::InternalServer)] + #[case::api_key_not_valid("API key not valid.", PublicError::Authentication)] + #[case::response_blocked("The response was blocked.", PublicError::ContentPolicyViolation)] #[case::output_blocked( - 400, "Output blocked by content filtering policy", - with_debug(failure( - kept( - StatusClass::ContentPolicyViolation, - "Output blocked by content filtering policy" - ), - "Vertex_aiException ContentPolicyViolationError - Output blocked by content filtering policy", - "vertex_ai", - )) + PublicError::ContentPolicyViolation )] - #[case::quota_marker( - 400, - "Quota exceeded for aiplatform", - with_debug(failure( - kept(StatusClass::RateLimit, "Quota exceeded for aiplatform"), - "litellm.RateLimitError: vertex_aiException - Quota exceeded for aiplatform", - "vertex_ai", - )) + #[case::quota_exceeded_429("429 Quota exceeded", PublicError::RateLimit)] + #[case::quota_exceeded_for("Quota exceeded for aiplatform", PublicError::RateLimit)] + #[case::resource_exhausted("Resource exhausted", PublicError::RateLimit)] + #[case::out_of_capacity( + "429 Unable to submit request because the service is temporarily out of capacity.", + PublicError::RateLimit )] - #[case::wrapped_429( - 503, - r#"{"error": {"code": "429"}}"#, - with_debug(failure( - kept(StatusClass::RateLimit, r#"{"error": {"code": "429"}}"#), - r#"litellm.RateLimitError: vertex_aiException - {"error": {"code": "429"}}"#, - "vertex_ai", - )) - )] - #[case::overloaded( - 400, - "The model is overloaded.", - with_debug(failure( - kept(StatusClass::InternalServer, "The model is overloaded."), - "litellm.InternalServerError: vertex_aiException - The model is overloaded.", - "vertex_ai", - )) - )] - #[case::internal_server_text( - 400, - "500 Internal Server Error", - with_debug(failure( - kept(StatusClass::InternalServer, "500 Internal Server Error"), - "litellm.InternalServerError: vertex_aiException - 500 Internal Server Error", - "vertex_ai", - )) - )] - fn each_text_rule_maps_by_the_body( - #[case] status_code: u16, - #[case] body: &str, - #[case] expected: PublicFailure, - ) { - assert_eq!(mapped(&http(status_code, body)), Some(expected)); + #[case::internal_server_text("500 Internal Server Error", PublicError::InternalServer)] + #[case::overloaded("The model is overloaded.", PublicError::InternalServer)] + fn each_text_rule_claims_its_marker(#[case] text: &str, #[case] expected: PublicError) { + assert_eq!(classified(Some(400), text), Some(expected)); } #[rstest::rstest] - #[case::bad_request( - 400, - with_debug(failure( - kept(StatusClass::BadRequest, "rejected"), - "Vertex_aiException BadRequestError - rejected", - "vertex_ai" - )) - )] - #[case::authentication( - 401, - failure( - kept(StatusClass::Authentication, "rejected"), - "Vertex_aiException - rejected", - "vertex_ai" - ) - )] - #[case::permission_denied( - 403, - failure( - kept(StatusClass::PermissionDenied, "rejected"), - "Vertex_aiException - rejected", - "vertex_ai" - ) - )] - #[case::not_found( - 404, - failure( - kept(StatusClass::NotFound, "rejected"), - "Vertex_aiException - rejected", - "vertex_ai" - ) - )] - #[case::request_timeout(408, failure(PublicKind::Timeout { status: None }, "Vertex_aiException - rejected", "vertex_ai"))] - #[case::rate_limited( - 429, - with_debug(failure( - kept(StatusClass::RateLimit, "rejected"), - "litellm.RateLimitError: Vertex_aiException - rejected", - "vertex_ai" - )) - )] - #[case::internal_server( - 500, - with_debug(failure( - kept(StatusClass::InternalServer, "rejected"), - "Vertex_aiException InternalServerError - rejected", - "vertex_ai" - )) - )] - #[case::bad_gateway( - 502, - failure( - PublicKind::ApiConnection, - "Vertex_aiException - rejected", - "vertex_ai" - ) - )] - #[case::service_unavailable( - 503, - failure( - kept(StatusClass::ServiceUnavailable, "rejected"), - "Vertex_aiException - rejected", - "vertex_ai" - ) - )] - fn each_status_rule_maps_by_the_status( - #[case] status_code: u16, - #[case] expected: PublicFailure, + #[case::server_error_wrapping_a_429(Some(503), Some(PublicError::RateLimit))] + #[case::lowest_server_error(Some(500), Some(PublicError::RateLimit))] + #[case::highest_server_error(Some(599), Some(PublicError::RateLimit))] + #[case::client_error(Some(400), None)] + #[case::no_status(None, None)] + fn a_wrapped_429_is_a_rate_limit_only_behind_a_server_error( + #[case] status: Option, + #[case] expected: Option, ) { - assert_eq!