From 33457d3ac925c5e97c7a0c4765c2ca79940a899f Mon Sep 17 00:00:00 2001 From: zoveki <349028329+zoveki@users.noreply.github.com> Date: Mon, 2 Nov 2026 09:48:59 -0800 Subject: [PATCH] fix(proxy): check the model in every OpenAI websocket passthrough frame (#48979) * test(proxy): cover per-frame model checks and session pricing on the OpenAI websocket passthrough * fix(proxy): check the model in every OpenAI websocket passthrough frame and price the session * fix(proxy): bill each OpenAI websocket turn on its own row and run full auth per frame Each finished turn (completed, incomplete, or failed) now logs its own spend row under the model and access groups that allowed it, so group budgets see spend as turns end instead of at close. Every model a frame names goes through the key's full auth, and a stored prompt reference with no model is refused for a key held to some models. A frame still in flight when the client leaves is recorded before the relay gives up on the send * fix(proxy): keep the stored prompt refusal's first sentence inside the websocket close reason The close reason carries the refusal's first sentence, and the stored prompt sentence ran past the 123 bytes a close frame allows, so it was cut mid-word. The sentence is shorter now and the rest of the explanation stays in the error event. * fix(proxy): bill pipelined websocket turns to their own frame and close the review gaps * fix(proxy): log no extra websocket spend row for frames after the last turn * fix(proxy): price websocket Responses turns by served tier and hosted tools, and bill turns a client walked out of A websocket Responses turn is now priced through the same Responses cost path the HTTP passthrough uses, so a priority-tier answer and a hosted web search cost what they cost over HTTP. When the client leaves mid-answer, the gateway keeps reading the provider's socket until that turn ends, up to 10 minutes, so the turn the provider still bills reaches spend * fix(proxy): bill websocket turns by lane, skip turns OpenAI refuses, and stop frame checks counting the session's own reservation An error event refuses a queued response.create, so the next turn is no longer charged to the refused frame's model. Turns on separate stream_id lanes are charged to the frame that started each one. The frame model gate settles the connect reservation before it re-runs auth, so a key with less budget left than one turn's max cost is no longer refused on its first frame. * fix(proxy): bill each websocket lane cut off at session end on its own row When a client leaves while turns on several stream_id lanes are still running, each unfinished turn now gets its own spend row under the groups and model its own frame asked for, instead of one merged row. Also pins the PR's new single-assigned test variables as Final and swaps datetime.now() in the logging-handler and success-handler fixtures for fixed timestamps. * test(proxy): mark the remaining new websocket test variables Final Builds the logging mocks with model_call_details in the constructor instead of setting it afterwards. * fix(proxy): check every model the realtime URL names before the OpenAI websocket relay connects * fix(proxy): keep reading running answers after a refused websocket frame so they are billed * fix(proxy): bill realtime transcription sessions on the OpenAI websocket passthrough A transcription session names its model only in session.updated under audio.input.transcription, and reports usage on each completed transcription instead of response.done, so its row logged unknown at $0 and a budgeted key never spent. * test(proxy): mark the turn-taking client's delivered list Final * fix(proxy): hold a websocket session's parallel slot until the socket closes Each finished turn logged its spend row under the session's own call id, so the first turn's success callback freed the key's max_parallel_requests slot while the socket was still open, and a key limited to one session could open several at once. The session now claims its slot under its own call id, every turn row logs under a call id of its own, and the slot is released when the session ends, as before Priced turn rows also take their row's call id as their id, so two turns that start in the same instant no longer share a spend log id and drop one row * fix(proxy): bill realtime transcription clips that finish outside an answer A conversation.item.input_audio_transcription.completed frame that arrived after a session's last response.done was never billed, because the session end only owes a row for a turn still running. A clip that completes while no answer is running now logs its own row, and a clip inside a running answer stays in that answer's row Each row also carries the session's latest session.updated frame, since later rows lost the transcription model it names and priced clips against the realtime model. A newer session.updated replaces the older one, so the stale model is never read first It also drops the docstrings this PR had added, since the names already say what each helper does * fix(cost): price a realtime transcription clip at the transcription model the session set A realtime conversation session opens with session.created naming the realtime model and no transcription model, and the client sets the transcription model later in session.update. The cost code took the first session frame and fell back to its session model, so a clip was priced at the realtime model. That model has no per-second rate, so a gpt-transcribe clip cost $0, and a token-billed clip used the realtime model's audio rate instead of the transcription model's. Now the newest session frame that names a transcription model sets the price, and the first session model is only the fallback when no frame names one, as before * fix(cost): price each realtime transcription clip at the model set when it ran A session that changes its transcription model mid-way kept every logged clip on the newest model. Each clip now takes the transcription model from the last session event before it, and falls back to the session's own model * fix(passthrough): keep the session update a buffered clip ran under until its answer is billed A transcription that completes while an answer streams waits in the buffer for that answer's response.done. A session.updated in between used to drop the earlier update, so the clip was priced at the opening configuration. An earlier update now stays while an unbilled clip follows it, and drops at the next update once that clip is billed. * fix(passthrough): keep reading a websocket turn the client left before OpenAI started it A client that sends response.create and hangs up before OpenAI answers with response.created made the walkout checks see no turn in flight, so the gateway closed the provider socket and the turn OpenAI still ran and billed never reached spend. The relay now counts every forwarded create until the provider starts or refuses it, and keeps draining while one is owed * fix(passthrough): settle a websocket turn that ends without its start frame A turn the provider ended without sending response.created first left its create counted as not started, so a client leaving later held the provider socket for the whole walkout drain with nothing left to bill. A turn's end now settles one waiting create when its lane saw no start frame * fix(proxy): refuse unparseable and unroutable websocket frames and hold legacy slots per socket A client frame too deep or too long for the JSON parser closed the provider socket mid-turn, so the turn in flight was never billed. The gate now refuses that frame with an error event and keeps draining the open turn. A stream_id OpenAI rejects comes back as an error that names no lane, which dropped another turn's authorization. The gate now refuses those frames before they reach OpenAI. With the legacy parallel request limiter, each per-turn spend row freed the key's slot while the socket stayed open. Websocket passthrough rows now free nothing, and the socket frees its slot once when it closes. --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/cost_calculator.py | 76 +- litellm/litellm_core_utils/core_helpers.py | 6 +- litellm/proxy/auth/auth_checks.py | 73 - litellm/proxy/common_utils/realtime_utils.py | 3 +- .../proxy/hooks/parallel_request_limiter.py | 11 +- .../llm_passthrough_endpoints.py | 197 ++- ...i_websocket_passthrough_logging_handler.py | 262 +++ .../pass_through_endpoints.py | 734 ++++++-- .../pass_through_endpoints/success_handler.py | 70 +- .../spend_tracking/budget_reservation.py | 2 +- .../pass_through_endpoints.py | 14 +- ...st_auth_checks_object_access_and_lookup.py | 133 -- .../proxy/common_utils/test_realtime_cache.py | 9 +- ...i_websocket_passthrough_logging_handler.py | 520 ++++++ .../test_pass_through_endpoints.py | 1535 ++++++++++++++++- .../test_success_handler.py | 53 + .../test_openai_ws_passthrough_routes.py | 439 ++++- tests/unit/test_cost_calculator.py | 32 + 18 files changed, 3642 insertions(+), 527 deletions(-) create mode 100644 litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_websocket_passthrough_logging_handler.py create mode 100644 tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_openai_websocket_passthrough_logging_handler.py diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index e66be233be4..3c028278b2d 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -6,6 +6,7 @@ import time from collections.abc import Mapping, Sequence from datetime import datetime from functools import lru_cache +from itertools import accumulate from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Literal, TypeAlias, cast @@ -3850,44 +3851,65 @@ def handle_realtime_transcription_cost_calculation( - {"type": "duration", "seconds": } → priced via input_cost_per_second - {"type": "tokens", "input_tokens": ...} → priced via input/audio token cost """ - completed_events: Final = [ - cast(dict, result) for result in results if result.get("type") == _TRANSCRIPTION_COMPLETED_EVENT_TYPE - ] - if not completed_events: + configured_models: Final = tuple(accumulate(map(_session_transcription_model, results), _newer_model)) + clips: Final = tuple( + (as_str_mapping(result.get("usage")) or {}, configured) + for result, configured in zip(results, configured_models, strict=True) + if result.get("type") == _TRANSCRIPTION_COMPLETED_EVENT_TYPE + ) + if not clips: return 0.0 - model_name: Final = _get_transcription_model_name_from_results(results) or litellm_model_name - model_info: Final = _cost_map_model_info(model_name, custom_llm_provider) + fallback_model: Final = _get_transcription_model_name_from_results(results) or litellm_model_name + clip_models: Final = tuple((usage, configured or fallback_model) for usage, configured in clips) + model_infos: Final = MappingProxyType( + {model: _cost_map_model_info(model, custom_llm_provider) for model in frozenset(m for _, m in clip_models)} + ) override_info: Final = ( _cost_map_model_info(custom_pricing_model, custom_llm_provider) if custom_pricing_model is not None else None ) + return sum( + (_transcription_usage_cost(dict(usage), model_infos[model], override_info) for usage, model in clip_models), + 0.0, + ) - total_cost = 0.0 - for event in completed_events: - usage = event.get("usage") or {} - total_cost += _transcription_usage_cost(usage, model_info, override_info) - return total_cost + +_TRANSCRIPTION_SESSION_EVENT_TYPES: Final = frozenset( + {"transcription_session.created", "transcription_session.updated", "session.created", "session.updated"} +) + + +def _str_field(mapping: Mapping[str, object] | None, key: str) -> str | None: + value: Final = None if mapping is None else mapping.get(key) + return value if isinstance(value, str) and value else None + + +def _configured_transcription_model(session: Mapping[str, object]) -> str | None: + audio_input: Final = as_str_mapping((as_str_mapping(session.get("audio")) or {}).get("input")) + nested: Final = _str_field(as_str_mapping((audio_input or {}).get("transcription")), "model") + return nested or _str_field(as_str_mapping(session.get("input_audio_transcription")), "model") + + +def _session_of(result: Mapping[str, object]) -> Mapping[str, object] | None: + return as_str_mapping(result.get("session")) if result.get("type") in _TRANSCRIPTION_SESSION_EVENT_TYPES else None + + +def _session_transcription_model(result: Mapping[str, object]) -> str | None: + session: Final = _session_of(result) + return None if session is None else _configured_transcription_model(session) + + +def _newer_model(current: str | None, newer: str | None) -> str | None: + return newer or current def _get_transcription_model_name_from_results( results: OpenAIRealtimeStreamList, ) -> str | None: - """Resolve the ASR model from a transcription_session.* / session.* event.""" - for result in results: - if result.get("type") in ( - "transcription_session.created", - "transcription_session.updated", - "session.created", - "session.updated", - ): - session = cast(dict, result).get("session", {}) or {} - transcription = ((session.get("audio", {}) or {}).get("input", {}) or {}).get( - "transcription", {} - ) or session.get("input_audio_transcription", {}) - model = (transcription or {}).get("model") or session.get("model") - if model: - return model - return None + return next( + (model for result in results if (model := _str_field(_session_of(result), "model"))), + None, + ) _TranscriptionRateKey: TypeAlias = Literal[ diff --git a/litellm/litellm_core_utils/core_helpers.py b/litellm/litellm_core_utils/core_helpers.py index 58189140a98..6fd6e91a627 100644 --- a/litellm/litellm_core_utils/core_helpers.py +++ b/litellm/litellm_core_utils/core_helpers.py @@ -4,7 +4,7 @@ import copy import logging import math import re -from collections.abc import Collection, Iterable, Iterator, Mapping +from collections.abc import Collection, Iterable, Iterator, Mapping, Sequence from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, TypeVar, cast, runtime_checkable @@ -104,6 +104,10 @@ def as_str_mapping(value: object) -> Mapping[str, object] | None: return value if isinstance(value, Mapping) else None +def as_object_list(value: object) -> Sequence[object] | None: + return value if isinstance(value, list) else None + + def decoded_json_object(value: object) -> Mapping[str, object]: mapping: Final = as_str_mapping(value) if mapping is not None: diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py index e2ae25686da..f5d3a80ee91 100644 --- a/litellm/proxy/auth/auth_checks.py +++ b/litellm/proxy/auth/auth_checks.py @@ -5378,79 +5378,6 @@ async def _granted_model_lists( ) -async def _user_object_or_none( - valid_token: UserAPIKeyAuth, - prisma_client: PrismaClient, - user_api_key_cache: UserApiKeyCache, - proxy_logging_obj: ProxyLogging, -) -> LiteLLM_UserTable | None: - try: - return await get_user_object( - user_id=valid_token.user_id, - prisma_client=prisma_client, - user_api_key_cache=user_api_key_cache, - user_id_upsert=False, - proxy_logging_obj=proxy_logging_obj, - ) - except UserNotFoundError: - return None - - -async def enforced_model_allowlists( - valid_token: UserAPIKeyAuth, - prisma_client: PrismaClient | None, - user_api_key_cache: UserApiKeyCache, - proxy_logging_obj: ProxyLogging, -) -> tuple[Sequence[str], ...]: - """One model allowlist per level that ``common_checks`` enforces on a request from this identity.""" - key_models: Final = _resolve_key_models_for_auth_check(valid_token=valid_token) - if prisma_client is None: - return (key_models, tuple(valid_token.team_models or ())) - team_object: Final = ( - None - if valid_token.team_id is None - else await get_team_object( - team_id=valid_token.team_id, - prisma_client=prisma_client, - user_api_key_cache=user_api_key_cache, - proxy_logging_obj=proxy_logging_obj, - ) - ) - user_object: Final = ( - None - if team_object is not None - else await _user_object_or_none( - valid_token=valid_token, - prisma_client=prisma_client, - user_api_key_cache=user_api_key_cache, - proxy_logging_obj=proxy_logging_obj, - ) - ) - project_object: Final = ( - None - if valid_token.project_id is None - else await get_project_object( - project_id=valid_token.project_id, - prisma_client=prisma_client, - user_api_key_cache=user_api_key_cache, - proxy_logging_obj=proxy_logging_obj, - ) - ) - return ( - key_models, - team_object.models if team_object is not None else (), - await _team_member_granted_models( - valid_token=valid_token, - team_object=team_object, - prisma_client=prisma_client, - user_api_key_cache=user_api_key_cache, - proxy_logging_obj=proxy_logging_obj, - ), - user_object.models if user_object is not None else (), - project_object.models if project_object is not None else (), - ) - - async def collect_matched_model_access_groups( model: str | Sequence[str] | None, valid_token: UserAPIKeyAuth | None, diff --git a/litellm/proxy/common_utils/realtime_utils.py b/litellm/proxy/common_utils/realtime_utils.py index c641748843e..e96685bdeae 100644 --- a/litellm/proxy/common_utils/realtime_utils.py +++ b/litellm/proxy/common_utils/realtime_utils.py @@ -1,3 +1,4 @@ +import json from functools import lru_cache from typing import Final @@ -10,7 +11,7 @@ def realtime_request_body(model: str | None) -> bytes: Generate the realtime websocket request body. Cached with LRU semantics to avoid repeated string formatting work while keeping memory usage bounded. """ - return f'{{"model": "{model or ""}"}}'.encode() + return json.dumps({"model": model or ""}).encode() _realtime_request_body: Final = realtime_request_body diff --git a/litellm/proxy/hooks/parallel_request_limiter.py b/litellm/proxy/hooks/parallel_request_limiter.py index cd01d15d834..f9e85c31d46 100644 --- a/litellm/proxy/hooks/parallel_request_limiter.py +++ b/litellm/proxy/hooks/parallel_request_limiter.py @@ -13,7 +13,7 @@ from litellm._internal_context import with_service_target from litellm._logging import verbose_proxy_logger from litellm.exceptions import RateLimitType from litellm.integrations.custom_logger import CustomLogger -from litellm.litellm_core_utils.core_helpers import get_parent_otel_span_from_kwargs +from litellm.litellm_core_utils.core_helpers import as_str_mapping, get_parent_otel_span_from_kwargs from litellm.proxy._types import CommonProxyErrors, CurrentItemRateLimit, UserAPIKeyAuth from litellm.proxy.auth.auth_utils import ( get_key_model_rpm_limit, @@ -53,9 +53,16 @@ def _response_total_tokens(kwargs: Mapping[str, object], response_obj: object) - _CONNECTION_SCOPED_CALL_TYPES: Final = frozenset({CallTypes.aresponses_websocket.value}) +def _is_websocket_passthrough_row(kwargs: Mapping[str, object]) -> bool: + passthrough_payload: Final = as_str_mapping(kwargs.get("passthrough_logging_payload")) + return passthrough_payload is not None and passthrough_payload.get("request_method") == "WEBSOCKET" + + def _slots_released(kwargs: Mapping[str, object]) -> int: """A websocket turn callback releases nothing; the socket's slot is released once at disconnect.""" - return 0 if kwargs.get("call_type") in _CONNECTION_SCOPED_CALL_TYPES else 1 + if kwargs.get("call_type") in _CONNECTION_SCOPED_CALL_TYPES or _is_websocket_passthrough_row(kwargs): + return 0 + return 1 class CacheObject(TypedDict): diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index c7983f40aec..be9f80694b8 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -15,11 +15,11 @@ import os import posixpath import re import sys -from collections.abc import AsyncGenerator, Callable, Mapping, Sequence +from collections.abc import AsyncGenerator, Awaitable, Callable, Mapping, Sequence from dataclasses import dataclass from functools import partial from types import MappingProxyType -from typing import TYPE_CHECKING, Annotated, Final, Literal, Protocol, cast +from typing import TYPE_CHECKING, Annotated, Final, Literal, Protocol, TypeAlias, cast import httpx import openai @@ -51,6 +51,7 @@ from litellm.constants import ( from litellm.exceptions import BlockedPiiEntityError, GuardrailRaisedException from litellm.integrations.custom_guardrail import guardrail_request_data_with_streaming, is_guardrail_intervention from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix +from litellm.litellm_core_utils.core_helpers import as_object_list, as_str_mapping, attribute_or_none from litellm.litellm_core_utils.fallback_generalizations import match_routing_generalization from litellm.llms.anthropic.common_utils import ( AnthropicModelInfo, @@ -85,7 +86,12 @@ from litellm.llms.vertex_ai.vertex_llm_base import VertexBase from litellm.llms.vllm.common_utils import endpoint_under_api_base from litellm.passthrough.main import AsyncPassthroughStreamingResponse, is_provider_error from litellm.proxy._types import * -from litellm.proxy.auth.auth_checks import enforced_model_allowlists +from litellm.proxy.auth.auth_checks import ( + can_agent_call_model, + can_key_call_resolved_model, + key_model_aliases_for_auth_check, + team_model_aliases_for_auth_check, +) from litellm.proxy.auth.handle_jwt import JWTHandler from litellm.proxy.auth.route_checks import RouteChecks from litellm.proxy.auth.user_api_key_auth import ( # noqa: F401 # legacy module exports @@ -93,6 +99,7 @@ from litellm.proxy.auth.user_api_key_auth import ( # noqa: F401 # legacy modul get_bearer_token, is_no_auth_dev_mode, user_api_key_auth, + user_api_key_auth_accepted_websocket_for_model, user_api_key_auth_websocket, user_api_key_auth_websocket_for_model, ) @@ -114,6 +121,7 @@ from litellm.proxy.common_utils.http_parsing_utils import ( # noqa: F401 # leg safe_get_request_headers, safe_set_request_parsed_body, ) +from litellm.proxy.common_utils.model_listing_utils import alias_map from litellm.proxy.common_utils.openai_error_payload import LITELLM_CALL_ID_HEADER from litellm.proxy.common_utils.resource_ownership import is_proxy_admin from litellm.proxy.common_utils.sse_keepalive import ( @@ -129,16 +137,26 @@ from litellm.proxy.pass_through_endpoints.injected_stream_usage import ( stream_usage_relay, without_injected_usage_frames, ) +from litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_websocket_passthrough_logging_handler import ( + openai_websocket_frame_kind, +) from litellm.proxy.pass_through_endpoints.pass_through_endpoints import ( + STORED_PROMPT_PROBE_MODEL, HttpPassThroughEndpointHelpers, + WebsocketFrameModelAllowed, + WebsocketFrameModelGate, + WebsocketFrameModelRefused, + WebsocketFrameModelVerdict, + WebsocketUpstreamFrameClassifier, YieldsToConfiguredPassThroughRoute, create_pass_through_route, create_websocket_passthrough_route, + refuse_websocket, truncated_close_reason, websocket_passthrough_request, ) from litellm.proxy.utils import ProxyLogging as ProxyLoggingType -from litellm.proxy.utils import is_known_model +from litellm.proxy.utils import is_known_model, normalize_route_for_root_path from litellm.proxy.vector_store_endpoints.utils import ( assert_proxy_admin_for_vector_store_index_management, assert_user_can_access_vector_store, @@ -198,7 +216,7 @@ def is_passthrough_request_using_router_model(request_body: dict, llm_router: li """ try: model: Final = request_body.get("model") - return is_known_model(model, llm_router) + return isinstance(model, str) and is_known_model(model, llm_router) except Exception: return False @@ -228,9 +246,8 @@ def is_passthrough_request_streaming(request_body: object) -> bool: A JSON body need not be an object, so a list or scalar can reach here; it carries no streaming flag. """ - if not isinstance(request_body, dict): - return False - return bool(request_body.get("stream", False)) + body: Final = as_str_mapping(request_body) + return body is not None and bool(body.get("stream", False)) def _optional_str(value: object) -> str | None: @@ -3387,19 +3404,6 @@ def _join_url_paths(base_url: httpx.URL, path: str, custom_llm_provider: litellm return joined_path_str -_OPENAI_WS_ALL_MODEL_ACCESS: Final = frozenset( - { - SpecialModelNames.all_proxy_models.value, - SpecialModelNames.all_team_models.value, - "*", - } -) - - -def _has_model_restrictions(model_allowlists: tuple[Sequence[str], ...]) -> bool: - return any(str(model) not in _OPENAI_WS_ALL_MODEL_ACCESS for allowlist in model_allowlists for model in allowlist) - - @dataclass(frozen=True, slots=True) class _OpenAIWebsocketRefusal: close_reason: str @@ -3416,6 +3420,16 @@ class _OpenAIWebsocketErrorFrame(TypedDict): error: ReadOnly[_OpenAIWebsocketErrorDetail] +_OPENAI_WS_UNAVAILABLE_AUTH_REFUSAL: Final = ( + "This gateway could not authorize the model named in this frame, so the session was closed. Try again, " + "and ask a proxy admin to check the gateway logs if it keeps happening." +) + +_OPENAI_WS_ALIASED_MODEL_REFUSAL: Final = ( + "This key cannot use a model alias over OpenAI websocket passthrough. The passthrough sends each frame to " + "OpenAI unchanged, so name the model this key may use directly." +) + _OPENAI_WS_DISABLED_REFUSAL: Final = _OpenAIWebsocketRefusal( close_reason="OpenAI websocket passthrough is disabled", message=( @@ -3424,14 +3438,6 @@ _OPENAI_WS_DISABLED_REFUSAL: Final = _OpenAIWebsocketRefusal( ), ) -_OPENAI_WS_MODEL_RESTRICTED_REFUSAL: Final = _OpenAIWebsocketRefusal( - close_reason="Keys with model restrictions cannot use OpenAI websocket passthrough", - message=( - "Keys with model restrictions cannot use OpenAI websocket passthrough, because this route " - "relays frames to the provider without reading