From 7e1f44f0cc15e639d8fe95fc8831c5a0b25cc76b Mon Sep 17 00:00:00 2001 From: mateo Date: Thu, 30 Jul 2026 02:38:28 +0000 Subject: [PATCH 001/220] feat(proxy): add admin toggle to block requests for models without pricing Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/__init__.py | 1 + litellm/proxy/_types.py | 2 + litellm/proxy/auth/auth_checks.py | 33 +++++ .../cost_tracking_settings.py | 65 ++++++++ .../proxy/auth/test_auth_checks.py | 139 +++++++++++++++++- .../test_cost_tracking_settings.py | 56 +++++++ .../_components/cost_tracking_settings.tsx | 47 +++++- .../_components/use_block_unpriced_config.ts | 63 ++++++++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 81 ++++++++++ 9 files changed, 482 insertions(+), 5 deletions(-) create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/use_block_unpriced_config.ts diff --git a/litellm/__init__.py b/litellm/__init__.py index 3f8c742c5a2..d14f41ad49c 100644 --- a/litellm/__init__.py +++ b/litellm/__init__.py @@ -443,6 +443,7 @@ max_end_user_budget_id: Optional[str] = None # backwards compatibility — arbitrary client-supplied identifiers still # pass through unchanged. validate_end_user_id_in_db: bool = False +block_requests_for_models_without_pricing: bool = False disable_end_user_cost_tracking: Optional[bool] = None disable_end_user_cost_tracking_prometheus_only: Optional[bool] = None enable_end_user_cost_tracking_prometheus_only: Optional[bool] = None diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index c6d4ee1120a..450868edf06 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -3540,6 +3540,8 @@ class ProxyErrorTypes(str, enum.Enum): Project does not have access to the model """ + model_cost_map_missing = "model_cost_map_missing" + expired_key = "expired_key" """ Key has expired diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py index c46bc110ca8..11a45e78410 100644 --- a/litellm/proxy/auth/auth_checks.py +++ b/litellm/proxy/auth/auth_checks.py @@ -274,6 +274,22 @@ def _is_cost_explicitly_configured(model: str, llm_router: "Router") -> bool: return False +def model_has_no_cost_mapping(model: Optional[str], llm_router: Optional[Router]) -> bool: + if not model or llm_router is None: + return False + + model_group_info = llm_router.get_model_group_info(model_group=model) + if model_group_info is None: + return False + + input_cost = model_group_info.input_cost_per_token or 0 + output_cost = model_group_info.output_cost_per_token or 0 + if input_cost > 0 or output_cost > 0: + return False + + return not _is_cost_explicitly_configured(model, llm_router) + + async def _run_project_checks( project_object: Optional[LiteLLM_ProjectTableCachedObj], _model: Optional[Union[str, List[str]]], @@ -534,6 +550,23 @@ async def common_checks( if route in MODEL_DISCOVERY_ROUTES: skip_budget_checks = True + if ( + litellm.block_requests_for_models_without_pricing + and isinstance(_model, str) + and RouteChecks.is_llm_api_route(route=route) + and model_has_no_cost_mapping(model=_model, llm_router=llm_router) + ): + raise ProxyException( + message=( + f"Model '{_model}' has no pricing in the cost map, so its spend would be tracked as $0. " + "Requests for unpriced models are blocked because 'block_requests_for_models_without_pricing' " + "is enabled. Add pricing for this model (input_cost_per_token/output_cost_per_token) to allow it." + ), + type=ProxyErrorTypes.model_cost_map_missing, + param="model", + code=status.HTTP_403_FORBIDDEN, + ) + # 1. If team is blocked if team_object is not None and team_object.blocked is True: raise Exception(f"Team={team_object.team_id} is blocked. Update via `/team/unblock` if you're an admin.") diff --git a/litellm/proxy/management_endpoints/cost_tracking_settings.py b/litellm/proxy/management_endpoints/cost_tracking_settings.py index cd2c5704778..2c18de6b903 100644 --- a/litellm/proxy/management_endpoints/cost_tracking_settings.py +++ b/litellm/proxy/management_endpoints/cost_tracking_settings.py @@ -13,6 +13,7 @@ POST /cost/estimate - Estimate cost for a given model and token counts from typing import Dict, Optional, Tuple, Union from fastapi import APIRouter, Depends, HTTPException +from pydantic import BaseModel import litellm from litellm._logging import verbose_proxy_logger @@ -407,6 +408,70 @@ async def update_cost_margin_config( ) +class BlockUnpricedModelsRequest(BaseModel): + enabled: bool + + +class BlockUnpricedModelsResponse(BaseModel): + enabled: bool + + +@router.get( + "/config/block_requests_for_models_without_pricing", + tags=["Cost Tracking"], + dependencies=[Depends(user_api_key_auth)], + response_model=BlockUnpricedModelsResponse, +) +async def get_block_requests_for_models_without_pricing() -> BlockUnpricedModelsResponse: + return BlockUnpricedModelsResponse(enabled=bool(litellm.block_requests_for_models_without_pricing)) + + +@router.patch( + "/config/block_requests_for_models_without_pricing", + tags=["Cost Tracking"], + dependencies=[Depends(user_api_key_auth)], + response_model=BlockUnpricedModelsResponse, +) +async def update_block_requests_for_models_without_pricing( + request: BlockUnpricedModelsRequest, +) -> BlockUnpricedModelsResponse: + from litellm.proxy.proxy_server import ( + prisma_client, + proxy_config, + store_model_in_db, + ) + + if prisma_client is None: + raise HTTPException( + status_code=500, + detail={"error": CommonProxyErrors.db_not_connected_error.value}, + ) + + if store_model_in_db is not True: + raise HTTPException( + status_code=500, + detail={"error": "Set `'STORE_MODEL_IN_DB='True'` in your env to enable this feature."}, + ) + + try: + config = await proxy_config.get_config() + if "litellm_settings" not in config: + config["litellm_settings"] = {} + config["litellm_settings"]["block_requests_for_models_without_pricing"] = request.enabled + await proxy_config.save_config(new_config=config) + + litellm.block_requests_for_models_without_pricing = request.enabled + verbose_proxy_logger.info(f"Updated block_requests_for_models_without_pricing: {request.enabled}") + + return BlockUnpricedModelsResponse(enabled=request.enabled) + except Exception as e: + verbose_proxy_logger.error(f"Error updating block_requests_for_models_without_pricing: {str(e)}") + raise HTTPException( + status_code=500, + detail={"error": f"Failed to update setting: {str(e)}"}, + ) + + @router.post( "/cost/estimate", tags=["Cost Tracking"], diff --git a/tests/test_litellm/proxy/auth/test_auth_checks.py b/tests/test_litellm/proxy/auth/test_auth_checks.py index 34a353966bf..aec0ddc55f8 100644 --- a/tests/test_litellm/proxy/auth/test_auth_checks.py +++ b/tests/test_litellm/proxy/auth/test_auth_checks.py @@ -2,6 +2,7 @@ import asyncio import json import os import sys +from typing import Optional from unittest.mock import AsyncMock, MagicMock, patch sys.path.insert( @@ -5164,4 +5165,140 @@ async def test_get_project_object_db_fetch_returns_cached_obj(): assert isinstance(result, LiteLLM_ProjectTableCachedObj) assert result.project_id == "p-1" - assert result.project_alias == "proj" + + +UNPRICED_UNDERLYING_MODEL = "openai/unpriced-model-lit4984-xyz" + + +def _router_with_priced_and_unpriced_models() -> "Router": + from litellm.router import Router + + return Router( + model_list=[ + { + "model_name": "priced-group", + "litellm_params": {"model": "gpt-3.5-turbo", "api_key": "sk-test"}, + }, + { + "model_name": "unpriced-group", + "litellm_params": {"model": UNPRICED_UNDERLYING_MODEL, "api_key": "sk-test"}, + }, + ] + ) + + +def test_model_has_no_cost_mapping_priced_model_is_false(): + from litellm.proxy.auth.auth_checks import model_has_no_cost_mapping + + router = _router_with_priced_and_unpriced_models() + + assert model_has_no_cost_mapping(model="priced-group", llm_router=router) is False + + +def test_model_has_no_cost_mapping_unpriced_model_is_true(): + from litellm.proxy.auth.auth_checks import model_has_no_cost_mapping + + router = _router_with_priced_and_unpriced_models() + + assert model_has_no_cost_mapping(model="unpriced-group", llm_router=router) is True + + +def test_model_has_no_cost_mapping_no_model_or_router_is_false(): + from litellm.proxy.auth.auth_checks import model_has_no_cost_mapping + + router = _router_with_priced_and_unpriced_models() + + assert model_has_no_cost_mapping(model=None, llm_router=router) is False + assert model_has_no_cost_mapping(model="unpriced-group", llm_router=None) is False + + +async def _run_common_checks( + model: Optional[str], llm_router: Optional["Router"], route: str = "/chat/completions" +) -> bool: + from fastapi import Request + + from litellm.proxy.auth.auth_checks import common_checks + + return await common_checks( + request_body={"model": model, "messages": [{"role": "user", "content": "hi"}]}, + team_object=None, + user_object=None, + end_user_object=None, + global_proxy_spend=None, + general_settings={}, + route=route, + llm_router=llm_router, + proxy_logging_obj=MagicMock(), + valid_token=UserAPIKeyAuth(token="test-token"), + request=MagicMock(spec=Request), + ) + + +@pytest.mark.asyncio +async def test_common_checks_blocks_unpriced_model_when_enabled(monkeypatch): + monkeypatch.setattr(litellm, "block_requests_for_models_without_pricing", True) + router = _router_with_priced_and_unpriced_models() + + with pytest.raises(ProxyException) as exc_info: + await _run_common_checks(model="unpriced-group", llm_router=router) + + assert exc_info.value.code == "403" + assert exc_info.value.type == ProxyErrorTypes.model_cost_map_missing + assert exc_info.value.param == "model" + assert "unpriced-group" in exc_info.value.message + assert "pricing" in exc_info.value.message.lower() + + +@pytest.mark.asyncio +async def test_common_checks_allows_unpriced_model_when_disabled(monkeypatch): + monkeypatch.setattr(litellm, "block_requests_for_models_without_pricing", False) + router = _router_with_priced_and_unpriced_models() + + result = await _run_common_checks(model="unpriced-group", llm_router=router) + + assert result is True + + +@pytest.mark.asyncio +async def test_common_checks_allows_priced_model_when_enabled(monkeypatch): + monkeypatch.setattr(litellm, "block_requests_for_models_without_pricing", True) + router = _router_with_priced_and_unpriced_models() + + result = await _run_common_checks(model="priced-group", llm_router=router) + + assert result is True + + +@pytest.mark.asyncio +async def test_common_checks_ignores_non_llm_route_when_enabled(monkeypatch): + monkeypatch.setattr(litellm, "block_requests_for_models_without_pricing", True) + router = _router_with_priced_and_unpriced_models() + + result = await _run_common_checks( + model="unpriced-group", llm_router=router, route="/model/new" + ) + + assert result is True + + +@pytest.mark.asyncio +async def test_common_checks_blocks_alias_resolving_to_unpriced_model(monkeypatch): + from litellm.router import Router + + monkeypatch.setattr(litellm, "block_requests_for_models_without_pricing", True) + router = Router( + model_list=[ + { + "model_name": "billed-underlying-group", + "litellm_params": {"model": UNPRICED_UNDERLYING_MODEL, "api_key": "sk-test"}, + } + ], + model_group_alias={"public-alias": "billed-underlying-group"}, + ) + + with pytest.raises(ProxyException) as exc_info: + await _run_common_checks(model="public-alias", llm_router=router) + + assert exc_info.value.code == "403" + assert exc_info.value.type == ProxyErrorTypes.model_cost_map_missing + assert "public-alias" in exc_info.value.message diff --git a/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py b/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py index bc463d5e75d..4fb90e9fb2d 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py +++ b/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py @@ -500,3 +500,59 @@ class TestResolveModelForCostLookup: assert resolved_model == "openai/gpt-4" assert provider is None + + +class TestBlockRequestsForModelsWithoutPricing: + """Test suite for the block_requests_for_models_without_pricing toggle endpoints""" + + @pytest.mark.asyncio + async def test_get_reflects_in_memory_flag(self): + with patch.object(litellm, "block_requests_for_models_without_pricing", True): + response = client.get( + "/config/block_requests_for_models_without_pricing", + headers={"Authorization": "Bearer sk-1234"}, + ) + + assert response.status_code == 200 + assert response.json() == {"enabled": True} + + @pytest.mark.asyncio + async def test_patch_persists_and_updates_flag(self): + mock_proxy_config = AsyncMock() + mock_proxy_config.get_config = AsyncMock(return_value={"litellm_settings": {}}) + mock_proxy_config.save_config = AsyncMock() + + with ( + patch("litellm.proxy.proxy_server.prisma_client", MagicMock()), + patch("litellm.proxy.proxy_server.proxy_config", mock_proxy_config), + patch("litellm.proxy.proxy_server.store_model_in_db", True), + patch.object(litellm, "block_requests_for_models_without_pricing", False), + ): + response = client.patch( + "/config/block_requests_for_models_without_pricing", + headers={"Authorization": "Bearer sk-1234"}, + json={"enabled": True}, + ) + + assert response.status_code == 200 + assert response.json() == {"enabled": True} + assert litellm.block_requests_for_models_without_pricing is True + + saved_config = mock_proxy_config.save_config.call_args.kwargs["new_config"] + assert saved_config["litellm_settings"]["block_requests_for_models_without_pricing"] is True + + @pytest.mark.asyncio + async def test_patch_requires_store_model_in_db(self): + with ( + patch("litellm.proxy.proxy_server.prisma_client", MagicMock()), + patch("litellm.proxy.proxy_server.proxy_config", AsyncMock()), + patch("litellm.proxy.proxy_server.store_model_in_db", False), + ): + response = client.patch( + "/config/block_requests_for_models_without_pricing", + headers={"Authorization": "Bearer sk-1234"}, + json={"enabled": True}, + ) + + assert response.status_code == 500 + assert "error" in response.json()["detail"] diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/cost_tracking_settings.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/cost_tracking_settings.tsx index b32e7afd756..f6a3d487ade 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/cost_tracking_settings.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/cost_tracking_settings.tsx @@ -12,7 +12,7 @@ import { TabPanels, TabPanel, } from "@tremor/react"; -import { Modal, Form } from "antd"; +import { Modal, Form, Switch } from "antd"; import { CostTrackingSettingsProps } from "./types"; import ProviderDiscountTable from "./provider_discount_table"; import AddProviderForm from "./add_provider_form"; @@ -24,6 +24,7 @@ import { DocsMenu } from "@/components/HelpLink"; import HowItWorks from "./how_it_works"; import { useDiscountConfig } from "./use_discount_config"; import { useMarginConfig } from "./use_margin_config"; +import { useBlockUnpricedConfig } from "./use_block_unpriced_config"; import { fetchAvailableModels, ModelGroup } from "@/components/llm_calls/fetch_models"; const DOCS_LINKS = [ @@ -65,9 +66,16 @@ const CostTrackingSettings: React.FC = ({ userID, use handleMarginChange, } = useMarginConfig({ accessToken }); + const { + blockUnpriced, + isUpdating: isUpdatingBlockUnpriced, + fetchBlockUnpriced, + setBlockUnpriced, + } = useBlockUnpricedConfig({ accessToken }); + useEffect(() => { if (accessToken) { - Promise.all([fetchDiscountConfig(), fetchMarginConfig()]).finally(() => { + Promise.all([fetchDiscountConfig(), fetchMarginConfig(), fetchBlockUnpriced()]).finally(() => { setIsFetching(false); }); @@ -82,7 +90,7 @@ const CostTrackingSettings: React.FC = ({ userID, use }; loadModels(); } - }, [accessToken, fetchDiscountConfig, fetchMarginConfig]); + }, [accessToken, fetchDiscountConfig, fetchMarginConfig, fetchBlockUnpriced]); const handleAddProvider = async () => { const success = await addProvider(selectedProvider, newDiscount); @@ -293,7 +301,38 @@ const CostTrackingSettings: React.FC = ({ userID, use )} - {/* Accordion 3: Pricing Calculator - Available to all roles */} + {isProxyAdmin && ( + + +
+ Block Unpriced Models + + Reject requests for models that have no pricing in the cost map instead of logging them as $0 spend + +
+ + +
+
+
+ Block requests for models without pricing + + When enabled, a request whose resolved model has no cost mapping is rejected with a 403 so an + admin can add pricing for it. Off by default + +
+ setBlockUnpriced(checked)} + /> +
+
+ + + )} + + {/* Accordion 4: Pricing Calculator - Available to all roles */}
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/use_block_unpriced_config.ts b/ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/use_block_unpriced_config.ts new file mode 100644 index 00000000000..f4a110c7878 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/use_block_unpriced_config.ts @@ -0,0 +1,63 @@ +import { useState, useCallback } from "react"; +import { apiClient } from "@/components/networking"; +import NotificationsManager from "@/components/molecules/notifications_manager"; + +export interface UseBlockUnpricedConfigProps { + accessToken: string | null; +} + +export interface UseBlockUnpricedConfigReturn { + blockUnpriced: boolean; + isUpdating: boolean; + fetchBlockUnpriced: () => Promise; + setBlockUnpriced: (enabled: boolean) => Promise; +} + +interface BlockUnpricedResponse { + enabled: boolean; +} + +const ENDPOINT = "/config/block_requests_for_models_without_pricing"; + +export function useBlockUnpricedConfig({ accessToken }: UseBlockUnpricedConfigProps): UseBlockUnpricedConfigReturn { + const [blockUnpriced, setBlockUnpricedState] = useState(false); + const [isUpdating, setIsUpdating] = useState(false); + + const fetchBlockUnpriced = useCallback(async () => { + if (!accessToken) return; + try { + const data = await apiClient.get(ENDPOINT, { accessToken }); + setBlockUnpricedState(Boolean(data?.enabled)); + } catch (error) { + console.error("Error fetching block-unpriced-models setting:", error); + } + }, [accessToken]); + + const setBlockUnpriced = useCallback( + async (enabled: boolean) => { + if (!accessToken) return; + setIsUpdating(true); + try { + const data = await apiClient.patch(ENDPOINT, { accessToken, body: { enabled } }); + setBlockUnpricedState(Boolean(data?.enabled)); + NotificationsManager.success( + enabled + ? "Requests for models without pricing will now be blocked" + : "Requests for models without pricing are now allowed", + ); + } catch (error) { + console.error("Error updating block-unpriced-models setting:", error); + } finally { + setIsUpdating(false); + } + }, + [accessToken], + ); + + return { + blockUnpriced, + isUpdating, + fetchBlockUnpriced, + setBlockUnpriced, + }; +} diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index ed975c6be0a..5abe3b4d587 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -1844,6 +1844,24 @@ export interface paths { patch?: never; trace?: never; }; + "/config/block_requests_for_models_without_pricing": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + /** Get Block Requests For Models Without Pricing */ + get: operations["get_block_requests_for_models_without_pricing_config_block_requests_for_models_without_pricing_get"]; + put?: never; + post?: never; + delete?: never; + options?: never; + head?: never; + /** Update Block Requests For Models Without Pricing */ + patch: operations["update_block_requests_for_models_without_pricing_config_block_requests_for_models_without_pricing_patch"]; + trace?: never; + }; "/config/callback/delete": { parameters: { query?: never; @@ -21247,6 +21265,16 @@ export interface components { /** Team Id */ team_id: string; }; + /** BlockUnpricedModelsRequest */ + BlockUnpricedModelsRequest: { + /** Enabled */ + enabled: boolean; + }; + /** BlockUnpricedModelsResponse */ + BlockUnpricedModelsResponse: { + /** Enabled */ + enabled: boolean; + }; /** BlockUsers */ BlockUsers: { /** User Ids */ @@ -37097,6 +37125,59 @@ export interface operations { }; }; }; + get_block_requests_for_models_without_pricing_config_block_requests_for_models_without_pricing_get: { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["BlockUnpricedModelsResponse"]; + }; + }; + }; + }; + update_block_requests_for_models_without_pricing_config_block_requests_for_models_without_pricing_patch: { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + requestBody: { + content: { + "application/json": components["schemas"]["BlockUnpricedModelsRequest"]; + }; + }; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["BlockUnpricedModelsResponse"]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; delete_callback_config_callback_delete_post: { parameters: { query?: never; From 84c41dcdc3e55c180255433538f6df8969a56297 Mon Sep 17 00:00:00 2001 From: mateo Date: Fri, 31 Jul 2026 03:02:10 +0000 Subject: [PATCH 002/220] fix(proxy): treat non-token pricing as priced and propagate the unpriced-model toggle across workers Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/constants.py | 1 + litellm/proxy/auth/auth_checks.py | 58 +++++++++++++++++-- .../proxy/auth/test_auth_checks.py | 45 ++++++++++++++ .../test_cost_tracking_settings.py | 14 +++++ 4 files changed, 112 insertions(+), 6 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index 1014b472c61..b0ff992931f 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1525,6 +1525,7 @@ LITELLM_SETTINGS_SAFE_DB_OVERRIDES = [ "public_model_groups_links", "cost_discount_config", "cost_margin_config", + "block_requests_for_models_without_pricing", "budget_exceeded_throttle_percentage", # Every field editable from the Admin UI (proxy_server._GENERAL_SETTINGS_UI_LITELLM_FIELDS) # must be listed here so a DB write from one worker overrides the live litellm attribute on diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py index 11a45e78410..136b8d19c04 100644 --- a/litellm/proxy/auth/auth_checks.py +++ b/litellm/proxy/auth/auth_checks.py @@ -13,7 +13,18 @@ import asyncio import math import re import time -from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional, Type, Union, cast +from typing import ( + TYPE_CHECKING, + Any, + Dict, + List, + Literal, + Mapping, + Optional, + Type, + Union, + cast, +) from fastapi import HTTPException, Request, status from pydantic import BaseModel @@ -274,17 +285,52 @@ def _is_cost_explicitly_configured(model: str, llm_router: "Router") -> bool: return False +def _has_positive_cost(value: object) -> bool: + if isinstance(value, bool): + return False + if isinstance(value, (int, float)): + return value > 0 + if isinstance(value, dict): + return any(_has_positive_cost(nested) for nested in value.values()) + return False + + +def _entry_has_priced_metric(entry: Mapping[str, object]) -> bool: + return any("cost_per" in key and _has_positive_cost(value) for key, value in entry.items()) + + +def _model_group_has_pricing(model: str, llm_router: "Router") -> bool: + """ + Check every deployment behind a model group for a positive price on any billed + metric (tokens, characters, seconds, pages, images, queries, ...), so models that + are billed by a non-token metric are not treated as unpriced. + """ + for deployment in llm_router.get_model_list(model_name=model) or []: + litellm_params = deployment.get("litellm_params") or {} + if _entry_has_priced_metric(litellm_params): + return True + + model_id = (deployment.get("model_info") or {}).get("id") + if model_id is None: + continue + + model_info = llm_router.get_deployment_model_info( + model_id=model_id, model_name=litellm_params.get("model") or "" + ) + if model_info is not None and _entry_has_priced_metric(model_info): + return True + + return False + + def model_has_no_cost_mapping(model: Optional[str], llm_router: Optional[Router]) -> bool: if not model or llm_router is None: return False - model_group_info = llm_router.get_model_group_info(model_group=model) - if model_group_info is None: + if llm_router.get_model_group_info(model_group=model) is None: return False - input_cost = model_group_info.input_cost_per_token or 0 - output_cost = model_group_info.output_cost_per_token or 0 - if input_cost > 0 or output_cost > 0: + if _model_group_has_pricing(model=model, llm_router=llm_router): return False return not _is_cost_explicitly_configured(model, llm_router) diff --git a/tests/test_litellm/proxy/auth/test_auth_checks.py b/tests/test_litellm/proxy/auth/test_auth_checks.py index aec0ddc55f8..a078e041aa7 100644 --- a/tests/test_litellm/proxy/auth/test_auth_checks.py +++ b/tests/test_litellm/proxy/auth/test_auth_checks.py @@ -5165,6 +5165,7 @@ async def test_get_project_object_db_fetch_returns_cached_obj(): assert isinstance(result, LiteLLM_ProjectTableCachedObj) assert result.project_id == "p-1" + assert result.project_alias == "proj" UNPRICED_UNDERLYING_MODEL = "openai/unpriced-model-lit4984-xyz" @@ -5212,6 +5213,50 @@ def test_model_has_no_cost_mapping_no_model_or_router_is_false(): assert model_has_no_cost_mapping(model="unpriced-group", llm_router=None) is False +@pytest.mark.parametrize( + "underlying_model", + [ + "azure/speech/azure-tts", + "mistral/mistral-ocr-latest", + "vertex_ai/imagen-3.0-generate-001", + ], +) +def test_model_has_no_cost_mapping_non_token_priced_model_is_false(underlying_model): + from litellm.proxy.auth.auth_checks import model_has_no_cost_mapping + from litellm.router import Router + + router = Router( + model_list=[ + { + "model_name": "non-token-priced-group", + "litellm_params": {"model": underlying_model, "api_key": "sk-test"}, + } + ] + ) + + assert model_has_no_cost_mapping(model="non-token-priced-group", llm_router=router) is False + + +def test_model_has_no_cost_mapping_non_token_price_from_litellm_params_is_false(): + from litellm.proxy.auth.auth_checks import model_has_no_cost_mapping + from litellm.router import Router + + router = Router( + model_list=[ + { + "model_name": "custom-tts", + "litellm_params": { + "model": UNPRICED_UNDERLYING_MODEL, + "api_key": "sk-test", + "input_cost_per_second": 0.0001, + }, + } + ] + ) + + assert model_has_no_cost_mapping(model="custom-tts", llm_router=router) is False + + async def _run_common_checks( model: Optional[str], llm_router: Optional["Router"], route: str = "/chat/completions" ) -> bool: diff --git a/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py b/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py index 4fb90e9fb2d..1124a4a31d4 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py +++ b/tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py @@ -541,6 +541,20 @@ class TestBlockRequestsForModelsWithoutPricing: saved_config = mock_proxy_config.save_config.call_args.kwargs["new_config"] assert saved_config["litellm_settings"]["block_requests_for_models_without_pricing"] is True + def test_peer_workers_pick_up_persisted_flag_on_config_reload(self): + """A PATCH only mutates the flag on the worker that served it; peer workers must pick the + persisted value up when they reload litellm_settings from the DB.""" + from litellm.proxy.proxy_server import ProxyConfig + + with patch.object(litellm, "block_requests_for_models_without_pricing", False): + ProxyConfig()._update_config_fields( + current_config={}, + param_name="litellm_settings", + db_param_value={"block_requests_for_models_without_pricing": True}, + ) + + assert litellm.block_requests_for_models_without_pricing is True + @pytest.mark.asyncio async def test_patch_requires_store_model_in_db(self): with ( From 074b37b4f987f239725a165565be16ff5e1f9686 Mon Sep 17 00:00:00 2001 From: mateo Date: Fri, 31 Jul 2026 03:08:29 +0000 Subject: [PATCH 003/220] refactor(proxy): flatten the pricing-metric check to avoid recursion Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/auth/auth_checks.py | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-) diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py index 136b8d19c04..3639ef245cf 100644 --- a/litellm/proxy/auth/auth_checks.py +++ b/litellm/proxy/auth/auth_checks.py @@ -285,18 +285,19 @@ def _is_cost_explicitly_configured(model: str, llm_router: "Router") -> bool: return False -def _has_positive_cost(value: object) -> bool: - if isinstance(value, bool): - return False - if isinstance(value, (int, float)): - return value > 0 - if isinstance(value, dict): - return any(_has_positive_cost(nested) for nested in value.values()) - return False +def _is_positive_cost(value: object) -> bool: + return isinstance(value, (int, float)) and not isinstance(value, bool) and value > 0 def _entry_has_priced_metric(entry: Mapping[str, object]) -> bool: - return any("cost_per" in key and _has_positive_cost(value) for key, value in entry.items()) + for key, value in entry.items(): + if "cost_per" not in key: + continue + if _is_positive_cost(value): + return True + if isinstance(value, dict) and any(_is_positive_cost(nested) for nested in value.values()): + return True + return False def _model_group_has_pricing(model: str, llm_router: "Router") -> bool: From 9c922f4aa4f6ab3a261326332d01b74e3717e63c Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Sat, 15 Aug 2026 02:49:10 +0000 Subject: [PATCH 004/220] fix(bedrock): forward provider response headers on chat completions Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../litellm_core_utils/streaming_handler.py | 2 +- litellm/llms/bedrock/chat/converse_handler.py | 18 ++++-- litellm/llms/bedrock/chat/invoke_handler.py | 8 +-- .../base_invoke_transformation.py | 64 +++++++++---------- litellm/llms/custom_httpx/llm_http_handler.py | 2 + litellm/types/utils.py | 7 +- .../llm_translation/test_bedrock_moonshot.py | 19 ++---- .../llms/bedrock/chat/test_invoke_handler.py | 25 ++++++++ .../llms/chat/test_converse_handler.py | 55 ++++++++++++++++ 9 files changed, 143 insertions(+), 57 deletions(-) diff --git a/litellm/litellm_core_utils/streaming_handler.py b/litellm/litellm_core_utils/streaming_handler.py index 99b1c1a2ab7..fe107d349c6 100644 --- a/litellm/litellm_core_utils/streaming_handler.py +++ b/litellm/litellm_core_utils/streaming_handler.py @@ -164,7 +164,7 @@ class CustomStreamWrapper: custom_llm_provider: str | None = None, stream_options=None, make_call: Callable | None = None, - _response_headers: dict | None = None, + _response_headers: dict | httpx.Headers | None = None, ): self.model = model self.make_call = make_call diff --git a/litellm/llms/bedrock/chat/converse_handler.py b/litellm/llms/bedrock/chat/converse_handler.py index 25e544f4521..c9ea06d37dc 100644 --- a/litellm/llms/bedrock/chat/converse_handler.py +++ b/litellm/llms/bedrock/chat/converse_handler.py @@ -33,7 +33,7 @@ def make_sync_call( json_mode: bool | None = False, fake_stream: bool = False, stream_chunk_size: int | None = None, -): +) -> tuple[Any, httpx.Headers]: if client is None: client = _get_httpx_client() # Create a new client if none provided @@ -74,7 +74,7 @@ def make_sync_call( additional_args={"complete_input_dict": data}, ) - return completion_stream + return completion_stream, response.headers class BedrockConverseLLM(BaseAWSLLM): @@ -132,7 +132,7 @@ class BedrockConverseLLM(BaseAWSLLM): }, ) - completion_stream: Final = await make_call( + completion_stream, response_headers = await make_call( client=client, api_base=api_base, headers=dict(prepped.headers), @@ -149,6 +149,7 @@ class BedrockConverseLLM(BaseAWSLLM): model=model, custom_llm_provider="bedrock", logging_obj=logging_obj, + _response_headers=response_headers, ) return streaming_response @@ -225,7 +226,7 @@ class BedrockConverseLLM(BaseAWSLLM): except httpx.TimeoutException: raise BedrockError(status_code=408, message="Timeout error occurred.") - return litellm.AmazonConverseConfig()._transform_response( + transformed_response: Final = litellm.AmazonConverseConfig()._transform_response( model=model, response=response, model_response=model_response, @@ -237,6 +238,8 @@ class BedrockConverseLLM(BaseAWSLLM): optional_params=optional_params, encoding=encoding, ) + transformed_response.set_provider_response_headers(response.headers) + return transformed_response def completion( self, @@ -440,7 +443,7 @@ class BedrockConverseLLM(BaseAWSLLM): client = client if stream is not None and stream is True: - completion_stream: Final = make_sync_call( + completion_stream, response_headers = make_sync_call( client=(client if client is not None and isinstance(client, HTTPHandler) else None), api_base=proxy_endpoint_url, headers=prepped.headers, @@ -457,6 +460,7 @@ class BedrockConverseLLM(BaseAWSLLM): model=model, custom_llm_provider="bedrock", logging_obj=logging_obj, + _response_headers=response_headers, ) return streaming_response @@ -477,7 +481,7 @@ class BedrockConverseLLM(BaseAWSLLM): except httpx.TimeoutException: raise BedrockError(status_code=408, message="Timeout error occurred.") - return litellm.AmazonConverseConfig()._transform_response( + sync_transformed_response: Final = litellm.AmazonConverseConfig()._transform_response( model=model, response=response, model_response=model_response, @@ -489,3 +493,5 @@ class BedrockConverseLLM(BaseAWSLLM): optional_params=optional_params, encoding=encoding, ) + sync_transformed_response.set_provider_response_headers(response.headers) + return sync_transformed_response diff --git a/litellm/llms/bedrock/chat/invoke_handler.py b/litellm/llms/bedrock/chat/invoke_handler.py index 8d2b3dae71b..21892af3aae 100644 --- a/litellm/llms/bedrock/chat/invoke_handler.py +++ b/litellm/llms/bedrock/chat/invoke_handler.py @@ -163,7 +163,7 @@ async def make_call( json_mode: bool | None = False, bedrock_invoke_provider: litellm.BEDROCK_INVOKE_PROVIDERS_LITERAL | None = None, stream_chunk_size: int | None = None, -): +) -> tuple[Any, httpx.Headers]: try: if client is None: client = get_async_httpx_client( @@ -225,7 +225,7 @@ async def make_call( additional_args={"complete_input_dict": data}, ) - return completion_stream + return completion_stream, response.headers except httpx.HTTPStatusError as err: error_code: Final = err.response.status_code raise BedrockError(status_code=error_code, message=err.response.text) @@ -248,7 +248,7 @@ def make_sync_call( json_mode: bool | None = False, bedrock_invoke_provider: litellm.BEDROCK_INVOKE_PROVIDERS_LITERAL | None = None, stream_chunk_size: int | None = None, -): +) -> tuple[Any, httpx.Headers]: try: if client is None: client = _get_httpx_client( @@ -309,7 +309,7 @@ def make_sync_call( additional_args={"complete_input_dict": data}, ) - return completion_stream + return completion_stream, response.headers except httpx.HTTPStatusError as err: error_code: Final = err.response.status_code raise BedrockError(status_code=error_code, message=err.response.text) diff --git a/litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py b/litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py index 430d0a92b51..0b5855ccbd4 100644 --- a/litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py +++ b/litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py @@ -1,7 +1,6 @@ import copy import json import time -from functools import partial from typing import TYPE_CHECKING, Any, Final, cast, get_args import httpx @@ -444,24 +443,24 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM): json_mode: bool | None = None, signed_json_body: bytes | None = None, ) -> CustomStreamWrapper: + completion_stream, response_headers = await make_call( + client=client, + api_base=api_base, + headers=headers, + data=json.dumps(data), + model=model, + messages=messages, + logging_obj=logging_obj, + fake_stream=True if "ai21" in api_base else False, + bedrock_invoke_provider=self.get_bedrock_invoke_provider(model), + json_mode=json_mode, + ) streaming_response: Final = CustomStreamWrapper( - completion_stream=None, - make_call=partial( - make_call, - client=client, - api_base=api_base, - headers=headers, - data=json.dumps(data), - model=model, - messages=messages, - logging_obj=logging_obj, - fake_stream=True if "ai21" in api_base else False, - bedrock_invoke_provider=self.get_bedrock_invoke_provider(model), - json_mode=json_mode, - ), + completion_stream=completion_stream, model=model, custom_llm_provider="bedrock", logging_obj=logging_obj, + _response_headers=response_headers, ) return streaming_response @@ -479,27 +478,28 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM): json_mode: bool | None = None, signed_json_body: bytes | None = None, ) -> CustomStreamWrapper: - if client is None or isinstance(client, AsyncHTTPHandler): - client = _get_httpx_client(params={}) + sync_client: Final = ( + _get_httpx_client(params={}) if client is None or isinstance(client, AsyncHTTPHandler) else client + ) + completion_stream, response_headers = make_sync_call( + client=sync_client, + api_base=api_base, + headers=headers, + data=json.dumps(data), + signed_json_body=signed_json_body, + model=model, + messages=messages, + logging_obj=logging_obj, + fake_stream=True if "ai21" in api_base else False, + bedrock_invoke_provider=self.get_bedrock_invoke_provider(model), + json_mode=json_mode, + ) streaming_response: Final = CustomStreamWrapper( - completion_stream=None, - make_call=partial( - make_sync_call, - client=client, - api_base=api_base, - headers=headers, - data=json.dumps(data), - signed_json_body=signed_json_body, - model=model, - messages=messages, - logging_obj=logging_obj, - fake_stream=True if "ai21" in api_base else False, - bedrock_invoke_provider=self.get_bedrock_invoke_provider(model), - json_mode=json_mode, - ), + completion_stream=completion_stream, model=model, custom_llm_provider="bedrock", logging_obj=logging_obj, + _response_headers=response_headers, ) return streaming_response diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index 721b9545ac1..36bcf58a243 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -608,6 +608,7 @@ class BaseLLMHTTPHandler: model=model, custom_llm_provider=custom_llm_provider, logging_obj=logging_obj, + _response_headers=headers, ) if client is None or not isinstance(client, HTTPHandler): @@ -771,6 +772,7 @@ class BaseLLMHTTPHandler: model=model, custom_llm_provider=custom_llm_provider, logging_obj=logging_obj, + _response_headers=_response_headers, ) return streamwrapper diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 220826ccbca..f49b5735f04 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -11,6 +11,7 @@ from typing import ( get_args, ) +import httpx from openai._models import BaseModel as OpenAIObject from openai.types.audio.transcription_create_params import ( FileTypes as FileTypes, @@ -49,7 +50,7 @@ from litellm.types.llms.base import ( ) from litellm.types.mcp import MCPServerCostInfo -from ..litellm_core_utils.core_helpers import map_finish_reason +from ..litellm_core_utils.core_helpers import map_finish_reason, process_response_headers from .agents import LiteLLMSendMessageResponse from .guardrails import GuardrailEventHooks from .llms.anthropic_messages.anthropic_response import AnthropicMessagesResponse @@ -1896,6 +1897,10 @@ class ModelResponseBase(OpenAIObject): _response_headers: dict | None = None + def set_provider_response_headers(self, headers: httpx.Headers) -> None: + """Surface a provider's raw response headers to the caller as `llm_provider-*` headers.""" + self._hidden_params["additional_headers"] = process_response_headers(headers) + def model_dump(self, **kwargs): """Default to exclude_unset to avoid Pydantic serializer warnings for OpenAIObject-derived types.""" if "exclude_unset" not in kwargs and "exclude_none" not in kwargs: diff --git a/tests/llm_translation/test_bedrock_moonshot.py b/tests/llm_translation/test_bedrock_moonshot.py index a9f4a86b3b6..c777305d562 100644 --- a/tests/llm_translation/test_bedrock_moonshot.py +++ b/tests/llm_translation/test_bedrock_moonshot.py @@ -209,14 +209,12 @@ class TestBedrockMoonshotInvoke(BaseLLMChatTest): not exercised here — moonshot streaming delegates to the OpenAI parser and is covered by the OpenAI test suite. - Note: bedrock invoke streaming cannot be intercepted by patching - the caller-supplied client, because ``CustomStreamWrapper.fetch_sync_stream`` - at streaming_handler.py invokes the stored ``make_call`` partial with - ``client=litellm.module_level_client``, which overrides any client the - caller passed. Patch ``make_sync_call`` at its import site in - ``base_invoke_transformation`` so we observe the exact kwargs the - partial was built with at stream-wrapper construction time. + Patch ``make_sync_call`` at its import site in + ``base_invoke_transformation`` so we observe the exact kwargs it is + called with at stream-wrapper construction time. """ + import httpx + from litellm.utils import CustomStreamWrapper captured: dict = {} @@ -225,7 +223,7 @@ class TestBedrockMoonshotInvoke(BaseLLMChatTest): captured.update(kwargs) # Return an empty iterator so the stream wrapper's iteration # doesn't try to parse real bytes. - return iter([]) + return iter([]), httpx.Headers() with patch( "litellm.llms.bedrock.chat.invoke_transformations." @@ -246,11 +244,6 @@ class TestBedrockMoonshotInvoke(BaseLLMChatTest): aws_region_name="us-west-2", ) assert isinstance(response, CustomStreamWrapper) - # Trigger fetch_sync_stream → make_call(...) → fake_make_sync_call. - try: - next(iter(response)) - except StopIteration: - pass assert captured, "make_sync_call was never invoked" assert captured["api_base"].endswith("/invoke-with-response-stream") diff --git a/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py b/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py index ee50b9db015..f211cdac475 100644 --- a/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py +++ b/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py @@ -2,17 +2,20 @@ import os import sys from unittest.mock import AsyncMock, MagicMock +import httpx import pytest sys.path.insert( 0, os.path.abspath("../../../../..") ) # Adds the parent directory to the system path +import litellm from litellm.llms.bedrock.chat.invoke_handler import ( AWSEventStreamDecoder, make_call, make_sync_call, ) +from litellm.llms.custom_httpx.http_handler import HTTPHandler def test_transform_thinking_blocks_with_redacted_content(): @@ -293,3 +296,25 @@ def test_make_sync_call_honors_explicit_stream_chunk_size(): response.iter_bytes.assert_called_once_with(chunk_size=2048) + +def test_invoke_streaming_forwards_bedrock_response_headers(): + """Streaming callers need `x-amzn-requestid` to correlate a LiteLLM request with AWS support.""" + response = MagicMock() + response.status_code = 200 + response.iter_bytes = MagicMock(return_value=iter([])) + response.headers = httpx.Headers({"x-amzn-requestid": "req-789"}) + client = HTTPHandler() + client.post = MagicMock(return_value=response) + + stream = litellm.completion( + model="bedrock/invoke/anthropic.claude-haiku-4-5-20251001-v1:0", + messages=[{"role": "user", "content": "hi"}], + stream=True, + client=client, + aws_access_key_id="fake", + aws_secret_access_key="fake", + aws_region_name="us-east-1", + ) + + assert stream._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-789" + diff --git a/tests/test_litellm/llms/chat/test_converse_handler.py b/tests/test_litellm/llms/chat/test_converse_handler.py index 2a3db5982ef..89ab29122d7 100644 --- a/tests/test_litellm/llms/chat/test_converse_handler.py +++ b/tests/test_litellm/llms/chat/test_converse_handler.py @@ -1,7 +1,9 @@ +import json import os import sys from unittest.mock import MagicMock +import httpx import pytest import litellm @@ -202,6 +204,59 @@ def test_make_sync_call_honors_explicit_stream_chunk_size(): response.iter_bytes.assert_called_once_with(chunk_size=2048) +def _converse_response_body() -> dict: + return { + "output": {"message": {"role": "assistant", "content": [{"text": "hi"}]}}, + "stopReason": "end_turn", + "usage": {"inputTokens": 1, "outputTokens": 1, "totalTokens": 2}, + } + + +def test_converse_completion_forwards_bedrock_response_headers(): + """Bedrock returns x-amzn-requestid on every converse call, which customers need to + correlate proxy requests with AWS support cases, so it must reach the caller as + llm_provider-x-amzn-requestid.""" + mock_response = MagicMock() + mock_response.status_code = 200 + mock_response.json = MagicMock(return_value=_converse_response_body()) + mock_response.text = json.dumps(_converse_response_body()) + mock_response.headers = httpx.Headers({"x-amzn-requestid": "req-123"}) + client = HTTPHandler() + client.post = MagicMock(return_value=mock_response) + + response = litellm.completion( + model="bedrock/converse/anthropic.claude-haiku-4-5-20251001-v1:0", + messages=[{"role": "user", "content": "hi"}], + client=client, + aws_access_key_id="fake", + aws_secret_access_key="fake", + aws_region_name="us-east-1", + ) + + assert