(mapped(&http(status_code, "rejected")), Some(expected)); - } - - #[rstest::rstest] - #[case::unmapped_status(409)] - #[case::gateway_timeout(504)] - fn statuses_without_a_rule_fall_through(#[case] status_code: u16) { - assert_eq!(mapped(&http(status_code, "rejected")), None); - } - - #[rstest::rstest] - #[case::stub_without_an_upstream_response( - OriginalException::Response { message: "got a 403".into() }, - status(StatusClass::BadRequest, Some(ResponseArg::Stub(HttpStub { status: 403, method: "POST", url: VERTEX_URL_WITH_SPACE, content: None }))) - )] - #[case::stub_for_a_synthesized_status( - OriginalException::Connection { message: "got a 403".into() }, - status(StatusClass::BadRequest, Some(ResponseArg::Stub(HttpStub { status: 403, method: "POST", url: VERTEX_URL_WITH_SPACE, content: None }))) - )] - fn the_rule_response_stays_when_there_is_no_real_upstream_response( - #[case] original: OriginalException, - #[case] kind: PublicKind, - ) { - assert_eq!(mapped(&original).map(|failure| failure.kind), Some(kind)); - } - - #[test] - fn a_synthesized_500_keeps_the_internal_server_stub() { - let original = OriginalException::Connection { - message: "refused".into(), - }; assert_eq!( - mapped(&original), - Some(with_debug(failure( - status( - StatusClass::InternalServer, - Some(ResponseArg::Stub(HttpStub { - status: 500, - method: "completion", - url: "https://github.com/BerriAI/litellm", - content: Some("refused".into()), - })) - ), - "Vertex_aiException InternalServerError - refused", - "vertex_ai" - ))) + classified(status, r#"{"error": {"code": "429"}}"#), + expected ); } #[rstest::rstest] #[case::project_before_payload_size( "Unable to find your project 400 Request payload size exceeds", - StatusClass::BadRequest, - "litellm.BadRequestError: vertex_aiException - Unable to find your project 400 Request payload size exceeds", - true + PublicError::BadRequest )] - #[case::payload_size_before_context_window( - "400 Request payload size exceeds; This model's maximum context length is 10", - StatusClass::ContextWindowExceeded, - "Vertex_aiException - 400 Request payload size exceeds; This model's maximum context length is 10", - false + #[case::context_window_before_unknown_error( + "This model's maximum context length is 10 None Unknown Error.", + PublicError::ContextWindowExceeded )] - #[case::api_key_before_forbidden( - "API key not valid. 403", - StatusClass::Authentication, - "Vertex_aiException - API key not valid. 403", - true + #[case::unknown_error_before_api_key( + "Content has no parts. API key not valid.", + PublicError::InternalServer )] - #[case::forbidden_before_blocked( - "403 The response was blocked.", - StatusClass::BadRequest, - "Vertex_aiException BadRequestError - 403 The response was blocked.", - true + #[case::api_key_before_blocked( + "API key not valid. The response was blocked.", + PublicError::Authentication )] #[case::blocked_before_quota( "The response was blocked. Resource exhausted", - StatusClass::ContentPolicyViolation, - "Vertex_aiException ContentPolicyViolationError - The response was blocked. Resource exhausted", - true + PublicError::ContentPolicyViolation )] #[case::quota_before_overloaded( "Resource exhausted The model is overloaded.", - StatusClass::RateLimit, - "litellm.RateLimitError: vertex_aiException - Resource exhausted The model is overloaded.", - true + PublicError::RateLimit )] - fn the_earlier_rule_wins_when_two_apply( - #[case] body: &str, - #[case] class: StatusClass, - #[case] message: &str, - #[case] debug: bool, - ) { - let expected = failure(kept(class, body), message, "vertex_ai"); - assert_eq!( - mapped(&http(401, body)), - Some(if debug { - with_debug(expected) - } else { - expected - }) - ); + fn the_earlier_rule_wins_when_two_apply(#[case] text: &str, #[case] expected: PublicError) { + assert_eq!(classified(Some(400), text), Some(expected)); + } + + #[rstest::rstest] + #[case::a_403_in_the_text("got a 403 from 4031 tokens")] + #[case::python_client_crash("IndexError: list index out of range")] + #[case::unmarked("rejected")] + fn text_without_a_marker_is_left_to_the_status_table(#[case] text: &str) { + assert_eq!