which model they ask for." - ), -) - def _is_openai_websocket_passthrough_enabled(general_settings: Mapping[str, object]) -> bool: setting: Final = general_settings.get("enable_openai_websocket_passthrough") @@ -3440,19 +3446,9 @@ def _is_openai_websocket_passthrough_enabled(general_settings: Mapping[str, obje return setting is True -class _OpenAIWebsocketModelAllowlists(Protocol): - async def __call__(self, valid_token: UserAPIKeyAuth, /) -> tuple[Sequence[str], ...]: ... - - -async def _openai_websocket_refusal( - user_api_key_dict: UserAPIKeyAuth, - general_settings: Mapping[str, object], - model_allowlists: _OpenAIWebsocketModelAllowlists, -) -> _OpenAIWebsocketRefusal | None: +def _openai_websocket_refusal(general_settings: Mapping[str, object]) -> _OpenAIWebsocketRefusal | None: if not _is_openai_websocket_passthrough_enabled(general_settings): return _OPENAI_WS_DISABLED_REFUSAL - if _has_model_restrictions(await model_allowlists(user_api_key_dict)): - return _OPENAI_WS_MODEL_RESTRICTED_REFUSAL return None @@ -3467,6 +3463,8 @@ class _WebsocketRelay(Protocol): forward_headers: bool, endpoint: str, accept_websocket: bool, + client_frame_model_gate: WebsocketFrameModelGate | None = None, + upstream_frame_kind: WebsocketUpstreamFrameClassifier | None = None, meter: WebSocketSessionMeter | None = None, ) -> None: ... @@ -3475,18 +3473,95 @@ def _websocket_relay() -> _WebsocketRelay: return websocket_passthrough_request -def _proxy_model_allowlists() -> _OpenAIWebsocketModelAllowlists: - from litellm.proxy.proxy_server import prisma_client, proxy_logging_obj, user_api_key_cache +def _proxy_error_type(denial: ProxyException) -> str: + return denial.type.value if isinstance(denial.type, ProxyErrorTypes) else denial.type - async def resolve(valid_token: UserAPIKeyAuth, /) -> tuple[Sequence[str], ...]: - return await enforced_model_allowlists( - valid_token=valid_token, - prisma_client=prisma_client, - user_api_key_cache=user_api_key_cache, - proxy_logging_obj=proxy_logging_obj, + +FrameModelAuthorizer: TypeAlias = Callable[[WebSocket, str], Awaitable[UserAPIKeyAuth]] + + +def _frame_model_authorizer() -> FrameModelAuthorizer: + return user_api_key_auth_accepted_websocket_for_model + + +def _proxy_model_list() -> Sequence[object] | None: + from litellm.proxy import proxy_server + + return as_object_list(attribute_or_none(proxy_server, "llm_model_list")) + + +def _gateway_remaps_model(model: str, frame_auth: UserAPIKeyAuth, llm_router: Router | None) -> bool: + alias_maps: Final = ( + key_model_aliases_for_auth_check(frame_auth) or {}, + team_model_aliases_for_auth_check(frame_auth) or {}, + alias_map(litellm.model_alias_map), + ) + targets: Final = ( + *(aliases.get(model) for aliases in alias_maps), + llm_router.registry().resolve(model, kind="alias") if llm_router is not None else None, + ) + return any(target not in (None, model) for target in targets) + + +def _aliased_model_verdict( + allowed: WebsocketFrameModelAllowed, probe: WebsocketFrameModelVerdict +) -> WebsocketFrameModelVerdict: + match probe: + case WebsocketFrameModelRefused(error_type=error_type) if error_type.endswith("_model_access_denied"): + return WebsocketFrameModelRefused(message=_OPENAI_WS_ALIASED_MODEL_REFUSAL, error_type=error_type) + case WebsocketFrameModelRefused(): + return probe + case WebsocketFrameModelAllowed(): + return allowed + + +def _proxy_frame_model_gate( + authorize_frame_model: Annotated[FrameModelAuthorizer, Depends(_frame_model_authorizer)], +) -> WebsocketFrameModelGate: + from litellm.proxy import proxy_server + from litellm.proxy.spend_tracking.budget_reservation import release_or_invalidate_budget_reservation + + async def authorize(websocket: WebSocket, model: str) -> WebsocketFrameModelVerdict: + connect_reservation: Final = attribute_or_none(websocket.state, "budget_reservation") + await release_or_invalidate_budget_reservation( + budget_reservation=connect_reservation if isinstance(connect_reservation, dict) else None ) + frame_auth: Final = await authorize_frame_model(websocket, model) + access_groups: Final = tuple(frame_auth.matched_model_access_groups or ()) + llm_router: Final = proxy_server.llm_router + try: + await can_key_call_resolved_model( + model=model, llm_model_list=_proxy_model_list(), valid_token=frame_auth, llm_router=llm_router + ) + await can_agent_call_model( + model=model, + valid_token=frame_auth, + llm_router=llm_router, + prisma_client=proxy_server.prisma_client, + user_api_key_cache=proxy_server.user_api_key_cache, + proxy_logging_obj=proxy_server.proxy_logging_obj, + ) + finally: + await release_or_invalidate_budget_reservation(budget_reservation=frame_auth.budget_reservation) + allowed: Final = WebsocketFrameModelAllowed(access_groups=access_groups) + if not _gateway_remaps_model(model, frame_auth, llm_router): + return allowed + return _aliased_model_verdict(allowed, await gate(websocket, STORED_PROMPT_PROBE_MODEL)) - return resolve + async def gate(websocket: WebSocket, model: str, /) -> WebsocketFrameModelVerdict: + try: + return await authorize(websocket, model) + except ProxyException as denial: + return WebsocketFrameModelRefused(message=denial.message, error_type=_proxy_error_type(denial)) + except HTTPException as denial: + return WebsocketFrameModelRefused(message=str(denial.detail), error_type="invalid_request_error") + except litellm.BudgetExceededError as exhausted: + return WebsocketFrameModelRefused(message=exhausted.message, error_type="budget_exceeded") + except Exception: # noqa: BLE001 # fail closed: a gate that cannot answer refuses, and says so in band + verbose_proxy_logger.exception("OpenAI websocket passthrough: the frame model gate failed on %s", model) + return WebsocketFrameModelRefused(message=_OPENAI_WS_UNAVAILABLE_AUTH_REFUSAL, error_type="server_error") + + return gate def _negotiated_websocket_subprotocol(websocket: WebSocket) -> str | None: @@ -3498,6 +3573,14 @@ def _negotiated_websocket_subprotocol(websocket: WebSocket) -> str | None: return requested_subprotocols[0] if requested_subprotocols else None +async def _opening_model_refusal( + websocket: WebSocket, query_string: str, frame_model_gate: WebsocketFrameModelGate +) -> WebsocketFrameModelRefused | None: + opening_models: Final = dict.fromkeys(httpx.QueryParams(query_string).get_list("model")) + verdicts: Final = tuple([await frame_model_gate(websocket, model) for model in opening_models]) + return next((verdict for verdict in verdicts if isinstance(verdict, WebsocketFrameModelRefused)), None) + + @router.websocket("/openai_passthrough/{endpoint:path}") @router.websocket("/openai/{endpoint:path}") async def openai_websocket_proxy_route( @@ -3506,12 +3589,12 @@ async def openai_websocket_proxy_route( user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth_websocket)], general_settings: Annotated[Mapping[str, object], Depends(_proxy_general_settings)], relay: Annotated[_WebsocketRelay, Depends(_websocket_relay)], - model_allowlists: Annotated[_OpenAIWebsocketModelAllowlists, Depends(_proxy_model_allowlists)], + frame_model_gate: Annotated[WebsocketFrameModelGate, Depends(_proxy_frame_model_gate)], ) -> None: """WebSocket passthrough for OpenAI prefixes (realtime / responses.connect).""" negotiated_subprotocol: Final = _negotiated_websocket_subprotocol(websocket) - refusal: Final = await _openai_websocket_refusal(user_api_key_dict, general_settings, model_allowlists) + refusal: Final = _openai_websocket_refusal(general_settings) if refusal is not None: await websocket.accept(subprotocol=negotiated_subprotocol) error_frame: Final[_OpenAIWebsocketErrorFrame] = { @@ -3557,14 +3640,21 @@ async def openai_websocket_proxy_route( await websocket.accept(subprotocol=negotiated_subprotocol) + opening_refusal: Final = await _opening_model_refusal(websocket, query_string, frame_model_gate) + if opening_refusal is not None: + await refuse_websocket(websocket, opening_refusal) + return + await relay( websocket=websocket, target=wss_target, custom_headers=custom_headers, user_api_key_dict=user_api_key_dict, forward_headers=False, - endpoint=websocket.url.path, + endpoint=normalize_route_for_root_path(websocket.url.path) or websocket.url.path, accept_websocket=False, + client_frame_model_gate=frame_model_gate, + upstream_frame_kind=openai_websocket_frame_kind, ) @@ -3882,7 +3972,8 @@ async def _tinyfish_json_object_field_names(request: Request) -> frozenset[str] parsed: Final = _DECODED_JSON.validate_python(json.loads(raw_body)) except (json.JSONDecodeError, UnicodeDecodeError): return None - return frozenset(parsed) if isinstance(parsed, dict) else None + fields: Final = as_str_mapping(parsed) + return None if fields is None else frozenset(fields) def _tinyfish_route_timeout() -> float | None: diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_websocket_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_websocket_passthrough_logging_handler.py new file mode 100644 index 00000000000..f3bad54ffb2 --- /dev/null +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_websocket_passthrough_logging_handler.py @@ -0,0 +1,262 @@ +from collections.abc import Mapping, Sequence +from dataclasses import dataclass +from datetime import datetime +from types import MappingProxyType +from typing import Final, Literal, TypeAlias, cast # noqa: TID251 # cost helpers type raw frames as a TypedDict list + +from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError + +import litellm +from litellm._logging import verbose_proxy_logger +from litellm.cost_calculator import RealtimeAPITokenUsageProcessor, handle_realtime_stream_cost_calculation +from litellm.litellm_core_utils.core_helpers import as_str_mapping +from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.proxy._types import PassThroughEndpointLoggingTypedDict +from litellm.responses.utils import ResponseAPILoggingUtils +from litellm.types.llms.openai import OpenAIRealtimeStreamList, ResponsesAPIResponse +from litellm.types.passthrough_endpoints.pass_through_endpoints import WebsocketUpstreamFrameKind +from litellm.types.utils import LlmProviders, ModelResponse, Usage + +WebsocketMessages: TypeAlias = Sequence[Mapping[str, object]] + +OPENAI_WEBSOCKET_PROVIDER: Final = LlmProviders.OPENAI.value + +_SESSION_CREATED: Final = "session.created" +_SESSION_UPDATED: Final = "session.updated" +_RESPONSE_CREATED: Final = "response.created" +_ERROR: Final = "error" +_REALTIME_RESPONSE_DONE: Final = "response.done" +_TRANSCRIPTION_COMPLETED: Final = "conversation.item.input_audio_transcription.completed" +_RESPONSES_TURN_ENDS: Final = frozenset({"response.completed", "response.incomplete", "response.failed"}) +_TURN_ENDS: Final = _RESPONSES_TURN_ENDS | {_REALTIME_RESPONSE_DONE} +_MODEL_NAMING_EVENTS: Final = _TURN_ENDS | {_SESSION_CREATED, _SESSION_UPDATED, _RESPONSE_CREATED} +_REALTIME_BILLED_EVENTS: Final = frozenset({_REALTIME_RESPONSE_DONE, _TRANSCRIPTION_COMPLETED}) +_RESPONSE_COST_HEADER: Final = "llm_provider-x-litellm-response-cost" +_FRAME_KINDS: Final[Mapping[str, WebsocketUpstreamFrameKind]] = MappingProxyType( + { + _SESSION_CREATED: "opens_session", + _SESSION_UPDATED: "updates_session", + _RESPONSE_CREATED: "starts_turn", + _TRANSCRIPTION_COMPLETED: "bills_usage", + _ERROR: "refuses_turn", + } +) + + +class _NamedModel(BaseModel): + model_config = ConfigDict(extra="ignore", frozen=True) + + model: str | None = None + + +class _AudioInput(BaseModel): + model_config = ConfigDict(extra="ignore", frozen=True) + + transcription: _NamedModel | None = None + + +class _SessionAudio(BaseModel): + model_config = ConfigDict(extra="ignore", frozen=True) + + input: _AudioInput | None = None + + +class _NamedSession(BaseModel): + model_config = ConfigDict(extra="ignore", frozen=True) + + model: str | None = None + audio: _SessionAudio | None = None + + +class _ModelNamingEvent(BaseModel): + model_config = ConfigDict(extra="ignore", frozen=True) + + session: _NamedSession | None = None + response: _NamedModel | None = None + + +class _ResponsesTurnEndEvent(BaseModel): + model_config = ConfigDict(extra="ignore", frozen=True) + + type: Literal["response.completed", "response.incomplete", "response.failed"] + response: ResponsesAPIResponse + + +_MODEL_NAMING_EVENT: Final = TypeAdapter(_ModelNamingEvent) +_RESPONSES_TURN_END_EVENT: Final = TypeAdapter(_ResponsesTurnEndEvent) + + +@dataclass(frozen=True, slots=True) +class _Billing: + usage: Usage + cost: float + + +@dataclass(frozen=True, slots=True) +class _CompletedTurn: + model: str + response: ResponsesAPIResponse + usage: Usage + + +def _event_model(event: Mapping[str, object]) -> str | None: + if event.get("type") not in _MODEL_NAMING_EVENTS: + return None + try: + naming: Final = _MODEL_NAMING_EVENT.validate_python(event) + except ValidationError: + return None + session: Final = naming.session or _NamedSession() + response: Final = naming.response or _NamedModel() + audio_input: Final = (session.audio or _SessionAudio()).input or _AudioInput() + transcription: Final = audio_input.transcription or _NamedModel() + return session.model or response.model or transcription.model or None + + +def observed_websocket_model(messages: WebsocketMessages) -> str | None: + return next((model for model in map(_event_model, messages) if model is not None), None) + + +def openai_websocket_frame_kind(frame: Mapping[str, object]) -> WebsocketUpstreamFrameKind: + frame_type: Final = frame.get("type") + if frame_type in _TURN_ENDS: + return "ends_turn" + if not isinstance(frame_type, str): + return "relayed" + return _FRAME_KINDS.get(frame_type, "relayed") + + +def _realtime_billing(messages: WebsocketMessages, model: str, logging_obj: LiteLLMLoggingObj) -> _Billing | None: + if not any(event.get("type") in _REALTIME_BILLED_EVENTS for event in messages): + return None + results: Final = cast( # cast-ok: the realtime cost helpers read raw provider frames through their TypedDict view + OpenAIRealtimeStreamList, tuple(event for event in messages if "type" in event) + ) + usage: Final = RealtimeAPITokenUsageProcessor.collect_and_combine_usage_from_realtime_stream_results(results) + cost: Final = handle_realtime_stream_cost_calculation( + results=results, + combined_usage_object=usage, + custom_llm_provider=OPENAI_WEBSOCKET_PROVIDER, + litellm_model_name=model, + litellm_logging_obj=logging_obj, + ) + return _Billing(usage=usage, cost=cost) + + +def _responses_turn(event: Mapping[str, object], session_model: str) -> _CompletedTurn | None: + if event.get("type") not in _RESPONSES_TURN_ENDS: + return None + try: + ended: Final = _RESPONSES_TURN_END_EVENT.validate_python(event) + except ValidationError: + verbose_proxy_logger.warning( + "OpenAI websocket passthrough: a %s frame carried no readable usage, so it is not billed", event.get("type") + ) + return None + if ended.response.usage is None: + return None + return _CompletedTurn( + model=ended.response.model or session_model, + response=ended.response, + usage=ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage( # pyright: ignore[reportPrivateUsage] # the one shared Responses usage converter + ended.response.usage + ), + ) + + +def _responses_billing(messages: WebsocketMessages, model: str, requested_model: str | None) -> _Billing | None: + turns: Final = tuple(turn for turn in (_responses_turn(event, model) for event in messages) if turn is not None) + if not turns: + return None + usage: Final = RealtimeAPITokenUsageProcessor.combine_usage_objects([turn.usage for turn in turns]) + return _Billing( + usage=usage, + cost=sum(_responses_cost(turn.response, (turn.model, requested_model)) for turn in turns), + ) + + +def _model_cost(response: ResponsesAPIResponse, model: str) -> float | None: + try: + return litellm.completion_cost( + completion_response=response, + model=model, + custom_llm_provider=OPENAI_WEBSOCKET_PROVIDER, + call_type="responses", + ) + except Exception: # noqa: BLE001 # completion_cost raises a bare Exception for an unpriced model, and the turn must still log + return None + + +def _responses_cost(response: ResponsesAPIResponse, price_as: Sequence[str | None]) -> float: + costs: Final = (_model_cost(response, model) for model in dict.fromkeys(price_as) if model is not None) + cost: Final = next((cost for cost in costs if cost), None) + if cost is None: + verbose_proxy_logger.warning( + "OpenAI websocket passthrough: could not price a Responses turn on %s, logging it at zero", price_as + ) + return 0.0 + return cost + + +def _requested_model(kwargs: Mapping[str, object]) -> str | None: + litellm_params: Final = as_str_mapping(kwargs.get("litellm_params")) + metadata: Final = None if litellm_params is None else as_str_mapping(litellm_params.get("metadata")) + model_group: Final = None if metadata is None else metadata.get("model_group") + return model_group if isinstance(model_group, str) else None + + +def _priced_result(model: str, billing: _Billing, start_time: datetime, call_id: str) -> ModelResponse: + result: Final = ModelResponse( + id=call_id, + object="chat.completion", + created=int(start_time.timestamp()), + model=model, + usage=billing.usage, + ) + hidden: Final = {"additional_headers": MappingProxyType({_RESPONSE_COST_HEADER: billing.cost})} + result._hidden_params = hidden # pyright: ignore[reportPrivateUsage] # the one settled-cost channel + return result + + +class OpenAIWebsocketPassthroughLoggingHandler: + def openai_websocket_passthrough_handler( + self, + websocket_messages: WebsocketMessages, + logging_obj: LiteLLMLoggingObj, + url_route: str, + start_time: datetime, + kwargs: Mapping[str, object], + ) -> PassThroughEndpointLoggingTypedDict: + model: Final = observed_websocket_model(websocket_messages) + if model is None: + verbose_proxy_logger.debug("OpenAI websocket passthrough (%s): no model named, skipping cost", url_route) + unpriced: Final[PassThroughEndpointLoggingTypedDict] = {"result": None, "kwargs": {**kwargs}} + return unpriced + + logging_obj.model = model # rebind-ok: the spend logger reads model and cost off the shared logging object + logging_obj.model_call_details.update( + MappingProxyType({"model": model, "custom_llm_provider": OPENAI_WEBSOCKET_PROVIDER}) + ) + + billing: Final = _realtime_billing(websocket_messages, model, logging_obj) or _responses_billing( + websocket_messages, model, _requested_model(kwargs) + ) + if billing is None: + verbose_proxy_logger.debug("OpenAI websocket passthrough (%s): no usage reported, skipping cost", url_route) + unbilled: Final[PassThroughEndpointLoggingTypedDict] = { + "result": None, + "kwargs": {**kwargs, "model": model, "custom_llm_provider": OPENAI_WEBSOCKET_PROVIDER}, + } + return unbilled + + logging_obj.model_call_details.update(MappingProxyType({"response_cost": billing.cost})) + priced: Final[PassThroughEndpointLoggingTypedDict] = { + "result": _priced_result(model, billing, start_time, logging_obj.litellm_call_id), + "kwargs": { + **kwargs, + "model": model, + "custom_llm_provider": OPENAI_WEBSOCKET_PROVIDER, + "response_cost": billing.cost, + }, + } + return priced diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py index 970cd466337..7f5983d5799 100644 --- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py @@ -6,15 +6,17 @@ import posixpath import re import traceback from base64 import b64encode -from collections.abc import AsyncGenerator, AsyncIterator, Callable, Iterable, Mapping, Sequence +from collections.abc import AsyncGenerator, AsyncIterator, Callable, Iterable, Mapping, MutableMapping, Sequence from dataclasses import dataclass from datetime import datetime -from itertools import chain, count, groupby +from functools import reduce +from itertools import chain, count, groupby, takewhile from types import MappingProxyType -from typing import TYPE_CHECKING, Any, Final, TypedDict, cast -from urllib.parse import urlencode, urlparse +from typing import TYPE_CHECKING, Any, Final, Protocol, TypeAlias, TypedDict, TypeVar, cast +from urllib.parse import parse_qs, urlencode, urlparse import httpx +import orjson from fastapi import ( APIRouter, Depends, @@ -24,6 +26,7 @@ from fastapi import ( Response, UploadFile, WebSocket, + WebSocketDisconnect, status, ) from fastapi.responses import StreamingResponse @@ -34,7 +37,7 @@ from starlette.routing import BaseRoute, Match, Route from starlette.types import Scope from starlette.websockets import WebSocketState from typing_extensions import ReadOnly -from websockets.asyncio.client import connect +from websockets.asyncio.client import ClientConnection, connect from websockets.exceptions import ( ConnectionClosedError, ConnectionClosedOK, @@ -60,6 +63,8 @@ from litellm.integrations.custom_guardrail import ( from litellm.integrations.custom_logger import CustomLogger from litellm.integrations.otel.model.semconv import PASSTHROUGH_OPERATION_KEY, GenAIOperation from litellm.litellm_core_utils.core_helpers import ( + as_object_list, + as_str_mapping, bind_budget_reservation_to_callbacks, get_metadata_variable_name_from_kwargs, get_or_create_metadata_bucket, @@ -69,6 +74,7 @@ from litellm.litellm_core_utils.internal_call_metadata import MODEL_ACCESS_GROUP 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.realtime_errors import realtime_error_event from litellm.litellm_core_utils.redact_messages import should_redact_message_logging from litellm.litellm_core_utils.safe_json_dumps import safe_dumps from litellm.llms.anthropic.common_utils import ( @@ -154,6 +160,7 @@ from litellm.types.passthrough_endpoints.pass_through_endpoints import ( PassthroughStandardLoggingPayload, WebSocketSessionMeter, WebSocketSessionUsage, + WebsocketUpstreamFrameKind, ) from litellm.types.utils import TRUSTED_CALLBACK_VARS_FIELD, CallTypes, Usage @@ -2574,11 +2581,16 @@ def _resolved_vertex_live_setup( return {**setup_data, "model": setup_model_rewriter(setup_model)} +def _decoded_frame(frame: str | bytes) -> object: + try: + return _DECODED_JSON.validate_python(orjson.loads(frame)) + except orjson.JSONDecodeError: + return _DECODED_JSON.validate_python(json.loads(frame if isinstance(frame, str) else frame.decode("utf-8"))) + + def _json_object_frame(frame: str | bytes) -> dict[str, object] | None: try: - decoded: Final[object] = _DECODED_JSON.validate_python( - json.loads(frame if isinstance(frame, str) else frame.decode("utf-8")) - ) + decoded: Final = _decoded_frame(frame) except (json.JSONDecodeError, UnicodeDecodeError): return None return decoded if isinstance(decoded, dict) else None @@ -2637,6 +2649,359 @@ def _with_trace_context(headers: Mapping[str, str], parent_span: object) -> dict return inject_trace_context(headers, parent_span=parent_span) +@dataclass(frozen=True, slots=True) +class WebsocketFrameModelAllowed: + access_groups: tuple[str, ...] + + +@dataclass(frozen=True, slots=True) +class WebsocketFrameModelRefused: + message: str + error_type: str + + +WebsocketFrameModelVerdict: TypeAlias = WebsocketFrameModelAllowed | WebsocketFrameModelRefused + + +class WebsocketFrameModelGate(Protocol): + async def __call__(self, websocket: WebSocket, model: str, /) -> WebsocketFrameModelVerdict: ... + + +WebsocketUpstreamFrameClassifier: TypeAlias = Callable[[Mapping[str, object]], WebsocketUpstreamFrameKind] +_SESSION_FRAMES_EVERY_ROW_CARRIES: Final[frozenset[WebsocketUpstreamFrameKind]] = frozenset( + {"opens_session", "updates_session"} +) +_FrameT: Final = TypeVar("_FrameT", bound=Mapping[str, object]) + + +@dataclass(frozen=True, slots=True) +class _TurnAuthorization: + requested_model: str | None + access_groups: tuple[str, ...] + + +@dataclass(frozen=True, slots=True) +class _ForwardClientFrame: + starts_turn: bool + turn: _TurnAuthorization | None + lane: str | None + + +@dataclass(frozen=True, slots=True) +class _RefuseClientFrame: + refusal: WebsocketFrameModelRefused + + +@dataclass(frozen=True, slots=True) +class _ClientFrameRequests: + starts_turn: bool + lane: str | None + turn_model: str | None + models: tuple[str, ...] + unverifiable_model_refusal: str | None + + +_NO_CLIENT_FRAME_REQUESTS: Final = _ClientFrameRequests( + starts_turn=False, lane=None, turn_model=None, models=(), unverifiable_model_refusal=None +) +_FORWARD_WITHOUT_TURN: Final = _ForwardClientFrame(starts_turn=False, turn=None, lane=None) +_UNREADABLE_CLIENT_FRAME: Final = _RefuseClientFrame( + refusal=WebsocketFrameModelRefused( + message="The gateway cannot parse this frame, so it cannot check which model it asks for.", + error_type="invalid_request_error", + ) +) +_TURN_STARTING_FRAME: Final = "response.create" +_FRAME_TEXT_WORTH_PARSING: Final = (f'"{_TURN_STARTING_FRAME}"', '"model"', '"prompt"', '"tools"', "\\") +_TURN_MODEL_PLACEMENTS: Final = (("model",), ("response", "model"), ("session", "model")) +_TRANSCRIPTION_MODEL_PLACEMENTS: Final = ( + ("session", "input_audio_transcription", "model"), + ("session", "audio", "input", "transcription", "model"), +) +_STORED_PROMPT_PLACEMENTS: Final = (("prompt",), ("response", "prompt"), ("session", "prompt")) +_TOOL_LIST_PLACEMENTS: Final = (("tools",), ("response", "tools"), ("session", "tools")) +_TOOL_THAT_RUNS_A_MODEL: Final = "image_generation" +STORED_PROMPT_PROBE_MODEL: Final = "litellm-unverifiable-stored-prompt-model" +_STORED_PROMPT_REFUSAL: Final = ( + "This key is limited to specific models, so it cannot run a stored prompt. " + "The gateway cannot check which model a stored prompt runs." +) +_UNNAMED_TOOL_MODEL_REFUSAL: Final = ( + "This key is limited to specific models, so an image_generation tool has to name its model. " + "The gateway cannot check which model an unnamed one runs." +) +_FIRST_SENTENCE: Final = re.compile(r".