response._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-123" + + +def test_converse_streaming_forwards_bedrock_response_headers(): + mock_response = MagicMock() + mock_response.status_code = 200 + mock_response.iter_bytes = MagicMock(return_value=iter([])) + mock_response.headers = httpx.Headers({"x-amzn-requestid": "req-456"}) + client = HTTPHandler() + client.post = MagicMock(return_value=mock_response) + + response = litellm.completion( + model="bedrock/converse/anthropic.claude-haiku-4-5-20251001-v1:0", + messages=[{"role": "user", "content": "hi"}], + stream=True, + client=client, + aws_access_key_id="fake", + aws_secret_access_key="fake", + aws_region_name="us-east-1", + ) + + assert response._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-456" + + def test_completion_plumbs_stream_chunk_size_through_converse(): iter_bytes_spy = _stream_completion_with_spied_iter_bytes( model="bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0" From 726db1a4c118a2ab92214ba741efacdce7fe25b5 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Sat, 15 Aug 2026 03:07:14 +0000 Subject: [PATCH 005/220] test(bedrock): cover async header forwarding for converse and invoke Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../llm_translation/test_bedrock_moonshot.py | 3 +- .../llms/bedrock/chat/test_invoke_handler.py | 29 +++++++++- .../llms/chat/test_converse_handler.py | 55 +++++++++++++++++-- 3 files changed, 78 insertions(+), 9 deletions(-) diff --git a/tests/llm_translation/test_bedrock_moonshot.py b/tests/llm_translation/test_bedrock_moonshot.py index c777305d562..61364cf2caa 100644 --- a/tests/llm_translation/test_bedrock_moonshot.py +++ b/tests/llm_translation/test_bedrock_moonshot.py @@ -12,6 +12,7 @@ This test suite verifies: """ from base_llm_unit_tests import BaseLLMChatTest +import httpx import pytest import sys import os @@ -213,8 +214,6 @@ class TestBedrockMoonshotInvoke(BaseLLMChatTest): ``base_invoke_transformation`` so we observe the exact kwargs it is called with at stream-wrapper construction time. """ - import httpx - from litellm.utils import CustomStreamWrapper captured: dict = {} diff --git a/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py b/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py index f211cdac475..e8964910c69 100644 --- a/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py +++ b/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py @@ -15,7 +15,7 @@ from litellm.llms.bedrock.chat.invoke_handler import ( make_call, make_sync_call, ) -from litellm.llms.custom_httpx.http_handler import HTTPHandler +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler def test_transform_thinking_blocks_with_redacted_content(): @@ -298,7 +298,6 @@ def test_make_sync_call_honors_explicit_stream_chunk_size(): def test_invoke_streaming_forwards_bedrock_response_headers(): - """Streaming callers need `x-amzn-requestid` to correlate a LiteLLM request with AWS support.""" response = MagicMock() response.status_code = 200 response.iter_bytes = MagicMock(return_value=iter([])) @@ -318,3 +317,29 @@ def test_invoke_streaming_forwards_bedrock_response_headers(): assert stream._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-789" + +@pytest.mark.asyncio +async def test_async_invoke_streaming_forwards_bedrock_response_headers(): + async def _no_bytes(chunk_size=None): + return + yield b"" + + response = MagicMock() + response.status_code = 200 + response.aiter_bytes = _no_bytes + response.headers = httpx.Headers({"x-amzn-requestid": "req-987"}) + client = AsyncHTTPHandler() + client.post = AsyncMock(return_value=response) + + stream = await litellm.acompletion( + model="bedrock/invoke/anthropic.claude-haiku-4-5-20251001-v1:0", + messages=[{"role": "user", "content": "hi"}], + stream=True, + client=client, + aws_access_key_id="fake", + aws_secret_access_key="fake", + aws_region_name="us-east-1", + ) + + assert stream._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-987" + diff --git a/tests/test_litellm/llms/chat/test_converse_handler.py b/tests/test_litellm/llms/chat/test_converse_handler.py index 89ab29122d7..6f8a2788c38 100644 --- a/tests/test_litellm/llms/chat/test_converse_handler.py +++ b/tests/test_litellm/llms/chat/test_converse_handler.py @@ -1,7 +1,7 @@ import json import os import sys -from unittest.mock import MagicMock +from unittest.mock import AsyncMock, MagicMock import httpx import pytest @@ -10,7 +10,7 @@ import litellm from litellm.llms.bedrock.chat import BedrockConverseLLM from litellm.llms.bedrock.chat.converse_handler import make_sync_call from litellm.llms.bedrock.common_utils import _get_all_bedrock_regions -from litellm.llms.custom_httpx.http_handler import HTTPHandler +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler sys.path.insert( 0, os.path.abspath("../../../../..") @@ -213,9 +213,6 @@ def _converse_response_body() -> dict: def test_converse_completion_forwards_bedrock_response_headers(): - """Bedrock returns x-amzn-requestid on every converse call, which customers need to - correlate proxy requests with AWS support cases, so it must reach the caller as - llm_provider-x-amzn-requestid.""" mock_response = MagicMock() mock_response.status_code = 200 mock_response.json = MagicMock(return_value=_converse_response_body()) @@ -257,6 +254,54 @@ def test_converse_streaming_forwards_bedrock_response_headers(): assert response._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-456" +@pytest.mark.asyncio +async def test_async_converse_completion_forwards_bedrock_response_headers(): + mock_response = MagicMock() + mock_response.status_code = 200 + mock_response.json = MagicMock(return_value=_converse_response_body()) + mock_response.text = json.dumps(_converse_response_body()) + mock_response.headers = httpx.Headers({"x-amzn-requestid": "req-abc"}) + client = AsyncHTTPHandler() + client.post = AsyncMock(return_value=mock_response) + + response = await litellm.acompletion( + model="bedrock/converse/anthropic.claude-haiku-4-5-20251001-v1:0", + messages=[{"role": "user", "content": "hi"}], + client=client, + aws_access_key_id="fake", + aws_secret_access_key="fake", + aws_region_name="us-east-1", + ) + + assert response._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-abc" + + +@pytest.mark.asyncio +async def test_async_converse_streaming_forwards_bedrock_response_headers(): + async def _no_bytes(chunk_size=None): + return + yield b"" + + mock_response = MagicMock() + mock_response.status_code = 200 + mock_response.aiter_bytes = _no_bytes + mock_response.headers = httpx.Headers({"x-amzn-requestid": "req-def"}) + client = AsyncHTTPHandler() + client.post = AsyncMock(return_value=mock_response) + + response = await litellm.acompletion( + model="bedrock/converse/anthropic.claude-haiku-4-5-20251001-v1:0", + messages=[{"role": "user", "content": "hi"}], + stream=True, + client=client, + aws_access_key_id="fake", + aws_secret_access_key="fake", + aws_region_name="us-east-1", + ) + + assert response._hidden_params["additional_headers"]["llm_provider-x-amzn-requestid"] == "req-def" + + def test_completion_plumbs_stream_chunk_size_through_converse(): iter_bytes_spy = _stream_completion_with_spied_iter_bytes( model="bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0" From ffa37d05b7cf16c9874101f3c737da19b4154aca Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Sun, 16 Aug 2026 14:28:50 -0400 Subject: [PATCH 006/220] feat(mistral): add zai-glm-5-2 model pricing and metadata --- .../model_prices_and_context_window_backup.json | 14 ++++++++++++++ model_prices_and_context_window.json | 14 ++++++++++++++ 2 files changed, 28 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index e6c6cab0631..b73feae90d3 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -29045,6 +29045,20 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "mistral/zai-glm-5-2": { + "input_cost_per_token": 1.4e-06, + "litellm_provider": "mistral", + "max_input_tokens": 1000000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://docs.mistral.ai/models/zai-glm-5-2", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, "mistral/magistral-medium-2506": { "deprecation_date": "2025-11-30", "input_cost_per_token": 2e-06, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index e6c6cab0631..b73feae90d3 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -29045,6 +29045,20 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "mistral/zai-glm-5-2": { + "input_cost_per_token": 1.4e-06, + "litellm_provider": "mistral", + "max_input_tokens": 1000000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://docs.mistral.ai/models/zai-glm-5-2", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, "mistral/magistral-medium-2506": { "deprecation_date": "2025-11-30", "input_cost_per_token": 2e-06, From 539a61be080b9ed8100ef00fca77f5b6175d8853 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Sun, 16 Aug 2026 14:44:09 -0400 Subject: [PATCH 007/220] feat(perplexity): add Agent API third-party models (DeepSeek V4 Flash, GLM 5.2, Kimi K3, Kimi K2.7 Code) --- ...odel_prices_and_context_window_backup.json | 44 +++++++++++++++++++ model_prices_and_context_window.json | 44 +++++++++++++++++++ 2 files changed, 88 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index e6c6cab0631..b786a84ada7 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -34286,6 +34286,50 @@ "supports_reasoning": false, "supports_function_calling": true }, + "perplexity/perplexity/deepseek-v4-flash-0731": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 1.3e-07, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 2.6e-07, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": true, + "supports_function_calling": true + }, + "perplexity/perplexity/glm-5.2": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 4.4e-06, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": true, + "supports_function_calling": true + }, + "perplexity/perplexity/kimi-k3": { + "cache_read_input_token_cost": 3e-07, + "input_cost_per_token": 3e-06, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 1.5e-05, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": true, + "supports_function_calling": true + }, + "perplexity/perplexity/kimi-k2.7-code": { + "cache_read_input_token_cost": 1.9e-07, + "input_cost_per_token": 9.5e-07, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 4e-06, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": false, + "supports_function_calling": true + }, "perplexity/pplx-embed-v1-0.6b": { "input_cost_per_token": 4e-09, "litellm_provider": "perplexity", diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index e6c6cab0631..b786a84ada7 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -34286,6 +34286,50 @@ "supports_reasoning": false, "supports_function_calling": true }, + "perplexity/perplexity/deepseek-v4-flash-0731": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 1.3e-07, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 2.6e-07, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": true, + "supports_function_calling": true + }, + "perplexity/perplexity/glm-5.2": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 4.4e-06, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": true, + "supports_function_calling": true + }, + "perplexity/perplexity/kimi-k3": { + "cache_read_input_token_cost": 3e-07, + "input_cost_per_token": 3e-06, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 1.5e-05, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": true, + "supports_function_calling": true + }, + "perplexity/perplexity/kimi-k2.7-code": { + "cache_read_input_token_cost": 1.9e-07, + "input_cost_per_token": 9.5e-07, + "litellm_provider": "perplexity", + "mode": "responses", + "output_cost_per_token": 4e-06, + "source": "https://docs.perplexity.ai/docs/agent-api/models", + "supports_web_search": true, + "supports_reasoning": false, + "supports_function_calling": true + }, "perplexity/pplx-embed-v1-0.6b": { "input_cost_per_token": 4e-09, "litellm_provider": "perplexity", From 782746553c109bdec6aa5ecddf8e674f5167f575 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Sun, 16 Aug 2026 14:57:11 -0400 Subject: [PATCH 008/220] fix(perplexity): accept float usage.cost in cost_per_token, not just dict ResponseAPIUsage.parse_cost already flattens Perplexity's usage.cost.total_cost dict down to a float before it reaches the perplexity cost calculator, so the isinstance(cost_info, dict) check was always False on that path. Every Responses-mode Perplexity model was silently falling back to manual token-rate calculation and recording $0 spend whenever static per-token rates were missing. --- litellm/llms/perplexity/cost_calculator.py | 19 ++++++++------ .../test_perplexity_cost_calculator.py | 25 +++++++++++++++++++ 2 files changed, 37 insertions(+), 7 deletions(-) diff --git a/litellm/llms/perplexity/cost_calculator.py b/litellm/llms/perplexity/cost_calculator.py index 337fa8e630d..27835ecbfe8 100644 --- a/litellm/llms/perplexity/cost_calculator.py +++ b/litellm/llms/perplexity/cost_calculator.py @@ -21,14 +21,19 @@ def cost_per_token(model: str, usage: Usage) -> tuple[float, float]: Tuple[float, float] - prompt_cost_in_usd, completion_cost_in_usd """ ## USE PRE-CALCULATED COST FROM PERPLEXITY IF AVAILABLE - ## Perplexity returns accurate cost in usage.cost.total_cost including request fees + ## Perplexity returns accurate cost in usage.cost.total_cost including request fees. + ## By the time it reaches here, ResponseAPIUsage.parse_cost has already flattened + ## that dict down to a float, so both shapes must be accepted. cost_info: Final = getattr(usage, "cost", None) - if cost_info is not None and isinstance(cost_info, dict): - total_cost: Final = cost_info.get("total_cost") - if total_cost is not None: - # Return total cost as completion_cost (prompt_cost=0) since Perplexity - # doesn't break down by input/output in their cost object - return (0.0, float(total_cost)) + total_cost: float | None = None + if isinstance(cost_info, dict): + total_cost = cost_info.get("total_cost") + elif isinstance(cost_info, (int, float)) and not isinstance(cost_info, bool): + total_cost = float(cost_info) + if total_cost is not None: + # Return total cost as completion_cost (prompt_cost=0) since Perplexity + # doesn't break down by input/output in their cost object + return (0.0, float(total_cost)) ## FALLBACK: Calculate cost manually if Perplexity doesn't provide it ## GET MODEL INFO diff --git a/tests/test_litellm/llms/perplexity/test_perplexity_cost_calculator.py b/tests/test_litellm/llms/perplexity/test_perplexity_cost_calculator.py index 46c1e457d7c..71ccb494cc7 100644 --- a/tests/test_litellm/llms/perplexity/test_perplexity_cost_calculator.py +++ b/tests/test_litellm/llms/perplexity/test_perplexity_cost_calculator.py @@ -400,6 +400,31 @@ class TestPerplexityCostCalculator: assert completion_cost == 0.008 assert prompt_cost + completion_cost == 0.008 + def test_uses_perplexity_provided_cost_when_normalized_to_float(self): + """ + Regression: for Responses API / Agent API models, `ResponseAPIUsage.parse_cost` + (litellm/types/llms/openai.py) already flattens Perplexity's + `usage.cost.total_cost` dict down to a plain float before + `_transform_response_api_usage_to_chat_usage` (litellm/responses/utils.py) copies + it onto the chat `Usage` object. So `usage.cost` arrives here as a float, not a + dict, on that path. + + Pre-fix, the `isinstance(cost_info, dict)` check was always False for a float, + so the pre-calculated cost branch was dead code for every Responses-mode + Perplexity model and it silently fell back to manual token-rate calculation, + recording $0 for any model missing static per-token rates (e.g. + perplexity/openai/gpt-5.2 before rates existed). + """ + usage = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=150) + usage.cost = 0.008 + + prompt_cost, completion_cost = perplexity_cost_per_token( + model="sonar-pro", usage=usage + ) + + assert prompt_cost == 0.0 + assert completion_cost == 0.008 + def test_falls_back_to_manual_calculation_when_no_cost_provided(self): """ Test that manual cost calculation is used when Perplexity doesn't From 5d5dc4523fb950e131a235bea9a4f767ba7e0e17 Mon Sep 17 00:00:00 2001 From: mateo Date: Wed, 19 Aug 2026 03:56:32 +0000 Subject: [PATCH 009/220] fix(cost): price streamed Messages usage via calculate_usage and the logging obj Streamed `/v1/messages` `usage.cost` disagreed with the cost the logging callback recorded in three ways: `input_tokens` was read as the whole prompt total, but Anthropic reports it excluding cache tokens, so the non-cached input went unbilled on cache hits; the `cache_creation` 5m/1h split was dropped, billing 1h writes at the 5m rate; and costing by model name alone ignored the deployment's custom pricing, so a negotiated discount still streamed sticker price. Anthropic usage now goes through `AnthropicConfig.calculate_usage`, the same transformation the non-streaming path uses, and the chunk is priced through the call's logging object when there is one so it inherits `custom_pricing`, `custom_llm_provider`, `base_model` and `router_model_id`, falling back to `completion_cost` by model name. `calculate_usage` only reads its `usage_object`, so it now takes a `Mapping`. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/llms/anthropic/chat/transformation.py | 2 +- litellm/proxy/common_request_processing.py | 130 ++++++++++------ .../streaming_handler.py | 2 +- .../proxy/test_common_request_processing.py | 140 ++++++++++++++++++ 4 files changed, 226 insertions(+), 48 deletions(-) diff --git a/litellm/llms/anthropic/chat/transformation.py b/litellm/llms/anthropic/chat/transformation.py index ef4ad7011c5..b4f040b0a8c 100644 --- a/litellm/llms/anthropic/chat/transformation.py +++ b/litellm/llms/anthropic/chat/transformation.py @@ -2152,7 +2152,7 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig): def calculate_usage( self, - usage_object: dict, + usage_object: Mapping[str, Any], reasoning_content: str | None, completion_response: dict | None = None, speed: str | None = None, diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index a0b69ecb0bf..f5299b273b4 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -158,7 +158,7 @@ ProxyRouteType: TypeAlias = Literal[ "acancel_run", "adelete_run", ] -from litellm.types.utils import ServerToolUse +from litellm.llms.anthropic.chat.transformation import AnthropicConfig # Type alias for streaming chunk serializer (chunk after hooks + cost injection -> wire format) StreamChunkSerializer = Callable[[Any], str] @@ -3321,7 +3321,9 @@ class ProxyBaseLLMRequestProcessing: str_so_far += str(chunk.get("content", "")) model_name = request_data.get("model", "") - chunk = ProxyBaseLLMRequestProcessing._process_chunk_with_cost_injection(chunk, model_name) + chunk = ProxyBaseLLMRequestProcessing._process_chunk_with_cost_injection( + chunk, model_name, request_data.get("litellm_logging_obj") + ) # Set before the yield: an async generator suspends at the yield, # so a GeneratorExit on client disconnect is raised there and any @@ -3418,20 +3420,28 @@ class ProxyBaseLLMRequestProcessing: @overload @staticmethod - def _process_chunk_with_cost_injection(chunk: bytes, model_name: str) -> bytes: ... + def _process_chunk_with_cost_injection( + chunk: bytes, model_name: str, litellm_logging_obj: LiteLLMLoggingObj | None = None + ) -> bytes: ... @overload @staticmethod - def _process_chunk_with_cost_injection(chunk: object, model_name: str) -> object: ... + def _process_chunk_with_cost_injection( + chunk: object, model_name: str, litellm_logging_obj: LiteLLMLoggingObj | None = None + ) -> object: ... @staticmethod - def _process_chunk_with_cost_injection(chunk: object, model_name: str) -> object: + def _process_chunk_with_cost_injection( + chunk: object, model_name: str, litellm_logging_obj: LiteLLMLoggingObj | None = None + ) -> object: """ Process a streaming chunk and inject cost information if enabled. Args: chunk: The streaming chunk (dict, str, bytes, or bytearray) model_name: Model name for cost calculation + litellm_logging_obj: The call's logging object, used to price the chunk with + the same custom/deployment pricing as the logging callback Returns: The processed chunk with cost information injected if applicable @@ -3441,21 +3451,27 @@ class ProxyBaseLLMRequestProcessing: try: if isinstance(chunk, dict): - maybe_modified: Final = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(chunk, model_name) + maybe_modified: Final = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict( + chunk, model_name, litellm_logging_obj + ) if maybe_modified is not None: return maybe_modified elif isinstance(chunk, (bytes, bytearray)): try: s: Final = chunk.decode("utf-8") if s.endswith(("\n\n", "\r\n\r\n")): - maybe_mod = ProxyBaseLLMRequestProcessing._inject_cost_into_sse_frame_str(s, model_name) + maybe_mod = ProxyBaseLLMRequestProcessing._inject_cost_into_sse_frame_str( + s, model_name, litellm_logging_obj + ) if maybe_mod is not None: return maybe_mod.encode("utf-8") except Exception: pass elif isinstance(chunk, str): # Try to parse SSE frame and inject cost into the data line - maybe_mod = ProxyBaseLLMRequestProcessing._inject_cost_into_sse_frame_str(chunk, model_name) + maybe_mod = ProxyBaseLLMRequestProcessing._inject_cost_into_sse_frame_str( + chunk, model_name, litellm_logging_obj + ) if maybe_mod is not None: # Ensure trailing frame separator return maybe_mod if maybe_mod.endswith("\n\n") else (maybe_mod + "\n\n") @@ -3466,13 +3482,16 @@ class ProxyBaseLLMRequestProcessing: return chunk @staticmethod - def _inject_cost_into_sse_frame_str(frame_str: str, model_name: str) -> str | None: + def _inject_cost_into_sse_frame_str( + frame_str: str, model_name: str, litellm_logging_obj: LiteLLMLoggingObj | None = None + ) -> str | None: """ Inject cost information into an SSE frame string by modifying the JSON in the 'data:' line. Args: frame_str: SSE frame string that may contain multiple lines model_name: Model name for cost calculation + litellm_logging_obj: The call's logging object, forwarded for pricing Returns: Modified SSE frame string with cost injected, or None if no modification needed @@ -3486,7 +3505,9 @@ class ProxyBaseLLMRequestProcessing: json_part = stripped_ln.split("data:", 1)[1].strip() if