(classified(Some(400), text), None); } } diff --git a/litellm-rust/crates/core-utils/src/lib.rs b/litellm-rust/crates/core-utils/src/lib.rs index 0c26aa50cc3..fcb232d8980 100644 --- a/litellm-rust/crates/core-utils/src/lib.rs +++ b/litellm-rust/crates/core-utils/src/lib.rs @@ -4,7 +4,6 @@ pub mod exception_mapping_utils; pub mod get_llm_provider_logic; pub mod params; pub mod prompt_templates; -pub mod python_repr; pub mod secret_redaction; pub mod serde_compat; pub mod url_utils; diff --git a/litellm-rust/crates/core-utils/src/python_repr.rs b/litellm-rust/crates/core-utils/src/python_repr.rs deleted file mode 100644 index 7ccfb826377..00000000000 --- a/litellm-rust/crates/core-utils/src/python_repr.rs +++ /dev/null @@ -1,93 +0,0 @@ -/// `repr()` of a Python `str`: single quotes unless the text holds a single quote and no -/// double quote, with backslashes, the chosen quote and control characters escaped. -pub fn python_str_repr(value: &str) -> String { - let quote = if value.contains('\'') && !value.contains('"') { - '"' - } else { - '\'' - }; - let escaped: String = value - .chars() - .map(|character| match character { - '\\' => "\\\\".to_string(), - '\t' => "\\t".to_string(), - '\n' => "\\n".to_string(), - '\r' => "\\r".to_string(), - character if character == quote => format!("\\{character}"), - character - if (character as u32) < 0x20 || (0x7f..0xa0).contains(&(character as u32)) => - { - format!("\\x{:02x}", character as u32) - } - character => character.to_string(), - }) - .collect(); - format!("{quote}{escaped}{quote}") -} - -/// `repr()` of the Python value a JSON value decodes to. -pub fn python_value_repr(value: &serde_json::Value) -> String { - use serde_json::Value; - match value { - Value::Null => "None".to_string(), - Value::Bool(true) => "True".to_string(), - Value::Bool(false) => "False".to_string(), - Value::Number(number) => number.to_string(), - Value::String(text) => python_str_repr(text), - Value::Array(items) => format!( - "[{}]", - items - .iter() - .map(python_value_repr) - .collect::>() - .join(", ") - ), - Value::Object(fields) => format!( - "{{{}}}", - fields - .iter() - .map(|(key, value)| format!( - "{}: {}", - python_str_repr(key), - python_value_repr(value) - )) - .collect::>() - .join(", ") - ), - } -} - -#[cfg(test)] -mod tests { - use serde_json::json; - - use super::{python_str_repr, python_value_repr}; - - #[rstest::rstest] - #[case::null(json!(null), "None")] - #[case::true_(json!(true), "True")] - #[case::false_(json!(false), "False")] - #[case::integer(json!(5), "5")] - #[case::float(json!(1.5), "1.5")] - #[case::string(json!("it's"), "\"it's\"")] - #[case::list(json!(["a", 1]), "['a', 1]")] - #[case::dict(json!({"format": "native"}), "{'format': 'native'}")] - #[case::empty_list(json!([]), "[]")] - fn value_repr_matches_python(#[case] value: serde_json::Value, #[case] expected: &str) { - assert_eq!(python_value_repr(&value), expected); - } - - #[rstest::rstest] - #[case::plain("native", "'native'")] - #[case::single_quote("it's", "\"it's\"")] - #[case::both_quotes("it's \"x\"", "'it\\'s \"x\"'")] - #[case::double_quote("say \"x\"", "'say \"x\"'")] - #[case::backslash("a\\b", "'a\\\\b'")] - #[case::whitespace("a\tb\nc\rd", "'a\\tb\\nc\\rd'")] - #[case::control("a\u{1}b\u{7f}c\u{85}", "'a\\x01b\\x7fc\\x85'")] - #[case::unicode("café", "'café'")] - #[case::empty("", "''")] - fn matches_python_repr(#[case] value: &str, #[case] expected: &str) { - assert_eq!(python_str_repr(value), expected); - } -} diff --git a/litellm-rust/crates/core-utils/src/secret_redaction.rs b/litellm-rust/crates/core-utils/src/secret_redaction.rs index 32922d7430c..e3caee4799a 100644 --- a/litellm-rust/crates/core-utils/src/secret_redaction.rs +++ b/litellm-rust/crates/core-utils/src/secret_redaction.rs @@ -1,5 +1,3 @@ -use std::sync::LazyLock; - use fancy_regex::Regex; pub const REDACTED: &str = "REDACTED"; @@ -51,21 +49,32 @@ fn secret_patterns(minimum_custom_key_length: usize) -> String { .join("|") } -static SECRET_RE: LazyLock = LazyLock::new(|| { - Regex::new(&format!( - "(?i){}", - secret_patterns(minimum_custom_key_length()) - )) - .expect("secret redaction patterns compile") -}); - -pub fn redact_string(value: &str) -> String { - SECRET_RE.replace_all(value, REDACTED).into_owned() +/// Python's `_ENABLE_SECRET_REDACTION` pattern set, compiled once per configuration. +#[derive(Clone, Debug)] +pub struct SecretRedactor { + pattern: Regex, } -pub fn secret_redaction_enabled() -> bool { - !std::env::var("LITELLM_DISABLE_REDACT_SECRETS") - .is_ok_and(|value| value.eq_ignore_ascii_case("true")) +impl SecretRedactor { + pub fn new(minimum_custom_key_length: usize) -> Self { + let pattern = Regex::new(&format!