*?[.!?](?=\s|$)") +_LANE_NAME: Final = re.compile(r"[A-Za-z0-9_.-]{1,256}") +_UNROUTABLE_LANE_REFUSAL: Final = _RefuseClientFrame( + refusal=WebsocketFrameModelRefused( + message=( + "This stream_id is not a lane name the provider accepts. " + "The provider's error for it names no lane, so the gateway cannot tell which turn it refuses." + ), + error_type="invalid_request_error", + ) +) + + +def _may_name_a_model(text_data: str) -> bool: + return any(marker in text_data for marker in _FRAME_TEXT_WORTH_PARSING) + + +def _frame_lane(frame: Mapping[str, object]) -> str | None: + stream_id: Final = frame.get("stream_id") + return stream_id if isinstance(stream_id, str) else None + + +def _names_an_unroutable_lane(frame: Mapping[str, object]) -> bool: + stream_id: Final = frame.get("stream_id") + return stream_id is not None and not (isinstance(stream_id, str) and _LANE_NAME.fullmatch(stream_id) is not None) + + +def _frames_of_lane(lane: str | None) -> Callable[[Mapping[str, object]], bool]: + def in_lane(frame: Mapping[str, object]) -> bool: + return _frame_lane(frame) == lane + + return in_lane + + +def _every_frame(_frame: Mapping[str, object]) -> bool: + return True + + +def _without_earlier_session_updates( + upstream_frame_kind: WebsocketUpstreamFrameClassifier, + frames: Sequence[_FrameT], + latest_update: _FrameT, +) -> tuple[_FrameT, ...]: + return tuple( + frame + for index, frame in enumerate(frames) + if frame is latest_update + or upstream_frame_kind(frame) != "updates_session" + or _unbilled_usage_ran_under(map(upstream_frame_kind, frames[index + 1 :])) + ) + + +def _unbilled_usage_ran_under(later_kinds: Iterable[WebsocketUpstreamFrameKind]) -> bool: + return "bills_usage" in takewhile(lambda kind: kind != "updates_session", later_kinds) + + +def _nested_frame(frame: Mapping[str, object] | None, key: str) -> Mapping[str, object] | None: + return None if frame is None else as_str_mapping(frame.get(key)) + + +def _placed_value(frame: Mapping[str, object], placement: Sequence[str]) -> object: + *parents, leaf = placement + container: Final = reduce(_nested_frame, parents, frame) + return None if container is None else container.get(leaf) + + +def _placed_models(frame: Mapping[str, object], placements: Iterable[Sequence[str]]) -> tuple[str, ...]: + placed: Final = (_placed_value(frame, placement) for placement in placements) + return tuple(value for value in placed if isinstance(value, str)) + + +def _placed_tools(frame: Mapping[str, object]) -> tuple[Mapping[str, object], ...]: + tool_lists: Final = (_placed_value(frame, placement) for placement in _TOOL_LIST_PLACEMENTS) + tools: Final = chain.from_iterable(as_object_list(tool_list) or () for tool_list in tool_lists) + return tuple(tool for tool in map(as_str_mapping, tools) if tool is not None) + + +def _unverifiable_model_refusal(frame: Mapping[str, object], tools: Sequence[Mapping[str, object]]) -> str | None: + prompts: Final = (_placed_value(frame, placement) for placement in _STORED_PROMPT_PLACEMENTS) + if any(isinstance(prompt, Mapping) for prompt in prompts): + return _STORED_PROMPT_REFUSAL + if any(tool.get("type") == _TOOL_THAT_RUNS_A_MODEL and not isinstance(tool.get("model"), str) for tool in tools): + return _UNNAMED_TOOL_MODEL_REFUSAL + return None + + +def _client_frame_requests(frame_data: str | bytes) -> _ClientFrameRequests | _RefuseClientFrame: + if isinstance(frame_data, str) and not _may_name_a_model(frame_data): + return _NO_CLIENT_FRAME_REQUESTS + try: + frame: Final = _json_object_frame(frame_data) + except (RecursionError, ValueError): + return _UNREADABLE_CLIENT_FRAME + if frame is None: + return _NO_CLIENT_FRAME_REQUESTS + if _names_an_unroutable_lane(frame): + return _UNROUTABLE_LANE_REFUSAL + turn_models: Final = _placed_models(frame, _TURN_MODEL_PLACEMENTS) + tools: Final = _placed_tools(frame) + tool_models: Final = (model for model in (tool.get("model") for tool in tools) if isinstance(model, str)) + return _ClientFrameRequests( + starts_turn=frame.get("type") == _TURN_STARTING_FRAME, + lane=_frame_lane(frame), + turn_model=next(iter(turn_models), None), + models=tuple( + dict.fromkeys((*turn_models, *_placed_models(frame, _TRANSCRIPTION_MODEL_PLACEMENTS), *tool_models)) + ), + unverifiable_model_refusal=_unverifiable_model_refusal(frame, tools), + ) + + +def _unverifiable_model_probe_refusal(probe: WebsocketFrameModelRefused, message: str) -> WebsocketFrameModelRefused: + if probe.error_type.endswith("_model_access_denied"): + return WebsocketFrameModelRefused(message=message, error_type=probe.error_type) + return probe + + +async def _client_frame_verdict( + websocket: WebSocket, frame_data: str | bytes, gate: WebsocketFrameModelGate | None +) -> _ForwardClientFrame | _RefuseClientFrame: + if gate is None: + return _FORWARD_WITHOUT_TURN + requests: Final = _client_frame_requests(frame_data) + if isinstance(requests, _RefuseClientFrame): + return requests + verdicts: Final = tuple([await gate(websocket, model) for model in requests.models]) + refusal: Final = next((verdict for verdict in verdicts if isinstance(verdict, WebsocketFrameModelRefused)), None) + if refusal is not None: + return _RefuseClientFrame(refusal=refusal) + if requests.unverifiable_model_refusal is not None and isinstance( + probe := await gate(websocket, STORED_PROMPT_PROBE_MODEL), WebsocketFrameModelRefused + ): + return _RefuseClientFrame(refusal=_unverifiable_model_probe_refusal(probe, requests.unverifiable_model_refusal)) + allowed: Final = tuple(verdict for verdict in verdicts if isinstance(verdict, WebsocketFrameModelAllowed)) + return _ForwardClientFrame( + starts_turn=requests.starts_turn, + lane=requests.lane, + turn=( + _TurnAuthorization(requested_model=requests.turn_model, access_groups=allowed[0].access_groups) + if requests.turn_model is not None and allowed + else None + ), + ) + + +def _refusal_close_reason(message: str) -> str: + first_sentence: Final = _FIRST_SENTENCE.match(message) + return truncated_close_reason(message if first_sentence is None else first_sentence.group(0)) + + +async def refuse_websocket(websocket: WebSocket, refusal: WebsocketFrameModelRefused) -> None: + await websocket.send_text(realtime_error_event(refusal.message, error_type=refusal.error_type)) + await websocket.close(code=CloseCode.POLICY_VIOLATION, reason=_refusal_close_reason(refusal.message)) + + +def _stamp_billed_turn(kwargs: Mapping[str, object], turn: _TurnAuthorization) -> None: + litellm_params: Final = as_str_mapping(kwargs.get("litellm_params")) + metadata: Final = None if litellm_params is None else litellm_params.get("metadata") + if not isinstance(metadata, MutableMapping): + return + metadata[MODEL_ACCESS_GROUP_METADATA_KEY] = list(turn.access_groups) + if turn.requested_model is not None: + metadata["model_group"] = turn.requested_model + + +def _opening_model(target: str) -> str | None: + return next(iter(parse_qs(urlparse(target).query).get("model", ())), None) + + +class _WebsocketTurnLedger: + def __init__(self, opening: _TurnAuthorization, started: datetime) -> None: + self.opening = opening + self.queued_turns: Mapping[str | None, tuple[_TurnAuthorization | None, ...]] = MappingProxyType({}) + self.session_opened = False + self.started = started + self.logged_turns = 0 + self.unstarted_turns = 0 + + def current(self, lane: str | None) -> _TurnAuthorization: + oldest: Final = next(iter(self.queued_turns.get(lane, ())), None) + return self.opening if oldest is None else oldest + + def queue_turn(self, lane: str | None, turn: _TurnAuthorization | None) -> None: + self.unstarted_turns += 1 + if not self.session_opened: + self.queued_turns = MappingProxyType({**self.queued_turns, lane: (*self.queued_turns.get(lane, ()), turn)}) + + def open_session(self) -> None: + self.session_opened = True + self.queued_turns = MappingProxyType({}) + + def close_turn(self, lane: str | None, end_time: datetime) -> None: + self.logged_turns += 1 + self.started = end_time + self._forget_turn(lane, 0) + + def start_turn(self) -> None: + self.unstarted_turns = max(0, self.unstarted_turns - 1) + + def drop_refused_turn(self, lane: str | None, lane_is_running_a_turn: bool) -> None: + self.start_turn() + self._forget_turn(lane, 1 if lane_is_running_a_turn else 0) + + def _forget_turn(self, lane: str | None, position: int) -> None: + turns: Final = self.queued_turns.get(lane, ()) + self.queued_turns = MappingProxyType({**self.queued_turns, lane: (*turns[:position], *turns[position + 1 :])}) + + +def _session_end_needs_a_row( + upstream_frame_kind: WebsocketUpstreamFrameClassifier | None, + unlogged_frames: Sequence[Mapping[str, object]], + ledger: _WebsocketTurnLedger, +) -> bool: + if upstream_frame_kind is None: + return True + kinds: Final = frozenset(map(upstream_frame_kind, unlogged_frames)) + return "starts_turn" in kinds or (ledger.logged_turns == 0 and bool(kinds - {"opens_session"})) + + +def _turn_in_flight( + upstream_frame_kind: WebsocketUpstreamFrameClassifier | None, unlogged_frames: Sequence[Mapping[str, object]] +) -> bool: + return upstream_frame_kind is not None and any( + upstream_frame_kind(frame) == "starts_turn" for frame in unlogged_frames + ) + + +def _turn_owed( + upstream_frame_kind: WebsocketUpstreamFrameClassifier | None, + unlogged_frames: Sequence[Mapping[str, object]], + ledger: _WebsocketTurnLedger, +) -> bool: + return _turn_in_flight(upstream_frame_kind, unlogged_frames) or ( + upstream_frame_kind is not None and ledger.unstarted_turns > 0 + ) + + +def _lanes_in_flight( + upstream_frame_kind: WebsocketUpstreamFrameClassifier, unlogged_frames: Sequence[Mapping[str, object]] +) -> tuple[str | None, ...]: + started: Final = (frame for frame in unlogged_frames if upstream_frame_kind(frame) == "starts_turn") + return tuple(dict.fromkeys(map(_frame_lane, started))) + + +async def _relay_client_frame( + websocket: WebSocket, + upstream_ws: ClientConnection, + checked_frame: str | bytes, + forwarded_frame: str | bytes, + gate: WebsocketFrameModelGate | None, + ledger: _WebsocketTurnLedger, +) -> _ForwardClientFrame | _RefuseClientFrame: + verdict: Final = await _client_frame_verdict(websocket, checked_frame, gate) + match verdict: + case _RefuseClientFrame(refusal=refusal): + await refuse_websocket(websocket, refusal) + case _ForwardClientFrame(starts_turn=True, turn=turn, lane=lane): + ledger.queue_turn(lane, turn) + await upstream_ws.send(forwarded_frame) + case _ForwardClientFrame(): + await upstream_ws.send(forwarded_frame) + return verdict + + +_EXPLICIT_SUCCESS_KWARGS: Final = frozenset({"logging_obj", "passthrough_logging_payload", "custom_llm_provider"}) +WALKOUT_TURN_DRAIN_SECONDS: Final = 600.0 + + +def _metered_client_text(meter: WebSocketSessionMeter | None, client_text: str) -> int: + return 0 if meter is None else meter.measure(client_text) + + def _log_session_usage(logging_obj: LiteLLMLoggingObj, usage: WebSocketSessionUsage) -> None: logging_obj.model = usage.model # rebind-ok: the spend logger reads model and cost off the shared logging object logging_obj.model_call_details["model"] = usage.model # rebind-ok: same shared logging object @@ -2654,6 +3019,9 @@ async def websocket_passthrough_request( cost_per_request: float | None = None, accept_websocket: bool = True, setup_model_rewriter: Callable[[str], str] | None = None, + client_frame_model_gate: WebsocketFrameModelGate | None = None, + upstream_frame_kind: WebsocketUpstreamFrameClassifier | None = None, + walkout_turn_drain_seconds: float = WALKOUT_TURN_DRAIN_SECONDS, builtin_provider_route: bool = False, model: str | None = None, meter: WebSocketSessionMeter | None = None, @@ -2670,6 +3038,11 @@ async def websocket_passthrough_request( endpoint: The endpoint path (for logging purposes) cost_per_request: Optional field - cost per request to the target endpoint setup_model_rewriter: Optional rewrite of the setup frame's model before it reaches the upstream + client_frame_model_gate: Optional per-frame check of every model a client frame names + upstream_frame_kind: Optional sort of upstream frames into the one that ends a billable turn and the ones + that open the session, so each turn logs its own spend and keeps the session's provider-named model + walkout_turn_drain_seconds: How long to keep reading a turn the provider is still running after the client + left or was refused, so the provider's usage for it still reaches spend builtin_provider_route: Whether a built-in provider route serves the connection, so its URL names the provider model: The model the caller named before connecting, logged on a failure before the setup frame arrives meter: Prices the session from the client text frames delivered upstream, for an upstream that reports no usage @@ -2709,26 +3082,31 @@ async def websocket_passthrough_request( proxy_config=proxy_config, route_description="websocket_passthrough", ) - logging_obj: Final = Logging( - model="unknown", - messages=[{"role": "user", "content": "WebSocket connection"}], - stream=True, # WebSockets are inherently streaming - call_type="pass_through_endpoint", - start_time=start_time, - litellm_call_id=litellm_call_id, - function_id="websocket_passthrough", - dynamic_success_callbacks=team_callbacks.success_callbacks, - dynamic_failure_callbacks=team_callbacks.failure_callbacks, - kwargs=team_callbacks.logging_kwargs, - ) - # Create passthrough logging payload - passthrough_logging_payload: Final = PassthroughStandardLoggingPayload( - url=target, - request_body={}, # WebSocket doesn't have a traditional request body - request_method="WEBSOCKET", - cost_per_request=cost_per_request, - ) + def new_call_logging(call_start: datetime, call_id: str) -> LiteLLMLoggingObj: + return Logging( + model="unknown", + messages=[{"role": "user", "content": "WebSocket connection"}], + stream=True, + call_type="pass_through_endpoint", + start_time=call_start, + litellm_call_id=call_id, + function_id="websocket_passthrough", + dynamic_success_callbacks=team_callbacks.success_callbacks, + dynamic_failure_callbacks=team_callbacks.failure_callbacks, + kwargs=team_callbacks.logging_kwargs, + ) + + def new_call_payload() -> PassthroughStandardLoggingPayload: + return PassthroughStandardLoggingPayload( + url=target, + request_body={}, + request_method="WEBSOCKET", + cost_per_request=cost_per_request, + ) + + logging_obj: Final = new_call_logging(start_time, litellm_call_id) + passthrough_logging_payload: Final = new_call_payload() # Create a dummy request object for WebSocket connections to maintain compatibility # with the existing _init_kwargs_for_pass_through_endpoint function @@ -2750,38 +3128,136 @@ async def websocket_passthrough_request( ) # Initialize kwargs for logging using the same pattern as HTTP passthrough - kwargs: Final = HttpPassThroughEndpointHelpers.init_kwargs_for_pass_through_endpoint( - user_api_key_dict=user_api_key_dict, - _parsed_body={}, # WebSocket doesn't have a traditional request body - passthrough_logging_payload=passthrough_logging_payload, - litellm_call_id=litellm_call_id, - request=dummy_request, - logging_obj=logging_obj, + def init_call_kwargs( + token: UserAPIKeyAuth, + call_payload: PassthroughStandardLoggingPayload, + call_id: str, + call_logging: LiteLLMLoggingObj, + ) -> dict[str, object]: # mutable-ok: the Vertex relay stamps the model from setup frames into these kwargs + return HttpPassThroughEndpointHelpers.init_kwargs_for_pass_through_endpoint( + user_api_key_dict=token, + _parsed_body={}, # WebSocket doesn't have a traditional request body + passthrough_logging_payload=call_payload, + litellm_call_id=call_id, + request=dummy_request, + logging_obj=call_logging, + ) + + kwargs: Final = init_call_kwargs(user_api_key_dict, passthrough_logging_payload, litellm_call_id, logging_obj) + + def start_call_logging(call_logging: LiteLLMLoggingObj, call_kwargs: Mapping[str, object]) -> None: + call_logging.update_environment_variables( + model="unknown", + user="unknown", + optional_params={}, + litellm_params=dict(as_str_mapping(call_kwargs.get("litellm_params")) or {}), + call_type="pass_through_endpoint", + ) + call_logging.pre_call( + input=[{"role": "user", "content": "WebSocket connection"}], + api_key="", + additional_args={ + "complete_input_dict": {}, + "api_base": target, + "headers": upstream_headers, + }, + ) + + start_call_logging(logging_obj, kwargs) + + class MockWebSocketRequest: + def __init__(self, target_url: str): + self.method = "WEBSOCKET" + self.url = target_url + + class MockWebSocketResponse: + def __init__(self, target_url: str): + self.status_code = 200 + self.text = "WebSocket connection successful" + self.headers: dict[str, str] = {} + self.request = MockWebSocketRequest(target_url) + + ledger: Final = _WebsocketTurnLedger( + opening=_TurnAuthorization( + requested_model=_opening_model(target), + access_groups=tuple(user_api_key_dict.matched_model_access_groups or ()), + ), + started=start_time, ) - # Update logging environment variables - logging_obj.update_environment_variables( - model="unknown", - user="unknown", - optional_params={}, - litellm_params=dict(kwargs.get("litellm_params", {})), - call_type="pass_through_endpoint", - ) - logging_obj.model_call_details["litellm_call_id"] = litellm_call_id + def log_frames( + call_logging: LiteLLMLoggingObj, + call_payload: PassthroughStandardLoggingPayload, + call_kwargs: Mapping[str, object], + end_time: datetime, + lane: str | None, + in_row: Callable[[Mapping[str, object]], bool], + ) -> None: + all_frames: Final = list(websocket_messages) + frames: Final = [frame for frame in all_frames if in_row(frame)] + websocket_messages[:] = ( + () + if upstream_frame_kind is None + else ( + frame + for frame in all_frames + if not in_row(frame) or upstream_frame_kind(frame) in _SESSION_FRAMES_EVERY_ROW_CARRIES + ) + ) + logged_payload: Final[PassthroughStandardLoggingPayload] = { + **call_payload, + "response_body": frames, + "end_time": end_time, + } + if client_frame_model_gate is not None: + _stamp_billed_turn(call_kwargs, ledger.current(lane)) + success_kwargs: Final = { + key: value for key, value in call_kwargs.items() if key not in _EXPLICIT_SUCCESS_KWARGS + } + custom_llm_provider: Final = call_kwargs.get("custom_llm_provider") + GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue( + async_coroutine=pass_through_endpoint_logging.pass_through_async_success_handler( + httpx_response=MockWebSocketResponse(target), + response_body=frames, + url_route=endpoint or "", + result="websocket_connection_successful", + start_time=ledger.started, + end_time=end_time, + logging_obj=call_logging, + cache_hit=False, + request_body={}, + passthrough_logging_payload=logged_payload, + custom_llm_provider=custom_llm_provider if isinstance(custom_llm_provider, str) else None, + finalize_before_dispatch=False, + **success_kwargs, + ) + ) + bind_budget_reservation_to_callbacks(call_logging.litellm_params) - # Pre-call logging - logging_obj.pre_call( - input=[{"role": "user", "content": "WebSocket connection"}], - api_key="", - additional_args={ - "complete_input_dict": {}, - "api_base": target, - "headers": upstream_headers, - }, - ) + def log_turn(lane: str | None, in_row: Callable[[Mapping[str, object]], bool]) -> None: + end_time: Final = datetime.now() + if upstream_frame_kind is None: + log_frames(logging_obj, passthrough_logging_payload, kwargs, end_time, lane, in_row) + else: + turn_call_id: Final = str(uuid.uuid4()) + turn_logging: Final = new_call_logging(ledger.started, turn_call_id) + turn_payload: Final = new_call_payload() + turn_kwargs: Final = init_call_kwargs( + user_api_key_dict + if ledger.logged_turns == 0 + else user_api_key_dict.model_copy(update={"budget_reservation": None}), + turn_payload, + turn_call_id, + turn_logging, + ) + start_call_logging(turn_logging, turn_kwargs) + log_frames(turn_logging, turn_payload, turn_kwargs, end_time, lane, in_row) + ledger.close_turn(lane, end_time) ### CALL HOOKS ### - modify incoming data / reject request before calling the model - websocket_hook_data: Final = guardrail_request_data_with_streaming(MappingProxyType({}), is_streaming=True) + websocket_hook_data: Final = guardrail_request_data_with_streaming( + MappingProxyType({"litellm_call_id": litellm_call_id}), is_streaming=True + ) await proxy_logging_obj.pre_call_hook( user_api_key_dict=user_api_key_dict, data=websocket_hook_data, @@ -2808,7 +3284,6 @@ async def websocket_passthrough_request( message = await websocket.receive() message_type = message.get("type") if message_type == "websocket.disconnect": - await upstream_ws.close() break text_data: str | None = message.get("text") @@ -2871,11 +3346,29 @@ async def websocket_passthrough_request( ) # Not a JSON message or doesn't contain setup data - await upstream_ws.send(_rewrite_vertex_live_setup_model(text_data, setup_model_rewriter)) - if meter is not None: - metered_total += meter.measure(text_data) + if isinstance( + await _relay_client_frame( + websocket, + upstream_ws, + text_data, + _rewrite_vertex_live_setup_model(text_data, setup_model_rewriter), + client_frame_model_gate, + ledger, + ), + _RefuseClientFrame, + ): + break + metered_total += _metered_client_text(meter, text_data) elif bytes_data is not None: - await upstream_ws.send(bytes_data) + if isinstance( + await _relay_client_frame( + websocket, upstream_ws, bytes_data, bytes_data, client_frame_model_gate, ledger + ), + _RefuseClientFrame, + ): + break + if not _turn_owed(upstream_frame_kind, websocket_messages, ledger): + await upstream_ws.close() except asyncio.CancelledError: raise except Exception: @@ -2902,11 +3395,7 @@ async def websocket_passthrough_request( is_vertex_live: Final = bool(endpoint and "/vertex_ai/live" in endpoint) json_frame_ordinal: Final = count() - async def relay_upstream_frame(upstream_message: str | bytes) -> None: - if isinstance(upstream_message, bytes): - await websocket.send_bytes(upstream_message) - else: - await websocket.send_text(upstream_message) + def record_upstream_frame(upstream_message: str | bytes) -> None: message_data: Final = _json_object_frame(upstream_message) if message_data is None: return @@ -2914,6 +3403,43 @@ async def websocket_passthrough_request( _extract_vertex_live_model_from_setup_response(message_data) return websocket_messages.append(message_data) + match None if upstream_frame_kind is None else upstream_frame_kind(message_data): + case "opens_session": + ledger.open_session() + case "starts_turn": + ledger.start_turn() + case "updates_session" if upstream_frame_kind is not None: + websocket_messages[:] = _without_earlier_session_updates( + upstream_frame_kind, websocket_messages, message_data + ) + case "ends_turn" if upstream_frame_kind is not None: + if _frame_lane(message_data) not in _lanes_in_flight(upstream_frame_kind, websocket_messages): + ledger.start_turn() + log_turn(_frame_lane(message_data), _frames_of_lane(_frame_lane(message_data))) + case "bills_usage" if not _turn_in_flight(upstream_frame_kind, websocket_messages): + log_turn(_frame_lane(message_data), _frames_of_lane(_frame_lane(message_data))) + case "refuses_turn" if upstream_frame_kind is not None: + lane: Final = _frame_lane(message_data) + ledger.drop_refused_turn( + lane, lane in _lanes_in_flight(upstream_frame_kind, websocket_messages) + ) + case _: + pass + + async def relay_upstream_frame(upstream_message: str | bytes) -> None: + record_upstream_frame(upstream_message) + if not _client_socket_is_open(websocket): + if not _turn_owed(upstream_frame_kind, websocket_messages, ledger): + await upstream_ws.close() + return + try: + if isinstance(upstream_message, bytes): + await websocket.send_bytes(upstream_message) + else: + await websocket.send_text(upstream_message) + except WebSocketDisconnect: + if not _turn_owed(upstream_frame_kind, websocket_messages, ledger): + raise async def forward_upstream_to_client() -> Close | None: try: @@ -2922,6 +3448,11 @@ async def websocket_passthrough_request( except (ConnectionClosedOK, ConnectionClosedError) as e: verbose_proxy_logger.debug("Upstream WebSocket connection closed: %s", e) return e.rcvd + except WebSocketDisconnect: + verbose_proxy_logger.debug( + "WebSocket passthrough (%s): client left while upstream was still sending", endpoint + ) + return None except asyncio.CancelledError: verbose_proxy_logger.debug("asyncio.CancelledError in forward_upstream_to_client") raise @@ -2933,12 +3464,15 @@ async def websocket_passthrough_request( raise # Create tasks for bidirectional message forwarding - tasks: Final = [ - asyncio.create_task(forward_client_to_upstream()), - asyncio.create_task(forward_upstream_to_client()), - ] + client_task: Final = asyncio.create_task(forward_client_to_upstream()) + upstream_task: Final = asyncio.create_task(forward_upstream_to_client()) + tasks: Final = (client_task, upstream_task) - done, pending = await asyncio.wait(tasks, return_when=asyncio.FIRST_COMPLETED) + await asyncio.wait(tasks, return_when=asyncio.FIRST_COMPLETED) + if client_task.done() and _turn_owed(upstream_frame_kind, websocket_messages, ledger): + await asyncio.wait((upstream_task,), timeout=walkout_turn_drain_seconds) + done: Final = tuple(task for task in tasks if task.done()) + pending: Final = tuple(task for task in