json_part and json_part != "[DONE]": obj = json.loads(json_part) - maybe_modified = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(obj, model_name) + maybe_modified = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict( + obj, model_name, litellm_logging_obj + ) if maybe_modified is not None: lines[idx] = "data: " + safe_dumps(maybe_modified) + ("\r" if ln.endswith("\r") else "") return "\n".join(lines) @@ -3494,34 +3515,6 @@ class ProxyBaseLLMRequestProcessing: except Exception: return None - @staticmethod - def _anthropic_stream_usage_kwargs(usage: Mapping[str, Any]) -> Mapping[str, Any]: - prompt_tokens: Final = int(usage.get("input_tokens", 0) or 0) - completion_tokens: Final = int(usage.get("output_tokens", 0) or 0) - total_tokens: Final = int( - usage.get("total_tokens", prompt_tokens + completion_tokens) or (prompt_tokens + completion_tokens) - ) - web_search_requests: Final = usage.get("web_search_requests") - server_tool_use: Final = ( - ServerToolUse(web_search_requests=web_search_requests) if web_search_requests is not None else None - ) - return MappingProxyType( - { - key: value - for key, value in ( - ("prompt_tokens", prompt_tokens), - ("completion_tokens", completion_tokens), - ("total_tokens", total_tokens), - ("completion_tokens_details", usage.get("completion_tokens_details")), - ("prompt_tokens_details", usage.get("prompt_tokens_details")), - ("cache_creation_input_tokens", usage.get("cache_creation_input_tokens")), - ("cache_read_input_tokens", usage.get("cache_read_input_tokens")), - ("server_tool_use", server_tool_use), - ) - if value is not None - } - ) - @staticmethod def _openai_stream_usage_kwargs(usage: Mapping[str, Any]) -> Mapping[str, Any]: prompt_tokens: Final = int(usage.get("prompt_tokens", 0) or 0) @@ -3544,11 +3537,19 @@ class ProxyBaseLLMRequestProcessing: ) @staticmethod - def _stream_usage_kwargs_for_event(obj: Mapping[str, object], usage: Mapping[str, Any]) -> Mapping[str, Any] | None: + def _stream_usage_for_event(obj: Mapping[str, object], usage: Mapping[str, Any]) -> Usage | None: + """ + Build the ``Usage`` to price a streamed usage event. + + Anthropic goes through ``AnthropicConfig.calculate_usage``, the same transformation + the non-streaming path uses, so ``prompt_tokens`` is the full input total and the + cache read plus 5m/1h cache creation split land in ``prompt_tokens_details`` where + the pricer looks for them. + """ if obj.get("type") == "message_delta": - return ProxyBaseLLMRequestProcessing._anthropic_stream_usage_kwargs(usage) + return AnthropicConfig().calculate_usage(usage_object=usage, reasoning_content=None) if obj.get("object") == "chat.completion.chunk": - return ProxyBaseLLMRequestProcessing._openai_stream_usage_kwargs(usage) + return Usage(**ProxyBaseLLMRequestProcessing._openai_stream_usage_kwargs(usage)) return None @staticmethod @@ -3563,7 +3564,41 @@ class ProxyBaseLLMRequestProcessing: return None @staticmethod - def _inject_cost_into_usage_dict(obj: dict, model_name: str) -> dict | None: + def _logging_obj_cost_or_none( + model_response: ModelResponse, litellm_logging_obj: LiteLLMLoggingObj + ) -> float | None: + try: + cost: Final = litellm_logging_obj._response_cost_calculator(result=model_response) # pyright: ignore[reportPrivateUsage] # reuse the call's own cost calc for pricing parity with the logging callback + except Exception: + return None + return float(cost) if isinstance(cost, (int, float)) and not isinstance(cost, bool) else None + + @staticmethod + def _streamed_usage_cost( + model_response: ModelResponse, + model_name: str, + service_tier: str | None, + litellm_logging_obj: LiteLLMLoggingObj | None, + ) -> float | None: + """ + Price a streamed usage response through the call's logging object when there is one, + so the streamed cost picks up the same custom/deployment pricing (``custom_pricing``, + ``custom_llm_provider``, ``base_model``, ``router_model_id``) as the logging callback + rather than the model's sticker price; fall back to ``completion_cost`` by model name. + """ + cost_from_logging_obj: Final = ( + ProxyBaseLLMRequestProcessing._logging_obj_cost_or_none(model_response, litellm_logging_obj) + if litellm_logging_obj is not None + else None + ) + if cost_from_logging_obj is not None: + return cost_from_logging_obj + return ProxyBaseLLMRequestProcessing._completion_cost_or_none(model_response, model_name, service_tier) + + @staticmethod + def _inject_cost_into_usage_dict( + obj: dict, model_name: str, litellm_logging_obj: LiteLLMLoggingObj | None = None + ) -> dict | None: """ Inject cost information into the usage object of a streamed usage event (Anthropic ``message_delta`` or OpenAI ``chat.completion.chunk``). @@ -3571,6 +3606,8 @@ class ProxyBaseLLMRequestProcessing: Args: obj: Dictionary containing the SSE event data model_name: Model name for cost calculation + litellm_logging_obj: The call's logging object, used so the injected cost + matches the cost the logging callback records Returns: Modified dictionary with cost injected, or None if no modification needed @@ -3578,14 +3615,15 @@ class ProxyBaseLLMRequestProcessing: usage: Final = obj.get("usage") if not isinstance(usage, dict): return None - usage_kwargs: Final = ProxyBaseLLMRequestProcessing._stream_usage_kwargs_for_event(obj, usage) - if usage_kwargs is None: + stream_usage: Final = ProxyBaseLLMRequestProcessing._stream_usage_for_event(obj, usage) + if stream_usage is None: return None service_tier: Final = obj.get("service_tier") - cost_val: Final = ProxyBaseLLMRequestProcessing._completion_cost_or_none( - ModelResponse(usage=Usage(**usage_kwargs)), + cost_val: Final = ProxyBaseLLMRequestProcessing._streamed_usage_cost( + ModelResponse(usage=stream_usage), model_name, service_tier if isinstance(service_tier, str) else None, + litellm_logging_obj, ) if cost_val is None: return None diff --git a/litellm/proxy/pass_through_endpoints/streaming_handler.py b/litellm/proxy/pass_through_endpoints/streaming_handler.py index 697eb7b96eb..b71622fc33d 100644 --- a/litellm/proxy/pass_through_endpoints/streaming_handler.py +++ b/litellm/proxy/pass_through_endpoints/streaming_handler.py @@ -106,7 +106,7 @@ class PassThroughStreamingHandler: ) # rebind-ok: SSE frame reassembly buffer across transport chunks if complete_frames: yield ProxyBaseLLMRequestProcessing._process_chunk_with_cost_injection( - complete_frames, resolved_model_name + complete_frames, resolved_model_name, litellm_logging_obj ) if pending: yield pending diff --git a/tests/test_litellm/proxy/test_common_request_processing.py b/tests/test_litellm/proxy/test_common_request_processing.py index 355c6d27eb2..083982778b9 100644 --- a/tests/test_litellm/proxy/test_common_request_processing.py +++ b/tests/test_litellm/proxy/test_common_request_processing.py @@ -6082,6 +6082,121 @@ class TestInjectCostIntoUsageDict: injected = json.loads(result.split("\n")[0].split("data:", 1)[1].strip()) assert injected["usage"]["cost"] == pytest.approx(self._expected_cost("gpt-4o-mini", 11, 4)) + def test_message_delta_cost_charges_the_non_cached_input_tokens(self): + """Anthropic reports ``input_tokens`` excluding cache tokens, so reading it as the whole + prompt total drops the non-cached input from the bill on every cache hit.""" + model = "claude-haiku-4-5" + pricing = litellm.model_cost[model] + event = { + "type": "message_delta", + "delta": {"stop_reason": "end_turn"}, + "usage": { + "input_tokens": 14, + "output_tokens": 8, + "cache_read_input_tokens": 3202, + "cache_creation_input_tokens": 0, + }, + } + + result = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(event, model) + + assert result is not None + expected = ( + 14 * pricing["input_cost_per_token"] + + 3202 * pricing["cache_read_input_token_cost"] + + 8 * pricing["output_cost_per_token"] + ) + dropped_input = expected - 14 * pricing["input_cost_per_token"] + assert result["usage"]["cost"] == pytest.approx(expected) + assert result["usage"]["cost"] > dropped_input + + def test_message_delta_prices_1h_cache_creation_above_the_5m_rate(self): + """The ``cache_creation`` 5m/1h split has to survive into ``prompt_tokens_details``, + otherwise a 1h write is billed at the cheaper 5m rate.""" + model = "claude-haiku-4-5" + pricing = litellm.model_cost[model] + event = { + "type": "message_delta", + "delta": {"stop_reason": "end_turn"}, + "usage": { + "input_tokens": 14, + "output_tokens": 8, + "cache_read_input_tokens": 0, + "cache_creation_input_tokens": 2000, + "cache_creation": {"ephemeral_5m_input_tokens": 0, "ephemeral_1h_input_tokens": 2000}, + }, + } + + result = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(event, model) + + assert result is not None + base = 14 * pricing["input_cost_per_token"] + 8 * pricing["output_cost_per_token"] + expected_1h = base + 2000 * pricing["cache_creation_input_token_cost_above_1hr"] + flat_5m = base + 2000 * pricing["cache_creation_input_token_cost"] + assert expected_1h != pytest.approx(flat_5m) + assert result["usage"]["cost"] == pytest.approx(expected_1h) + + def test_message_delta_prices_through_the_logging_obj_so_custom_pricing_applies(self): + """Costing by model name alone yields sticker price, so a deployment with a negotiated + discount streamed a ``usage.cost`` that disagreed with the callback's ``response_cost``.""" + + class _StubLoggingObj: + def __init__(self, cost): + self._cost = cost + self.captured_result = None + + def _response_cost_calculator(self, result): + self.captured_result = result + return self._cost + + model = "claude-haiku-4-5" + discounted_cost = 0.00099 + stub = _StubLoggingObj(discounted_cost) + event = { + "type": "message_delta", + "delta": {"stop_reason": "end_turn"}, + "usage": { + "input_tokens": 14, + "output_tokens": 8, + "cache_read_input_tokens": 3202, + "cache_creation_input_tokens": 500, + "cache_creation": {"ephemeral_5m_input_tokens": 100, "ephemeral_1h_input_tokens": 400}, + }, + } + + result = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(event, model, stub) + + assert result is not None + assert result["usage"]["cost"] == discounted_cost + assert result["usage"]["cost"] != pytest.approx(self._expected_cost(model, 14 + 500 + 3202, 8)) + usage = stub.captured_result.usage + assert usage.prompt_tokens == 14 + 500 + 3202 + details = usage.prompt_tokens_details.cache_creation_token_details + assert details.ephemeral_5m_input_tokens == 100 + assert details.ephemeral_1h_input_tokens == 400 + + def test_message_delta_falls_back_to_model_pricing_when_the_logging_obj_returns_no_cost(self): + class _StubLoggingObj: + def _response_cost_calculator(self, result): + return None + + model = "claude-haiku-4-5" + pricing = litellm.model_cost[model] + event = { + "type": "message_delta", + "delta": {"stop_reason": "end_turn"}, + "usage": {"input_tokens": 14, "output_tokens": 8, "cache_read_input_tokens": 3202}, + } + + result = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(event, model, _StubLoggingObj()) + + assert result is not None + assert result["usage"]["cost"] == pytest.approx( + 14 * pricing["input_cost_per_token"] + + 3202 * pricing["cache_read_input_token_cost"] + + 8 * pricing["output_cost_per_token"] + ) + class TestProcessChunkWithCostInjection: def test_complete_usage_frame_chunk_is_injected(self, monkeypatch): @@ -6116,6 +6231,31 @@ class TestProcessChunkWithCostInjection: assert ProxyBaseLLMRequestProcessing._process_chunk_with_cost_injection(chunk, "gpt-4o-mini") == chunk + def test_message_delta_frame_is_priced_with_the_logging_obj(self, monkeypatch): + """Pins that the logging object reaches the pricer through the byte-frame entry point, + which is how the proxy actually calls this on a streamed Messages API request.""" + monkeypatch.setattr(litellm, "include_cost_in_streaming_usage", True) + + class _StubLoggingObj: + def _response_cost_calculator(self, result): + return 0.00042 + + chunk = ( + b"event: message_delta\n" + b'data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},' + b'"usage":{"input_tokens":14,"output_tokens":8,"cache_read_input_tokens":3202}}\n\n' + ) + + result = ProxyBaseLLMRequestProcessing._process_chunk_with_cost_injection( + chunk, "claude-haiku-4-5", _StubLoggingObj() + ) + + assert result != chunk + data_line = next(ln for ln in result.decode("utf-8").splitlines() if ln.startswith("data:")) + payload = json.loads(data_line.split("data:", 1)[1].strip()) + assert payload["usage"]["cost"] == 0.00042 + assert payload["usage"]["cache_read_input_tokens"] == 3202 + # --------------------------------------------------------------------------- # SSE keepalive during the time-to-first-token (issue #34819) From 6f4844bfd9dd50901e3e4542f296602c9c8c9e56 Mon Sep 17 00:00:00 2001 From: mateo Date: Wed, 19 Aug 2026 03:58:24 +0000 Subject: [PATCH 010/220] refactor(cost): trim streamed cost helper docstrings to the non-obvious bits Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/common_request_processing.py | 24 ++++++---------------- 1 file changed, 6 insertions(+), 18 deletions(-) diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index f5299b273b4..d533df8616d 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -3440,8 +3440,7 @@ class ProxyBaseLLMRequestProcessing: Args: chunk: The streaming chunk (dict, str, bytes, or bytearray) model_name: Model name for cost calculation - litellm_logging_obj: The call's logging object, used to price the chunk with - the same custom/deployment pricing as the logging callback + litellm_logging_obj: The call's logging object, used for pricing Returns: The processed chunk with cost information injected if applicable @@ -3538,14 +3537,8 @@ class ProxyBaseLLMRequestProcessing: @staticmethod def _stream_usage_for_event(obj: Mapping[str, object], usage: Mapping[str, Any]) -> Usage | None: - """ - Build the ``Usage`` to price a streamed usage event. - - Anthropic goes through ``AnthropicConfig.calculate_usage``, the same transformation - the non-streaming path uses, so ``prompt_tokens`` is the full input total and the - cache read plus 5m/1h cache creation split land in ``prompt_tokens_details`` where - the pricer looks for them. - """ + # Anthropic reports input_tokens excluding cache tokens, so reuse the non-streaming + # transformation to total the prompt and keep the 5m/1h cache creation split if obj.get("type") == "message_delta": return AnthropicConfig().calculate_usage(usage_object=usage, reasoning_content=None) if obj.get("object") == "chat.completion.chunk": @@ -3580,12 +3573,8 @@ class ProxyBaseLLMRequestProcessing: service_tier: str | None, litellm_logging_obj: LiteLLMLoggingObj | None, ) -> float | None: - """ - Price a streamed usage response through the call's logging object when there is one, - so the streamed cost picks up the same custom/deployment pricing (``custom_pricing``, - ``custom_llm_provider``, ``base_model``, ``router_model_id``) as the logging callback - rather than the model's sticker price; fall back to ``completion_cost`` by model name. - """ + # Pricing via the logging object inherits the deployment's custom pricing, so the + # streamed cost matches what the logging callback records instead of sticker price cost_from_logging_obj: Final = ( ProxyBaseLLMRequestProcessing._logging_obj_cost_or_none(model_response, litellm_logging_obj) if litellm_logging_obj is not None @@ -3606,8 +3595,7 @@ class ProxyBaseLLMRequestProcessing: Args: obj: Dictionary containing the SSE event data model_name: Model name for cost calculation - litellm_logging_obj: The call's logging object, used so the injected cost - matches the cost the logging callback records + litellm_logging_obj: The call's logging object, used for pricing Returns: Modified dictionary with cost injected, or None if no modification needed From aa832d81e91bb17e0f5ab081431cb03d2ef4083f Mon Sep 17 00:00:00 2001 From: ljogeiger <39981740+ljogeiger@users.noreply.github.com> Date: Wed, 19 Aug 2026 22:08:42 +0000 Subject: [PATCH 011/220] fix(vertex_ai): only fall back to a placeholder thought signature on the first parallel function call Gemini returns a thoughtSignature on the first function call of a parallel batch and leaves the siblings bare. When replaying that assistant turn, litellm gave every unsigned call the skip_thought_signature_validator placeholder, so a three-call turn went back with three signatures where Gemini had produced one. Keep the placeholder for the first call only and forward the siblings with whatever signature they actually carry, which is usually none. --- .../prompt_templates/factory.py | 23 ++-- .../test_vertex_ai_gemini_transformation.py | 115 ++++++++++++++++++ 2 files changed, 126 insertions(+), 12 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/factory.py b/litellm/litellm_core_utils/prompt_templates/factory.py index 0ed15c43ccf..0a9d7c427b4 100644 --- a/litellm/litellm_core_utils/prompt_templates/factory.py +++ b/litellm/litellm_core_utils/prompt_templates/factory.py @@ -1200,13 +1200,14 @@ def _encode_tool_call_id_with_signature(tool_call_id: str, thought_signature: st return tool_call_id -def _get_thought_signature_from_tool(tool: dict, model: str | None = None) -> str | None: +def _get_thought_signature_from_tool(tool: dict) -> str | None: """Extract thought signature from tool call's provider_specific_fields. If not provided try to extract thought signature from tool call id Checks both tool.provider_specific_fields and tool.function.provider_specific_fields. - If no signature is found and model is gemini-3, returns a dummy signature. + Returns None when the tool call carries no signature; callers decide whether a + placeholder signature is needed. """ # First check tool's provider_specific_fields provider_fields: Final = tool.get("provider_specific_fields") or {} @@ -1236,13 +1237,6 @@ def _get_thought_signature_from_tool(tool: dict, model: str | None = None) -> st if len(parts) == 2: _, signature = parts return signature - # If no signature found and model is gemini-3, return dummy signature - from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import ( - VertexGeminiConfig, - ) - - if model and VertexGeminiConfig._is_gemini_3_or_newer(model): - return _get_dummy_thought_signature() return None @@ -1312,8 +1306,10 @@ def convert_to_gemini_tool_call_invoke( VertexGeminiConfig, ) + needs_dummy_signature: Final = model is not None and VertexGeminiConfig._is_gemini_3_or_newer(model) + if tool_calls is not None: - for idx, tool in enumerate(tool_calls): + for tool in tool_calls: if "function" in tool: gemini_function_call: VertexFunctionCall | None = _gemini_tool_call_invoke_helper( function_call_params=tool["function"], @@ -1321,7 +1317,10 @@ def convert_to_gemini_tool_call_invoke( ) if gemini_function_call is not None: part_dict: VertexPartType = {"function_call": gemini_function_call} - thought_signature = _get_thought_signature_from_tool(dict(tool), model=model) + thought_signature = _get_thought_signature_from_tool(dict(tool)) + is_first_function_call = len(_parts_list) == 0 + if not thought_signature and is_first_function_call and needs_dummy_signature: + thought_signature = _get_dummy_thought_signature() if thought_signature: part_dict["thoughtSignature"] = thought_signature @@ -1344,7 +1343,7 @@ def convert_to_gemini_tool_call_invoke( thought_signature = provider_fields.get("thought_signature") # If no signature found and model is gemini-3, use dummy signature - if not thought_signature and model and VertexGeminiConfig._is_gemini_3_or_newer(model): + if not thought_signature and needs_dummy_signature: thought_signature = _get_dummy_thought_signature() if thought_signature: diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py index 8ee8186f6bb..28c551ab824 100644 --- a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py +++ b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py @@ -1,3 +1,5 @@ +import base64 + from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_gemini_tool_call_result, ) @@ -784,6 +786,119 @@ def test_dummy_signature_with_function_call_mode(): assert gemini_parts[0]["thoughtSignature"] == expected_dummy +def _parallel_tool_calls(*signatures): + return [ + { + "id": f"call_{idx}", + "type": "function", + "function": { + "name": f"tool_{idx}", + "arguments": '{"location": "Paris"}', + **( + {"provider_specific_fields": {"thought_signature": signature}} + if signature is not None + else {} + ), + }, + "index": idx, + } + for idx, signature in enumerate(signatures) + ] + + +REAL_THOUGHT_SIGNATURE = "Co4CAdHtim/rWgXbz2Ghp4tShzLeMASrPw6JJyYIC3cbVyZnKzU3uv8/wVzyS2sKRPL2m8QQHHXbNQhEEz500G7n" + + +def test_dummy_signature_only_on_first_parallel_tool_call(): + """Gemini only returns a thought signature on the first of N parallel function calls. + + The sibling calls carry no signature, so replaying them must not fabricate one. + """ + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls(None, None, None), + }, + model="gemini-3-pro-preview", + ) + + expected_dummy = base64.b64encode(b"skip_thought_signature_validator").decode( + "utf-8" + ) + assert len(gemini_parts) == 3 + assert gemini_parts[0]["thoughtSignature"] == expected_dummy + assert "thoughtSignature" not in gemini_parts[1] + assert "thoughtSignature" not in gemini_parts[2] + + +def test_real_signature_on_first_parallel_tool_call_leaves_siblings_empty(): + """The real signature from Gemini rides on the first call; siblings stay signature-free.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls(REAL_THOUGHT_SIGNATURE, None, None), + }, + model="gemini-3-pro-preview", + ) + + assert len(gemini_parts) == 3 + assert gemini_parts[0]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE + assert "thoughtSignature" not in gemini_parts[1] + assert "thoughtSignature" not in gemini_parts[2] + + +def test_real_signature_on_later_parallel_tool_call_is_preserved(): + """A signature attached to a non-first call is still forwarded as-is.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls(None, REAL_THOUGHT_SIGNATURE), + }, + model="gemini-3-pro-preview", + ) + + expected_dummy = base64.b64encode(b"skip_thought_signature_validator").decode( + "utf-8" + ) + assert len(gemini_parts) == 2 + assert gemini_parts[0]["thoughtSignature"] == expected_dummy + assert gemini_parts[1]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE + + +def test_no_signatures_on_parallel_tool_calls_for_gemini_2_5(): + """Non-gemini-3 models never get a placeholder signature, on any call.