( + "(?i){}", + secret_patterns(minimum_custom_key_length) + )) + .expect("secret redaction patterns compile"); + Self { pattern } + } + + /// `None` when `LITELLM_DISABLE_REDACT_SECRETS` turns redaction off. + pub fn from_env() -> Option { + let disabled = std::env::var("LITELLM_DISABLE_REDACT_SECRETS") + .is_ok_and(|value| value.eq_ignore_ascii_case("true")); + (!disabled).then(|| Self::new(minimum_custom_key_length())) + } + + pub fn redact(&self, value: &str) -> String { + self.pattern.replace_all(value, REDACTED).into_owned() + } } #[cfg(test)] @@ -85,13 +94,16 @@ mod tests { #[case::password_needs_word_boundary("db_password=hunter2", "REDACTED")] #[case::plain_text_is_kept(r#"{"message": "rejected"}"#, r#"{"message": "rejected"}"#)] fn redacts_the_same_spans_as_the_python_patterns(#[case] input: &str, #[case] expected: &str) { - assert_eq!(redact_string(input), expected); + assert_eq!( + SecretRedactor::new(DEFAULT_MINIMUM_CUSTOM_KEY_LENGTH).redact(input), + expected + ); } #[test] fn sk_threshold_follows_the_minimum_custom_key_length() { - let patterns = Regex::new(&format!("(?i){}", secret_patterns(8))).unwrap(); - assert_eq!(patterns.replace_all("sk-abcde", REDACTED), REDACTED); - assert_eq!(patterns.replace_all("sk-abcd", REDACTED), "sk-abcd"); + let redactor = SecretRedactor::new(8); + assert_eq!(redactor.redact("sk-abcde"), REDACTED); + assert_eq!(redactor.redact("sk-abcd"), "sk-abcd"); } } diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/api.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/api.json deleted file mode 100644 index ae629114589..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/api.json +++ /dev/null @@ -1,13 +0,0 @@ -{ - "kind": { - "type": "api", - "status": 409, - "request_url": "https://docs.litellm.ai/docs" - }, - "message": "MistralException - api", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": null, - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json deleted file mode 100644 index ab400ed9e01..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/api_connection.json +++ /dev/null @@ -1,11 +0,0 @@ -{ - "kind": { - "type": "api_connection" - }, - "message": "MistralException - api_connection", - "model": "ocr-model", - "llm_provider": null, - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": null, - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json deleted file mode 100644 index 057392575a1..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_authentication.json +++ /dev/null @@ -1,13 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "authentication", - "response": null - }, - "message": "MistralException - status_authentication", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": null, - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json deleted file mode 100644 index abb1425f686..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_gateway.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "bad_gateway", - "response": { - "type": "upstream", - "status": 502, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_bad_gateway", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json deleted file mode 100644 index 171a994cd35..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_bad_request.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "bad_request", - "response": { - "type": "upstream", - "status": 400, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_bad_request", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json deleted file mode 100644 index ae9c1e145de..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_content_policy_violation.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "content_policy_violation", - "response": { - "type": "upstream", - "status": 400, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_content_policy_violation", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json deleted file mode 100644 index 61e1a56a622..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_context_window_exceeded.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "context_window_exceeded", - "response": { - "type": "upstream", - "status": 400, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_context_window_exceeded", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json deleted file mode 100644 index b3c5c51a785..