tasks if not task.done()) # Cancel remaining tasks for task in pending: @@ -2961,64 +3495,17 @@ async def websocket_passthrough_request( reason=truncated_close_reason(upstream_close.reason), ) - end_time: Final = datetime.now() - - # Update passthrough logging payload with response data - passthrough_logging_payload["response_body"] = websocket_messages - passthrough_logging_payload["end_time"] = end_time - - # Remove logging_obj from kwargs to avoid duplicate keyword argument - success_kwargs: Final = kwargs.copy() - success_kwargs.pop("logging_obj", None) - - # # Add user authentication context for database logging - # if user_api_key_dict: - # success_kwargs.setdefault('litellm_params', {}) - # success_kwargs['litellm_params'].update({ - # 'proxy_server_request': { - # 'body': { - # 'user': user_api_key_dict.user_id, - # 'team_id': user_api_key_dict.team_id, - # 'end_user_id': user_api_key_dict.end_user_id, - # } - # } - # }) - # # Also add the user_api_key for direct access - # success_kwargs['user_api_key'] = user_api_key_dict.api_key - - # Create a dummy httpx.Response for WebSocket connections - class MockWebSocketResponse: - def __init__(self, target_url: str): - self.status_code = 200 - self.text = "WebSocket connection successful" - self.headers: dict[str, str] = {} - self.request = MockWebSocketRequest(target_url) - - class MockWebSocketRequest: - def __init__(self, target_url: str): - self.method = "WEBSOCKET" - self.url = target_url - - mock_response: Final = MockWebSocketResponse(target) if meter is not None: _log_session_usage(logging_obj, meter.price(metered_total)) - # Use the same success handler as HTTP passthrough endpoints - GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue( - async_coroutine=pass_through_endpoint_logging.pass_through_async_success_handler( - httpx_response=mock_response, - response_body=websocket_messages, - url_route=endpoint or "", - result="websocket_connection_successful", - start_time=start_time, - end_time=end_time, - logging_obj=logging_obj, - cache_hit=False, - request_body={}, - **success_kwargs, + if _session_end_needs_a_row(upstream_frame_kind, websocket_messages, ledger): + lanes_in_flight: Final = ( + () if upstream_frame_kind is None else _lanes_in_flight(upstream_frame_kind, websocket_messages) ) - ) - bind_budget_reservation_to_callbacks(logging_obj.litellm_params) + *earlier_lanes, last_lane = lanes_in_flight or (None,) + for lane in earlier_lanes: + log_turn(lane, _frames_of_lane(lane)) + log_turn(last_lane, _every_frame) # Call the proxy logging success hook if proxy_logging_obj: @@ -3076,6 +3563,7 @@ async def websocket_passthrough_request( if _client_socket_is_open(websocket): await websocket.close(code=1011, reason="WebSocket passthrough error") finally: + await proxy_logging_obj.arelease_responses_websocket_slot(user_api_key_dict) if _client_socket_is_open(websocket): await websocket.close() diff --git a/litellm/proxy/pass_through_endpoints/success_handler.py b/litellm/proxy/pass_through_endpoints/success_handler.py index 8b09b1cf7cd..8f55906498b 100644 --- a/litellm/proxy/pass_through_endpoints/success_handler.py +++ b/litellm/proxy/pass_through_endpoints/success_handler.py @@ -1,5 +1,6 @@ import asyncio import json +from collections.abc import Mapping, Sequence from contextlib import suppress from datetime import datetime from types import MappingProxyType @@ -13,7 +14,7 @@ from litellm._logging import verbose_proxy_logger from litellm.constants import AZURE_SPEECH_CUSTOM_LLM_PROVIDER from litellm.litellm_core_utils.core_helpers import bind_budget_reservation_to_callbacks from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj -from litellm.proxy._types import PassThroughEndpointLoggingResultValues +from litellm.proxy._types import PassThroughEndpointLoggingResultValues, PassThroughEndpointLoggingTypedDict from litellm.types.passthrough_endpoints.pass_through_endpoints import ( PassthroughStandardLoggingPayload, ) @@ -70,6 +71,14 @@ def _safe_response_text(httpx_response: httpx.Response) -> str: return "" +def _websocket_frames( + response_body: Mapping[str, object] | Sequence[Mapping[str, object]] | None, +) -> tuple[Mapping[str, object], ...]: + return tuple( + message for message in (response_body if isinstance(response_body, list) else ()) if isinstance(message, dict) + ) + + class PassThroughEndpointLogging: def __init__( self, @@ -120,6 +129,7 @@ class PassThroughEndpointLogging: # Vertex AI Live API WebSocket self.TRACKED_VERTEX_AI_LIVE_ROUTES = ["/vertex_ai/live"] + self.TRACKED_OPENAI_WEBSOCKET_ROUTES: Final = ("/openai/", "/openai_passthrough/") @property def _log_dispatch(self) -> PassThroughLogDispatch: @@ -422,21 +432,18 @@ class PassThroughEndpointLogging: standard_logging_response_object = vertex_ai_live_handler_result["result"] kwargs = vertex_ai_live_handler_result["kwargs"] - elif DeepgramListenPassthroughLoggingHandler.is_deepgram_listen_route(url_route): - deepgram_handler_result: Final = ( - DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( - websocket_messages=tuple( - message - for message in (response_body if isinstance(response_body, list) else ()) - if isinstance(message, dict) - ), - logging_obj=logging_obj, - upstream_url=str(httpx_response.request.url), - kwargs=kwargs, - ) + elif ( + websocket_relay_result := self._websocket_relay_logging_result( + url_route=url_route, + websocket_frames=_websocket_frames(response_body), + logging_obj=logging_obj, + httpx_response=httpx_response, + start_time=start_time, + kwargs=kwargs, ) - standard_logging_response_object = deepgram_handler_result["result"] # rebind-ok: elif-chain - kwargs = deepgram_handler_result["kwargs"] # rebind-ok: elif-chain contract + ) is not None: + standard_logging_response_object = websocket_relay_result["result"] # rebind-ok: elif-chain + kwargs = websocket_relay_result["kwargs"] # rebind-ok: elif-chain contract elif FalAIPassthroughLoggingHandler.is_fal_ai_route(url_route, custom_llm_provider): fal_ai_handler_result: Final = FalAIPassthroughLoggingHandler().fal_ai_passthrough_handler( response_body=response_body if isinstance(response_body, dict) else MappingProxyType({}), @@ -452,6 +459,36 @@ class PassThroughEndpointLogging: return_dict["kwargs"] = kwargs return return_dict + def _websocket_relay_logging_result( + self, + url_route: str, + websocket_frames: tuple[Mapping[str, object], ...], + logging_obj: LiteLLMLoggingObj, + httpx_response: httpx.Response, + start_time: datetime, + kwargs: Mapping[str, object], + ) -> PassThroughEndpointLoggingTypedDict | None: + if self.is_openai_websocket_route(url_route): + from .llm_provider_handlers.openai_websocket_passthrough_logging_handler import ( + OpenAIWebsocketPassthroughLoggingHandler, + ) + + return OpenAIWebsocketPassthroughLoggingHandler().openai_websocket_passthrough_handler( + websocket_messages=websocket_frames, + logging_obj=logging_obj, + url_route=url_route, + start_time=start_time, + kwargs=kwargs, + ) + if DeepgramListenPassthroughLoggingHandler.is_deepgram_listen_route(url_route): + return DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=websocket_frames, + logging_obj=logging_obj, + upstream_url=str(httpx_response.request.url), + kwargs=kwargs, + ) + return None + async def pass_through_async_success_handler( self, httpx_response: httpx.Response, @@ -625,6 +662,9 @@ class PassThroughEndpointLogging: return True return False + def is_openai_websocket_route(self, url_route: str) -> bool: + return url_route.startswith(self.TRACKED_OPENAI_WEBSOCKET_ROUTES) + def is_cursor_route(self, url_route: str, custom_llm_provider: str | None = None): """Check if the URL route is a Cursor Cloud Agents API route.""" if custom_llm_provider == "cursor": diff --git a/litellm/proxy/spend_tracking/budget_reservation.py b/litellm/proxy/spend_tracking/budget_reservation.py index fc6fb266d04..dcfd6148f54 100644 --- a/litellm/proxy/spend_tracking/budget_reservation.py +++ b/litellm/proxy/spend_tracking/budget_reservation.py @@ -432,7 +432,7 @@ async def invalidate_budget_reservation_counters( async def release_or_invalidate_budget_reservation( - budget_reservation: dict | None, # mutable-ok: stamps finalized on the caller's shared reservation dict + budget_reservation: dict[str, object] | None, # mutable-ok: stamps finalized on the caller's reservation ) -> None: """Reconcile a still-open reservation on a terminal path that settles no cost. diff --git a/litellm/types/passthrough_endpoints/pass_through_endpoints.py b/litellm/types/passthrough_endpoints/pass_through_endpoints.py index 2f269dc87d9..b6bab77ab33 100644 --- a/litellm/types/passthrough_endpoints/pass_through_endpoints.py +++ b/litellm/types/passthrough_endpoints/pass_through_endpoints.py @@ -1,9 +1,11 @@ +from collections.abc import Mapping, Sequence from dataclasses import dataclass +from datetime import datetime from enum import Enum -from typing import Final, Protocol +from typing import Final, Literal, Protocol, TypeAlias from pydantic import Field -from typing_extensions import TypedDict +from typing_extensions import ReadOnly, TypedDict from litellm.types.llms.base import LiteLLMPydanticObjectBase @@ -28,6 +30,10 @@ LITELLM_PASS_THROUGH_ENDPOINT_MARKER: Final = "__litellm_pass_through_endpoint__ SETTLES_BACKGROUND_RESPONSE_KWARG: Final = "settles_background_response" +WebsocketUpstreamFrameKind: TypeAlias = Literal[ + "starts_turn", "ends_turn", "opens_session", "updates_session", "bills_usage", "refuses_turn", "relayed" +] + class BackgroundResponseSettlement(str, Enum): SETTLES = "settles" @@ -63,7 +69,7 @@ class PassthroughStandardLoggingPayload(TypedDict, total=False): """ The body of the request """ - response_body: dict | None # only tracked for non-streaming responses + response_body: dict | Sequence[Mapping[str, object]] | None # only tracked for non-streaming responses """ The body of the response """ @@ -75,6 +81,8 @@ class PassthroughStandardLoggingPayload(TypedDict, total=False): Optional field, we use this for cost tracking only if it's set. """ + end_time: ReadOnly[datetime] + @dataclass(frozen=True, slots=True) class WebSocketSessionUsage: diff --git a/tests/unit/proxy/auth/test_auth_checks_object_access_and_lookup.py b/tests/unit/proxy/auth/test_auth_checks_object_access_and_lookup.py index 8e9e87f370a..7a060a214f7 100644 --- a/tests/unit/proxy/auth/test_auth_checks_object_access_and_lookup.py +++ b/tests/unit/proxy/auth/test_auth_checks_object_access_and_lookup.py @@ -10998,139 +10998,6 @@ async def test_key_budget_error_keeps_the_masked_key_name(key_name): assert f"Key=prod-key ({key_name}) Current cost" in message -class _UntouchedPrisma: - def __getattr__(self, name: str) -> object: - raise AssertionError(f"database reached through {name}") - - -class _MissingUserPrisma: - class db: - class litellm_usertable: - @staticmethod - async def find_unique(where: dict[str, str], include: dict[str, bool]) -> None: - return None - - -@pytest.mark.asyncio -async def test_enforced_model_allowlists_treats_a_missing_user_row_as_unrestricted(): - from litellm.proxy.auth.auth_checks import enforced_model_allowlists - from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache - from litellm.proxy.utils import ProxyLogging - - cache = UserApiKeyCache() - scopes = await enforced_model_allowlists( - valid_token=UserAPIKeyAuth(token="hashed-fake", user_id="default_user_id"), - prisma_client=_MissingUserPrisma(), - user_api_key_cache=cache, - proxy_logging_obj=ProxyLogging(user_api_key_cache=cache), - ) - - assert [list(scope) for scope in scopes] == [[], [], [], [], []] - - -class _UnreachableUserPrisma: - class db: - class litellm_usertable: - @staticmethod - async def find_unique(where: dict[str, str], include: dict[str, bool]) -> None: - raise RuntimeError("database gone") - - -@pytest.mark.asyncio -async def test_enforced_model_allowlists_surfaces_a_failed_user_lookup(): - from litellm.proxy.auth.auth_checks import enforced_model_allowlists - from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache - from litellm.proxy.utils import ProxyLogging - - cache = UserApiKeyCache() - with pytest.raises(ValueError, match="database gone"): - await enforced_model_allowlists( - valid_token=UserAPIKeyAuth(token="hashed-fake", user_id="user-fake"), - prisma_client=_UnreachableUserPrisma(), - user_api_key_cache=cache, - proxy_logging_obj=ProxyLogging(user_api_key_cache=cache), - ) - - -@pytest.mark.asyncio -async def test_enforced_model_allowlists_reads_every_level_from_cache(): - from litellm.proxy._types import ( - LiteLLM_BudgetTable, - LiteLLM_ProjectTableCachedObj, - LiteLLM_TeamMembership, - LiteLLM_TeamTableCachedObj, - LiteLLM_UserTable, - ) - from litellm.proxy.auth.auth_checks import enforced_model_allowlists - from litellm.proxy.common_utils.user_api_key_cache import ( - UserApiKeyCache, - team_membership_reservation_cache_key, - ) - from litellm.proxy.utils import ProxyLogging - - cache = UserApiKeyCache() - proxy_logging_obj = ProxyLogging(user_api_key_cache=cache) - await cache.async_set_cache( - key="team_id:team-fake", value=LiteLLM_TeamTableCachedObj(team_id="team-fake", models=["gpt-4o"]) - ) - await cache.async_set_cache( - key=team_membership_reservation_cache_key(user_id="user-fake", team_id="team-fake"), - value=LiteLLM_TeamMembership( - user_id="user-fake", - team_id="team-fake", - litellm_budget_table=LiteLLM_BudgetTable(allowed_models=["gpt-4o-mini"]), - ), - ) - await cache.async_set_cache( - key="project_id:project-fake", - value=LiteLLM_ProjectTableCachedObj(project_id="project-fake", models=["gpt-4.1"]), - ) - await cache.async_set_cache(key="user-fake", value=LiteLLM_UserTable(user_id="user-fake", models=["o3"])) - prisma_client = _UntouchedPrisma() - - team_scoped = await enforced_model_allowlists( - valid_token=UserAPIKeyAuth( - token="hashed-fake", - models=["all-team-models"], - team_models=["gpt-4o", "gpt-4o-mini"], - user_id="user-fake", - team_id="team-fake", - project_id="project-fake", - ), - prisma_client=prisma_client, - user_api_key_cache=cache, - proxy_logging_obj=proxy_logging_obj, - ) - personal = await enforced_model_allowlists( - valid_token=UserAPIKeyAuth(token="hashed-fake", user_id="user-fake"), - prisma_client=prisma_client, - user_api_key_cache=cache, - proxy_logging_obj=proxy_logging_obj, - ) - without_database = await enforced_model_allowlists( - valid_token=UserAPIKeyAuth( - token="hashed-fake", - models=["gpt-4o"], - team_models=["gpt-4o-mini"], - user_id="user-fake", - team_id="team-fake", - ), - prisma_client=None, - user_api_key_cache=cache, - proxy_logging_obj=proxy_logging_obj, - ) - - assert [list(scope) for scope in team_scoped] == [ - ["gpt-4o", "gpt-4o-mini"], - ["gpt-4o"], - ["gpt-4o-mini"], - [], - ["gpt-4.1"], - ] - assert [list(scope) for scope in personal] == [[], [], [], ["o3"], []] - assert [list(scope) for scope in without_database] == [["gpt-4o"], ["gpt-4o-mini"]] - - @pytest.mark.asyncio @pytest.mark.parametrize("channel", ["team", "key"]) async def test_access_group_model_fallback_uses_the_injected_database(channel: str) -> None: diff --git a/tests/unit/proxy/common_utils/test_realtime_cache.py b/tests/unit/proxy/common_utils/test_realtime_cache.py index 6f5751fa923..9e76caff054 100644 --- a/tests/unit/proxy/common_utils/test_realtime_cache.py +++ b/tests/unit/proxy/common_utils/test_realtime_cache.py @@ -1,4 +1,5 @@ -from typing import Any, cast +import json +from typing import Any, Final, cast import pytest @@ -40,6 +41,12 @@ def test_realtime_request_body_caches_each_model_separately(): assert gpt4o_body_first is not gpt4o_mini_body +def test_realtime_request_body_keeps_a_model_name_with_quotes_inside_its_own_field(): + body: Final = realtime_request_body('gpt-4o", "user": "someone-else') + + assert json.loads(body) == {"model": 'gpt-4o", "user": "someone-else'} + + def test_realtime_query_params_template_caches_each_pair_separately(): params_with_intent_first = _realtime_query_params_template("gpt-4o", "intent-a") params_with_intent_second = _realtime_query_params_template("gpt-4o", "intent-a") diff --git a/tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_openai_websocket_passthrough_logging_handler.py b/tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_openai_websocket_passthrough_logging_handler.py new file mode 100644 index 00000000000..e7b4d087ed3 --- /dev/null +++ b/tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_openai_websocket_passthrough_logging_handler.py @@ -0,0 +1,520 @@ +import uuid +from collections.abc import Mapping, Sequence +from datetime import datetime +from types import MappingProxyType +from typing import Final +from unittest.mock import MagicMock + +import httpx +import pytest + +import litellm +from litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_passthrough_logging_handler import ( + OpenAIPassthroughLoggingHandler, +) +from litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_websocket_passthrough_logging_handler import ( + OpenAIWebsocketPassthroughLoggingHandler, + openai_websocket_frame_kind, +) +from litellm.proxy.spend_tracking.spend_tracking_utils import get_spend_logs_id +from litellm.types.utils import CallTypes + +REALTIME_MODEL: Final = "gpt-realtime-2.1" +RESPONSES_MODEL: Final = "gpt-5.4-mini" +REASONING_MODEL: Final = "gpt-5.4" +TRANSCRIPTION_MODEL: Final = "gpt-4o-mini-transcribe" +SESSION_START: Final = datetime(2026, 1, 1, 12, 0, 0) + + +def _logging_obj() -> MagicMock: + logging_obj: Final = MagicMock(model_call_details={}, litellm_call_id=str(uuid.uuid4())) + return logging_obj + + +def _session_created(model: str) -> dict[str, object]: + return {"type": "session.created", "session": {"id": "sess_fake", "model": model}} + + +def _realtime_response_done(text_in: int, audio_in: int, text_out: int, audio_out: int) -> dict[str, object]: + return { + "type": "response.done", + "response": { + "id": "resp_fake", + "status": "completed", + "usage": { + "input_tokens": text_in + audio_in, + "output_tokens": text_out + audio_out, + "total_tokens": text_in + audio_in + text_out + audio_out, + "input_token_details": {"text_tokens": text_in, "audio_tokens": audio_in, "cached_tokens": 0}, + "output_token_details": {"text_tokens": text_out, "audio_tokens": audio_out}, + }, + }, + } + + +def _responses_completed( + model: str, + input_tokens: int, + output_tokens: int, + event: str = "response.completed", + served: Mapping[str, object] = MappingProxyType({}), +) -> dict[str, object]: + return { + "type": event, + "response": { + "id": "resp_fake", + "object": "response", + "created_at": 1761700000, + "model": model, + "status": "completed", + "output": [], + **served, + "usage": { + "input_tokens": input_tokens, + "output_tokens": output_tokens, + "total_tokens": input_tokens + output_tokens, + "input_tokens_details": {"cached_tokens": 0}, + "output_tokens_details": {"reasoning_tokens": 0}, + }, + }, + } + + +def _handle( + messages: Sequence[Mapping[str, object]], + kwargs: Mapping[str, object] = MappingProxyType({}), + logging_obj: MagicMock | None = None, + url_route: str = "/openai/v1/realtime", +): + return OpenAIWebsocketPassthroughLoggingHandler().openai_websocket_passthrough_handler( + websocket_messages=messages, + logging_obj=logging_obj if logging_obj is not None else _logging_obj(), + url_route=url_route, + start_time=SESSION_START, + kwargs=kwargs, + ) + + +def _realtime_cost(model: str, text_in: int, audio_in: int, text_out: int, audio_out: int) -> float: + prices: Final = litellm.model_cost[model] + return ( + text_in * prices["input_cost_per_token"] + + audio_in * prices["input_cost_per_audio_token"] + + text_out * prices["output_cost_per_token"] + + audio_out * prices["output_cost_per_audio_token"] + ) + + +def test_realtime_session_is_priced_on_the_model_the_session_reported(): + logging_obj: Final = _logging_obj() + messages: Final = [ + _session_created(REALTIME_MODEL), + {"type": "response.created", "response": {"id": "resp_fake"}}, + _realtime_response_done(text_in=33, audio_in=120, text_out=16, audio_out=240), + ] + + handled: Final = _handle(messages, logging_obj=logging_obj) + + expected_cost: Final = _realtime_cost(REALTIME_MODEL, 33, 120, 16, 240) + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert handled["kwargs"]["model"] == REALTIME_MODEL + assert handled["kwargs"]["custom_llm_provider"] == "openai" + assert handled["result"].model == REALTIME_MODEL + assert handled["result"].usage.total_tokens == 409 + assert logging_obj.model_call_details["response_cost"] == pytest.approx(expected_cost) + assert logging_obj.model_call_details["model"] == REALTIME_MODEL + + +def test_realtime_cost_covers_every_turn_of_the_session(): + one_turn: Final = _handle([_session_created(REALTIME_MODEL), _realtime_response_done(10, 0, 5, 0)]) + two_turns: Final = _handle( + [ + _session_created(REALTIME_MODEL), + _realtime_response_done(10, 0, 5, 0), + _realtime_response_done(10, 0, 5, 0), + ] + ) + + assert two_turns["kwargs"]["response_cost"] == pytest.approx(2 * one_turn["kwargs"]["response_cost"]) + assert two_turns["result"].usage.total_tokens == 30 + + +def test_responses_turn_is_priced_on_the_model_the_response_reported(): + handled: Final = _handle( + [_responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5)], + url_route="/openai_passthrough/v1/responses", + ) + + prices: Final = litellm.model_cost[RESPONSES_MODEL] + expected_cost: Final = 13 * prices["input_cost_per_token"] + 5 * prices["output_cost_per_token"] + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert handled["kwargs"]["model"] == RESPONSES_MODEL + assert handled["result"].usage.prompt_tokens == 13 + assert handled["result"].usage.completion_tokens == 5 + + +def test_responses_cost_covers_every_turn_of_the_connection(): + handled: Final = _handle( + [ + _responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5), + _responses_completed(RESPONSES_MODEL, input_tokens=20, output_tokens=7), + ], + url_route="/openai_passthrough/v1/responses", + ) + + prices: Final = litellm.model_cost[RESPONSES_MODEL] + expected_cost: Final = 33 * prices["input_cost_per_token"] + 12 * prices["output_cost_per_token"] + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert handled["result"].usage.total_tokens == 45 + + +def test_a_model_the_caller_asked_for_never_prices_a_session_the_provider_never_named(): + """ + The caller picks what to ask for, and the provider picks what to run. Only + the provider's own frames may set the price of a session, so a model the + caller put in the request cannot stand in for one the provider never named. + """ + logging_obj: Final = _logging_obj() + + handled: Final = _handle( + [_realtime_response_done(text_in=10, audio_in=0, text_out=5, audio_out=0)], + kwargs=MappingProxyType({"model": REALTIME_MODEL}), + logging_obj=logging_obj, + ) + + assert handled["result"] is None + assert handled["kwargs"] == {"model": REALTIME_MODEL} + assert logging_obj.model_call_details == {} + + +def test_each_responses_turn_is_priced_on_the_model_that_turn_ran(): + """ + One socket can run several models one after another, so a connection that + starts on a big model and finishes on a small one has to pay the big + model's price for the turn that used it. + """ + handled: Final = _handle( + [ + _responses_completed(REASONING_MODEL, input_tokens=13, output_tokens=5), + _responses_completed(RESPONSES_MODEL, input_tokens=20, output_tokens=7), + ], + url_route="/openai_passthrough/v1/responses", + ) + + first: Final = litellm.model_cost[REASONING_MODEL] + second: Final = litellm.model_cost[RESPONSES_MODEL] + expected_cost: Final = ( + 13 * first["input_cost_per_token"] + + 5 * first["output_cost_per_token"] + + 20 * second["input_cost_per_token"] + + 7 * second["output_cost_per_token"] + ) + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert handled["result"].usage.total_tokens == 45 + + +def test_a_responses_turn_that_names_no_model_is_priced_on_the_session_model(): + handled: Final = _handle( + [ + _session_created(RESPONSES_MODEL), + _responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5, served={"model": None}), + ], + url_route="/openai_passthrough/v1/responses", + ) + + prices: Final = litellm.model_cost[RESPONSES_MODEL] + expected_cost: Final = 13 * prices["input_cost_per_token"] + 5 * prices["output_cost_per_token"] + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + + +def test_a_session_nobody_named_a_model_for_is_left_unpriced(): + logging_obj: Final = _logging_obj() + handled: Final = _handle( + [{"type": "response.done", "response": {"usage": {"input_tokens": 10, "output_tokens": 5}}}], + kwargs=MappingProxyType({"litellm_call_id": "call-fake"}), + logging_obj=logging_obj, + ) + + assert handled["result"] is None + assert handled["kwargs"] == {"litellm_call_id": "call-fake"} + assert logging_obj.model_call_details == {} + + +def test_a_session_that_reported_no_usage_still_records_its_model(): + logging_obj: Final = _logging_obj() + + handled: Final = _handle([_session_created(REALTIME_MODEL)], logging_obj=logging_obj) + + assert handled["result"] is None + assert handled["kwargs"] == {"model": REALTIME_MODEL, "custom_llm_provider": "openai"} + assert logging_obj.model_call_details["model"] == REALTIME_MODEL + assert logging_obj.model_call_details["custom_llm_provider"] == "openai" + + +def test_a_responses_turn_without_readable_usage_is_not_billed(): + handled: Final = _handle( + [{"type": "response.completed", "response": {"model": RESPONSES_MODEL, "status": "completed"}}], + url_route="/openai_passthrough/v1/responses", + ) + + assert handled["result"] is None + assert handled["kwargs"] == {"model": RESPONSES_MODEL, "custom_llm_provider": "openai"} + + +def test_the_price_of_a_multi_model_connection_reaches_the_spend_row(): + """ + The logging layer prices whatever result it is handed, and one result can + only name one model, so a connection that ran two models has to carry the + price it already worked out or the cheaper model pays for both turns. The + row it lands in names the model the connection opened on. + """ + handled: Final = _handle( + [ + _responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5), + _responses_completed(REASONING_MODEL, input_tokens=20, output_tokens=7), + ], + url_route="/openai_passthrough/v1/responses", + ) + + logged_cost: Final = litellm.response_cost_calculator( + response_object=handled["result"], + model=handled["kwargs"]["model"], + custom_llm_provider="openai", + call_type="completion", + optional_params={}, + ) + + session_prices: Final = litellm.model_cost[RESPONSES_MODEL] + priced_on_the_session_model: Final = ( + 33 * session_prices["input_cost_per_token"] + 12 * session_prices["output_cost_per_token"] + ) + assert logged_cost == pytest.approx(handled["kwargs"]["response_cost"]) + assert logged_cost > priced_on_the_session_model + assert handled["kwargs"]["model"] == RESPONSES_MODEL + + +def test_turns_that_start_in_the_same_instant_each_keep_their_own_spend_row(): + """ + Spend logs key each row by the id its result carries and skip a repeated id + without a word, so two turns that start in the same instant, on one socket + or on two, each need an id of their own or one of them is never billed. + """ + turns: Final = [ + _handle( + [_responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5)], + url_route="/openai_passthrough/v1/responses", + ) + for _ in range(2) + ] + + spend_log_ids: Final = { + get_spend_logs_id(CallTypes.pass_through.value, turn["result"].model_dump(), turn["kwargs"]) for turn in turns + } + assert len(spend_log_ids) == 2 + assert None not in spend_log_ids + + +def test_a_turn_the_caller_walked_out_of_still_names_the_model_it_ran(): + """ + A client that closes before response.completed leaves no usage behind to + bill, but response.created already named the model, so the row says which + model answered instead of falling back to unknown. + """ + logging_obj: Final = _logging_obj() + + handled: Final = _handle( + [{"type": "response.created", "response": {"id": "resp_fake", "model": REASONING_MODEL}}], + logging_obj=logging_obj, + url_route="/openai_passthrough/v1/responses", + ) + + assert handled["result"] is None + assert handled["kwargs"] == {"model": REASONING_MODEL, "custom_llm_provider": "openai"} + assert logging_obj.model_call_details["model"] == REASONING_MODEL + + +def _transcription_session(model: str) -> tuple[dict[str, object], ...]: + return ( + {"type": "session.created", "session": {"type": "transcription", "audio": {"input": {"transcription": None}}}}, + { + "type": "session.updated", + "session": {"type": "transcription", "audio": {"input": {"transcription": {"model": model}}}}, + }, + ) + + +def _transcription_completed(usage: Mapping[str, object]) -> dict[str, object]: + return { + "type": "conversation.item.input_audio_transcription.completed", + "item_id": "item_fake", + "transcript": "hello", + "usage": usage, + } + + +@pytest.mark.parametrize( + ("usage", "billed_units"), + [ + ({"type": "duration", "seconds": 3}, ((3, "input_cost_per_second"),)), + ( + { + "type": "tokens", + "input_tokens": 30, + "output_tokens": 11, + "input_token_details": {"text_tokens": 0, "audio_tokens": 30}, + }, + ((30, "input_cost_per_audio_token"), (11, "output_cost_per_token")), + ), + ], +) +def test_a_transcription_session_is_billed_for_every_clip_it_transcribed(usage, billed_units): + handled: Final = _handle( + [ + *_transcription_session(TRANSCRIPTION_MODEL), + _transcription_completed(usage), + _transcription_completed(usage), + ], + url_route="/openai_passthrough/v1/realtime", + ) + + logged_cost: Final = litellm.response_cost_calculator( + response_object=handled["result"], + model=handled["kwargs"]["model"], + custom_llm_provider="openai", + call_type="completion", + optional_params={}, + ) + + prices: Final = litellm.model_cost[TRANSCRIPTION_MODEL] + expected_cost: Final = 2 * sum(units * prices[rate] for units, rate in billed_units) + assert expected_cost > 0 + assert handled["kwargs"]["model"] == TRANSCRIPTION_MODEL + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert logged_cost == pytest.approx(expected_cost) + + +def test_a_conversation_that_also_transcribes_the_caller_is_logged_on_its_own_model(): + handled: Final = _handle( + [ + { + "type": "session.created", + "session": { + "model": REALTIME_MODEL, + "audio": {"input": {"transcription": {"model": TRANSCRIPTION_MODEL}}}, + }, + }, + _realtime_response_done(text_in=10, audio_in=0, text_out=5, audio_out=0), + ] + ) + + assert handled["kwargs"]["model"] == REALTIME_MODEL + + +@pytest.mark.parametrize("event", ["response.incomplete", "response.failed"]) +def test_a_responses_turn_that_ended_early_is_billed_for_the_tokens_it_used(event): + """ + A turn cut short by max_output_tokens or a provider failure still reports + the tokens it burned, and OpenAI bills them, so the row has to as well. + """ + handled: Final = _handle( + [_responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5, event=event)], + url_route="/openai_passthrough/v1/responses", + ) + + prices: Final = litellm.model_cost[RESPONSES_MODEL] + expected_cost: Final = 13 * prices["input_cost_per_token"] + 5 * prices["output_cost_per_token"] + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert handled["result"].usage.total_tokens == 18 + + +def test_a_failed_responses_turn_that_reported_no_usage_is_not_billed(): + handled: Final = _handle( + [{"type": "response.failed", "response": {"model": RESPONSES_MODEL, "status": "failed", "usage": None}}], + url_route="/openai_passthrough/v1/responses", + ) + + assert handled["result"] is None + assert handled["kwargs"] == {"model": RESPONSES_MODEL, "custom_llm_provider": "openai"} + + +@pytest.mark.parametrize( + ("turn_model", "requested_model", "priced_as"), + [ + ("gpt-unreleased-qa-snapshot", RESPONSES_MODEL, RESPONSES_MODEL), + ("gpt-5.4-mini-qa-preview", RESPONSES_MODEL, RESPONSES_MODEL), + (RESPONSES_MODEL, REASONING_MODEL, RESPONSES_MODEL), + ], + ids=["unmapped_model", "model_the_cost_map_prices_at_zero", "priced_turn_model"], +) +def test_a_responses_turn_falls_back_to_the_requested_model_only_when_its_own_has_no_price( + turn_model: str, requested_model: str, priced_as: str +): + """ + The provider can answer under a model name the cost map does not price + yet. That turn still ran the model the key was allowed to ask for, so it is + billed at that model's price instead of $0, while a turn whose own model + has a price keeps it. + """ + handled: Final = _handle( + [_responses_completed(turn_model, input_tokens=13, output_tokens=5)], + kwargs=MappingProxyType({"litellm_params": {"metadata": {"model_group": requested_model}}}), + url_route="/openai_passthrough/v1/responses", + ) + + prices: Final = litellm.model_cost[priced_as] + expected_cost: Final = 13 * prices["input_cost_per_token"] + 5 * prices["output_cost_per_token"] + assert expected_cost > 0 + assert handled["kwargs"]["response_cost"] == pytest.approx(expected_cost) + assert handled["kwargs"]["model"] == turn_model + + +def _http_passthrough_cost(response: Mapping[str, object]) -> float: + http_response: Final = httpx.Response( + 200, json=response, request=httpx.Request("POST", "https://api.openai.com/v1/responses") + ) + _, cost = OpenAIPassthroughLoggingHandler._build_responses_api_response_and_cost( + RESPONSES_MODEL, http_response, _logging_obj(), "openai" + ) + return cost + + +@pytest.mark.parametrize( + "served", + [ + {"service_tier": "priority"}, + {"output": [{"type": "web_search_call", "id": "ws_fake", "status": "completed", "action": {"type": "search"}}]}, + ], + ids=["priority_tier", "hosted_web_search"], +) +def test_a_responses_turn_costs_what_the_http_passthrough_charges_for_the_same_response(served): + """ + The provider bills a turn by the tier that served it and the hosted tools it + called, not only its tokens, so a websocket turn has to cost exactly what the + same response costs over HTTP, never the cheaper token-only price. + """ + turn: Final = _responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5, served=served) + plain_turn: Final = _responses_completed(RESPONSES_MODEL, input_tokens=13, output_tokens=5) + + handled: Final = _handle([turn], url_route="/openai_passthrough/v1/responses") + + assert handled["kwargs"]["response_cost"] == pytest.approx(_http_passthrough_cost(turn["response"])) + assert handled["kwargs"]["response_cost"] > _http_passthrough_cost(plain_turn["response"]) + + +@pytest.mark.parametrize( + ("frame_type", "kind"), + [ + ("response.completed", "ends_turn"), + ("response.incomplete", "ends_turn"), + ("response.failed", "ends_turn"), + ("response.done", "ends_turn"), + ("session.created", "opens_session"), + ("session.updated", "updates_session"), + ("conversation.item.input_audio_transcription.completed", "bills_usage"), + ("response.created", "starts_turn"), + ("response.output_text.delta", "relayed"), + ("error", "refuses_turn"), + ], +) +def test_each_frame_is_classified_by_what_it_does_to_a_turn(frame_type, kind): + assert openai_websocket_frame_kind({"type": frame_type}) == kind diff --git a/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py index 4fe358c38a3..824f04aee75 100644 --- a/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -1,5 +1,6 @@ import asyncio import base64 +import contextvars import gzip import json import logging @@ -7,20 +8,20 @@ import os import sys import uuid import zlib -from collections.abc import AsyncIterator, Callable, Mapping +from collections.abc import AsyncIterator, Callable, Mapping, Sequence from contextlib import ExitStack, contextmanager from dataclasses import dataclass from datetime import datetime from io import BytesIO from types import MappingProxyType, ModuleType, SimpleNamespace from typing import Final, Optional, cast -from unittest.mock import AsyncMock, MagicMock, patch +from unittest.mock import AsyncMock, MagicMock, call, patch from uuid import UUID import httpx import pytest import respx -from fastapi import APIRouter, FastAPI, HTTPException, Request, Response, UploadFile +from fastapi import APIRouter, FastAPI, HTTPException, Request, Response, UploadFile, WebSocketDisconnect from fastapi.responses import StreamingResponse from fastapi.routing import APIRoute from fastapi.testclient import TestClient @@ -35,9 +36,13 @@ from litellm.constants import ANTHROPIC_COMPATIBLE_MAX_USER_ID_LENGTH, ANTHROPIC from litellm.integrations.custom_guardrail import CustomGuardrail from litellm.integrations.custom_logger import CustomLogger from litellm.integrations.prometheus import PrometheusLogger +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.proxy._lazy_features import LazyFeature, attach_lazy_features from litellm.proxy._types import ProxyException, UserAPIKeyAuth +from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError +from litellm.proxy.hooks.parallel_request_limiter import PROXY_MaxParallelRequestsHandler +from litellm.proxy.hooks.parallel_request_limiter_v3 import PROXY_MaxParallelRequestsHandler_v3 from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import router as llm_passthrough_router from litellm.proxy.pass_through_endpoints.llm_provider_handlers.vertex_passthrough_logging_handler import ( @@ -46,9 +51,15 @@ from litellm.proxy.pass_through_endpoints.llm_provider_handlers.vertex_passthrou from litellm.proxy.pass_through_endpoints.pass_through_endpoints import ( DEFAULT_PASS_THROUGH_REQUEST_TIMEOUT_SECONDS, LITELLM_PASS_THROUGH_CUSTOM_BODY_STATE_KEY, + STORED_PROMPT_PROBE_MODEL, + WALKOUT_TURN_DRAIN_SECONDS, HttpPassThroughEndpointHelpers, InitPassThroughEndpointHelpers, SafeRouteAdder, + WebsocketFrameModelAllowed, + WebsocketFrameModelGate, + WebsocketFrameModelRefused, + WebsocketFrameModelVerdict, _build_passthrough_failure_request_payload, _registered_pass_through_routes, _truncate_upstream_error_body, @@ -63,12 +74,15 @@ from litellm.proxy.pass_through_endpoints.pass_through_endpoints import ( set_env_variables_in_header, websocket_passthrough_request, ) +from litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_websocket_passthrough_logging_handler import ( + openai_websocket_frame_kind, +) from litellm.proxy.pass_through_endpoints.success_handler import ( PassThroughEndpointLogging, ) from litellm.proxy.route_llm_request import ProxyModelNotFoundError from litellm.proxy.spend_tracking.spend_tracking_utils import get_spend_logs_id -from litellm.proxy.utils import ProxyLogging +from litellm.proxy.utils import InternalUsageCache, ProxyLogging, hash_token from litellm.types import utils as types_utils from litellm.types.guardrails import GuardrailEventHooks from litellm.types.utils import CallTypes @@ -7382,6 +7396,7 @@ async def test_websocket_passthrough_forwards_non_ascii_first_frame(): mock_proxy_logging.pre_call_hook = AsyncMock(return_value={}) mock_proxy_logging.post_call_success_hook = AsyncMock() mock_proxy_logging.post_call_failure_hook = AsyncMock() + mock_proxy_logging.arelease_responses_websocket_slot = AsyncMock() mock_worker.ensure_initialized_and_enqueue = MagicMock( side_effect=lambda async_coroutine: async_coroutine.close() ) @@ -7434,6 +7449,7 @@ async def test_websocket_passthrough_propagates_active_trace_context( mock_proxy_logging.pre_call_hook = AsyncMock(return_value={}) mock_proxy_logging.post_call_success_hook = AsyncMock() mock_proxy_logging.post_call_failure_hook = AsyncMock() + mock_proxy_logging.arelease_responses_websocket_slot = AsyncMock() mock_worker = MagicMock() mock_worker.ensure_initialized_and_enqueue = MagicMock(side_effect=lambda async_coroutine: async_coroutine.close()) monkeypatch.setattr("litellm.proxy.proxy_server.proxy_logging_obj", mock_proxy_logging) @@ -7500,6 +7516,7 @@ async def test_websocket_passthrough_never_forwards_caller_credentials_upstream( mock_proxy_logging.pre_call_hook = AsyncMock(return_value={}) mock_proxy_logging.post_call_success_hook = AsyncMock() mock_proxy_logging.post_call_failure_hook = AsyncMock() + mock_proxy_logging.arelease_responses_websocket_slot = AsyncMock() mock_worker = MagicMock() mock_worker.ensure_initialized_and_enqueue = MagicMock(side_effect=lambda async_coroutine: async_coroutine.close()) monkeypatch.setattr("litellm.proxy.proxy_server.proxy_logging_obj", mock_proxy_logging) @@ -7593,10 +7610,11 @@ def _patched_websocket_passthrough_environment(upstream_ws): mock_proxy_logging.pre_call_hook = AsyncMock(return_value={}) mock_proxy_logging.post_call_success_hook = AsyncMock() mock_proxy_logging.post_call_failure_hook = AsyncMock() + mock_proxy_logging.arelease_responses_websocket_slot = AsyncMock() mock_worker.ensure_initialized_and_enqueue = MagicMock( side_effect=lambda async_coroutine: async_coroutine.close() ) - yield + yield mock_proxy_logging async def _pending_receive(): @@ -7901,6 +7919,1487 @@ async def test_websocket_passthrough_does_not_close_twice_when_success_logging_f websocket.close.assert_awaited_once_with(code=1008, reason=upstream_reason) +SUCCESS_HANDLER: Final = ( + "litellm.proxy.pass_through_endpoints.pass_through_endpoints." + "pass_through_endpoint_logging.pass_through_async_success_handler" +) + + +KEY_MODEL_REFUSAL: Final = "key not allowed to access model. This key can only access models=['gpt-5.4-mini']." +KEY_MODEL_CLOSE_REASON: Final = "key not allowed to access model." + + +class RecordingFrameModelGate: + def __init__( + self, + refusals: Mapping[str, str] = MappingProxyType({}), + groups: Mapping[str, tuple[str, ...]] = MappingProxyType({}), + error_type: str = "key_model_access_denied", + ): + self.refusals = refusals + self.groups = groups + self.error_type = error_type + self.checked: list[str] = [] + + async def __call__(self, websocket: object, model: str, /) -> WebsocketFrameModelVerdict: + self.checked.append(model) + refusal: Final = self.refusals.get(model) + if refusal is not None: + return WebsocketFrameModelRefused(message=refusal, error_type=self.error_type) + return WebsocketFrameModelAllowed(access_groups=self.groups.get(model, ())) + + +def _received_frame(frame: str | bytes) -> Mapping[str, str | bytes]: + if isinstance(frame, bytes): + return {"type": "websocket.receive", "bytes": frame} + return {"type": "websocket.receive", "text": frame} + + +async def _relay_client_frames( + frames: Sequence[str | bytes], + gate: WebsocketFrameModelGate | None, + user_api_key_dict: UserAPIKeyAuth | None = None, +) -> SimpleNamespace: + upstream_ws: Final = RecordingUpstreamWebSocket() + websocket: Final = _client_websocket( + AsyncMock(side_effect=[_received_frame(frame) for frame in frames] + [{"type": "websocket.disconnect"}]) + ) + + with ( + _patched_websocket_passthrough_environment(upstream_ws), + patch(SUCCESS_HANDLER, new_callable=AsyncMock) as success_handler, + ): + await websocket_passthrough_request( + websocket=websocket, + target="wss://api.openai.com/v1/realtime?model=gpt-realtime-2.1", + custom_headers={"Authorization": "Bearer sk-provider"}, + user_api_key_dict=user_api_key_dict or UserAPIKeyAuth(), + forward_headers=False, + endpoint="/openai/v1/realtime", + accept_websocket=False, + client_frame_model_gate=gate, + ) + + return SimpleNamespace(upstream=upstream_ws, client=websocket, success_handler=success_handler) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "frame, expected_model", + [ + ({"type": "response.create", "model": "gpt-5.4-mini"}, "gpt-5.4-mini"), + ({"type": "response.create", "response": {"model": "gpt-5.4"}}, "gpt-5.4"), + ({"type": "session.update", "session": {"model": "gpt-realtime-2.1"}}, "gpt-realtime-2.1"), + ( + {"type": "session.update", "session": {"input_audio_transcription": {"model": "gpt-4o-transcribe"}}}, + "gpt-4o-transcribe", + ), + ( + { + "type": "session.update", + "session": {"audio": {"input": {"transcription": {"model": "gpt-transcribe"}}}}, + }, + "gpt-transcribe", + ), + ], +) +async def test_websocket_passthrough_checks_the_model_a_client_frame_asks_for(frame, expected_model): + gate: Final = RecordingFrameModelGate() + + relayed: Final = await _relay_client_frames([json.dumps(frame)], gate) + + assert gate.checked == [expected_model] + assert [call.args[0] for call in relayed.upstream.send.await_args_list] == [json.dumps(frame)] + assert relayed.client.close.await_args_list == [call()] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_refuses_a_frame_naming_a_model_the_key_cannot_call(): + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-realtime-2.1": KEY_MODEL_REFUSAL})) + frame: Final = json.dumps({"type": "session.update", "session": {"model": "gpt-realtime-2.1"}}) + + relayed: Final = await _relay_client_frames([frame], gate) + + assert relayed.upstream.send.await_args_list == [] + assert json.loads(relayed.client.send_text.await_args.args[0]) == { + "type": "error", + "error": {"type": "key_model_access_denied", "message": KEY_MODEL_REFUSAL}, + } + assert relayed.client.close.await_args_list == [call(code=1008, reason=KEY_MODEL_CLOSE_REASON)] + relayed.upstream.close.assert_awaited() + + +@pytest.mark.asyncio +async def test_websocket_passthrough_closes_with_the_whole_first_sentence_of_the_refusal(): + """ + The close frame only has room for 123 bytes, so a refusal cut at that limit + reads as a truncated number. The first sentence names the reason whole, and + the error event before the close still carries every word with its type. + """ + refusal: Final = "Budget has been exceeded! Key=" + "f" * 64 + " Spend=1.23456789, Budget=1.0" + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-5.4": refusal}), error_type="budget_exceeded") + + relayed: Final = await _relay_client_frames([json.dumps({"type": "response.create", "model": "gpt-5.4"})], gate) + + assert relayed.client.close.await_args_list == [call(code=1008, reason="Budget has been exceeded!")] + assert json.loads(relayed.client.send_text.await_args.args[0])["error"] == { + "type": "budget_exceeded", + "message": refusal, + } + + +@pytest.mark.asyncio +async def test_websocket_passthrough_refuses_the_whole_frame_when_any_named_model_is_denied(): + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-4o-transcribe": KEY_MODEL_REFUSAL})) + frame: Final = json.dumps( + { + "type": "session.update", + "session": {"model": "gpt-realtime-2.1", "input_audio_transcription": {"model": "gpt-4o-transcribe"}}, + } + ) + + relayed: Final = await _relay_client_frames([frame], gate) + + assert gate.checked == ["gpt-realtime-2.1", "gpt-4o-transcribe"] + assert relayed.upstream.send.await_args_list == [] + assert relayed.client.close.await_args_list == [call(code=1008, reason=KEY_MODEL_CLOSE_REASON)] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_stops_relaying_after_a_refused_frame(): + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-realtime-2.1": KEY_MODEL_REFUSAL})) + frames: Final = [ + json.dumps({"type": "session.update", "session": {"model": "gpt-realtime-2.1"}}), + json.dumps({"type": "response.create"}), + ] + + relayed: Final = await _relay_client_frames(frames, gate) + + assert relayed.upstream.send.await_args_list == [] + assert gate.checked == ["gpt-realtime-2.1"] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "frame", + [ + '{"type": "input_audio_buffer.append", "audio": "AAAA"}', + '{"type": "response.create", "response": {"modalities": ["audio"]}}', + '{"type": "session.update", "session": {"model": 42}}', + 'not json at all, but it does say "model"', + '{"type": "session.update", "session": "model"}', + ], +) +async def test_websocket_passthrough_forwards_frames_that_name_no_model(frame): + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-realtime-2.1": "denied"})) + + relayed: Final = await _relay_client_frames([frame], gate) + + assert gate.checked == [] + assert [call_.args[0] for call_ in relayed.upstream.send.await_args_list] == [frame] + assert relayed.client.close.await_args_list == [call()] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_never_bills_the_model_a_client_frame_named(): + """ + A frame is what the caller asked for, not what the provider ran. A caller + who names a cheap model in a frame the provider ignores must not move the + session onto that model's price, so only the provider's own frames decide + what a session is billed as. + """ + gate: Final = RecordingFrameModelGate() + frames: Final = [ + json.dumps({"type": "session.update", "session": {"model": "gpt-realtime-2.1"}}), + json.dumps({"type": "response.create", "response": {"model": "gpt-realtime-2.1-mini"}}), + ] + + relayed: Final = await _relay_client_frames(frames, gate) + + assert gate.checked == ["gpt-realtime-2.1", "gpt-realtime-2.1-mini"] + assert "model" not in relayed.success_handler.call_args.kwargs + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "frame, expected_model", + [ + ('{"type": "response.create", "\\u006dodel": "gpt-5.4"}', "gpt-5.4"), + ('{"type": "session.update", "session": {"\\u006d\\u006fdel": "gpt-realtime-2.1"}}', "gpt-realtime-2.1"), + ], +) +async def test_websocket_passthrough_checks_a_model_named_with_escaped_json(frame, expected_model): + """ + JSON lets a caller spell the same key several ways, so a frame carrying + "\\u006dodel" asks for a model just as plainly as one carrying "model" and + has to face the same check. + """ + gate: Final = RecordingFrameModelGate(MappingProxyType({expected_model: KEY_MODEL_REFUSAL})) + + relayed: Final = await _relay_client_frames([frame], gate) + + assert gate.checked == [expected_model] + assert relayed.upstream.send.await_args_list == [] + assert relayed.client.close.await_args_list == [call(code=1008, reason=KEY_MODEL_CLOSE_REASON)] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_checks_the_model_a_binary_frame_asks_for(): + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-realtime-2.1": KEY_MODEL_REFUSAL})) + frame: Final = json.dumps({"type": "session.update", "session": {"model": "gpt-realtime-2.1"}}).encode() + + relayed: Final = await _relay_client_frames([frame], gate) + + assert gate.checked == ["gpt-realtime-2.1"] + assert relayed.upstream.send.await_args_list == [] + assert relayed.client.close.await_args_list == [call(code=1008, reason=KEY_MODEL_CLOSE_REASON)] + + +STORED_PROMPT_FRAMES: Final = ( + {"type": "response.create", "prompt": {"id": "pmpt_fake", "version": "2"}}, + {"type": "response.create", "response": {"prompt": {"id": "pmpt_fake"}}}, + {"type": "session.update", "session": {"prompt": {"id": "pmpt_fake"}}}, +) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("frame", STORED_PROMPT_FRAMES) +async def test_websocket_passthrough_refuses_a_stored_prompt_for_a_key_limited_to_some_models(frame): + """ + A stored prompt carries its own model on the provider side, so a key held to + a model list could reach any model through one. The gateway cannot see that + model, so a key that is not allowed every model cannot run a stored prompt. + """ + gate: Final = RecordingFrameModelGate(MappingProxyType({STORED_PROMPT_PROBE_MODEL: KEY_MODEL_REFUSAL})) + + relayed: Final = await _relay_client_frames([json.dumps(frame)], gate) + + error_message: Final = json.loads(relayed.client.send_text.await_args.args[0])["error"]["message"] + close_reason: Final = relayed.client.close.await_args.kwargs["reason"] + assert gate.checked == [STORED_PROMPT_PROBE_MODEL] + assert relayed.upstream.send.await_args_list == [] + assert relayed.client.close.await_args.kwargs["code"] == 1008 + assert "stored prompt" in close_reason + assert close_reason.endswith(".") and error_message.startswith(close_reason) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("frame", STORED_PROMPT_FRAMES) +async def test_websocket_passthrough_forwards_a_stored_prompt_for_a_key_allowed_every_model(frame): + gate: Final = RecordingFrameModelGate() + + relayed: Final = await _relay_client_frames([json.dumps(frame)], gate) + + assert [call_.args[0] for call_ in relayed.upstream.send.await_args_list] == [json.dumps(frame)] + assert relayed.client.close.await_args_list == [call()] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_keeps_the_budget_reason_when_a_stored_prompt_is_refused_for_spend(): + refusal: Final = "Budget has been exceeded! Key=hashed Spend=2.0, Budget=1.0" + gate: Final = RecordingFrameModelGate( + MappingProxyType({STORED_PROMPT_PROBE_MODEL: refusal}), error_type="budget_exceeded" + ) + + relayed: Final = await _relay_client_frames([json.dumps(STORED_PROMPT_FRAMES[0])], gate) + + assert json.loads(relayed.client.send_text.await_args.args[0])["error"] == { + "type": "budget_exceeded", + "message": refusal, + } + + +@pytest.mark.asyncio +async def test_websocket_passthrough_checks_the_model_a_hosted_tool_names(): + """A hosted tool such as image generation runs its own model, so that model gets the same check as the turn's""" + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-image-2": KEY_MODEL_REFUSAL})) + frame: Final = { + "type": "response.create", + "model": "gpt-5.4-mini", + "tools": [{"type": "web_search"}, {"type": "image_generation", "model": "gpt-image-2"}], + } + + relayed: Final = await _relay_client_frames([json.dumps(frame)], gate) + + assert gate.checked == ["gpt-5.4-mini", "gpt-image-2"] + assert relayed.upstream.send.await_args_list == [] + assert relayed.client.close.await_args.kwargs["code"] == 1008 + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "frame", + [ + {"type": "response.create", "model": "gpt-5.4-mini", "tools": [{"type": "image_generation"}]}, + {"type": "response.create", "response": {"tools": [{"type": "image_generation", "model": None}]}}, + ], +) +async def test_websocket_passthrough_refuses_an_image_tool_without_a_model_for_a_key_limited_to_some_models(frame): + gate: Final = RecordingFrameModelGate(MappingProxyType({STORED_PROMPT_PROBE_MODEL: KEY_MODEL_REFUSAL})) + + relayed: Final = await _relay_client_frames([json.dumps(frame)], gate) + + assert gate.checked[-1] == STORED_PROMPT_PROBE_MODEL + assert relayed.upstream.send.await_args_list == [] + assert "image_generation tool has to name its model" in relayed.client.close.await_args.kwargs["reason"] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_checks_a_frame_only_the_standard_json_parser_reads(): + """A lone surrogate escape is valid JSON syntax the fast parser refuses, so the frame is still checked""" + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-5.4": KEY_MODEL_REFUSAL})) + frame: Final = '{"type": "response.create", "model": "gpt-5.4", "metadata": {"note": "\\ud800"}}' + + relayed: Final = await _relay_client_frames([frame], gate) + + assert gate.checked == ["gpt-5.4"] + assert relayed.upstream.send.await_args_list == [] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_forwards_binary_audio_untouched(): + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-realtime-2.1": "denied"})) + audio: Final = bytes(range(256)) * 4 + + relayed: Final = await _relay_client_frames([audio], gate) + + assert gate.checked == [] + assert [call_.args[0] for call_ in relayed.upstream.send.await_args_list] == [audio] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_without_a_frame_gate_forwards_every_frame(): + frame: Final = json.dumps({"type": "session.update", "session": {"model": "gpt-realtime-2.1"}}) + + relayed: Final = await _relay_client_frames([frame], None) + + assert [call_.args[0] for call_ in relayed.upstream.send.await_args_list] == [frame] + assert "model" not in relayed.success_handler.call_args.kwargs + + +REALTIME_SESSION_CREATED: Final = json.dumps( + {"type": "session.created", "session": {"id": "sess_fake", "model": "gpt-realtime-2.1"}} +) +REALTIME_RESPONSE_DONE: Final = json.dumps( + { + "type": "response.done", + "response": {"status": "completed", "usage": {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15}}, + } +) + + +@pytest.mark.asyncio +async def test_websocket_passthrough_still_bills_a_session_the_client_walked_out_of(): + """ + A realtime client closes the moment it has the answer it asked for, so the + next frame the provider sends lands on a socket that is already gone. That + is how a session ends, not a gateway failure, and the turns the caller + already used still have to reach spend, including the last frame the client + never collected. + """ + upstream_ws: Final = FakeUpstreamWebSocket( + REALTIME_SESSION_CREATED, + REALTIME_RESPONSE_DONE, + json.dumps({"type": "response.output_text.delta", "delta": "too late"}), + ) + websocket: Final = _client_websocket(_pending_receive) + websocket.send_text = AsyncMock(side_effect=[None, None, WebSocketDisconnect(code=1006)]) + + with ( + _patched_websocket_passthrough_environment(upstream_ws), + patch(SUCCESS_HANDLER, new_callable=AsyncMock) as success_handler, + ): + await websocket_passthrough_request( + websocket=websocket, + target="wss://api.openai.com/v1/realtime?model=gpt-realtime-2.1", + custom_headers={"Authorization": "Bearer sk-provider"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/openai_passthrough/v1/realtime", + accept_websocket=False, + ) + + relayed_frames: Final = success_handler.call_args.kwargs["response_body"] + assert [frame["type"] for frame in relayed_frames] == [ + "session.created", + "response.done", + "response.output_text.delta", + ] + + +class ScriptedUpstreamWebSocket: + """Answers each client frame with its scripted reply""" + + def __init__(self, *opening: str, replies: Sequence[Sequence[str]] = ()): + self._frames: asyncio.Queue[str | None] = asyncio.Queue() + for frame in opening: + self._frames.put_nowait(frame) + self._replies = iter(replies) + self._closed = False + self.send = AsyncMock(side_effect=self._reply) + self.close = AsyncMock(side_effect=self._flush_and_close) + + async def _reply(self, frame: str | bytes) -> None: + for reply in next(self._replies, ()): + self._frames.put_nowait(reply) + + async def _flush_and_close(self, *args: object, **kwargs: object) -> None: + if self._closed: + return + self._closed = True + self._frames.put_nowait(None) + while not self._frames.empty(): + await asyncio.sleep(0) + + async def recv(self, decode: bool | None = None) -> str: + from websockets.exceptions import ConnectionClosedOK + from websockets.frames import Close + + frame: Final = await self._frames.get() + if frame is None: + raise ConnectionClosedOK(rcvd=Close(1000, ""), sent=Close(1000, ""), rcvd_then_sent=True) + return frame + + +def _turn_taking_client( + frames: Sequence[str], answered_by: str, answers_due: Sequence[int] | None = None +) -> MagicMock: + """A client that sends each frame, and then leaves, once the answers due by then reached it (by default the + answer to every earlier frame), and that Starlette refuses to send to once it is closed""" + from starlette.websockets import WebSocketState + + delivered: Final[list[str]] = [] + script: Final = iter(frames) + receives: Final = iter(range(len(frames) + 1) if answers_due is None else answers_due) + + async def receive(): + answers_due: Final = next(receives) + while sum(answered_by in text for text in delivered) < answers_due: + await asyncio.sleep(0) + frame: Final = next(script, None) + return {"type": "websocket.disconnect"} if frame is None else _received_frame(frame) + + websocket: Final = _client_websocket(receive) + + async def send_text(text: str) -> None: + if websocket.application_state == WebSocketState.DISCONNECTED: + raise RuntimeError('Cannot call "send" once a close message has been sent.') + delivered.append(text) + + websocket.send_text = AsyncMock(side_effect=send_text) + return websocket + + +async def _relay_turns( + client: MagicMock, + upstream_ws: ScriptedUpstreamWebSocket, + gate: WebsocketFrameModelGate, + endpoint: str = "/openai_passthrough/v1/responses", + target: str = "wss://api.openai.com/v1/responses", + opening_access_groups: Sequence[str] = (), + walkout_turn_drain_seconds: float = 0.05, +) -> SimpleNamespace: + with ( + _patched_websocket_passthrough_environment(upstream_ws) as proxy_logging, + patch(SUCCESS_HANDLER, new_callable=AsyncMock) as success_handler, + ): + await asyncio.wait_for( + websocket_passthrough_request( + websocket=client, + target=target, + custom_headers={"Authorization": "Bearer sk-provider"}, + user_api_key_dict=UserAPIKeyAuth( + api_key="hashed-key", matched_model_access_groups=list(opening_access_groups) + ), + forward_headers=False, + endpoint=endpoint, + accept_websocket=False, + client_frame_model_gate=gate, + upstream_frame_kind=openai_websocket_frame_kind, + walkout_turn_drain_seconds=walkout_turn_drain_seconds, + ), + timeout=10, + ) + return SimpleNamespace(success_handler=success_handler, proxy_logging=proxy_logging) + + +def _responses_turn(model: str) -> list[str]: + return [_responses_turn_started(model), _responses_turn_ended(model)] + + +def _responses_turn_started(model: str) -> str: + return json.dumps({"type": "response.created", "response": {"model": model}}) + + +def _responses_turn_ended(model: str) -> str: + return ( + + json.dumps( + { + "type": "response.completed", + "response": {"model": model, "usage": {"input_tokens": 9, "output_tokens": 3, "total_tokens": 12}}, + } + ) + ) + + +@pytest.mark.asyncio +async def test_websocket_passthrough_logs_each_turn_on_its_own_row_charged_to_the_groups_that_allowed_it(): + """ + One socket can run a model from one access group and then a model from + another. Each turn has to reach spend as it finishes, under its own call, + and be charged to the group that let its model through, so a group budget + sees its share during the session and never pays for the other group's turn. + """ + gate: Final = RecordingFrameModelGate( + groups=MappingProxyType({"gpt-5.4-mini": ("tier-a",), "gpt-5.4": ("tier-b",)}), + ) + client: Final = _turn_taking_client( + [ + json.dumps({"type": "response.create", "model": "gpt-5.4-mini"}), + json.dumps({"type": "response.create", "model": "gpt-5.4"}), + ], + answered_by='"response.completed"', + ) + upstream_ws: Final = ScriptedUpstreamWebSocket(replies=[_responses_turn("gpt-5.4-mini"), _responses_turn("gpt-5.4")]) + + relayed: Final = await _relay_turns(client, upstream_ws, gate) + + rows: Final = [logged.kwargs for logged in relayed.success_handler.call_args_list] + assert [[frame["response"]["model"] for frame in row["response_body"]] for row in rows] == [ + ["gpt-5.4-mini", "gpt-5.4-mini"], + ["gpt-5.4", "gpt-5.4"], + ] + assert [row["litellm_params"]["metadata"][MODEL_ACCESS_GROUP_METADATA_KEY] for row in rows] == [ + ["tier-a"], + ["tier-b"], + ] + assert [row["litellm_params"]["metadata"]["user_api_key"] for row in rows] == ["hashed-key", "hashed-key"] + assert rows[0]["litellm_call_id"] != rows[1]["litellm_call_id"] + assert rows[1]["start_time"] == rows[0]["end_time"] + relayed.proxy_logging.post_call_failure_hook.assert_not_awaited() + + +@pytest.mark.asyncio +async def test_websocket_passthrough_holds_the_keys_parallel_slot_until_the_session_ends(): + """ + A key allowed one parallel request holds that slot for as long as its + socket stays open. A turn that finishes and reaches spend mid-session must + not free it, or a second session gets in while the first is still running. + """ + limiter_cache: Final = DualCache() + limiter: Final = PROXY_MaxParallelRequestsHandler_v3(internal_usage_cache=InternalUsageCache(limiter_cache)) + key: Final = UserAPIKeyAuth(api_key=hash_token("sk-one-socket"), max_parallel_requests=1) + logged_rows: Final[list[asyncio.Task[None]]] = [] + + async def admit(user_api_key_dict: UserAPIKeyAuth, data: dict[str, object], call_type: str) -> None: + await limiter.async_pre_call_hook( + user_api_key_dict=user_api_key_dict, cache=limiter_cache, data=data, call_type=call_type + ) + + async def release(data: dict[str, object], user_api_key_dict: UserAPIKeyAuth, response: object) -> None: + await limiter.async_post_call_success_hook(data=data, user_api_key_dict=user_api_key_dict, response=response) + + async def log_row(**row: object) -> None: + await limiter.async_log_success_event( + kwargs=row, response_obj=None, start_time=row["start_time"], end_time=row["end_time"] + ) + + async def opens_another_session() -> bool: + await asyncio.gather(*logged_rows) + try: + await asyncio.create_task( + admit(key, {"litellm_call_id": str(uuid.uuid4())}, "pass_through_endpoint"), + context=contextvars.Context(), + ) + except ProxyRateLimitError: + return False + return True + + client: Final = _turn_taking_client( + [json.dumps({"type": "response.create", "model": "gpt-5.4-mini"})], answered_by='"response.completed"' + ) + take_frame: Final = client.receive + while_open: Final[list[bool]] = [] + + async def receive() -> Mapping[str, str | bytes]: + message: Final = await take_frame() + if message["type"] == "websocket.disconnect": + while_open.append(await opens_another_session()) + return message + + client.receive = receive + + with ( + _patched_websocket_passthrough_environment( + ScriptedUpstreamWebSocket(replies=[_responses_turn("gpt-5.4-mini")]) + ) as proxy_logging, + patch("litellm.proxy.pass_through_endpoints.pass_through_endpoints.GLOBAL_LOGGING_WORKER") as worker, + patch(SUCCESS_HANDLER, new_callable=AsyncMock, side_effect=log_row), + ): + proxy_logging.pre_call_hook = AsyncMock(side_effect=admit) + proxy_logging.post_call_success_hook = AsyncMock(side_effect=release) + worker.ensure_initialized_and_enqueue = MagicMock( + side_effect=lambda async_coroutine: logged_rows.append(asyncio.create_task(async_coroutine)) + ) + await asyncio.wait_for( + websocket_passthrough_request( + websocket=client, + target="wss://api.openai.com/v1/responses", + custom_headers={"Authorization": "Bearer sk-provider"}, + user_api_key_dict=key, + forward_headers=False, + endpoint="/openai_passthrough/v1/responses", + accept_websocket=False, + client_frame_model_gate=RecordingFrameModelGate(), + upstream_frame_kind=openai_websocket_frame_kind, + walkout_turn_drain_seconds=0.05, + ), + timeout=10, + ) + after_close: Final = await opens_another_session() + + assert len(logged_rows) == 1 + assert (while_open, after_close) == ([False], True) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "models, upstream_refuses, while_open", + [ + ((), False, [False]), + (("gpt-5.4-mini",), False, [False]), + (("gpt-5.4-mini", "gpt-5.4"), False, [False]), + ((), True, []), + ], + ids=["no-turn", "one-turn", "two-turns", "upstream-refused"], +) +async def test_websocket_passthrough_frees_one_legacy_parallel_slot_only_when_the_session_ends( + models: tuple[str, ...], upstream_refuses: bool, while_open: list[bool] +): + limiter: Final = PROXY_MaxParallelRequestsHandler( + internal_usage_cache=InternalUsageCache(DualCache()), clock=lambda: datetime(2026, 1, 31, 23, 59, 30) + ) + proxy_logging_obj: Final = ProxyLogging(user_api_key_cache=DualCache()) + proxy_logging_obj.proxy_hook_mapping["parallel_request_limiter"] = limiter + key: Final = UserAPIKeyAuth(api_key=hash_token("sk-legacy-socket"), max_parallel_requests=2) + logged_rows: Final[list[asyncio.Task[None]]] = [] + seen_while_open: Final[list[bool]] = [] + + async def admit(user_api_key_dict: UserAPIKeyAuth, data: Mapping[str, object], call_type: str) -> None: + await limiter.async_pre_call_hook( + user_api_key_dict=user_api_key_dict, cache=DualCache(), data=data, call_type=call_type + ) + + async def log_row(**row: object) -> None: + await limiter.async_log_success_event( + kwargs=row, response_obj=None, start_time=row["start_time"], end_time=row["end_time"] + ) + + async def opens_another_session() -> bool: + await asyncio.gather(*logged_rows) + try: + await admit(key, {}, "pass_through_endpoint") + except ProxyRateLimitError: + return False + return True + + client: Final = _turn_taking_client( + [json.dumps({"type": "response.create", "model": model}) for model in models], + answered_by='"response.completed"', + ) + take_frame: Final = client.receive + + async def receive() -> Mapping[str, str | bytes]: + message: Final = await take_frame() + if message["type"] == "websocket.disconnect": + seen_while_open.append(await opens_another_session()) + return message + + client.receive = receive + await admit(key, {}, "pass_through_endpoint") + + with ExitStack() as stack: + proxy_logging: Final = stack.enter_context( + _patched_websocket_passthrough_environment( + ScriptedUpstreamWebSocket(replies=[_responses_turn(model) for model in models]) + ) + ) + worker: Final = stack.enter_context( + patch("litellm.proxy.pass_through_endpoints.pass_through_endpoints.GLOBAL_LOGGING_WORKER") + ) + stack.enter_context(patch(SUCCESS_HANDLER, new_callable=AsyncMock, side_effect=log_row)) + if upstream_refuses: + stack.enter_context( + patch( + "litellm.proxy.pass_through_endpoints.pass_through_endpoints.connect", + side_effect=OSError("upstream refused the socket"), + ) + ) + proxy_logging.pre_call_hook = AsyncMock(side_effect=admit) + proxy_logging.arelease_responses_websocket_slot = proxy_logging_obj.arelease_responses_websocket_slot + worker.ensure_initialized_and_enqueue = MagicMock( + side_effect=lambda async_coroutine: logged_rows.append(asyncio.create_task(async_coroutine)) + ) + await asyncio.wait_for( + websocket_passthrough_request( + websocket=client, + target="wss://api.openai.com/v1/responses", + custom_headers={"Authorization": "Bearer sk-provider"}, + user_api_key_dict=key, + forward_headers=False, + endpoint="/openai_passthrough/v1/responses", + accept_websocket=False, + client_frame_model_gate=RecordingFrameModelGate(), + upstream_frame_kind=openai_websocket_frame_kind, + walkout_turn_drain_seconds=0.05, + ), + timeout=10, + ) + after_close: Final = (await opens_another_session(), await opens_another_session()) + + assert len(logged_rows) == len(models) + assert (seen_while_open, after_close) == (while_open, (True, False)) + + +@pytest.mark.asyncio +async def test_websocket_passthrough_charges_pipelined_turns_to_the_frame_that_asked_for_each(): + """ + A client can send its next response.create before the provider finished the + last one, and the provider then runs them one after the other. Each turn + still has to reach spend under the groups that let its own frame through, + and under the model that frame asked for, even when the provider answers + under a dated snapshot name a per-model budget was never set on. + """ + gate: Final = RecordingFrameModelGate( + groups=MappingProxyType({"gpt-5.4-mini": ("tier-a",), "gpt-5.4": ("tier-b",)}), + ) + client: Final = _turn_taking_client( + [ + json.dumps({"type": "response.create", "model": "gpt-5.4-mini"}), + json.dumps({"type": "response.create", "model": "gpt-5.4"}), + ], + answered_by='"response.completed"', + answers_due=(0, 0, 2), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + replies=[_responses_turn("gpt-5.4-mini-2099-01-01"), _responses_turn("gpt-5.4-2099-01-01")] + ) + + relayed: Final = await _relay_turns(client, upstream_ws, gate) + + rows: Final = [logged.kwargs for logged in relayed.success_handler.call_args_list] + assert [[frame["response"]["model"] for frame in row["response_body"]] for row in rows] == [ + ["gpt-5.4-mini-2099-01-01", "gpt-5.4-mini-2099-01-01"], + ["gpt-5.4-2099-01-01", "gpt-5.4-2099-01-01"], + ] + assert [row["litellm_params"]["metadata"][MODEL_ACCESS_GROUP_METADATA_KEY] for row in rows] == [ + ["tier-a"], + ["tier-b"], + ] + assert [row["litellm_params"]["metadata"]["model_group"] for row in rows] == ["gpt-5.4-mini", "gpt-5.4"] + + +def _responses_error(code: str) -> str: + return json.dumps({"type": "error", "status": 400, "error": {"type": "invalid_request_error", "code": code}}) + + +def _on_lane(frame: str, stream_id: str) -> str: + return json.dumps({**json.loads(frame), "stream_id": stream_id}) + + +@pytest.mark.asyncio +async def test_websocket_passthrough_charges_turns_after_a_refused_frame_to_the_frames_that_asked_for_them(): + """ + The provider answers a response.create it cannot run, one naming no model + or a stale previous_response_id, with an error event and keeps the socket + open. Those frames never run, so each later turn still has to reach spend + under the groups and model its own frame asked for. + """ + gate: Final = RecordingFrameModelGate( + groups=MappingProxyType({"gpt-5.4-mini": ("tier-a",), "gpt-5.4": ("tier-b",)}), + ) + client: Final = _turn_taking_client( + [ + json.dumps({"type": "response.create", "input": "hi"}), + json.dumps({"type": "response.create", "model": "gpt-5.4", "previous_response_id": "resp_gone"}), + json.dumps({"type": "response.create", "model": "gpt-5.4-mini"}), + json.dumps({"type": "response.create", "model": "gpt-5.4"}), + ], + answered_by='"response.completed"', + answers_due=(0, 0, 0, 0, 2), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + replies=[ + [_responses_error("missing_required_parameter")], + [_responses_error("previous_response_not_found")], + _responses_turn("gpt-5.4-mini-2099-01-01"), + _responses_turn("gpt-5.4-2099-01-01"), + ] + ) + + relayed: Final = await _relay_turns(client, upstream_ws, gate) + + rows: Final = [logged.kwargs for logged in relayed.success_handler.call_args_list] + assert [row["litellm_params"]["metadata"]["model_group"] for row in rows] == ["gpt-5.4-mini", "gpt-5.4"] + assert [row["litellm_params"]["metadata"][MODEL_ACCESS_GROUP_METADATA_KEY] for row in rows] == [ + ["tier-a"], + ["tier-b"], + ] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_charges_turns_on_parallel_lanes_to_the_frame_that_asked_for_each(): + """ + A client can put each response.create on its own stream_id lane, and the + provider runs lanes side by side, so a later frame's turn can end first. + Each turn still has to reach spend on its own row, under the groups and + model its own frame asked for. + """ + gate: Final = RecordingFrameModelGate( + groups=MappingProxyType({"gpt-5.4-mini": ("tier-a",), "gpt-5.4": ("tier-b",)}), + ) + client: Final = _turn_taking_client( + [ + json.dumps({"type": "response.create", "stream_id": "slow", "model": "gpt-5.4"}), + json.dumps({"type": "response.create", "stream_id": "fast", "model": "gpt-5.4-mini"}), + ], + answered_by='"response.completed"', + answers_due=(0, 0, 2), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + replies=[ + [_on_lane(_responses_turn_started("gpt-5.4"), "slow")], + [ + *(_on_lane(frame, "fast") for frame in _responses_turn("gpt-5.4-mini")), + _on_lane(_responses_turn_ended("gpt-5.4"), "slow"), + ], + ] + ) + + relayed: Final = await _relay_turns(client, upstream_ws, gate) + + rows: Final = [logged.kwargs for logged in relayed.success_handler.call_args_list] + assert [[(frame["stream_id"], frame["response"]["model"]) for frame in row["response_body"]] for row in rows] == [ + [("fast", "gpt-5.4-mini"), ("fast", "gpt-5.4-mini")], + [("slow", "gpt-5.4"), ("slow", "gpt-5.4")], + ] + assert [row["litellm_params"]["metadata"]["model_group"] for row in rows] == ["gpt-5.4-mini", "gpt-5.4"] + assert [row["litellm_params"]["metadata"][MODEL_ACCESS_GROUP_METADATA_KEY] for row in rows] == [ + ["tier-a"], + ["tier-b"], + ] + + +@pytest.mark.asyncio +@pytest.mark.parametrize("stream_id", ["", 7]) +async def test_websocket_passthrough_refuses_a_frame_on_a_lane_the_provider_error_cannot_name(stream_id): + """ + The provider refuses a frame with an invalid stream_id through an error that + names no lane (OpenAI WebSocket mode guide, "Errors to handle", read + 2026-11-02, which also says an empty stream_id is invalid). Relayed, that + error would refuse a turn waiting on the default lane instead, and that turn + would reach spend without the groups its own frame asked for. + """ + gate: Final = RecordingFrameModelGate() + allowed: Final = json.dumps({"type": "response.create", "model": "gpt-5.4-mini"}) + + relayed: Final = await _relay_client_frames( + [allowed, json.dumps({"type": "response.create", "stream_id": stream_id, "model": "gpt-5.4-mini"})], gate + ) + + assert [sent.args[0] for sent in relayed.upstream.send.await_args_list] == [allowed] + assert json.loads(relayed.client.send_text.await_args.args[0])["type"] == "error" + assert relayed.client.close.await_args.kwargs["code"] == 1008 + + +@pytest.mark.asyncio +async def test_websocket_passthrough_charges_realtime_turns_to_the_session_the_key_opened(): + """ + A realtime session runs the model it was opened on, so its turns stay + charged to the groups and the model the key opened it with, whatever model + a later session.update or response.create names. + """ + gate: Final = RecordingFrameModelGate(groups=MappingProxyType({"gpt-realtime-2.1-mini": ("tier-b",)})) + client: Final = _turn_taking_client( + [ + json.dumps({"type": "session.update", "session": {"model": "gpt-realtime-2.1-mini"}}), + json.dumps({"type": "response.create", "response": {"model": "gpt-realtime-2.1-mini"}}), + ], + answered_by='"type"', + answers_due=(1, 2, 4), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + REALTIME_SESSION_CREATED, + replies=[ + [json.dumps({"type": "session.updated"})], + [json.dumps({"type": "response.created", "response": {"status": "in_progress"}}), REALTIME_RESPONSE_DONE], + ], + ) + + relayed: Final = await _relay_turns( + client, + upstream_ws, + gate, + "/openai/v1/realtime", + target="wss://api.openai.com/v1/realtime?model=gpt-realtime-2.1", + opening_access_groups=("tier-a",), + ) + + assert gate.checked == ["gpt-realtime-2.1-mini", "gpt-realtime-2.1-mini"] + rows: Final = [logged.kwargs for logged in relayed.success_handler.call_args_list] + assert [row["litellm_params"]["metadata"][MODEL_ACCESS_GROUP_METADATA_KEY] for row in rows] == [["tier-a"]] + assert [row["litellm_params"]["metadata"]["model_group"] for row in rows] == ["gpt-realtime-2.1"] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_logs_no_spend_row_for_a_session_refused_before_any_answer(): + """A session refused at its first frame ran nothing, so its only record is the refusal, not an extra $0 row""" + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-5.4": KEY_MODEL_REFUSAL})) + client: Final = _turn_taking_client([json.dumps({"type": "response.create", "model": "gpt-5.4"})], answered_by='"error"') + + relayed: Final = await _relay_turns(client, ScriptedUpstreamWebSocket(), gate) + + assert relayed.success_handler.call_args_list == [] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_logs_the_unfinished_turn_a_client_walked_out_of(): + """ + A client can leave while a later turn is still running. That turn never + ends, but it was started and may be billed, so it still gets its own row. + """ + client: Final = _turn_taking_client( + [ + json.dumps({"type": "response.create", "model": "gpt-5.4-mini"}), + json.dumps({"type": "response.create", "model": "gpt-5.4"}), + ], + answered_by='"response.created"', + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + replies=[_responses_turn("gpt-5.4-mini"), [_responses_turn_started("gpt-5.4")]] + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate()) + + assert [ + [json.loads(frame)["type"] if isinstance(frame, str) else frame["type"] for frame in logged.kwargs["response_body"]] + for logged in relayed.success_handler.call_args_list + ] == [["response.created", "response.completed"], ["response.created"]] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_charges_each_lane_a_client_walked_out_of_to_the_frame_that_started_it(): + """ + A client can leave while turns on several stream_id lanes are still + running. Each unfinished turn still gets its own row, under the groups and + model its own frame asked for. + """ + gate: Final = RecordingFrameModelGate( + groups=MappingProxyType({"gpt-5.4-mini": ("tier-a",), "gpt-5.4": ("tier-b",)}), + ) + client: Final = _turn_taking_client( + [ + json.dumps({"type": "response.create", "stream_id": "slow", "model": "gpt-5.4"}), + json.dumps({"type": "response.create", "stream_id": "fast", "model": "gpt-5.4-mini"}), + ], + answered_by='"response.created"', + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + replies=[ + [_on_lane(_responses_turn_started("gpt-5.4"), "slow")], + [_on_lane(_responses_turn_started("gpt-5.4-mini"), "fast")], + ] + ) + + relayed: Final = await _relay_turns(client, upstream_ws, gate) + + rows: Final = [logged.kwargs for logged in relayed.success_handler.call_args_list] + assert [[frame["stream_id"] for frame in row["response_body"]] for row in rows] == [["slow"], ["fast"]] + assert [row["litellm_params"]["metadata"]["model_group"] for row in rows] == ["gpt-5.4", "gpt-5.4-mini"] + assert [row["litellm_params"]["metadata"][MODEL_ACCESS_GROUP_METADATA_KEY] for row in rows] == [ + ["tier-b"], + ["tier-a"], + ] + + +class UpstreamFinishingTheTurnAfterTheClientLeft: + """Starts a turn for each client frame, unless it only answers once the client left, sends the rest of the session + once the client left, and stops once closed""" + + def __init__( + self, + client_left: asyncio.Event, + started: str | None, + rest_of_the_turn: Sequence[str], + opening: Sequence[str] = (), + ): + self._frames: asyncio.Queue[str | None] = asyncio.Queue() + for frame in opening: + self._frames.put_nowait(frame) + self._client_left = client_left + self._started = started + self._rest_of_the_turn = rest_of_the_turn + self._closed = False + self._finishing: list[asyncio.Task[None]] = [] + self.send = AsyncMock(side_effect=self._start_turn) + self.close = AsyncMock(side_effect=self._close) + + def _close(self, *args: object, **kwargs: object) -> None: + self._closed = True + self._frames.put_nowait(None) + + async def _start_turn(self, frame: str | bytes) -> None: + if self._started is not None: + self._frames.put_nowait(self._started) + if not self._finishing: + self._finishing.append(asyncio.create_task(self._finish_once_the_client_left())) + + async def _finish_once_the_client_left(self) -> None: + await self._client_left.wait() + for frame in self._rest_of_the_turn: + await asyncio.sleep(0) + if self._closed: + return + self._frames.put_nowait(frame) + + async def recv(self, decode: bool | None = None) -> str: + from websockets.exceptions import ConnectionClosedOK + from websockets.frames import Close + + frame: Final = await self._frames.get() + if frame is None: + raise ConnectionClosedOK(rcvd=Close(1000, ""), sent=Close(1000, ""), rcvd_then_sent=True) + return frame + + +@pytest.mark.asyncio +@pytest.mark.parametrize("hang_up_read_before_the_next_send", [True, False], ids=["read_first", "found_by_a_send"]) +async def test_websocket_passthrough_bills_the_turn_a_client_walked_out_of_once_the_provider_ends_it( + hang_up_read_before_the_next_send: bool, +): + """ + Leaving mid-answer must not make a turn free. The provider keeps running and + bills it, so the gateway keeps reading until the turn ends and logs its usage, + whether it reads the hang-up first or finds it by failing to relay a frame. + """ + client_left: Final = asyncio.Event() + client: Final = _walking_out_client( + [json.dumps({"type": "response.create", "model": "gpt-5.4"})], + '"response.created"', + None, + client_left, + hang_up_read_before_the_next_send, + ) + upstream_ws: Final = UpstreamFinishingTheTurnAfterTheClientLeft( + client_left, + _responses_turn_started("gpt-5.4"), + (json.dumps({"type": "response.output_text.delta", "delta": "Hi"}), _responses_turn_ended("gpt-5.4")), + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), walkout_turn_drain_seconds=5) + + assert [ + [json.loads(frame)["type"] if isinstance(frame, str) else frame["type"] for frame in logged.kwargs["response_body"]] + for logged in relayed.success_handler.call_args_list + ] == [["response.created", "response.output_text.delta", "response.completed"]] + upstream_ws.close.assert_awaited() + + +def _walking_out_client( + frames: Sequence[str], + answered_by: str, + answers_due: Sequence[int] | None, + client_left: asyncio.Event, + hang_up_read_before_the_next_send: bool = True, +) -> MagicMock: + """A turn-taking client that sets client_left as it hangs up, which the gateway either reads at once or only finds + when its next send to the client fails""" + from starlette.websockets import WebSocketDisconnect, WebSocketState + + send_failed: Final = asyncio.Event() + client: Final = _turn_taking_client(frames, answered_by, answers_due) + receive: Final = client.receive + send_text: Final = client.send_text + + async def receive_until_leaving() -> Mapping[str, object]: + message: Final = await receive() + if message["type"] == "websocket.disconnect": + client_left.set() + if not hang_up_read_before_the_next_send: + await send_failed.wait() + client.client_state = WebSocketState.DISCONNECTED + return message + + async def send_text_until_leaving(text: str) -> None: + if client_left.is_set(): + send_failed.set() + raise WebSocketDisconnect(code=1006) + await send_text(text) + + client.receive = receive_until_leaving + client.send_text = AsyncMock(side_effect=send_text_until_leaving) + return client + + +@pytest.mark.asyncio +@pytest.mark.parametrize("turns", [1, 2], ids=["one_turn", "two_turns_asked_at_once"]) +async def test_websocket_passthrough_bills_the_turn_a_client_walked_out_of_before_the_provider_started_it(turns: int): + """ + A client can ask for turns and leave before the provider says they started. + The provider still runs and bills each one, so the gateway keeps reading + until the last one ends and logs each one's usage. + """ + client_left: Final = asyncio.Event() + client: Final = _walking_out_client( + [json.dumps({"type": "response.create", "model": "gpt-5.4"})] * turns, + '"response.created"', + (0,) * (turns + 1), + client_left, + ) + upstream_ws: Final = UpstreamFinishingTheTurnAfterTheClientLeft( + client_left, None, (_responses_turn_started("gpt-5.4"), _responses_turn_ended("gpt-5.4")) * turns + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), walkout_turn_drain_seconds=5) + + assert _logged_frame_types(relayed) == [["response.created", "response.completed"]] * turns + upstream_ws.close.assert_awaited() + + +@pytest.mark.asyncio +@pytest.mark.parametrize("hang_up_read_before_the_next_send", [True, False], ids=["read_first", "found_by_a_send"]) +async def test_websocket_passthrough_bills_a_realtime_turn_a_client_walked_out_of_while_the_provider_stored_its_input( + hang_up_read_before_the_next_send: bool, +): + """ + A realtime client can add an item, ask for a turn, and leave at once. The + provider confirms the item before it starts the turn, and that confirmation + must not end the session while the turn it was asked for is still owed. + """ + client_left: Final = asyncio.Event() + client: Final = _walking_out_client( + [ + json.dumps({"type": "conversation.item.create", "item": {"type": "message", "role": "user"}}), + json.dumps({"type": "response.create"}), + ], + '"session.created"', + (1, 1, 1), + client_left, + hang_up_read_before_the_next_send, + ) + upstream_ws: Final = UpstreamFinishingTheTurnAfterTheClientLeft( + client_left, + None, + ( + json.dumps({"type": "conversation.item.created", "item": {"type": "message", "role": "user"}}), + json.dumps({"type": "response.created", "response": {"status": "in_progress"}}), + REALTIME_RESPONSE_DONE, + ), + opening=(REALTIME_SESSION_CREATED,), + ) + + relayed: Final = await _relay_turns( + client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime", walkout_turn_drain_seconds=5 + ) + + assert _logged_frame_types(relayed) == [ + ["session.created", "conversation.item.created", "response.created", "response.done"] + ] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_bills_a_realtime_turn_a_client_walked_out_of_after_the_provider_started_one_itself(): + """ + A realtime provider can start a turn on its own when voice detection hears + the user stop. That turn answers no client frame, so a turn the client asks + for next and leaves before it starts is still owed and still billed. + """ + realtime_response_created: Final = json.dumps({"type": "response.created", "response": {"status": "in_progress"}}) + client_left: Final = asyncio.Event() + client: Final = _walking_out_client( + [json.dumps({"type": "response.create"})], '"response.done"', (1, 1), client_left + ) + upstream_ws: Final = UpstreamFinishingTheTurnAfterTheClientLeft( + client_left, + None, + (realtime_response_created, REALTIME_RESPONSE_DONE), + opening=(REALTIME_SESSION_CREATED, realtime_response_created, REALTIME_RESPONSE_DONE), + ) + + relayed: Final = await _relay_turns( + client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime", walkout_turn_drain_seconds=5 + ) + + assert _logged_frame_types(relayed) == [ + ["session.created", "response.created", "response.done"], + ["session.created", "response.created", "response.done"], + ] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "answer", + [ + _responses_turn("gpt-5.4"), + [_responses_turn_ended("gpt-5.4")], + [_responses_error("previous_response_not_found")], + ], + ids=["finished", "finished_with_no_start_frame", "refused"], +) +async def test_websocket_passthrough_lets_go_of_the_provider_at_once_when_a_client_leaves_owing_no_turn( + answer: Sequence[str], +): + """ + Once every turn a client asked for has ended or been refused, nothing more + can be billed, so a client leaving closes the provider socket right away + instead of holding it for the walkout drain. + """ + client: Final = _turn_taking_client( + [json.dumps({"type": "response.create", "model": "gpt-5.4"})], + answered_by=json.dumps(json.loads(answer[-1])["type"]), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket(replies=[answer]) + + relayed: Final = await _relay_turns( + client, upstream_ws, RecordingFrameModelGate(), walkout_turn_drain_seconds=WALKOUT_TURN_DRAIN_SECONDS + ) + + upstream_ws.close.assert_awaited() + assert len(relayed.success_handler.call_args_list) == 1 + + +@pytest.mark.asyncio +async def test_websocket_passthrough_logs_no_extra_row_for_what_the_provider_sends_after_the_last_turn(): + """ + A realtime provider follows each finished turn with frames that start no + turn, such as rate_limits.updated. The session ends with nothing running, + so its turn is its only row and the request count stays one per turn. + """ + client: Final = _turn_taking_client([json.dumps({"type": "response.create"})], answered_by='"rate_limits.updated"') + upstream_ws: Final = ScriptedUpstreamWebSocket( + REALTIME_SESSION_CREATED, + replies=[[REALTIME_RESPONSE_DONE, json.dumps({"type": "rate_limits.updated", "rate_limits": []})]], + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime") + + assert [ + [frame["type"] for frame in logged.kwargs["response_body"]] + for logged in relayed.success_handler.call_args_list + ] == [["session.created", "response.done"]] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_carries_the_session_the_provider_opened_into_every_later_turn(): + """ + A realtime provider names the session's model once, when it opens the + session, and never again. Each later turn logs on its own, so it has to + carry that opening frame or it would be priced as a model nobody named. + """ + client: Final = _turn_taking_client( + [json.dumps({"type": "response.create"}), json.dumps({"type": "response.create"})], + answered_by='"response.done"', + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + REALTIME_SESSION_CREATED, replies=[[REALTIME_RESPONSE_DONE], [REALTIME_RESPONSE_DONE]] + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime") + + assert [ + [frame["type"] for frame in logged.kwargs["response_body"]] + for logged in relayed.success_handler.call_args_list + ] == [["session.created", "response.done"], ["session.created", "response.done"]] + + +REALTIME_TRANSCRIPTION_COMPLETED: Final = json.dumps( + { + "type": "conversation.item.input_audio_transcription.completed", + "item_id": "item_fake", + "transcript": "hello", + "usage": {"type": "duration", "seconds": 2}, + } +) +TRANSCRIPTION_COMPLETED: Final = "conversation.item.input_audio_transcription.completed" + + +def _transcription_session_update(frame_type: str, model: str) -> str: + return json.dumps({"type": frame_type, "session": {"audio": {"input": {"transcription": {"model": model}}}}}) + + +def _logged_frame_types(relayed: SimpleNamespace) -> list[list[str]]: + return [ + [frame["type"] for frame in logged.kwargs["response_body"]] + for logged in relayed.success_handler.call_args_list + ] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_bills_a_clip_transcribed_after_the_last_answer(): + client: Final = _turn_taking_client( + [_transcription_session_update("session.update", "gpt-transcribe"), json.dumps({"type": "response.create"})], + answered_by='"type"', + answers_due=(1, 2, 4), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + REALTIME_SESSION_CREATED, + replies=[ + [_transcription_session_update("session.updated", "gpt-transcribe")], + [REALTIME_RESPONSE_DONE, REALTIME_TRANSCRIPTION_COMPLETED], + ], + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime") + + assert _logged_frame_types(relayed) == [ + ["session.created", "session.updated", "response.done"], + ["session.created", "session.updated", TRANSCRIPTION_COMPLETED], + ] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_bills_a_clip_transcribed_during_an_answer_in_that_answer_row(): + client: Final = _turn_taking_client( + [json.dumps({"type": "response.create"})], answered_by='"type"', answers_due=(1, 4) + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + REALTIME_SESSION_CREATED, + replies=[ + [ + json.dumps({"type": "response.created", "response": {"status": "in_progress"}}), + REALTIME_TRANSCRIPTION_COMPLETED, + REALTIME_RESPONSE_DONE, + ] + ], + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime") + + assert _logged_frame_types(relayed) == [ + ["session.created", "response.created", TRANSCRIPTION_COMPLETED, "response.done"] + ] + + +@pytest.mark.asyncio +async def test_websocket_passthrough_carries_only_the_latest_session_update_into_later_rows(): + client: Final = _turn_taking_client( + [ + _transcription_session_update("session.update", "gpt-transcribe-a"), + json.dumps({"type": "response.create"}), + _transcription_session_update("session.update", "gpt-transcribe-b"), + json.dumps({"type": "response.create"}), + ], + answered_by='"type"', + answers_due=(1, 2, 3, 4, 5), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + REALTIME_SESSION_CREATED, + replies=[ + [_transcription_session_update("session.updated", "gpt-transcribe-a")], + [REALTIME_RESPONSE_DONE], + [_transcription_session_update("session.updated", "gpt-transcribe-b")], + [REALTIME_RESPONSE_DONE], + ], + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime") + + last_row: Final = relayed.success_handler.call_args_list[-1].kwargs["response_body"] + assert [frame for frame in last_row if frame["type"] == "session.updated"] == [ + json.loads(_transcription_session_update("session.updated", "gpt-transcribe-b")) + ] + assert len(relayed.success_handler.call_args_list) == 2 + + +@pytest.mark.asyncio +async def test_websocket_passthrough_keeps_the_session_update_a_clip_ran_under_until_its_answer_is_billed(): + client: Final = _turn_taking_client( + [ + _transcription_session_update("session.update", "gpt-transcribe-a"), + json.dumps({"type": "response.create"}), + _transcription_session_update("session.update", "gpt-transcribe-b"), + ], + answered_by='"type"', + answers_due=(1, 2, 4, 6), + ) + upstream_ws: Final = ScriptedUpstreamWebSocket( + REALTIME_SESSION_CREATED, + replies=[ + [_transcription_session_update("session.updated", "gpt-transcribe-a")], + [ + json.dumps({"type": "response.created", "response": {"status": "in_progress"}}), + REALTIME_TRANSCRIPTION_COMPLETED, + ], + [_transcription_session_update("session.updated", "gpt-transcribe-b"), REALTIME_RESPONSE_DONE], + ], + ) + + relayed: Final = await _relay_turns(client, upstream_ws, RecordingFrameModelGate(), "/openai/v1/realtime") + + assert _logged_frame_types(relayed) == [ + [ + "session.created", + "session.updated", + "response.created", + TRANSCRIPTION_COMPLETED, + "session.updated", + "response.done", + ] + ] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "later_frame", + [ + json.dumps({"type": "response.create", "model": "gpt-5.4"}), + '{"type": "response.create", "model": "gpt-5.4-mini", "metadata": ' + "[" * 100_000 + "]" * 100_000 + "}", + ( + '{"type": "response.create", "model": "gpt-5.4-mini", "seed": ' + + "9" * (sys.get_int_max_str_digits() + 1) + + "}" + ), + ], + ids=["names_a_denied_model", "nests_too_deep_to_parse", "holds_a_number_too_long_to_parse"], +) +async def test_websocket_passthrough_still_bills_the_answer_in_flight_when_a_later_frame_is_refused(later_frame: str): + """ + A refused frame closes the client while the provider may still be sending the + answer to an earlier, allowed frame. That answer cannot go to a closed client, + but the provider bills it once it ends, so the gateway keeps reading until it + ends and its usage reaches spend. + """ + gate: Final = RecordingFrameModelGate(MappingProxyType({"gpt-5.4": KEY_MODEL_REFUSAL})) + client: Final = _turn_taking_client( + [json.dumps({"type": "response.create", "model": "gpt-5.4-mini"}), later_frame], + answered_by='"response.created"', + ) + refused: Final = asyncio.Event() + close_client: Final = client.close.side_effect + + def close_client_and_note_the_refusal(*args: object, **kwargs: object) -> None: + close_client(*args, **kwargs) + refused.set() + + client.close = AsyncMock(side_effect=close_client_and_note_the_refusal) + upstream_ws: Final = UpstreamFinishingTheTurnAfterTheClientLeft( + refused, _responses_turn_started("gpt-5.4-mini"), (_responses_turn_ended("gpt-5.4-mini"),) + ) + + relayed: Final = await _relay_turns(client, upstream_ws, gate, walkout_turn_drain_seconds=5) + + assert [json.loads(text.args[0])["type"] for text in client.send_text.await_args_list] == [ + "response.created", + "error", + ] + assert [ + [frame["type"] for frame in logged.kwargs["response_body"]] + for logged in relayed.success_handler.call_args_list + ] == [["response.created", "response.completed"]] + relayed.proxy_logging.post_call_failure_hook.assert_not_awaited() + + DEEPGRAM_LISTEN_TARGET = "wss://api.deepgram.com/v1/listen?model=nova-3&encoding=linear16&sample_rate=16000" DEEPGRAM_INTERIM_FRAME = json.dumps( { @@ -8962,6 +10461,32 @@ def test_passthrough_client_cannot_forge_session_id_omission(client_metadata_key ) +async def _billed_metadata(gate: WebsocketFrameModelGate) -> Mapping[str, object]: + served: Final = await _relay_client_frames( + [json.dumps({"type": "response.create", "model": "gpt-5.4-mini"})], + gate, + UserAPIKeyAuth(api_key="hashed-key"), + ) + return served.success_handler.call_args.kwargs["litellm_params"]["metadata"] + + +@pytest.mark.asyncio +async def test_websocket_session_is_charged_to_the_group_its_frames_were_authorized_through(): + """A websocket names its model in a frame, long after the relay built its logging metadata, so the + groups the gate matched have to be picked up at the end or the group pays for nothing it allowed.""" + metadata: Final = await _billed_metadata(RecordingFrameModelGate(groups=MappingProxyType({"gpt-5.4-mini": ("tier-a",)}))) + + assert metadata[MODEL_ACCESS_GROUP_METADATA_KEY] == ["tier-a"] + assert metadata["user_api_key"] == "hashed-key" + + +@pytest.mark.asyncio +async def test_websocket_session_that_matched_no_group_is_charged_to_none(): + metadata: Final = await _billed_metadata(RecordingFrameModelGate()) + + assert not metadata[MODEL_ACCESS_GROUP_METADATA_KEY] + + _GEMINI_FLASH_TARGET: Final = "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent" _VERTEX_FLASH_TARGET: Final = ( "https://us-east5-aiplatform.googleapis.com/v1/projects/p/locations/us-east5" diff --git a/tests/unit/proxy/pass_through_endpoints/test_success_handler.py b/tests/unit/proxy/pass_through_endpoints/test_success_handler.py index cf00545f331..4934573a777 100644 --- a/tests/unit/proxy/pass_through_endpoints/test_success_handler.py +++ b/tests/unit/proxy/pass_through_endpoints/test_success_handler.py @@ -652,6 +652,59 @@ def test_vertex_ai_live_route_without_usage_frames_yields_no_logging_response( assert normalized["kwargs"] == {"model": _LIVE_MODEL} +REALTIME_FRAMES: Final = [ + {"type": "session.created", "session": {"id": "sess_fake", "model": "gpt-realtime-2.1"}}, + { + "type": "response.done", + "response": { + "usage": { + "input_tokens": 33, + "output_tokens": 16, + "total_tokens": 49, + "input_token_details": {"text_tokens": 33, "audio_tokens": 0, "cached_tokens": 0}, + "output_token_details": {"text_tokens": 16, "audio_tokens": 0}, + } + }, + }, +] + + +def _normalize(url_route: str, response_body: dict[str, object] | list[dict[str, object]]) -> dict: + logging_obj: Final = MagicMock(model_call_details={}, litellm_call_id="call-fake") + return PassThroughEndpointLogging().normalize_llm_passthrough_logging_payload( + httpx_response=MagicMock(spec=httpx.Response), + response_body=response_body, + request_body={}, + logging_obj=logging_obj, + url_route=url_route, + result="websocket_connection_successful", + start_time=_START, + end_time=_END, + cache_hit=False, + litellm_call_id="call-fake", + ) + + +@pytest.mark.parametrize("url_route", ["/openai/v1/realtime", "/openai_passthrough/v1/realtime"]) +def test_openai_websocket_frames_reach_the_websocket_cost_handler(url_route): + normalized: Final = _normalize(url_route, REALTIME_FRAMES) + + assert normalized["kwargs"]["model"] == "gpt-realtime-2.1" + assert normalized["kwargs"]["custom_llm_provider"] == "openai" + prices: Final = litellm.model_cost["gpt-realtime-2.1"] + assert normalized["kwargs"]["response_cost"] == pytest.approx( + 33 * prices["input_cost_per_token"] + 16 * prices["output_cost_per_token"] + ) + assert normalized["standard_logging_response_object"].usage.total_tokens == 49 + + +def test_openai_http_passthrough_bodies_are_left_to_the_http_handlers(): + normalized: Final = _normalize("/openai/v1/responses", {"model": "gpt-5.4-mini", "usage": {"input_tokens": 13}}) + + assert normalized["standard_logging_response_object"] is None + assert