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls(None, None), + }, + model="gemini-2.5-flash", + ) + + assert len(gemini_parts) == 2 + assert all("thoughtSignature" not in part for part in gemini_parts) + + # Tests for media_resolution (detail parameter) handling - Issue #17084 class TestMediaResolution: """Tests for media_resolution handling in Gemini 2.x models""" From 677ef1e317080c20aec895a7ed25c058a1e18582 Mon Sep 17 00:00:00 2001 From: ljogeiger <39981740+ljogeiger@users.noreply.github.com> Date: Wed, 19 Aug 2026 22:21:31 +0000 Subject: [PATCH 012/220] docs(vertex_ai): drop stale note about the removed model argument --- litellm/llms/vertex_ai/gemini/transformation.py | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/litellm/llms/vertex_ai/gemini/transformation.py b/litellm/llms/vertex_ai/gemini/transformation.py index f2d318a9ffd..11c026010ee 100644 --- a/litellm/llms/vertex_ai/gemini/transformation.py +++ b/litellm/llms/vertex_ai/gemini/transformation.py @@ -645,10 +645,9 @@ def _collect_tool_call_thought_signatures( the text part as well would send two copies and double-bill the previous turn's reasoning tokens on gemini-3 and newer models. - Detection deliberately calls _get_thought_signature_from_tool without the - model argument: with a gemini-3 model that helper synthesizes a dummy - signature for unsigned tool calls, which must not suppress a real - text-part signature (e.g. replaying gemini-2.5 history to a newer model). + Only real signatures count here; a synthesized placeholder must not + suppress a genuine text-part signature (e.g. replaying gemini-2.5 history + to a newer model). """ signatures: tuple[str, ...] = () From d5af42717e9713771b8a014455b79d22b0756754 Mon Sep 17 00:00:00 2001 From: ljogeiger <39981740+ljogeiger@users.noreply.github.com> Date: Wed, 19 Aug 2026 22:32:24 +0000 Subject: [PATCH 013/220] test(vertex_ai): cover id-embedded, tool-level, and end-to-end parallel signature replay --- .../test_vertex_ai_gemini_transformation.py | 181 ++++++++++++++++++ 1 file changed, 181 insertions(+) diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py index 28c551ab824..562f27c11cf 100644 --- a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py +++ b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py @@ -806,6 +806,27 @@ def _parallel_tool_calls(*signatures): ] +def _parallel_tool_calls_signed_via_id(*signatures): + """Parallel tool calls in the shape LiteLLM actually hands back to clients. + + The signature rides in the tool call id behind __thought__, which is what an + OpenAI-format client echoes back on the next turn. + """ + from litellm.litellm_core_utils.prompt_templates.factory import ( + _encode_tool_call_id_with_signature, + ) + + return [ + { + "id": _encode_tool_call_id_with_signature(f"call_{idx}", signature), + "type": "function", + "function": {"name": f"tool_{idx}", "arguments": '{"location": "Paris"}'}, + "index": idx, + } + for idx, signature in enumerate(signatures) + ] + + REAL_THOUGHT_SIGNATURE = "Co4CAdHtim/rWgXbz2Ghp4tShzLeMASrPw6JJyYIC3cbVyZnKzU3uv8/wVzyS2sKRPL2m8QQHHXbNQhEEz500G7n" @@ -899,6 +920,166 @@ def test_no_signatures_on_parallel_tool_calls_for_gemini_2_5(): assert all("thoughtSignature" not in part for part in gemini_parts) +def test_signature_embedded_in_tool_call_id_only_on_first_parallel_call(): + """The production shape: the signature arrives inside the first call's id, siblings have bare ids.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls_signed_via_id( + REAL_THOUGHT_SIGNATURE, None, None + ), + }, + model="gemini-3-pro-preview", + ) + + assert len(gemini_parts) == 3 + assert gemini_parts[0]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE + assert "thoughtSignature" not in gemini_parts[1] + assert "thoughtSignature" not in gemini_parts[2] + + +def test_tool_level_provider_specific_fields_signature_leaves_siblings_empty(): + """A signature on the tool call itself, rather than on its function, behaves the same way.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + tool_calls = _parallel_tool_calls(None, None) + tool_calls[0]["provider_specific_fields"] = { + "thought_signature": REAL_THOUGHT_SIGNATURE + } + + gemini_parts = convert_to_gemini_tool_call_invoke( + {"role": "assistant", "content": None, "tool_calls": tool_calls}, + model="gemini-3-pro-preview", + ) + + assert len(gemini_parts) == 2 + assert gemini_parts[0]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE + assert "thoughtSignature" not in gemini_parts[1] + + +def test_placeholder_lands_on_first_emitted_part_not_first_tool_call_entry(): + """A non-function entry (e.g. an OpenAI custom tool call) emits no part, so it must not + consume the one placeholder slot and leave the real first function call bare.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + _get_dummy_thought_signature, + convert_to_gemini_tool_call_invoke, + ) + + tool_calls = [ + {"id": "call_custom", "type": "custom", "custom": {"name": "noop", "input": ""}} + ] + _parallel_tool_calls(None, None) + + gemini_parts = convert_to_gemini_tool_call_invoke( + {"role": "assistant", "content": None, "tool_calls": tool_calls}, + model="gemini-3-pro-preview", + ) + + assert len(gemini_parts) == 2 + assert gemini_parts[0]["thoughtSignature"] == _get_dummy_thought_signature() + assert "thoughtSignature" not in gemini_parts[1] + + +def test_no_placeholder_when_model_is_unknown(): + """Without a model there is nothing to prove the target needs a placeholder, so none is added.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls(None, None), + }, + ) + + assert len(gemini_parts) == 2 + assert all("thoughtSignature" not in part for part in gemini_parts) + + +def test_real_signature_forwarded_to_gemini_2_5_without_placeholder_siblings(): + """Older models still receive a real signature that a client replays, and still get no placeholder.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls(REAL_THOUGHT_SIGNATURE, None), + }, + model="gemini-2.5-flash", + ) + + assert len(gemini_parts) == 2 + assert gemini_parts[0]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE + assert "thoughtSignature" not in gemini_parts[1] + + +def test_parallel_tool_call_history_replayed_through_full_message_conversion(): + """End to end through the message-history converter, the path a real /chat/completions replay takes.""" + from litellm.llms.vertex_ai.gemini.transformation import ( + _gemini_convert_messages_with_history, + ) + + messages = [ + {"role": "user", "content": "Weather in Paris, London and Tokyo?"}, + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls_signed_via_id( + REAL_THOUGHT_SIGNATURE, None, None + ), + }, + ] + + contents = _gemini_convert_messages_with_history( + messages=messages, model="gemini-3-pro-preview" + ) + + model_parts = contents[1]["parts"] + assert len(model_parts) == 3 + assert model_parts[0]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE + assert "thoughtSignature" not in model_parts[1] + assert "thoughtSignature" not in model_parts[2] + + +def test_signed_text_part_survives_alongside_unsigned_parallel_tool_calls(): + """gemini-2.5 history with a signed text part and parallel unsigned calls: the text keeps its real + signature, only the first call gets the placeholder, and the siblings stay bare.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + _get_dummy_thought_signature, + ) + from litellm.llms.vertex_ai.gemini.transformation import ( + _gemini_convert_messages_with_history, + ) + + msg = { + "role": "assistant", + "content": "Checking all three cities.", + "provider_specific_fields": {"thought_signatures": ["real_25_signature"]}, + "tool_calls": _parallel_tool_calls(None, None, None), + } + + parts = _gemini_convert_messages_with_history( + messages=[msg], model="gemini-3-pro-preview" + )[0]["parts"] + + assert parts[0]["text"] == "Checking all three cities." + assert parts[0]["thoughtSignature"] == "real_25_signature" + assert parts[1]["thoughtSignature"] == _get_dummy_thought_signature() + assert "thoughtSignature" not in parts[2] + assert "thoughtSignature" not in parts[3] + + # Tests for media_resolution (detail parameter) handling - Issue #17084 class TestMediaResolution: """Tests for media_resolution handling in Gemini 2.x models""" From db50e123d5f23d475dab0bc62e33364ea19df817 Mon Sep 17 00:00:00 2001 From: ljogeiger <39981740+ljogeiger@users.noreply.github.com> Date: Wed, 19 Aug 2026 22:42:24 +0000 Subject: [PATCH 014/220] test(vertex_ai): parametrize placeholder scoping across gemini-3 model variants --- .../test_vertex_ai_gemini_transformation.py | 36 +++++++++++++++++++ 1 file changed, 36 insertions(+) diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py index 562f27c11cf..9ebf6db11de 100644 --- a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py +++ b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py @@ -1,5 +1,7 @@ import base64 +import pytest + from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_gemini_tool_call_result, ) @@ -1052,6 +1054,40 @@ def test_parallel_tool_call_history_replayed_through_full_message_conversion(): assert "thoughtSignature" not in model_parts[2] +@pytest.mark.parametrize( + "model", + [ + "gemini-3-pro-preview", + "gemini-3-flash-preview", + "gemini-3.1-pro-preview", + "gemini-3.6-flash", + "gemini-3.7-flash", + "vertex_ai/gemini-3.7-flash", + "gemini/gemini-3.7-flash", + ], +) +def test_placeholder_scoped_to_first_call_across_gemini_3_variants(model): + """Every gemini-3 family member, bare or provider-prefixed, gets one placeholder at most.""" + from litellm.litellm_core_utils.prompt_templates.factory import ( + _get_dummy_thought_signature, + convert_to_gemini_tool_call_invoke, + ) + + gemini_parts = convert_to_gemini_tool_call_invoke( + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls(None, None, None), + }, + model=model, + ) + + assert len(gemini_parts) == 3 + assert gemini_parts[0]["thoughtSignature"] == _get_dummy_thought_signature() + assert "thoughtSignature" not in gemini_parts[1] + assert "thoughtSignature" not in gemini_parts[2] + + def test_signed_text_part_survives_alongside_unsigned_parallel_tool_calls(): """gemini-2.5 history with a signed text part and parallel unsigned calls: the text keeps its real signature, only the first call gets the placeholder, and the siblings stay bare.""" From 579291774b0a8b5e98c33140ae89fe60a30360b0 Mon Sep 17 00:00:00 2001 From: ljogeiger <39981740+ljogeiger@users.noreply.github.com> Date: Thu, 20 Aug 2026 00:16:36 +0000 Subject: [PATCH 015/220] docs(vertex_ai): cite Google's thought signature rules for parallel calls Link the Gemini Enterprise Agent Platform docs at both places the behavior is decided. The docs state that only the first functionCall part of a parallel batch carries a thought_signature, and that setting skip_thought_signature_validator "should be a last resort as it will negatively impact model performance". --- .../litellm_core_utils/prompt_templates/factory.py | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/factory.py b/litellm/litellm_core_utils/prompt_templates/factory.py index 0a9d7c427b4..b676077ab0e 100644 --- a/litellm/litellm_core_utils/prompt_templates/factory.py +++ b/litellm/litellm_core_utils/prompt_templates/factory.py @@ -1245,10 +1245,14 @@ def _get_dummy_thought_signature() -> str: This is used when transferring conversation history from older models (like gemini-2.5-flash) to gemini-3, which requires thought_signature - for strict validation. + for strict validation. Google documents it as a last resort that "will + negatively impact model performance", so callers must only fall back to it + when no real signature is available. + + See: + https://ai.google.dev/gemini-api/docs/thought-signatures#faqs + https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/thinking/thought-signatures """ - # Return a base64-encoded dummy signature string - # Below dummy signature is recommended by google - https://ai.google.dev/gemini-api/docs/thought-signatures#faqs dummy_data: Final = b"skip_thought_signature_validator" return base64.b64encode(dummy_data).decode("utf-8") @@ -1318,6 +1322,9 @@ def convert_to_gemini_tool_call_invoke( if gemini_function_call is not None: part_dict: VertexPartType = {"function_call": gemini_function_call} thought_signature = _get_thought_signature_from_tool(dict(tool)) + # Gemini signs only the first functionCall part of a parallel batch, so scope the + # placeholder fallback to that part instead of fabricating one per sibling call: + # https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/thinking/thought-signatures#parallel_function_calling_example is_first_function_call = len(_parts_list) == 0 if not thought_signature and is_first_function_call and needs_dummy_signature: thought_signature = _get_dummy_thought_signature() From a5ad22b8a3f734b09ce5e05a68dab5c5205907a4 Mon Sep 17 00:00:00 2001 From: ljogeiger <39981740+ljogeiger@users.noreply.github.com> Date: Thu, 20 Aug 2026 00:16:36 +0000 Subject: [PATCH 016/220] test(vertex_ai): cover gemini-3.5-flash and drop assertion-echoing docstrings Add gemini-3.5-flash to the placeholder-scoping matrix and a regression test that a natively signed parallel turn replays with no skip_thought_signature_validator anywhere in the payload, the shape that was producing empty text responses on 3.5. Hoist the repeated placeholder expression into one constant and rewrite the docstrings that restated their own assertions to say why the case matters instead. --- .../test_vertex_ai_gemini_transformation.py | 83 +++++++++++++------ 1 file changed, 58 insertions(+), 25 deletions(-) diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py index 9ebf6db11de..8c1de12e7d9 100644 --- a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py +++ b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py @@ -830,13 +830,14 @@ def _parallel_tool_calls_signed_via_id(*signatures): REAL_THOUGHT_SIGNATURE = "Co4CAdHtim/rWgXbz2Ghp4tShzLeMASrPw6JJyYIC3cbVyZnKzU3uv8/wVzyS2sKRPL2m8QQHHXbNQhEEz500G7n" +PLACEHOLDER_SIGNATURE = base64.b64encode(b"skip_thought_signature_validator").decode( + "utf-8" +) def test_dummy_signature_only_on_first_parallel_tool_call(): - """Gemini only returns a thought signature on the first of N parallel function calls. - - The sibling calls carry no signature, so replaying them must not fabricate one. - """ + """Google documents the placeholder as a last resort that degrades quality, so an unsigned + parallel turn replayed to gemini-3 gets a budget of exactly one.""" from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_gemini_tool_call_invoke, ) @@ -850,17 +851,15 @@ def test_dummy_signature_only_on_first_parallel_tool_call(): model="gemini-3-pro-preview", ) - expected_dummy = base64.b64encode(b"skip_thought_signature_validator").decode( - "utf-8" - ) assert len(gemini_parts) == 3 - assert gemini_parts[0]["thoughtSignature"] == expected_dummy + assert gemini_parts[0]["thoughtSignature"] == PLACEHOLDER_SIGNATURE assert "thoughtSignature" not in gemini_parts[1] assert "thoughtSignature" not in gemini_parts[2] def test_real_signature_on_first_parallel_tool_call_leaves_siblings_empty(): - """The real signature from Gemini rides on the first call; siblings stay signature-free.""" + """Gemini signs only the first of N parallel function calls, so a faithful replay has + nothing to attach to the siblings.""" from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_gemini_tool_call_invoke, ) @@ -881,7 +880,8 @@ def test_real_signature_on_first_parallel_tool_call_leaves_siblings_empty(): def test_real_signature_on_later_parallel_tool_call_is_preserved(): - """A signature attached to a non-first call is still forwarded as-is.""" + """Clients may reorder or drop calls, so a signature that lands on a non-first call is + still the model's own and must survive the round trip.""" from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_gemini_tool_call_invoke, ) @@ -895,11 +895,8 @@ def test_real_signature_on_later_parallel_tool_call_is_preserved(): model="gemini-3-pro-preview", ) - expected_dummy = base64.b64encode(b"skip_thought_signature_validator").decode( - "utf-8" - ) assert len(gemini_parts) == 2 - assert gemini_parts[0]["thoughtSignature"] == expected_dummy + assert gemini_parts[0]["thoughtSignature"] == PLACEHOLDER_SIGNATURE assert gemini_parts[1]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE @@ -970,7 +967,6 @@ def test_placeholder_lands_on_first_emitted_part_not_first_tool_call_entry(): """A non-function entry (e.g. an OpenAI custom tool call) emits no part, so it must not consume the one placeholder slot and leave the real first function call bare.""" from litellm.litellm_core_utils.prompt_templates.factory import ( - _get_dummy_thought_signature, convert_to_gemini_tool_call_invoke, ) @@ -984,7 +980,7 @@ def test_placeholder_lands_on_first_emitted_part_not_first_tool_call_entry(): ) assert len(gemini_parts) == 2 - assert gemini_parts[0]["thoughtSignature"] == _get_dummy_thought_signature() + assert gemini_parts[0]["thoughtSignature"] == PLACEHOLDER_SIGNATURE assert "thoughtSignature" not in gemini_parts[1] @@ -1054,22 +1050,62 @@ def test_parallel_tool_call_history_replayed_through_full_message_conversion(): assert "thoughtSignature" not in model_parts[2] +@pytest.mark.parametrize( + "model", + ["gemini-3.5-flash", "vertex_ai/gemini-3.5-flash", "gemini/gemini-3.5-flash"], +) +def test_natively_signed_parallel_turn_never_carries_a_placeholder(model): + """A native gemini-3.5 parallel turn replays with zero skip_thought_signature_validator parts. + + Fabricating the placeholder alongside a real signature is what produced empty text responses + on gemini-3.5 parallel function calling, so the whole payload has to stay placeholder-free. + """ + import json + + from litellm.llms.vertex_ai.gemini.transformation import ( + _gemini_convert_messages_with_history, + ) + + messages = [ + {"role": "user", "content": "Weather in Paris, London and Tokyo?"}, + { + "role": "assistant", + "content": None, + "tool_calls": _parallel_tool_calls_signed_via_id( + REAL_THOUGHT_SIGNATURE, None, None + ), + }, + ] + + contents = _gemini_convert_messages_with_history(messages=messages, model=model) + + model_parts = contents[1]["parts"] + assert len(model_parts) == 3 + assert model_parts[0]["thoughtSignature"] == REAL_THOUGHT_SIGNATURE + assert "thoughtSignature" not in model_parts[1] + assert "thoughtSignature" not in model_parts[2] + assert PLACEHOLDER_SIGNATURE not in json.dumps(contents) + + @pytest.mark.parametrize( "model", [ "gemini-3-pro-preview", "gemini-3-flash-preview", "gemini-3.1-pro-preview", + "gemini-3.5-flash", "gemini-3.6-flash", "gemini-3.7-flash", + "vertex_ai/gemini-3.5-flash", "vertex_ai/gemini-3.7-flash", + "gemini/gemini-3.5-flash", "gemini/gemini-3.7-flash", ], ) def test_placeholder_scoped_to_first_call_across_gemini_3_variants(model): - """Every gemini-3 family member, bare or provider-prefixed, gets one placeholder at most.""" + """The gemini-3 gate is a substring match, so every family member and prefix form has to + land on the same one-placeholder budget rather than only the versions we happened to try.""" from litellm.litellm_core_utils.prompt_templates.factory import ( - _get_dummy_thought_signature, convert_to_gemini_tool_call_invoke, ) @@ -1083,17 +1119,14 @@ def test_placeholder_scoped_to_first_call_across_gemini_3_variants(model): ) assert len(gemini_parts) == 3 - assert gemini_parts[0]["thoughtSignature"] == _get_dummy_thought_signature() + assert gemini_parts[0]["thoughtSignature"] == PLACEHOLDER_SIGNATURE assert "thoughtSignature" not in gemini_parts[1] assert "thoughtSignature" not in gemini_parts[2] def test_signed_text_part_survives_alongside_unsigned_parallel_tool_calls(): - """gemini-2.5 history with a signed text part and parallel unsigned calls: the text keeps its real - signature, only the first call gets the placeholder, and the siblings stay bare.""" - from litellm.litellm_core_utils.prompt_templates.factory import ( - _get_dummy_thought_signature, - ) + """Text-part and function-call signatures are collected by separate code paths, so scoping the + placeholder must not disturb a real signature that arrived on the text part.""" from litellm.llms.vertex_ai.gemini.transformation import ( _gemini_convert_messages_with_history, ) @@ -1111,7 +1144,7 @@ def test_signed_text_part_survives_alongside_unsigned_parallel_tool_calls(): assert parts[0]["text"] == "Checking all three cities." assert parts[0]["thoughtSignature"] == "real_25_signature" - assert parts[1]["thoughtSignature"] == _get_dummy_thought_signature() + assert parts[1]["thoughtSignature"] == PLACEHOLDER_SIGNATURE assert "thoughtSignature" not in parts[2] assert "thoughtSignature" not in parts[3] From 367dd537b995d23421fd99bdb1093e443e98911e Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 18:39:15 -0700 Subject: [PATCH 017/220] feat(e2e): move record/replay to the provider edge (LIT-5745) Replaces the test-side fixture transport with an in-process provider-edge HTTP server the proxy's deployments point their api_base at. Record forwards provider calls verbatim and writes them to the bundle; replay answers them from the bundle with zero provider calls while key auth, routing, cost calculation, and spend-log writes still execute against the live proxy and database. Drift comes back as HTTP 599 naming the computed and closest recorded keys. Request headers are never stored and responses are kept byte-identical between modes from the proxy's side of the socket. --- tests/e2e/CLAUDE.md | 12 +- tests/e2e/CONTRIBUTING.md | 10 +- tests/e2e/conftest.py | 14 +- tests/e2e/e2e_config.py | 32 +- tests/e2e/e2e_http.py | 34 + tests/e2e/fixture_bundle.py | 133 +--- tests/e2e/fixture_mode.py | 132 ++++ tests/e2e/fixture_transport.py | 724 ------------------ tests/e2e/provider_edge.py | 546 +++++++++++++ tests/e2e/proxy_client.py | 16 +- .../test_provider_edge_spend_e2e.py | 50 ++ tests/e2e/test_fixture_bundle.py | 46 +- tests/e2e/test_fixture_mode.py | 114 +++ tests/e2e/test_fixture_transport.py | 676 ---------------- tests/e2e/test_provider_edge.py | 492 ++++++++++++ 15 files changed, 1453 