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_internal_server.json +++ /dev/null @@ -1,19 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "internal_server", - "response": { - "type": "stub", - "status": 500, - "method": "completion", - "url": "https://github.com/BerriAI/litellm", - "content": "upstream text" - } - }, - "message": "MistralException - status_internal_server", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": null, - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json deleted file mode 100644 index 26ffd872961..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_not_found.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "not_found", - "response": { - "type": "upstream", - "status": 404, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_not_found", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json deleted file mode 100644 index 42772f98830..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_permission_denied.json +++ /dev/null @@ -1,19 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "permission_denied", - "response": { - "type": "stub", - "status": 403, - "method": "POST", - "url": " https://cloud.google.com/vertex-ai/", - "content": null - } - }, - "message": "MistralException - status_permission_denied", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": null, - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json deleted file mode 100644 index c9b88822ebd..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_rate_limit.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "rate_limit", - "response": { - "type": "upstream", - "status": 429, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_rate_limit", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json deleted file mode 100644 index 5af65202ef1..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_service_unavailable.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "service_unavailable", - "response": { - "type": "upstream", - "status": 503, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_service_unavailable", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json deleted file mode 100644 index a1b318ce03c..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/status_unsupported_params.json +++ /dev/null @@ -1,28 +0,0 @@ -{ - "kind": { - "type": "status", - "status_class": "unsupported_params", - "response": { - "type": "upstream", - "status": 400, - "body": "{\"message\": \"rejected\"}", - "headers": [ - [ - "retry-after", - "7" - ] - ] - } - }, - "message": "MistralException - status_unsupported_params", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": [ - [ - "retry-after", - "7" - ] - ], - "print_banner": false -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json deleted file mode 100644 index 8b215bdb7e6..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_with_status.json +++ /dev/null @@ -1,12 +0,0 @@ -{ - "kind": { - "type": "timeout", - "status": 504 - }, - "message": "MistralException - timeout_with_status", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": "\nModel: ocr-model", - "litellm_response_headers": null, - "print_banner": true -} diff --git a/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json b/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json deleted file mode 100644 index 4a79169776e..00000000000 --- a/tests/test_litellm/rust_bridge/fixtures/public_failures/timeout_without_status.json +++ /dev/null @@ -1,12 +0,0 @@ -{ - "kind": { - "type": "timeout", - "status": null - }, - "message": "MistralException - timeout_without_status", - "model": "ocr-model", - "llm_provider": "mistral", - "litellm_debug_info": null, - "litellm_response_headers": null, - "print_banner": false -} From 9d134413c9b6ad28b693d844a99ac1f167599f8d Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Fri, 18 Sep 2026 14:00:47 -0700 Subject: [PATCH 141/141] test(rust): rename the standalone 429 test to match the rule --- .../crates/core-utils/src/exception_mapping_utils/openai.rs | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs index b4f4496dbf0..d45078c415e 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs @@ -183,7 +183,7 @@ mod tests { } #[test] - fn a_standalone_429_counts_only_with_a_429_status() { + fn a_standalone_429_counts_with_a_429_status() { assert_eq!( classified(Some(429), "got 429 back"), Some(PublicError::RateLimit)