normalized["kwargs"] == {"litellm_call_id": "call-fake"} + + @pytest.mark.asyncio @pytest.mark.parametrize("cancel_mode", ["once", "twice", "scope", "none"]) @pytest.mark.parametrize("cancel_phase", ["allocation", "lock"]) diff --git a/tests/unit/proxy/test_openai_ws_passthrough_routes.py b/tests/unit/proxy/test_openai_ws_passthrough_routes.py index 6a9cd972dc2..839ab49408f 100644 --- a/tests/unit/proxy/test_openai_ws_passthrough_routes.py +++ b/tests/unit/proxy/test_openai_ws_passthrough_routes.py @@ -1,31 +1,41 @@ -"""OpenAI passthrough WebSocket route: registration, opt-in gating, and refusals.""" +"""OpenAI passthrough WebSocket route: registration, opt-in gating, refusals, and per-frame model checks.""" import json -from collections.abc import Mapping, Sequence +from collections.abc import Mapping from dataclasses import dataclass from types import MappingProxyType, SimpleNamespace from typing import Final -from unittest.mock import patch +from unittest.mock import AsyncMock, patch import httpx import pytest import respx +from fastapi import HTTPException from starlette.routing import WebSocketRoute +from starlette.websockets import WebSocket import litellm from litellm.llms.openai.workload_identity import _workload_identity_auth -from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy._types import Litellm_EntityType, ProxyErrorTypes, ProxyException, UserAPIKeyAuth from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import ( + _OPENAI_WS_ALIASED_MODEL_REFUSAL, _OPENAI_WS_DISABLED_REFUSAL, - _OPENAI_WS_MODEL_RESTRICTED_REFUSAL, - _has_model_restrictions, + _OPENAI_WS_UNAVAILABLE_AUTH_REFUSAL, _openai_websocket_refusal, - _proxy_model_allowlists, + _proxy_frame_model_gate, openai_websocket_proxy_route, router, ) - -Scopes = tuple[Sequence[str], ...] +from litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_websocket_passthrough_logging_handler import ( + openai_websocket_frame_kind, +) +from litellm.proxy.pass_through_endpoints.pass_through_endpoints import ( + WebsocketFrameModelAllowed, + WebsocketFrameModelGate, + WebsocketFrameModelRefused, + WebsocketFrameModelVerdict, + WebsocketUpstreamFrameClassifier, +) ENABLED: Final = MappingProxyType({"enable_openai_websocket_passthrough": True}) DISABLED_SETTINGS: Final = ( @@ -76,6 +86,8 @@ class _RelayCall: forward_headers: bool endpoint: str accept_websocket: bool + client_frame_model_gate: WebsocketFrameModelGate | None + upstream_frame_kind: WebsocketUpstreamFrameClassifier | None class _FakeRelay: @@ -92,6 +104,8 @@ class _FakeRelay: forward_headers: bool, endpoint: str, accept_websocket: bool, + client_frame_model_gate: WebsocketFrameModelGate | None = None, + upstream_frame_kind: WebsocketUpstreamFrameClassifier | None = None, ) -> None: self.calls.append( _RelayCall( @@ -100,24 +114,26 @@ class _FakeRelay: forward_headers=forward_headers, endpoint=endpoint, accept_websocket=accept_websocket, + client_frame_model_gate=client_frame_model_gate, + upstream_frame_kind=upstream_frame_kind, ) ) -class _FakeModelAllowlists: - def __init__(self, scopes: Scopes) -> None: - self.scopes = scopes - self.calls: list[UserAPIKeyAuth] = [] +class _FakeFrameModelGate: + def __init__(self, refusals: Mapping[str, WebsocketFrameModelRefused] = MappingProxyType({})) -> None: + self.refusals = refusals + self.checked: list[str] = [] - async def __call__(self, valid_token: UserAPIKeyAuth, /) -> Scopes: - self.calls.append(valid_token) - return self.scopes + async def __call__(self, websocket: object, model: str, /) -> WebsocketFrameModelVerdict: + self.checked.append(model) + return self.refusals.get(model) or WebsocketFrameModelAllowed(access_groups=()) @dataclass(frozen=True, slots=True) class _Served: relay: _FakeRelay - allowlists: _FakeModelAllowlists + frame_model_gate: _FakeFrameModelGate async def _serve( @@ -125,16 +141,16 @@ async def _serve( endpoint: str, user_api_key_dict: UserAPIKeyAuth, general_settings: Mapping[str, object], - scopes: Scopes = (), + frame_model_gate: _FakeFrameModelGate | None = None, ) -> _Served: - served = _Served(relay=_FakeRelay(), allowlists=_FakeModelAllowlists(scopes)) + served: Final = _Served(relay=_FakeRelay(), frame_model_gate=frame_model_gate or _FakeFrameModelGate()) await openai_websocket_proxy_route( websocket=websocket, endpoint=endpoint, user_api_key_dict=user_api_key_dict, general_settings=general_settings, relay=served.relay, - model_allowlists=served.allowlists, + frame_model_gate=served.frame_model_gate, ) return served @@ -143,18 +159,20 @@ async def _serve( @pytest.mark.parametrize("prefix", ["openai", "openai_passthrough"]) async def test_openai_websocket_forwards_query_and_keeps_provider_auth(prefix, monkeypatch): monkeypatch.delenv("OPENAI_API_BASE", raising=False) - websocket = _FakeWebSocket(f"/{prefix}/v1/realtime", "model=gpt-4o-realtime-preview") + websocket: Final = _FakeWebSocket(f"/{prefix}/v1/realtime", "model=gpt-realtime-2.1") with patch(GET_CREDENTIALS, return_value="sk-provider"): served = await _serve(websocket, "v1/realtime", UserAPIKeyAuth(), ENABLED) assert served.relay.calls == [ _RelayCall( - target="wss://api.openai.com/v1/realtime?model=gpt-4o-realtime-preview", + target="wss://api.openai.com/v1/realtime?model=gpt-realtime-2.1", custom_headers=MappingProxyType({"Authorization": "Bearer sk-provider"}), forward_headers=False, endpoint=f"/{prefix}/v1/realtime", accept_websocket=False, + client_frame_model_gate=served.frame_model_gate, + upstream_frame_kind=openai_websocket_frame_kind, ) ] assert websocket.accepts == [None] @@ -162,11 +180,23 @@ async def test_openai_websocket_forwards_query_and_keeps_provider_auth(prefix, m assert websocket.closed is None +@pytest.mark.asyncio +async def test_openai_websocket_reports_its_route_without_the_proxy_root_path(monkeypatch): + """Route allowlists and spend rows key on the route the proxy serves, which a root path prefix is not part of""" + monkeypatch.setenv("SERVER_ROOT_PATH", "/litellm") + websocket: Final = _FakeWebSocket("/litellm/openai/v1/responses", "") + + with patch(GET_CREDENTIALS, return_value="sk-provider"): + served: Final = await _serve(websocket, "v1/responses", UserAPIKeyAuth(), ENABLED) + + assert [call.endpoint for call in served.relay.calls] == ["/openai/v1/responses"] + + @pytest.mark.asyncio async def test_openai_websocket_accepts_first_client_subprotocol(): websocket = _FakeWebSocket( "/openai/v1/realtime", - "model=gpt-4o-realtime-preview", + "model=gpt-realtime-2.1", subprotocols="realtime, openai-insecure-api-key.sk-abc, openai-beta.realtime-v1", ) @@ -239,7 +269,7 @@ async def test_openai_websocket_closes_cleanly_when_workload_identity_exchange_f @pytest.mark.asyncio async def test_openai_websocket_closes_cleanly_when_provider_credentials_missing(): - websocket = _FakeWebSocket("/openai/v1/realtime", "model=gpt-4o-realtime-preview") + websocket: Final = _FakeWebSocket("/openai/v1/realtime", "model=gpt-realtime-2.1") with patch(GET_CREDENTIALS, return_value=None): served = await _serve(websocket, "v1/realtime", UserAPIKeyAuth(), ENABLED) @@ -255,7 +285,7 @@ async def test_openai_websocket_closes_cleanly_when_provider_credentials_missing @pytest.mark.parametrize("prefix", ["openai", "openai_passthrough"]) @pytest.mark.parametrize("general_settings", DISABLED_SETTINGS) async def test_openai_websocket_refused_unless_explicitly_enabled(prefix, general_settings): - websocket = _FakeWebSocket(f"/{prefix}/v1/realtime", "model=gpt-4o-realtime-preview") + websocket: Final = _FakeWebSocket(f"/{prefix}/v1/realtime", "model=gpt-realtime-2.1") served = await _serve(websocket, "v1/realtime", UserAPIKeyAuth(), general_settings) @@ -265,25 +295,21 @@ async def test_openai_websocket_refused_unless_explicitly_enabled(prefix, genera assert served.relay.calls == [] -@pytest.mark.asyncio @pytest.mark.parametrize("general_settings", DISABLED_SETTINGS) -async def test_openai_websocket_refusal_is_disabled_for_falsy_settings(general_settings): - refusal = await _openai_websocket_refusal(UserAPIKeyAuth(), general_settings, _FakeModelAllowlists(())) - assert refusal is _OPENAI_WS_DISABLED_REFUSAL +def test_openai_websocket_refusal_is_disabled_for_falsy_settings(general_settings): + assert _openai_websocket_refusal(general_settings) is _OPENAI_WS_DISABLED_REFUSAL -@pytest.mark.asyncio @pytest.mark.parametrize("value", [True, "true", "True"]) -async def test_openai_websocket_refusal_is_none_for_truthy_settings(value): - settings = MappingProxyType({"enable_openai_websocket_passthrough": value}) - assert await _openai_websocket_refusal(UserAPIKeyAuth(), settings, _FakeModelAllowlists(())) is None +def test_openai_websocket_refusal_is_none_for_truthy_settings(value): + assert _openai_websocket_refusal(MappingProxyType({"enable_openai_websocket_passthrough": value})) is None @pytest.mark.asyncio async def test_openai_websocket_refusal_echoes_requested_subprotocol(): websocket = _FakeWebSocket( "/openai_passthrough/v1/realtime", - "model=gpt-4o-realtime-preview", + "model=gpt-realtime-2.1", subprotocols="realtime, openai-beta.realtime-v1", ) @@ -294,68 +320,303 @@ async def test_openai_websocket_refusal_echoes_requested_subprotocol(): assert served.relay.calls == [] -RESTRICTED_SCOPES: Final[tuple[Scopes, ...]] = ( - (("gpt-4o",),), - ((), ("gpt-4o-realtime-preview",)), - (("all-team-models",), ("gpt-4o",)), - ((), ("all-proxy-models",), ("gpt-4o",)), - ((), (), (), ("gpt-4o",)), - (("*",), (), (), (), ("gpt-4o",)), +@pytest.mark.asyncio +@pytest.mark.parametrize( + "key_models", + [["gpt-5.4-mini"], ["gpt-realtime-2.1", "gpt-5.4"], []], ) -UNRESTRICTED_SCOPES: Final[tuple[Scopes, ...]] = ( - (), - ((),), - (("all-proxy-models",),), - (("*",),), - (("all-team-models",), ("all-proxy-models",)), - ((), (), (), (), ()), - (("*",), ("all-proxy-models",), ("all-team-models",), (), ()), -) - - -@pytest.mark.asyncio -@pytest.mark.parametrize("scopes", RESTRICTED_SCOPES) -async def test_openai_websocket_rejects_model_restricted_identities(scopes): - websocket = _FakeWebSocket("/openai/v1/realtime", "model=gpt-4o-realtime-preview") - user_api_key_dict = UserAPIKeyAuth(token="hashed-fake", user_id="user-fake", team_id="team-fake") - - served = await _serve(websocket, "v1/realtime", user_api_key_dict, ENABLED, scopes) - - assert "model restrictions" in websocket.error_message() - assert websocket.closed == (1008, _OPENAI_WS_MODEL_RESTRICTED_REFUSAL.close_reason) - assert served.relay.calls == [] - assert served.allowlists.calls == [user_api_key_dict] - - -@pytest.mark.asyncio -@pytest.mark.parametrize("scopes", RESTRICTED_SCOPES) -async def test_openai_websocket_disabled_refusal_skips_allowlist_lookups(scopes): - allowlists = _FakeModelAllowlists(scopes) - - refusal = await _openai_websocket_refusal(UserAPIKeyAuth(), MappingProxyType({}), allowlists) - - assert refusal is _OPENAI_WS_DISABLED_REFUSAL - assert allowlists.calls == [] - - -@pytest.mark.asyncio -@pytest.mark.parametrize("scopes", UNRESTRICTED_SCOPES) -async def test_openai_websocket_allows_unrestricted_identities(scopes): - websocket = _FakeWebSocket("/openai/v1/responses", "") +async def test_openai_websocket_connects_model_restricted_keys(key_models): + websocket: Final = _FakeWebSocket("/openai/v1/responses", "") + user_api_key_dict: Final = UserAPIKeyAuth(models=key_models, token="hashed-fake", user_id="user-fake") with patch(GET_CREDENTIALS, return_value="sk-provider"): - served = await _serve(websocket, "v1/responses", UserAPIKeyAuth(), ENABLED, scopes) + served: Final = await _serve(websocket, "v1/responses", user_api_key_dict, ENABLED) assert len(served.relay.calls) == 1 + assert served.relay.calls[0].client_frame_model_gate is served.frame_model_gate assert websocket.sent == [] assert websocket.closed is None -@pytest.mark.asyncio -async def test_proxy_model_allowlists_reads_the_token_scopes_without_a_database(): - token: Final = UserAPIKeyAuth(models=[], team_id="team-fake", team_models=["gpt-4o"]) - with patch("litellm.proxy.proxy_server.prisma_client", None): - scopes = await _proxy_model_allowlists()(token) +ALIAS_REFUSAL: Final = WebsocketFrameModelRefused( + message=_OPENAI_WS_ALIASED_MODEL_REFUSAL, error_type=ProxyErrorTypes.key_model_access_denied.value +) - assert tuple(tuple(scope) for scope in scopes) == ((), ("gpt-4o",)) - assert _has_model_restrictions(scopes) + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "query", + [ + "model=gpt-realtime-2.1", + "model=gpt-realtime-2.1&model=gpt-realtime-2.1-mini", + "model=gpt-realtime-2.1-mini&model=gpt-realtime-2.1", + ], + ids=["one_model", "refused_model_first", "refused_model_last"], +) +async def test_openai_websocket_refuses_an_opening_model_the_frame_gate_refuses(query): + """ + The relay sends the connection URL to OpenAI unchanged, so every model it names has to pass the + same gate as a frame, or an alias the upgrade resolves to an allowed model opens the session on the + alias name itself + """ + websocket: Final = _FakeWebSocket("/openai/v1/realtime", query) + gate: Final = _FakeFrameModelGate(MappingProxyType({"gpt-realtime-2.1": ALIAS_REFUSAL})) + + with patch(GET_CREDENTIALS, return_value="sk-provider"): + served: Final = await _serve( + websocket, "v1/realtime", UserAPIKeyAuth(models=["gpt-realtime-2.1-mini"]), ENABLED, gate + ) + + assert websocket.error_message() == _OPENAI_WS_ALIASED_MODEL_REFUSAL + assert websocket.closed == (1008, "This key cannot use a model alias over OpenAI websocket passthrough.") + assert served.relay.calls == [] + assert "gpt-realtime-2.1" in gate.checked + + +@pytest.mark.asyncio +async def test_openai_websocket_relays_a_session_once_every_opening_model_passes_the_frame_gate(): + websocket: Final = _FakeWebSocket("/openai/v1/realtime", "model=gpt-realtime-2.1-mini") + + with patch(GET_CREDENTIALS, return_value="sk-provider"): + served: Final = await _serve( + websocket, "v1/realtime", UserAPIKeyAuth(models=["gpt-realtime-2.1-mini"]), ENABLED + ) + + assert served.frame_model_gate.checked == ["gpt-realtime-2.1-mini"] + assert len(served.relay.calls) == 1 + assert websocket.closed is None + + +class _FrameAuthorizer: + """Stands in for the full key, team, user and budget auth the gate re-runs for every model a frame names.""" + + def __init__(self, token: UserAPIKeyAuth | None = None, denial: Exception | None = None) -> None: + self.token = token or UserAPIKeyAuth() + self.denial = denial + self.authorized: list[str] = [] + + async def __call__(self, websocket: object, model: str) -> UserAPIKeyAuth: + self.authorized.append(model) + if self.denial is not None: + raise self.denial + return self.token + + +SOCKET: Final = WebSocket({"type": "websocket"}, receive=AsyncMock(), send=AsyncMock()) + + +async def _gate_verdict(authorizer: _FrameAuthorizer, model: str) -> WebsocketFrameModelVerdict: + with patch("litellm.proxy.proxy_server.prisma_client", None): # test-quality-ok: the model checks need no database + return await _proxy_frame_model_gate(authorizer)(SOCKET, model) + + +def _refusal(verdict: WebsocketFrameModelVerdict) -> WebsocketFrameModelRefused: + assert isinstance(verdict, WebsocketFrameModelRefused) + return verdict + + +@pytest.mark.asyncio +async def test_proxy_frame_model_gate_runs_full_auth_for_the_model_the_frame_names(): + authorizer: Final = _FrameAuthorizer(UserAPIKeyAuth(models=["gpt-realtime-2.1", "gpt-5.4-mini"])) + + verdict: Final = await _gate_verdict(authorizer, "gpt-realtime-2.1") + + assert verdict == WebsocketFrameModelAllowed(access_groups=()) + assert authorizer.authorized == ["gpt-realtime-2.1"] + + +@pytest.mark.asyncio +@pytest.mark.parametrize("requested_model", ["gpt-5.4", "gpt-realtime-2.1-mini"]) +async def test_proxy_frame_model_gate_refuses_a_model_the_key_lacks(requested_model): + authorizer: Final = _FrameAuthorizer(UserAPIKeyAuth(models=["gpt-realtime-2.1", "gpt-5.4-mini"])) + + refusal: Final = _refusal(await _gate_verdict(authorizer, requested_model)) + + assert requested_model in refusal.message + assert refusal.error_type == ProxyErrorTypes.key_model_access_denied.value + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "granted_models, refusal_names", + [(["gpt-5.4-mini"], "gpt-5.4"), ([], "no model grants")], + ids=["model_outside_the_grant", "no_grants"], +) +async def test_proxy_frame_model_gate_refuses_a_model_the_agent_was_not_granted(granted_models, refusal_names): + from litellm.types.agents import AgentResponse + + token: Final = UserAPIKeyAuth(models=[], agent_id="agent-fake") + token.managed_agent_policy = AgentResponse( + agent_id="agent-fake", + agent_name="Agent", + agent_card_params={}, + object_permission={"object_permission_id": "grant-fake", "models": granted_models}, + ) + + refusal: Final = _refusal(await _gate_verdict(_FrameAuthorizer(token), "gpt-5.4")) + + assert refusal_names in refusal.message + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "token", + [ + UserAPIKeyAuth(models=["gpt-5.4-mini"], aliases={"gpt-5.4": "gpt-5.4-mini"}), + UserAPIKeyAuth(models=["gpt-5.4-mini"], team_model_aliases={"gpt-5.4": "gpt-5.4-mini"}), + ], + ids=["key_alias", "team_alias"], +) +async def test_proxy_frame_model_gate_refuses_an_alias_that_would_reach_openai_unchanged(token): + """ + The passthrough sends each frame to OpenAI as written, so OpenAI runs the + alias name itself. An alias pointing at an allowed model must not let a key + limited to that model run the model the alias is named after. + """ + refusal: Final = _refusal(await _gate_verdict(_FrameAuthorizer(token), "gpt-5.4")) + + assert refusal == WebsocketFrameModelRefused( + message=_OPENAI_WS_ALIASED_MODEL_REFUSAL, error_type=ProxyErrorTypes.key_model_access_denied.value + ) + + +@pytest.mark.asyncio +async def test_proxy_frame_model_gate_refuses_a_proxy_wide_alias_for_a_key_limited_to_its_target(monkeypatch): + monkeypatch.setattr(litellm, "model_alias_map", {"gpt-5.4": "gpt-5.4-mini"}) + + refusal: Final = _refusal( + await _gate_verdict(_FrameAuthorizer(UserAPIKeyAuth(models=["gpt-5.4-mini"])), "gpt-5.4") + ) + + assert refusal.message == _OPENAI_WS_ALIASED_MODEL_REFUSAL + + +@pytest.mark.asyncio +async def test_proxy_frame_model_gate_lets_a_key_that_may_use_any_model_send_an_alias(): + token: Final = UserAPIKeyAuth(models=[], aliases={"fast": "gpt-5.4-mini"}) + + verdict: Final = await _gate_verdict(_FrameAuthorizer(token), "fast") + + assert verdict == WebsocketFrameModelAllowed(access_groups=()) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("unrestricted_models", [[], ["all-proxy-models"], ["*"]]) +async def test_proxy_frame_model_gate_allows_every_model_for_unrestricted_keys(unrestricted_models): + verdict: Final = await _gate_verdict(_FrameAuthorizer(UserAPIKeyAuth(models=unrestricted_models)), "gpt-5.4") + + assert verdict == WebsocketFrameModelAllowed(access_groups=()) + + +@pytest.mark.asyncio +async def test_proxy_frame_model_gate_hands_back_the_access_groups_that_authorized_the_model(): + """The turn a frame starts is billed to the groups that let its model through, so the gate has to report them.""" + token: Final = UserAPIKeyAuth(models=["gpt-5.4-mini"], matched_model_access_groups=["tier-a"]) + + verdict: Final = await _gate_verdict(_FrameAuthorizer(token), "gpt-5.4-mini") + + assert verdict == WebsocketFrameModelAllowed(access_groups=("tier-a",)) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "requested_model", + ["gpt-5.4-mini", "gpt-5.4"], + ids=["allowed", "refused_by_the_model_check"], +) +async def test_proxy_frame_model_gate_releases_the_budget_it_reserved_for_the_check(requested_model): + """ + Full auth reserves budget for the request it authorizes, but a frame check is + not a request, so a reservation the gate leaves open would hold the key's + spend counter above its real spend and refuse later calls on every route. + """ + reservation: Final[dict[str, object]] = {"reserved_cost": 0.5, "entries": []} + token: Final = UserAPIKeyAuth(models=["gpt-5.4-mini"], budget_reservation=reservation) + + await _gate_verdict(_FrameAuthorizer(token), requested_model) + + assert token.budget_reservation == {**reservation, "finalized": True} + + +@pytest.mark.asyncio +async def test_proxy_frame_model_gate_does_not_count_the_sessions_own_reservation_as_key_spend(): + """ + The upgrade reserves up to the key's remaining budget for the session, and the key budget check + counts open reservations as spend, so a frame checked while that reservation is open is refused + on a key with budget left, where an HTTP call with the same key goes through. + """ + connect_reservation: Final[dict[str, object]] = {"reserved_cost": 0.0001, "entries": []} + socket: Final = WebSocket( + {"type": "websocket", "state": {"budget_reservation": connect_reservation}}, receive=AsyncMock(), send=AsyncMock() + ) + + async def key_budget_check(websocket: object, model: str) -> UserAPIKeyAuth: + if connect_reservation.get("finalized") is not True: + raise litellm.BudgetExceededError(current_cost=0.0001, max_budget=0.0001) + return UserAPIKeyAuth(models=["gpt-5.4-mini"]) + + with patch("litellm.proxy.proxy_server.prisma_client", None): # test-quality-ok: the model checks need no database + verdict: Final = await _proxy_frame_model_gate(key_budget_check)(socket, "gpt-5.4-mini") + + assert verdict == WebsocketFrameModelAllowed(access_groups=()) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "denial, error_type, message", + [ + ( + ProxyException( + message="Budget has been exceeded! Key=hashed Spend=2.0, Budget=1.0", + type=ProxyErrorTypes.budget_exceeded, + param=None, + code=400, + ), + "budget_exceeded", + "Budget has been exceeded! Key=hashed", + ), + ( + litellm.BudgetExceededError( + current_cost=2.0, + max_budget=1.0, + message="Budget has been exceeded! Model access group=tier-a Current cost: 2.0, Max budget: 1.0", + entity_type=Litellm_EntityType.MODEL_ACCESS_GROUP.value, + entity_id="tier-a", + ), + "budget_exceeded", + "Model access group=tier-a", + ), + ( + ProxyException( + message="team not allowed to access model. This team can only access models=['gpt-5.4-mini']", + type=ProxyErrorTypes.team_model_access_denied, + param="model", + code=401, + ), + "team_model_access_denied", + "team not allowed to access model", + ), + (HTTPException(status_code=403, detail="Invalid Authorization header format"), "invalid_request_error", ""), + ], + ids=["key_budget", "group_budget", "team_allowlist", "malformed_auth"], +) +async def test_proxy_frame_model_gate_refuses_what_full_auth_denies_with_its_reason(denial, error_type, message): + """A frame is denied for the same reasons a request is, and the caller reads which reason in the error event.""" + refusal: Final = _refusal(await _gate_verdict(_FrameAuthorizer(denial=denial), "gpt-5.4-mini")) + + assert refusal.error_type == error_type + assert message in refusal.message + + +@pytest.mark.asyncio +async def test_proxy_frame_model_gate_refuses_out_loud_when_auth_cannot_answer(): + """ + A gate that cannot answer has to close the session with a reason the caller can read. + Letting the failure escape drops the socket with no frame at all, which the caller + cannot tell apart from the network going away. + """ + refusal: Final = _refusal( + await _gate_verdict(_FrameAuthorizer(denial=ConnectionError("connection pool is exhausted")), "gpt-5.4-mini") + ) + + assert refusal == WebsocketFrameModelRefused(message=_OPENAI_WS_UNAVAILABLE_AUTH_REFUSAL, error_type="server_error") diff --git a/tests/unit/test_cost_calculator.py b/tests/unit/test_cost_calculator.py index cca138e0bb5..aa759f2c4d9 100644 --- a/tests/unit/test_cost_calculator.py +++ b/tests/unit/test_cost_calculator.py @@ -1451,6 +1451,38 @@ def test_transcription_usage_cost_returns_zero_for_unknown_type(): assert _transcription_usage_cost({}, {}) == 0.0 +def test_realtime_transcription_prices_each_clip_at_the_transcription_model_set_when_it_ran(monkeypatch): + monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") + monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url="")) + + from litellm.cost_calculator import handle_realtime_transcription_cost_calculation + + realtime_model = "gpt-realtime-2.1-mini" + first_asr, second_asr = "gpt-4o-transcribe", "gpt-transcribe" + first_rate = litellm.model_cost[first_asr]["input_cost_per_second"] + second_rate = litellm.model_cost[second_asr]["input_cost_per_second"] + assert first_rate != second_rate and first_rate > 0 and second_rate > 0 + + def session_frame(frame_type: str, asr_model: str | None) -> dict[str, object]: + transcription = None if asr_model is None else {"model": asr_model} + return {"type": frame_type, "session": {"model": realtime_model, "audio": {"input": {"transcription": transcription}}}} + + clip = {"type": "conversation.item.input_audio_transcription.completed", "usage": {"type": "duration", "seconds": 5}} + results: OpenAIRealtimeStreamList = [ + session_frame("session.created", None), + session_frame("session.updated", first_asr), + clip, + session_frame("session.updated", second_asr), + clip, + ] + + cost = handle_realtime_transcription_cost_calculation( + results=results, custom_llm_provider="openai", litellm_model_name=realtime_model + ) + + assert cost == pytest.approx(5 * first_rate + 5 * second_rate, rel=1e-9) + + def test_get_transcription_model_falls_back_to_session_model(monkeypatch): """session.model is used when transcription-specific model fields are absent.""" monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")