insertions(+), 1578 deletions(-) create mode 100644 tests/e2e/fixture_mode.py delete mode 100644 tests/e2e/fixture_transport.py create mode 100644 tests/e2e/provider_edge.py create mode 100644 tests/e2e/quota_management/spend_tracking/test_provider_edge_spend_e2e.py create mode 100644 tests/e2e/test_fixture_mode.py delete mode 100644 tests/e2e/test_fixture_transport.py create mode 100644 tests/e2e/test_provider_edge.py diff --git a/tests/e2e/CLAUDE.md b/tests/e2e/CLAUDE.md index d7334552d0c..840a40a54cd 100644 --- a/tests/e2e/CLAUDE.md +++ b/tests/e2e/CLAUDE.md @@ -73,13 +73,17 @@ Mark live tests with `@pytest.mark.e2e` (on the class or the module). Pure cover ## Record and replay fixtures -`E2E_FIXTURE_MODE` selects the transport every client is built on: `live` (the default, and what an unset variable means: nothing changes), `record` (run against the live proxy and write every interaction to a fixture bundle), or `replay` (serve every interaction back from the bundle with no HTTP at all, so a replay run needs no proxy and cannot bill a provider). The seam is `select_transport` in `fixture_transport.py`, applied inside `build_proxy_client`; both transports fulfil the same `Transport` protocol, so no test or client changes shape in any mode +`E2E_FIXTURE_MODE` scopes the proxy's provider-bound traffic: `live` (the default, and what an unset variable means: nothing changes), `record` (the proxy's provider calls are forwarded to the real provider through a local edge server and written to a fixture bundle), or `replay` (the edge answers those calls from the bundle, so the run makes zero provider calls and spends nothing). Test-to-proxy traffic always goes over the wire in every mode: record and replay both need the live proxy and database, because the point is that key auth, routing, cost calculation, and spend-log writes execute for real while only the provider is swapped out. Breaking any of those in the proxy turns a replay run red -A bundle (default `tests/e2e/.fixtures`, override with `E2E_FIXTURE_DIR`) is a directory: `manifest.json` carries the record timestamp, harness git version, and format version, and each test gets a subdirectory holding one JSON file per transport call in call order (`0000-post-chat-completions.json`). Auth header values and credential request fields (`api_key`, `*_secret_key`, `static_headers`, and the like; the list is `fixture_canonical.py`'s) are redacted on write, and file uploads store a sha256 digest instead of the bytes; response bodies are stored verbatim (a /key/generate response keeps the ephemeral virtual key it minted), which is part of why bundles are gitignored. `fixture_bundle.py` owns the format +The seam is `provider_edge.py`: `start_provider_edge` boots an in-process HTTP server (one shared instance per pytest process, `e2e_config.provider_edge_base` is the accessor) that mounts each supported provider under a path prefix (`EDGE_MOUNTS`: `/openai` -> `https://api.openai.com`, `/anthropic` -> `https://api.anthropic.com`). A test participates by registering its deployment with `api_base=provider_edge_base("openai")` plus the provider's path suffix; `quota_management/spend_tracking/test_provider_edge_spend_e2e.py` is the reference. In live mode the accessor returns None and the deployment defaults to the real provider, so an edge-wired test runs in all three modes unchanged. Non-wired tests hit their providers live in every mode. The edge binds `E2E_PROVIDER_EDGE_BIND_HOST` (default 127.0.0.1) and advertises `E2E_PROVIDER_EDGE_ADVERTISE_HOST` in the api_base it hands out, for proxies running in containers -Replay matches calls per test by canonical key: `fixture_canonical.py` canonicalizes the recorded request (volatile headers and credential fields out, unique markers, generated ids, uuids, and timestamps replaced with fixed placeholders, object keys sorted) and the key is the method, path, and a content hash, so identity survives re-records and machine changes while any real content drift is a `ReplayMiss` that names the computed key, the closest recorded key with its file, and a content diff, and never falls through to a live call. Matching is order-independent across distinct keys (concurrent calls may interleave) and FIFO within one key (a poll loop replays its responses in recorded order); a passed test must also consume its whole recording, or teardown fails it naming a leftover key. Either way the fix is always to re-record with `E2E_FIXTURE_MODE=record`. Every rewrite rule lives in `fixture_canonical.py`, so a new volatile header, credential field name, or generated-id shape is one edit there. Record starts fresh every time: it wipes the previous bundle (refusing to wipe a directory that is not a bundle) and never reads it. A replay bundle whose manifest is older than seven days hard-fails at collection time naming the bundle's age, so replay can never certify against fixtures that have drifted more than a week from the live proxy +A bundle (default `tests/e2e/.fixtures`, override with `E2E_FIXTURE_DIR`) is a directory: `manifest.json` carries the record timestamp, harness git version, and format version, and each test gets a subdirectory holding one JSON file per provider call in call order (`0000-post-openai-v1-chat-completions.json`). Request headers are never stored (provider credentials never touch disk), non-JSON request bodies store a canonicalized sha256 digest instead of the bytes, and responses store status, filtered headers, and the verbatim body base64-encoded, which is part of why bundles are gitignored. `fixture_bundle.py` owns the format. Record serves the proxy the same filtered stored response replay will serve later, so the two modes are byte-identical from the proxy's side of the socket -Deliberately not here yet: streaming chunk fidelity (LIT-5742) and scoping record/replay to provider-bound traffic (LIT-5745) +Replay matches calls per test by canonical key: `fixture_canonical.py` canonicalizes the recorded request (volatile headers and credential fields out, unique markers, generated ids, uuids, and timestamps replaced with fixed placeholders, object keys sorted) and the key is the method, edge path, and a content hash, so identity survives re-records and machine changes while any real content drift comes back as an HTTP 599 naming the computed key, the closest recorded key with its file, and a content diff, and never falls through to a live call. Matching is order-independent across distinct keys (concurrent calls may interleave) and FIFO within one key (a retry loop replays its responses in recorded order); a passed test must also consume its whole recording, or teardown fails it naming a leftover key. Either way the fix is always to re-record with `E2E_FIXTURE_MODE=record`. Every rewrite rule lives in `fixture_canonical.py`, so a new volatile header, credential field name, or generated-id shape is one edit there. Record starts fresh every time: it wipes the previous bundle (refusing to wipe a directory that is not a bundle) and never reads it. A replay bundle whose manifest is older than seven days hard-fails at collection time naming the bundle's age, so replay can never certify against fixtures that have drifted more than a week from the live providers + +A replayed response carries the recorded provider response id, and `LiteLLM_SpendLogs.request_id` (the table's primary key) is that id, so a replay against a database that still holds the record run's rows silently dedupes its spend inserts and any spend assertion goes red with zero matching rows and nothing in the proxy log. Run both modes with `E2E_RESET_SPEND_LOGS=1` (plus `DATABASE_URL` in the runner env) so each session truncates the table after itself, or replay against a fresh database, which is the CI shape + +Current limits: streaming chunk fidelity is LIT-5742 (a streamed response records as one buffered body), CI wiring is LIT-5748, Bedrock cannot be mounted (SigV4 signs the Host header, so a rewritten api_base fails signature verification), multipart uploads have per-run random boundaries (the digest changes every run, so they always miss), and deployments baked into the proxy's config file cannot be edge-wired (only `/model/new` registrations can carry the edge api_base) ## Typing diff --git a/tests/e2e/CONTRIBUTING.md b/tests/e2e/CONTRIBUTING.md index 67da1be9562..9096050a45a 100644 --- a/tests/e2e/CONTRIBUTING.md +++ b/tests/e2e/CONTRIBUTING.md @@ -54,14 +54,16 @@ Some suites need extra services the bare proxy does not start. The `logging/` OT ### Record and replay -`E2E_FIXTURE_MODE=record` runs a suite against the live proxy as usual while writing every request/response pair to a fixture bundle (default `tests/e2e/.fixtures`, override with `E2E_FIXTURE_DIR`); `E2E_FIXTURE_MODE=replay` then runs the same suite entirely from that bundle, with no proxy traffic and no provider spend; the proxy liveness gate is skipped, so replay runs with no proxy up at all. Unset (or `live`) behaves exactly as before the knob existed +Record/replay scopes to the proxy's provider-bound traffic only. In `E2E_FIXTURE_MODE=record` the harness boots a local provider-edge server, edge-wired tests register their deployments with an `api_base` pointing at it, and every provider call the proxy makes is forwarded verbatim and written to a fixture bundle (default `tests/e2e/.fixtures`, override with `E2E_FIXTURE_DIR`). `E2E_FIXTURE_MODE=replay` runs the same tests against the same live proxy and database, but the edge answers the proxy's provider calls from the bundle instead of the provider, so the run makes zero provider calls and spends nothing while key auth, routing, cost calculation, and spend-log writes all still execute for real. Unset (or `live`) behaves exactly as before the knob existed. Both record and replay need the proxy up; only the provider is taken out of the loop ```bash -E2E_FIXTURE_MODE=record uv run pytest tests/e2e/llm_translation/ -v -E2E_FIXTURE_MODE=replay uv run pytest tests/e2e/llm_translation/ -v +E2E_FIXTURE_MODE=record uv run pytest tests/e2e/quota_management/spend_tracking/test_provider_edge_spend_e2e.py -v +E2E_FIXTURE_MODE=replay uv run pytest tests/e2e/quota_management/spend_tracking/test_provider_edge_spend_e2e.py -v ``` -Replay fails hard (`ReplayMiss`) when the tests drift from the recording, and a bundle older than seven days fails at collection time naming its age; either way the fix is to re-record. See `CLAUDE.md` in this directory for the bundle format and the transport seam +One sharp edge: a replayed response reuses the recorded provider response id, and that id is the primary key of `LiteLLM_SpendLogs`, so replaying against a database that still holds the record run's rows silently dedupes the spend writes and a spend assertion fails with zero rows. Run both commands above with `E2E_RESET_SPEND_LOGS=1` (and `DATABASE_URL` set in the pytest env) so each session truncates the spend log table after itself, or point replay at a fresh database + +Replay answers any provider call that drifted from the recording with an HTTP 599 whose body names the computed and closest recorded keys, so the test fails loudly instead of silently going live, and a bundle older than seven days fails at collection time naming its age; either way the fix is to re-record. Only tests that register edge-wired deployments participate: everything else hits its provider live in every mode, so record exactly the suite you replay. If the proxy runs in a container, set `E2E_PROVIDER_EDGE_ADVERTISE_HOST` (e.g. `host.docker.internal`) so the api_base the proxy stores can reach the edge on the pytest host, and `E2E_PROVIDER_EDGE_BIND_HOST=0.0.0.0` so the edge accepts it. See `CLAUDE.md` in this directory for the bundle format, the edge design, and the current limits (streaming, Bedrock, multipart) Tests marked `@pytest.mark.e2e` hard-fail when no proxy answers `/health/liveliness`, so a run that goes red with `No live proxy` at setup means the proxy isn't up; they never skip for a missing proxy, so an absent proxy can't be mistaken for a pass diff --git a/tests/e2e/conftest.py b/tests/e2e/conftest.py index da2a7da0bfa..dbe2d6e514e 100644 --- a/tests/e2e/conftest.py +++ b/tests/e2e/conftest.py @@ -23,12 +23,8 @@ import requests from e2e_config import CONTROL_PLANE_BASE_URL, FIXTURE_DIR, FIXTURE_MODE_RAW, PROXY_BASE_URL from e2e_db import RESET_OPT_IN_ENV, reset_spend_logs, run_spend_log_cleanup -from fixture_transport import ( - fixture_mode_collection_error, - fixture_report_lines, - parse_fixture_mode, - replay_leftover_error, -) +from fixture_mode import fixture_mode_collection_error, fixture_report_lines +from provider_edge import replay_leftover_error from junit_properties import attach_result_properties from lifecycle import ProxyClientProvider, ResourceManager from proxy_client import ProxyClient, build_proxy_client @@ -114,12 +110,10 @@ def _proxy_fail_reason() -> str | None: def pytest_runtest_setup(item: pytest.Item) -> None: """Hard-fail `e2e`-marked tests unless a proxy answers its liveness probe. Unmarked tests (unit coverage of the harness) don't touch the proxy, so they - run even when none is up. Never skip for a missing proxy. Replay mode serves - every call from the fixture bundle, so it needs no live proxy either.""" + run even when none is up. Never skip for a missing proxy. Replay mode needs + the proxy too: only provider-bound traffic replays from the bundle.""" if item.get_closest_marker("e2e") is None: return - if parse_fixture_mode(FIXTURE_MODE_RAW) == "replay": - return reason = _proxy_fail_reason() if reason is not None: pytest.fail(reason) diff --git a/tests/e2e/e2e_config.py b/tests/e2e/e2e_config.py index a5c3729f4be..8bf39f6021f 100644 --- a/tests/e2e/e2e_config.py +++ b/tests/e2e/e2e_config.py @@ -13,7 +13,8 @@ from pathlib import Path from dotenv import load_dotenv -from fixture_transport import deterministic_marker, parse_fixture_mode +from fixture_mode import deterministic_marker, parse_fixture_mode +from provider_edge import provider_edge_api_base # Local runs keep provider / DataDog keys in tests/e2e/.env (see CONTRIBUTING.md). # Compose injects them into the proxy container, but pytest on the host does not @@ -92,15 +93,24 @@ PROPAGATION_TIMEOUT = float(os.environ.get("E2E_PROPAGATION_TIMEOUT", "15")) EXPECT_RUST = os.environ.get("E2E_EXPECT_RUST", "").strip().lower() in ("1", "true", "yes") -# Record/replay fixture selection (see fixture_transport.py). The raw mode value -# is parsed and validated there; "live" (the default, also for empty values) -# means the harness behaves exactly as before this knob existed. +# Record/replay fixture selection (see fixture_mode.py and provider_edge.py). +# The raw mode value is parsed and validated there; "live" (the default, also +# for empty values) means the harness behaves exactly as before this knob +# existed. FIXTURE_MODE_RAW = os.environ.get("E2E_FIXTURE_MODE", "live") FIXTURE_DIR = Path( os.environ.get("E2E_FIXTURE_DIR", "").strip() or str(Path(__file__).resolve().parent / ".fixtures") ) +# Where the provider-edge server binds, and the host name edge api_base URLs +# advertise to the proxy. They differ when the proxy runs in a container and +# reaches the pytest host via a gateway name like host.docker.internal. +PROVIDER_EDGE_BIND_HOST = os.environ.get("E2E_PROVIDER_EDGE_BIND_HOST", "").strip() or "127.0.0.1" +PROVIDER_EDGE_ADVERTISE_HOST = ( + os.environ.get("E2E_PROVIDER_EDGE_ADVERTISE_HOST", "").strip() or PROVIDER_EDGE_BIND_HOST +) + # Deliberately modest concurrency. The suite shares its proxy with every other # suite in the run, and 750 users at spawn rate 50 saturated the request path hard # enough to distort latency-sensitive neighbours (and to spend real provider money @@ -157,6 +167,20 @@ def datadog_mcp_url(*, toolsets: str = "core") -> str: return f"{base}?toolsets={toolsets}" if toolsets else base +def provider_edge_base(mount: str) -> str | None: + """The api_base an edge-wired deployment should register with, using this + process's fixture-mode and edge-host configuration: None in live mode, the + shared edge server's mount URL in record and replay.""" + return provider_edge_api_base( + mount, + mode_raw=FIXTURE_MODE_RAW, + bundle_dir=FIXTURE_DIR, + bind_host=PROVIDER_EDGE_BIND_HOST, + advertise_host=PROVIDER_EDGE_ADVERTISE_HOST, + forward_timeout=REQUEST_TIMEOUT, + ) + + def unique_marker() -> str: """A short unique token per call/run, so concurrent runs and the shared response cache never collide on prompts, tags, or customer ids. In record diff --git a/tests/e2e/e2e_http.py b/tests/e2e/e2e_http.py index cb6fc7a01e5..03f201e946e 100644 --- a/tests/e2e/e2e_http.py +++ b/tests/e2e/e2e_http.py @@ -647,3 +647,37 @@ def download( content_type=_hdr(resp, "content-type"), body=resp.text, ) + + +class RawResponse(BaseModel): + """A verbatim upstream HTTP response for the provider edge (provider_edge.py): + status, lowercased headers, raw bytes. No Result classification because the + edge relays provider errors to the proxy untouched.""" + + status_code: int + headers: dict[str, str] + body: bytes + + +def forward( + method: str, + url: str, + *, + headers: dict[str, str], + body: bytes | None, + timeout: float = 60.0, +) -> RawResponse | NetworkError: + """Relay one provider-bound request verbatim for the provider edge's record + mode. No retries, no redirects, no schema: the proxy owns retry policy and + the recorded bundle must hold exactly what the provider returned.""" + try: + resp = requests.request( + method, url, headers=headers, data=body, timeout=timeout, allow_redirects=False + ) + except requests.RequestException as exc: + return NetworkError(message=str(exc)) + return RawResponse( + status_code=resp.status_code, + headers={name.lower(): value for name, value in resp.headers.items()}, + body=resp.content, + ) diff --git a/tests/e2e/fixture_bundle.py b/tests/e2e/fixture_bundle.py index 615ae8df1a4..6feb40fc8bc 100644 --- a/tests/e2e/fixture_bundle.py +++ b/tests/e2e/fixture_bundle.py @@ -1,17 +1,18 @@ -"""On-disk fixture bundle format for record/replay e2e runs (LIT-5729). +"""On-disk fixture bundle format for record/replay e2e runs (LIT-5729/LIT-5745). A bundle is a directory: one ``manifest.json`` (record timestamp + harness version + format version) plus one subdirectory per test, holding one JSON file -per transport interaction in call order. Bundles older than +per provider-bound interaction in call order. Bundles older than ``MAX_BUNDLE_AGE`` hard-fail replay at collection time (see conftest), so a green replay run can never certify against fixtures that have drifted more than -a week from the live proxy. +a week from the live providers. -This module owns the format only. The transports that produce and consume it -live in fixture_transport.py and the canonical match keys they compute live in -fixture_canonical.py (LIT-5741); streaming chunk fidelity and provider-scoping -are follow-ups (LIT-5742/5745). Every interaction file stores the full redacted -request because replay matches on its canonicalized content. +This module owns the format only. The provider-edge server that produces and +consumes it lives in provider_edge.py (LIT-5745) and the canonical match keys +it computes live in fixture_canonical.py (LIT-5741); streaming chunk fidelity +is a follow-up (LIT-5742). Every interaction file stores the full redacted +request because replay matches on its canonicalized content, and the response +as the raw HTTP status, filtered headers, and base64 body the provider sent. """ from __future__ import annotations @@ -23,29 +24,14 @@ import subprocess from dataclasses import dataclass, field from datetime import datetime, timedelta, timezone from pathlib import Path -from typing import Annotated, Final, Literal +from typing import Final -from pydantic import BaseModel, Field, JsonValue, TypeAdapter +from pydantic import BaseModel, JsonValue -from e2e_http import ( - BinaryStream, - NetworkError, - ProbeResult, - RateLimitedError, - Result, - StreamingResponse, - Success, - UnauthorizedError, - UnknownApiError, - ValidationError, -) - -BUNDLE_FORMAT_VERSION: Final = 1 +BUNDLE_FORMAT_VERSION: Final = 2 MAX_BUNDLE_AGE: Final = timedelta(days=7) MANIFEST_FILENAME: Final = "manifest.json" -_JSON: Final[TypeAdapter[JsonValue]] = TypeAdapter(JsonValue) - class Manifest(BaseModel): format_version: int @@ -54,13 +40,14 @@ class Manifest(BaseModel): class RecordedRequest(BaseModel): - """The request as the transport saw it, auth header values and credential - body/form fields redacted. + """The provider-bound request as the edge saw it, headers empty (SDK + telemetry headers vary run to run and auth material never touches disk). Replay matches on the canonical content key fixture_canonical.py computes - over ``method`` (the transport verb, not the HTTP verb), ``path``, and the - canonicalized headers, params, body, form, and file identity. File uploads - store a content digest instead of the bytes.""" + over ``method``, ``path`` (the edge path including the provider mount, + query string excluded), and the canonicalized headers, params, body, form, + and file identity. Non-JSON bodies store a canonicalized content digest + instead of the bytes.""" method: str path: str @@ -73,85 +60,19 @@ class RecordedRequest(BaseModel): file_bytes: int | None = None -class RecordedResult(BaseModel): - """A ``Result[R]`` flattened for disk. ``data`` holds the success payload as - raw JSON; replay re-validates it against the ``response_type`` the caller - passes, exactly like a live response body.""" +class RecordedHttpResponse(BaseModel): + """The provider's raw HTTP response: status, headers minus hop-by-hop and + volatile entries (see provider_edge.py), and the body as base64 so binary + payloads survive JSON.""" - shape: Literal["result"] = "result" - kind: Literal["success", "network", "unauthorized", "rate_limited", "validation", "unknown"] - status_code: int | None = None - data: JsonValue | None = None - message: str | None = None - body: str | None = None - retry_after_seconds: int | None = None - - -class RecordedStreaming(BaseModel): - shape: Literal["streaming"] = "streaming" - payload: StreamingResponse - - -class RecordedBinary(BaseModel): - shape: Literal["binary"] = "binary" - payload: BinaryStream - - -class RecordedProbe(BaseModel): - shape: Literal["probe"] = "probe" - payload: ProbeResult - - -type RecordedResponse = RecordedResult | RecordedStreaming | RecordedBinary | RecordedProbe + status_code: int + headers: dict[str, str] + body_b64: str class Interaction(BaseModel): request: RecordedRequest - response: Annotated[ - RecordedResult | RecordedStreaming | RecordedBinary | RecordedProbe, - Field(discriminator="shape"), - ] - - -def to_json_value(model: BaseModel) -> JsonValue: - return _JSON.validate_json(model.model_dump_json(by_alias=True)) - - -def from_result[R: BaseModel](result: Result[R]) -> RecordedResult: - match result: - case Success(status_code=status_code, data=data): - return RecordedResult(kind="success", status_code=status_code, data=to_json_value(data)) - case NetworkError(message=message): - return RecordedResult(kind="network", message=message) - case UnauthorizedError(): - return RecordedResult(kind="unauthorized") - case RateLimitedError(retry_after_seconds=retry_after_seconds, body=body): - return RecordedResult(kind="rate_limited", retry_after_seconds=retry_after_seconds, body=body) - case ValidationError(message=message): - return RecordedResult(kind="validation", message=message) - case UnknownApiError(status_code=status_code, body=body): - return RecordedResult(kind="unknown", status_code=status_code, body=body) - - -def to_result[R: BaseModel](recorded: RecordedResult, response_type: type[R]) -> Result[R]: - match recorded.kind: - case "success": - return Success( - status_code=recorded.status_code or 200, - data=response_type.model_validate(recorded.data), - ) - case "network": - return NetworkError(message=recorded.message or "") - case "unauthorized": - return UnauthorizedError() - case "rate_limited": - return RateLimitedError( - retry_after_seconds=recorded.retry_after_seconds, body=recorded.body or "" - ) - case "validation": - return ValidationError(message=recorded.message or "") - case "unknown": - return UnknownApiError(status_code=recorded.status_code or 0, body=recorded.body or "") + response: RecordedHttpResponse def slugify(raw: str, *, limit: int = 60) -> str: @@ -198,7 +119,7 @@ class BundleRecorder: root: Path _ordinals: dict[str, int] = field(default_factory=dict) - def record(self, *, test_key: str, request: RecordedRequest, response: RecordedResponse) -> None: + def record(self, *, test_key: str, request: RecordedRequest, response: RecordedHttpResponse) -> None: slug = slug_for_test(test_key) ordinal = self._ordinals.get(slug, 0) self._ordinals[slug] = ordinal + 1 diff --git a/tests/e2e/fixture_mode.py b/tests/e2e/fixture_mode.py new file mode 100644 index 00000000000..110f44380b4 --- /dev/null +++ b/tests/e2e/fixture_mode.py @@ -0,0 +1,132 @@ +"""Fixture-mode selection and per-test determinism for record/replay e2e runs. + +``E2E_FIXTURE_MODE`` is live (the default; nothing changes), record, or replay. +This module owns everything mode-shaped that is independent of the provider +edge itself: parsing the raw env value, the collection-time gate that aborts a +run whose mode can never work (unknown value, or replay against a missing or +stale bundle), the pytest report-header lines, the running test's node id, and +the deterministic per-test marker that lets a replay run regenerate exactly +the requests the record run sent. The provider-edge server that records and +serves provider traffic lives in provider_edge.py (LIT-5745). +""" + +from __future__ import annotations + +import hashlib +import os +from dataclasses import dataclass +from datetime import datetime +from pathlib import Path +from typing import Final, Literal, assert_never + +from fixture_bundle import ( + FreshBundle, + StaleBundle, + UnreadableBundle, + check_freshness, + format_age, +) + +type FixtureMode = Literal["live", "record", "replay"] + +FIXTURE_MODES: Final[tuple[FixtureMode, ...]] = ("live", "record", "replay") + +SESSION_TEST_KEY: Final = "session" + + +@dataclass(frozen=True, slots=True) +class InvalidFixtureMode: + value: str + + +def parse_fixture_mode(raw: str) -> FixtureMode | InvalidFixtureMode: + normalized = raw.strip().lower() or "live" + match normalized: + case "live" | "record" | "replay": + return normalized + case _: + return InvalidFixtureMode(value=raw) + + +def current_test_key() -> str: + """The pytest node id of the running test, from the PYTEST_CURRENT_TEST env + var pytest maintains (`` (setup|call|teardown)``); ``session`` for + calls outside any test (e.g. session-finish cleanup).""" + raw = os.environ.get("PYTEST_CURRENT_TEST", "") + if not raw: + return SESSION_TEST_KEY + return raw.rsplit(" (", 1)[0] + + +class ReplayMiss(AssertionError): + """Replay had no recorded interaction for a provider call the proxy made. + The suite drifted from the bundle (or the bundle from the suite): re-record.""" + + +_marker_ordinals: Final[dict[str, int]] = {} + + +def deterministic_marker() -> str: + """Stable stand-in for uuid-based unique markers in record and replay modes: + the Nth marker of a test is a pure function of the test's node id and N, so a + replay run regenerates exactly the model names, prompts, and tags the record + run sent and every recorded provider interaction still matches its key.""" + test_key = current_test_key() + ordinal = _marker_ordinals.get(test_key, 0) + _marker_ordinals[test_key] = ordinal + 1 + return hashlib.sha1(f"{test_key}#{ordinal}".encode()).hexdigest()[:12] + + +def fixture_mode_collection_error(mode_raw: str, bundle_dir: Path, *, now: datetime) -> str | None: + """Session-abort reason for a fixture-mode setup that can never work, or None. + Called at collection time (conftest pytest_sessionstart) so a stale or missing + bundle fails the whole run up front, naming the bundle age, instead of failing + every test individually.""" + mode = parse_fixture_mode(mode_raw) + match mode: + case InvalidFixtureMode(value=value): + return f"E2E_FIXTURE_MODE={value!r} is not one of {', '.join(FIXTURE_MODES)}" + case "live" | "record": + return None + case "replay": + freshness = check_freshness(bundle_dir, now=now) + match freshness: + case FreshBundle(): + return None + case StaleBundle(recorded_at=recorded_at, age=age, limit=limit): + return ( + f"fixture bundle at {bundle_dir} is stale: recorded {recorded_at.isoformat()}, " + f"age {format_age(age)} exceeds the {limit.days}-day limit; " + "re-record with E2E_FIXTURE_MODE=record" + ) + case UnreadableBundle(reason=reason): + return f"E2E_FIXTURE_MODE=replay cannot use bundle at {bundle_dir}: {reason}" + case _: + assert_never(freshness) + case _: + assert_never(mode) + + +def fixture_report_lines(mode_raw: str, bundle_dir: Path, *, now: datetime) -> list[str]: + """pytest report-header lines; empty in live mode so an unset + E2E_FIXTURE_MODE keeps today's output byte-identical.""" + mode = parse_fixture_mode(mode_raw) + match mode: + case InvalidFixtureMode() | "live": + return [] + case "record": + return [f"e2e fixture mode: record -> {bundle_dir}"] + case "replay": + freshness = check_freshness(bundle_dir, now=now) + match freshness: + case FreshBundle(manifest=manifest): + return [ + f"e2e fixture mode: replay <- {bundle_dir} " + f"(recorded {manifest.recorded_at.isoformat()}, harness {manifest.harness_version})" + ] + case StaleBundle() | UnreadableBundle(): + return [f"e2e fixture mode: replay <- {bundle_dir}"] + case _: + assert_never(freshness) + case _: + assert_never(mode) diff --git a/tests/e2e/fixture_transport.py b/tests/e2e/fixture_transport.py deleted file mode 100644 index ce4eec701ca..00000000000 --- a/tests/e2e/fixture_transport.py +++ /dev/null @@ -1,724 +0,0 @@ -"""Record/replay transports behind the same ``Transport`` protocol (LIT-5729). - -``RecordingTransport`` decorates the live transport: every call passes through -unchanged and its request/response pair is appended to the fixture bundle. -``ReplayTransport`` implements the protocol from a recorded bundle alone: no -HTTP, no proxy, no provider spend. Because both fulfil ``Transport``, no test -or client changes shape; ``build_proxy_client`` picks the transport from -``E2E_FIXTURE_MODE`` (live | record | replay, default live). - -Replay matches each call by test node id and canonical content key -(fixture_canonical.py, LIT-5741): volatile headers, credential fields, unique -markers, generated ids, and timestamps are canonicalized out before hashing, so -matching is order-independent across distinct keys, FIFO within a key, and a -miss fails hard (``ReplayMiss``) printing the computed key and the closest -recorded key without ever falling through to a live call. Streaming chunk -fidelity is LIT-5742; scoping record/replay to provider-bound traffic is -LIT-5745. -""" - -from __future__ import annotations - -import difflib -import functools -import hashlib -import os -from collections import deque -from dataclasses import dataclass, field -from datetime import datetime -from itertools import islice -from pathlib import Path -from typing import Final, Literal, assert_never - -from pydantic import BaseModel, JsonValue - -from e2e_http import AuthHeaders, BinaryStream, ProbeResult, Result, StreamingResponse -from fixture_bundle import ( - BundleRecorder, - FreshBundle, - Interaction, - LoadedBundle, - RecordedBinary, - RecordedProbe, - RecordedRequest, - RecordedResponse, - RecordedResult, - RecordedStreaming, - StaleBundle, - UnreadableBundle, - UnsafeBundleDir, - check_freshness, - format_age, - from_result, - interaction_filename, - load_bundle, - prepare_bundle, - slug_for_test, - to_json_value, - to_result, -) -from fixture_canonical import CanonicalRequest, canonicalize, is_secret_field -from transport import Transport - -type FixtureMode = Literal["live", "record", "replay"] - -FIXTURE_MODES: Final[tuple[FixtureMode, ...]] = ("live", "record", "replay") - -SESSION_TEST_KEY: Final = "session" - -REDACTED_HEADER_NAMES: Final[frozenset[str]] = frozenset({"authorization", "x-litellm-api-key"}) -REDACTED_VALUE: Final = "" - - -@dataclass(frozen=True, slots=True) -class InvalidFixtureMode: - value: str - - -def parse_fixture_mode(raw: str) -> FixtureMode | InvalidFixtureMode: - normalized = raw.strip().lower() or "live" - match normalized: - case "live" | "record" | "replay": - return normalized - case _: - return InvalidFixtureMode(value=raw) - - -def current_test_key() -> str: - """The pytest node id of the running test, from the PYTEST_CURRENT_TEST env - var pytest maintains (`` (setup|call|teardown)``); ``session`` for - calls outside any test (e.g. session-finish cleanup).""" - raw = os.environ.get("PYTEST_CURRENT_TEST", "") - if not raw: - return SESSION_TEST_KEY - return raw.rsplit(" (", 1)[0] - - -class ReplayMiss(AssertionError): - """Replay had no recorded interaction for a call the suite made. The test - drifted from the bundle (or the bundle from the suite): re-record.""" - - -_marker_ordinals: Final[dict[str, int]] = {} - - -def deterministic_marker() -> str: - """Stable stand-in for uuid-based unique markers in record and replay modes: - the Nth marker of a test is a pure function of the test's node id and N, so a - replay run regenerates exactly the model names, prompts, and tags the record - run sent and every recorded poll response still satisfies its predicate.""" - test_key = current_test_key() - ordinal = _marker_ordinals.get(test_key, 0) - _marker_ordinals[test_key] = ordinal + 1 - return hashlib.sha1(f"{test_key}#{ordinal}".encode()).hexdigest()[:12] - - -def _dump_flat(model: BaseModel | None) -> dict[str, str]: - if model is None: - return {} - dumped: dict[str, object] = model.model_dump(by_alias=True, exclude_none=True) - return {key: str(value) for key, value in dumped.items()} - - -def _redact(headers: dict[str, str]) -> dict[str, str]: - return { - name: REDACTED_VALUE if name.lower() in REDACTED_HEADER_NAMES else value - for name, value in headers.items() - } - - -def _redact_secret_fields(value: JsonValue) -> JsonValue: - match value: - case dict(): - return { - key: REDACTED_VALUE - if is_secret_field(key) and item is not None - else _redact_secret_fields(item) - for key, item in value.items() - } - case list(): - return [_redact_secret_fields(item) for item in value] - case _: - return value - - -def _redact_flat(fields: dict[str, str]) -> dict[str, str]: - return { - key: REDACTED_VALUE if is_secret_field(key) else value for key, value in fields.items() - } - - -def recorded_request( - method: str, - path: str, - *, - headers: BaseModel, - body: BaseModel | None = None, - params: BaseModel | None = None, - form: BaseModel | None = None, - file_name: str | None = None, - file_content: bytes | None = None, -) -> RecordedRequest: - return RecordedRequest( - method=method, - path=path, - headers=_redact(_dump_flat(headers)), - params=_redact_flat(_dump_flat(params)), - body=None if body is None else _redact_secret_fields(to_json_value(body)), - form=None if form is None else _redact_flat(_dump_flat(form)), - file_name=file_name, - file_sha256=None if file_content is None else hashlib.sha256(file_content).hexdigest(), - file_bytes=None if file_content is None else len(file_content), - ) - - -@dataclass(frozen=True, slots=True) -class RecordingTransport: - """Decorator over the live transport: forwards every call and appends the - interaction to the bundle, so a green live run leaves behind exactly the - traffic replay needs.""" - - inner: Transport - recorder: BundleRecorder - - def _record(self, request: RecordedRequest, response: RecordedResponse) -> None: - self.recorder.record(test_key=current_test_key(), request=request, response=response) - - def bearer(self, key: str) -> AuthHeaders: - return self.inner.bearer(key) - - @property - def master(self) -> AuthHeaders: - return self.inner.master - - def post[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - result = self.inner.post(path, headers=headers, json=json, response_type=response_type) - self._record(recorded_request("post", path, headers=headers, body=json), from_result(result)) - return result - - def get[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - params: BaseModel, - response_type: type[R], - timeout: float | None = None, - ) -> Result[R]: - result = self.inner.get( - path, headers=headers, params=params, response_type=response_type, timeout=timeout - ) - self._record(recorded_request("get", path, headers=headers, params=params), from_result(result)) - return result - - def delete[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - json: BaseModel, - response_type: type[R], - params: BaseModel | None = None, - ) -> Result[R]: - result = self.inner.delete( - path, headers=headers, json=json, response_type=response_type, params=params - ) - self._record( - recorded_request("delete", path, headers=headers, body=json, params=params), - from_result(result), - ) - return result - - def patch[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - result = self.inner.patch(path, headers=headers, json=json, response_type=response_type) - self._record(recorded_request("patch", path, headers=headers, body=json), from_result(result)) - return result - - def put[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - result = self.inner.put(path, headers=headers, json=json, response_type=response_type) - self._record(recorded_request("put", path, headers=headers, body=json), from_result(result)) - return result - - def stream(self, path: str, *, headers: BaseModel, json: BaseModel) -> StreamingResponse: - response = self.inner.stream(path, headers=headers, json=json) - self._record( - recorded_request("stream", path, headers=headers, body=json), - RecordedStreaming(payload=response), - ) - return response - - def stream_binary( - self, path: str, *, headers: BaseModel, json: BaseModel, chunk_size: int = 8192 - ) -> BinaryStream: - response = self.inner.stream_binary(path, headers=headers, json=json, chunk_size=chunk_size) - self._record( - recorded_request("stream_binary", path, headers=headers, body=json), - RecordedBinary(payload=response), - ) - return response - - def send( - self, - path: str, - *, - headers: BaseModel, - json: BaseModel, - params: BaseModel | None = None, - stream: bool = False, - ) -> StreamingResponse: - response = self.inner.send(path, headers=headers, json=json, params=params, stream=stream) - self._record( - recorded_request("send", path, headers=headers, body=json, params=params), - RecordedStreaming(payload=response), - ) - return response - - def probe(self, path: str, *, params: BaseModel) -> ProbeResult: - response = self.inner.probe(path, params=params) - self._record( - recorded_request("probe", path, headers=self.master, params=params), - RecordedProbe(payload=response), - ) - return response - - def upload[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - form: BaseModel, - filename: str, - content: bytes, - file_content_type: str = "application/jsonl", - file_field: str = "file", - params: BaseModel | None = None, - response_type: type[R], - ) -> Result[R]: - result = self.inner.upload( - path, - headers=headers, - form=form, - filename=filename, - content=content, - file_content_type=file_content_type, - file_field=file_field, - params=params, - response_type=response_type, - ) - self._record( - recorded_request( - "upload", - path, - headers=headers, - params=params, - form=form, - file_name=filename, - file_content=content, - ), - from_result(result), - ) - return result - - def download(self, path: str, *, headers: BaseModel) -> StreamingResponse: - response = self.inner.download(path, headers=headers) - self._record( - recorded_request("download", path, headers=headers), - RecordedStreaming(payload=response), - ) - return response - - -def _build_pool(recorded: tuple[Interaction, ...]) -> dict[str, deque[Interaction]]: - keys: Final = tuple(canonicalize(interaction.request).key for interaction in recorded) - return { - key: deque( - interaction - for candidate_key, interaction in zip(keys, recorded, strict=True) - if candidate_key == key - ) - for key in dict.fromkeys(keys) - } - - -def _closest_recorded( - canonical: CanonicalRequest, recorded: tuple[Interaction, ...] -) -> tuple[CanonicalRequest, str]: - candidates: Final = tuple(canonicalize(interaction.request) for interaction in recorded) - ratios: Final = tuple( - difflib.SequenceMatcher( - None, f"{canonical.method} {canonical.path}\n{canonical.content}", - f"{candidate.method} {candidate.path}\n{candidate.content}", - ).ratio() - for candidate in candidates - ) - best: Final = max(range(len(candidates)), key=lambda index: ratios[index]) - return candidates[best], interaction_filename(best, recorded[best].request) - - -def _miss_message(test_key: str, slug: str, canonical: CanonicalRequest, bundle: LoadedBundle) -> str: - recorded: Final = bundle.interactions.get(slug, ()) - if not recorded: - return ( - f"replay miss for {test_key}: computed key {canonical.key} but nothing is recorded " - f"under {slug}; re-record with E2E_FIXTURE_MODE=record" - ) - closest, closest_file = _closest_recorded(canonical, recorded) - diff: Final = "\n".join( - islice( - difflib.unified_diff( - closest.pretty_content().splitlines(), - canonical.pretty_content().splitlines(), - fromfile=f"closest recorded ({closest_file})", - tofile="test made", - lineterm="", - ), - 60, - ) - ) - return ( - f"replay miss for {test_key}: no recorded interaction matches key {canonical.key}; " - f"closest recorded key is {closest.key} ({closest_file})\n{diff}\n" - "re-record with E2E_FIXTURE_MODE=record" - ) - - -@dataclass(slots=True) -class ReplaySource: - """One shared pool per test over a loaded bundle, so every client built in - the session consumes the same recorded interactions. Every pool is built - once at construction and per-key consumption is a single atomic deque pop, - so concurrent replay calls never race. Calls match by canonical content - key: order-independent across distinct keys (concurrent tests interleave - calls nondeterministically), FIFO within one key (a poll loop replays its - recorded responses in recorded order).""" - - bundle: LoadedBundle - _pools: dict[str, dict[str, deque[Interaction]]] = field(init=False) - - def __post_init__(self) -> None: - self._pools = { - slug: _build_pool(recorded) for slug, recorded in self.bundle.interactions.items() - } - - def _pool(self, slug: str) -> dict[str, deque[Interaction]]: - return self._pools.get(slug, {}) - - def next_interaction(self, request: RecordedRequest) -> Interaction: - test_key: Final = current_test_key() - slug: Final = slug_for_test(test_key) - pool: Final = self._pool(slug) - canonical: Final = canonicalize(request) - queue: Final = pool.get(canonical.key) - if queue is None: - raise ReplayMiss(_miss_message(test_key, slug, canonical, self.bundle)) - try: - return queue.popleft() - except IndexError: - raise ReplayMiss( - f"replay exhausted for {test_key}: every recorded interaction for key " - f"{canonical.key} is already consumed; re-record with E2E_FIXTURE_MODE=record" - ) from None - - def leftover_error(self, test_key: str) -> str | None: - """Non-None when the test consumed fewer interactions than were recorded, - meaning a passing replay proved less than the bundle claims.""" - slug: Final = slug_for_test(test_key) - recorded: Final = self.bundle.interactions.get(slug, ()) - if not recorded: - return None - leftover: Final = tuple( - interaction for queue in self._pool(slug).values() for interaction in queue - ) - if not leftover: - return None - return ( - f"replay incomplete for {test_key}: {len(leftover)} of {len(recorded)} recorded " - f"interactions never consumed, e.g. {canonicalize(leftover[0].request).key}; " - "re-record with E2E_FIXTURE_MODE=record" - ) - - -def _expect_result(interaction: Interaction) -> RecordedResult: - match interaction.response: - case RecordedResult() as recorded: - return recorded - case RecordedStreaming() | RecordedBinary() | RecordedProbe(): - raise ReplayMiss( - f"recorded {interaction.request.method} {interaction.request.path} is not a typed result" - ) - - -def _expect_streaming(interaction: Interaction) -> StreamingResponse: - match interaction.response: - case RecordedStreaming(payload=payload): - return payload - case RecordedResult() | RecordedBinary() | RecordedProbe(): - raise ReplayMiss( - f"recorded {interaction.request.method} {interaction.request.path} is not a streaming response" - ) - - -@dataclass(frozen=True, slots=True) -class ReplayTransport: - """A ``Transport`` served entirely from a recorded bundle: never opens a - connection, so a replay run cannot bill a provider.""" - - source: ReplaySource - master_key: str - - def bearer(self, key: str) -> AuthHeaders: - return AuthHeaders(authorization=f"Bearer {key}") - - @property - def master(self) -> AuthHeaders: - return self.bearer(self.master_key) - - def post[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - return to_result( - _expect_result( - self.source.next_interaction(recorded_request("post", path, headers=headers, body=json)) - ), - response_type, - ) - - def get[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - params: BaseModel, - response_type: type[R], - timeout: float | None = None, - ) -> Result[R]: - return to_result( - _expect_result( - self.source.next_interaction(recorded_request("get", path, headers=headers, params=params)) - ), - response_type, - ) - - def delete[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - json: BaseModel, - response_type: type[R], - params: BaseModel | None = None, - ) -> Result[R]: - return to_result( - _expect_result( - self.source.next_interaction( - recorded_request("delete", path, headers=headers, body=json, params=params) - ) - ), - response_type, - ) - - def patch[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - return to_result( - _expect_result( - self.source.next_interaction(recorded_request("patch", path, headers=headers, body=json)) - ), - response_type, - ) - - def put[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - return to_result( - _expect_result( - self.source.next_interaction(recorded_request("put", path, headers=headers, body=json)) - ), - response_type, - ) - - def stream(self, path: str, *, headers: BaseModel, json: BaseModel) -> StreamingResponse: - return _expect_streaming( - self.source.next_interaction(recorded_request("stream", path, headers=headers, body=json)) - ) - - def stream_binary( - self, path: str, *, headers: BaseModel, json: BaseModel, chunk_size: int = 8192 - ) -> BinaryStream: - interaction = self.source.next_interaction( - recorded_request("stream_binary", path, headers=headers, body=json) - ) - match interaction.response: - case RecordedBinary(payload=payload): - return payload - case RecordedResult() | RecordedStreaming() | RecordedProbe(): - raise ReplayMiss( - f"recorded stream_binary {interaction.request.path} is not a binary stream" - ) - - def send( - self, - path: str, - *, - headers: BaseModel, - json: BaseModel, - params: BaseModel | None = None, - stream: bool = False, - ) -> StreamingResponse: - return _expect_streaming( - self.source.next_interaction( - recorded_request("send", path, headers=headers, body=json, params=params) - ) - ) - - def probe(self, path: str, *, params: BaseModel) -> ProbeResult: - interaction = self.source.next_interaction( - recorded_request("probe", path, headers=self.master, params=params) - ) - match interaction.response: - case RecordedProbe(payload=payload): - return payload - case RecordedResult() | RecordedStreaming() | RecordedBinary(): - raise ReplayMiss(f"recorded probe {interaction.request.path} is not a probe result") - - def upload[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - form: BaseModel, - filename: str, - content: bytes, - file_content_type: str = "application/jsonl", - file_field: str = "file", - params: BaseModel | None = None, - response_type: type[R], - ) -> Result[R]: - return to_result( - _expect_result( - self.source.next_interaction( - recorded_request( - "upload", - path, - headers=headers, - params=params, - form=form, - file_name=filename, - file_content=content, - ) - ) - ), - response_type, - ) - - def download(self, path: str, *, headers: BaseModel) -> StreamingResponse: - return _expect_streaming( - self.source.next_interaction(recorded_request("download", path, headers=headers)) - ) - - -@functools.lru_cache(maxsize=8) -def _shared_recorder(root: Path) -> BundleRecorder: - prepared = prepare_bundle(root) - if isinstance(prepared, UnsafeBundleDir): - raise ValueError(f"E2E_FIXTURE_DIR {prepared.path} {prepared.reason}") - return prepared - - -@functools.lru_cache(maxsize=8) -def _shared_replay_source(root: Path) -> ReplaySource: - loaded = load_bundle(root) - if isinstance(loaded, UnreadableBundle): - raise ValueError(f"cannot replay from {root}: {loaded.reason}") - return ReplaySource(bundle=loaded) - - -def replay_leftover_error(*, mode_raw: str, bundle_dir: Path, test_key: str) -> str | None: - """Teardown-time completeness check: in replay mode a passed test with - unconsumed recorded interactions must fail instead of passing against a - recording it no longer matches. Inert in every other mode.""" - if parse_fixture_mode(mode_raw) != "replay": - return None - return _shared_replay_source(bundle_dir).leftover_error(test_key) - - -def select_transport( - live: Transport, *, mode_raw: str, bundle_dir: Path, master_key: str -) -> Transport: - """The one seam every client build goes through: wraps (record), replaces - (replay), or passes through (live) the transport per E2E_FIXTURE_MODE. The - recorder and replay cursors are process-wide singletons per bundle dir, so - every client in a session shares one bundle and one recorded sequence.""" - mode = parse_fixture_mode(mode_raw) - match mode: - case InvalidFixtureMode(value=value): - raise ValueError(f"E2E_FIXTURE_MODE={value!r} is not one of {', '.join(FIXTURE_MODES)}") - case "live": - return live - case "record": - return RecordingTransport(inner=live, recorder=_shared_recorder(bundle_dir)) - case "replay": - return ReplayTransport(source=_shared_replay_source(bundle_dir), master_key=master_key) - case _: - assert_never(mode) - - -def fixture_mode_collection_error(mode_raw: str, bundle_dir: Path, *, now: datetime) -> str | None: - """Session-abort reason for a fixture-mode setup that can never work, or None. - Called at collection time (conftest pytest_sessionstart) so a stale or missing - bundle fails the whole run up front, naming the bundle age, instead of failing - every test individually.""" - mode = parse_fixture_mode(mode_raw) - match mode: - case InvalidFixtureMode(value=value): - return f"E2E_FIXTURE_MODE={value!r} is not one of {', '.join(FIXTURE_MODES)}" - case "live" | "record": - return None - case "replay": - freshness = check_freshness(bundle_dir, now=now) - match freshness: - case FreshBundle(): - return None - case StaleBundle(recorded_at=recorded_at, age=age, limit=limit): - return ( - f"fixture bundle at {bundle_dir} is stale: recorded {recorded_at.isoformat()}, " - f"age {format_age(age)} exceeds the {limit.days}-day limit; " - "re-record with E2E_FIXTURE_MODE=record" - ) - case UnreadableBundle(reason=reason): - return f"E2E_FIXTURE_MODE=replay cannot use bundle at {bundle_dir}: {reason}" - case _: - assert_never(freshness) - case _: - assert_never(mode) - - -def fixture_report_lines(mode_raw: str, bundle_dir: Path, *, now: datetime) -> list[str]: - """pytest report-header lines; empty in live mode so an unset - E2E_FIXTURE_MODE keeps today's output byte-identical.""" - mode = parse_fixture_mode(mode_raw) - match mode: - case InvalidFixtureMode() | "live": - return [] - case "record": - return [f"e2e fixture mode: record -> {bundle_dir}"] - case "replay": - freshness = check_freshness(bundle_dir, now=now) - match freshness: - case FreshBundle(manifest=manifest): - return [ - f"e2e fixture mode: replay <- {bundle_dir} " - f"(recorded {manifest.recorded_at.isoformat()}, harness {manifest.harness_version})" - ] - case StaleBundle() | UnreadableBundle(): - return [f"e2e fixture mode: replay <- {bundle_dir}"] - case _: - assert_never(freshness) - case _: - assert_never(mode) diff --git a/tests/e2e/provider_edge.py b/tests/e2e/provider_edge.py new file mode 100644 index 00000000000..ab0791e6b74 --- /dev/null +++ b/tests/e2e/provider_edge.py @@ -0,0 +1,546 @@ +"""Provider-edge record/replay server for e2e runs (LIT-5745). + +Record and replay scope to provider-bound traffic only: the proxy boots for +real, tests hit it for real, and only the hop from the proxy to the provider +is recorded or served from a bundle. Suites opt in per deployment by pointing +``litellm_params.api_base`` at ``provider_edge_api_base(mount)``, which is an +in-process HTTP server mounting each supported provider under a path prefix +(``http://127.0.0.1:/openai`` forwards to ``https://api.openai.com``). +In record mode the edge relays each request verbatim, stores the interaction, +and serves the proxy the same filtered response replay will serve later; in +replay mode it serves straight from the bundle and never opens a provider +connection, so a green replay run with a fake provider key proves the entire +proxy pipeline (auth, routing, spend logging) without provider spend. + +Request identity reuses fixture_canonical.py: interactions match by canonical +content key, order-independent across keys and FIFO within one. Edge requests +store no headers at all: SDK telemetry headers vary run to run and credential +headers must never touch disk. An unmatched replay call returns HTTP +``REPLAY_MISS_STATUS`` naming the closest recorded interaction, which the +proxy relays as a provider error the failing test surfaces. + +v1 limits: only the mounts in ``EDGE_MOUNTS`` (SigV4 providers like Bedrock +sign the Host header, so a forwarding edge breaks their signatures), JSON and +opaque single-part bodies (multipart boundaries are random per request), +streaming fidelity is LIT-5742, and CI wiring is LIT-5748. Suites that do not +wire the edge keep hitting providers live in every mode. +""" + +from __future__ import annotations + +import base64 +import difflib +import functools +import hashlib +import threading +from collections import deque +from collections.abc import Mapping +from dataclasses import dataclass, field +from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer +from itertools import islice +from pathlib import Path +from types import MappingProxyType +from typing import Final, Literal, assert_never +from urllib.parse import parse_qsl, urlsplit + +from pydantic import JsonValue, TypeAdapter + +from e2e_http import NetworkError, RawResponse, forward +from fixture_bundle import ( + BundleRecorder, + Interaction, + LoadedBundle, + RecordedHttpResponse, + RecordedRequest, + UnreadableBundle, + UnsafeBundleDir, + interaction_filename, + load_bundle, + prepare_bundle, + slug_for_test, +) +from fixture_canonical import CanonicalRequest, canonical_string, canonicalize +from fixture_mode import ( + FIXTURE_MODES, + InvalidFixtureMode, + ReplayMiss, + current_test_key, + parse_fixture_mode, +) + +EDGE_MOUNTS: Final[Mapping[str, str]] = MappingProxyType( + { + "openai": "https://api.openai.com", + "anthropic": "https://api.anthropic.com", + } +) + +REPLAY_MISS_STATUS: Final = 599 + +_HOP_BY_HOP_HEADERS: Final[frozenset[str]] = frozenset( + { + "connection", + "keep-alive", + "proxy-authenticate", + "proxy-authorization", + "te", + "trailers", + "transfer-encoding", + "upgrade", + } +) +_REQUEST_DROPPED_HEADERS: Final[frozenset[str]] = _HOP_BY_HOP_HEADERS | { + "host", + "content-length", + "accept-encoding", +} +_RESPONSE_DROPPED_HEADERS: Final[frozenset[str]] = _HOP_BY_HOP_HEADERS | { + "content-encoding", + "content-length", + "set-cookie", +} + +_JSON: Final[TypeAdapter[JsonValue]] = TypeAdapter(JsonValue) + + +def _edge_request(method: str, path: str, query: str, body: bytes | None) -> RecordedRequest: + """The identity replay matches on: the edge path (mount included), the query + as params, and the body as parsed JSON, or as a canonicalized content digest + when it is not JSON so opaque uploads still match across runs.""" + params: Final = dict(parse_qsl(query, keep_blank_values=True)) + if not body: + return RecordedRequest(method=method.lower(), path=path, headers={}, params=params) + decoded: Final = body.decode("utf-8", errors="replace") + try: + parsed: Final[JsonValue] = _JSON.validate_json(decoded) + except ValueError: + return RecordedRequest( + method=method.lower(), + path=path, + headers={}, + params=params, + file_sha256=hashlib.sha256(canonical_string(decoded).encode()).hexdigest(), + file_bytes=len(body), + ) + return RecordedRequest(method=method.lower(), path=path, headers={}, params=params, body=parsed) + + +def _build_pool(recorded: tuple[Interaction, ...]) -> dict[str, deque[Interaction]]: + keys: Final = tuple(canonicalize(interaction.request).key for interaction in recorded) + return { + key: deque( + interaction + for candidate_key, interaction in zip(keys, recorded, strict=True) + if candidate_key == key + ) + for key in dict.fromkeys(keys) + } + + +def _closest_recorded( + canonical: CanonicalRequest, recorded: tuple[Interaction, ...] +) -> tuple[CanonicalRequest, str]: + candidates: Final = tuple(canonicalize(interaction.request) for interaction in recorded) + ratios: Final = tuple( + difflib.SequenceMatcher( + None, f"{canonical.method} {canonical.path}\n{canonical.content}", + f"{candidate.method} {candidate.path}\n{candidate.content}", + ).ratio() + for candidate in candidates + ) + best: Final = max(range(len(candidates)), key=lambda index: ratios[index]) + return candidates[best], interaction_filename(best, recorded[best].request) + + +def _miss_message(test_key: str, slug: str, canonical: CanonicalRequest, bundle: LoadedBundle) -> str: + recorded: Final = bundle.interactions.get(slug, ()) + if not recorded: + return ( + f"replay miss for {test_key}: computed key {canonical.key} but nothing is recorded " + f"under {slug}; re-record with E2E_FIXTURE_MODE=record" + ) + closest, closest_file = _closest_recorded(canonical, recorded) + diff: Final = "\n".join( + islice( + difflib.unified_diff( + closest.pretty_content().splitlines(), + canonical.pretty_content().splitlines(), + fromfile=f"closest recorded ({closest_file})", + tofile="test made", + lineterm="", + ), + 60, + ) + ) + return ( + f"replay miss for {test_key}: no recorded interaction matches key {canonical.key}; " + f"closest recorded key is {closest.key} ({closest_file})\n{diff}\n" + "re-record with E2E_FIXTURE_MODE=record" + ) + + +@dataclass(slots=True) +class ReplaySource: + """One shared pool per test over a loaded bundle, so every provider call the + proxy makes in the session consumes from the same recorded interactions. + Every pool is built once at construction and per-key consumption is a single + atomic deque pop, so concurrent replay calls never race. Calls match by + canonical content key: order-independent across distinct keys (concurrent + tests interleave calls nondeterministically), FIFO within one key (a retry + or poll loop replays its recorded responses in recorded order).""" + + bundle: LoadedBundle + _pools: dict[str, dict[str, deque[Interaction]]] = field(init=False) + + def __post_init__(self) -> None: + self._pools = { + slug: _build_pool(recorded) for slug, recorded in self.bundle.interactions.items() + } + + def _pool(self, slug: str) -> dict[str, deque[Interaction]]: + return self._pools.get(slug, {}) + + def next_interaction(self, request: RecordedRequest) -> Interaction: + test_key: Final = current_test_key() + slug: Final = slug_for_test(test_key) + pool: Final = self._pool(slug) + canonical: Final = canonicalize(request) + queue: Final = pool.get(canonical.key) + if queue is None: + raise ReplayMiss(_miss_message(test_key, slug, canonical, self.bundle)) + try: + return queue.popleft() + except IndexError: + raise ReplayMiss( + f"replay exhausted for {test_key}: every recorded interaction for key " + f"{canonical.key} is already consumed; re-record with E2E_FIXTURE_MODE=record" + ) from None + + def leftover_error(self, test_key: str) -> str | None: + """Non-None when the test consumed fewer interactions than were recorded, + meaning a passing replay proved less than the bundle claims.""" + slug: Final = slug_for_test(test_key) + recorded: Final = self.bundle.interactions.get(slug, ()) + if not recorded: + return None + leftover: Final = tuple( + interaction for queue in self._pool(slug).values() for interaction in queue + ) + if not leftover: + return None + return ( + f"replay incomplete for {test_key}: {len(leftover)} of {len(recorded)} recorded " + f"interactions never consumed, e.g. {canonicalize(leftover[0].request).key}; " + "re-record with E2E_FIXTURE_MODE=record" + ) + + +@dataclass(frozen=True, slots=True) +class RecordEdge: + """Record backend: forward to the provider, persist, serve the filtered copy. + The lock serializes recorder writes because the edge server handles requests + on concurrent threads.""" + + recorder: BundleRecorder + lock: threading.Lock + + +@dataclass(frozen=True, slots=True) +class ReplayEdge: + source: ReplaySource + + +type EdgeBackend = RecordEdge | ReplayEdge + + +@dataclass(frozen=True, slots=True) +class EdgeReply: + status_code: int + headers: dict[str, str] + body: bytes + + +def _text_reply(status_code: int, message: str) -> EdgeReply: + return EdgeReply( + status_code=status_code, + headers={"content-type": "text/plain; charset=utf-8"}, + body=message.encode(), + ) + + +def _reply_from_recorded(response: RecordedHttpResponse) -> EdgeReply: + return EdgeReply( + status_code=response.status_code, + headers=dict(response.headers), + body=base64.b64decode(response.body_b64), + ) + + +def _recorded_response(outcome: RawResponse | NetworkError) -> RecordedHttpResponse: + match outcome: + case RawResponse(status_code=status_code, headers=headers, body=body): + return RecordedHttpResponse( + status_code=status_code, + headers={ + name: value + for name, value in headers.items() + if name not in _RESPONSE_DROPPED_HEADERS + }, + body_b64=base64.b64encode(body).decode("ascii"), + ) + case NetworkError(message=message): + return RecordedHttpResponse( + status_code=502, + headers={"content-type": "text/plain; charset=utf-8"}, + body_b64=base64.b64encode( + f"provider edge could not reach the provider: {message}".encode() + ).decode("ascii"), + ) + + +def _upstream_url(upstream_base: str, upstream_path: str, query: str) -> str: + url: Final = f"{upstream_base}/{upstream_path}" + return f"{url}?{query}" if query else url + + +def _handle_record( + backend: RecordEdge, + request: RecordedRequest, + *, + method: str, + url: str, + headers: Mapping[str, str], + body: bytes | None, + timeout: float, +) -> EdgeReply: + forwarded: Final = { + name: value for name, value in headers.items() if name.lower() not in _REQUEST_DROPPED_HEADERS + } + outcome: Final = forward(method, url, headers=forwarded, body=body, timeout=timeout) + response: Final = _recorded_response(outcome) + with backend.lock: + backend.recorder.record(test_key=current_test_key(), request=request, response=response) + return _reply_from_recorded(response) + + +def _handle_replay(source: ReplaySource, request: RecordedRequest) -> EdgeReply: + try: + interaction: Final = source.next_interaction(request) + except ReplayMiss as miss: + return _text_reply(REPLAY_MISS_STATUS, str(miss)) + return _reply_from_recorded(interaction.response) + + +def handle_edge_request( + backend: EdgeBackend, + mounts: Mapping[str, str], + method: str, + raw_path: str, + headers: Mapping[str, str], + body: bytes | None, + *, + timeout: float, +) -> EdgeReply: + """The edge's pure core, one HTTP exchange in and out: resolve the mount + prefix, then record (forward + persist) or replay (serve from the bundle). + Socket-free so unit tests exercise every branch without a server.""" + split: Final = urlsplit(raw_path) + mount, _, upstream_path = split.path.lstrip("/").partition("/") + upstream_base: Final = mounts.get(mount) + if upstream_base is None: + return _text_reply( + 404, f"unknown provider mount {mount!r}; known mounts: {', '.join(sorted(mounts))}" + ) + request: Final = _edge_request(method, split.path, split.query, body) + match backend: + case RecordEdge(): + return _handle_record( + backend, + request, + method=method, + url=_upstream_url(upstream_base, upstream_path, split.query), + headers=headers, + body=body, + timeout=timeout, + ) + case ReplayEdge(source=source): + return _handle_replay(source, request) + case _: + assert_never(backend) + + +class _EdgeHandler(BaseHTTPRequestHandler): + protocol_version = "HTTP/1.1" + + def do_GET(self) -> None: + self._handle() + + def do_POST(self) -> None: + self._handle() + + def do_PUT(self) -> None: + self._handle() + + def do_PATCH(self) -> None: + self._handle() + + def do_DELETE(self) -> None: + self._handle() + + def _handle(self) -> None: + edge_server: Final = self.server + assert isinstance(edge_server, _EdgeHTTPServer) + length: Final = int(self.headers.get("content-length") or "0") + body: Final = self.rfile.read(length) if length else None + reply: Final = handle_edge_request( + edge_server.backend, + edge_server.mounts, + self.command, + self.path, + {name.lower(): value for name, value in self.headers.items()}, + body, + timeout=edge_server.forward_timeout, + ) + self.send_response(reply.status_code) + for name, value in reply.headers.items(): + self.send_header(name, value) + self.send_header("content-length", str(len(reply.body))) + self.end_headers() + self.wfile.write(reply.body) + + def log_message(self, format: str, *args: object) -> None: + """Silence the per-request stderr line BaseHTTPRequestHandler emits.""" + + +class _EdgeHTTPServer(ThreadingHTTPServer): + daemon_threads = True + + def __init__( + self, + bind: tuple[str, int], + *, + backend: EdgeBackend, + mounts: Mapping[str, str], + forward_timeout: float, + ) -> None: + super().__init__(bind, _EdgeHandler) + self.backend: Final = backend + self.mounts: Final = mounts + self.forward_timeout: Final = forward_timeout + + +@dataclass(frozen=True, slots=True) +class ProviderEdge: + port: int + advertise_host: str + + def api_base(self, mount: str) -> str: + return f"http://{self.advertise_host}:{self.port}/{mount}" + + +@dataclass(frozen=True, slots=True) +class RunningEdge: + edge: ProviderEdge + server: _EdgeHTTPServer + + def shutdown(self) -> None: + self.server.shutdown() + self.server.server_close() + + +def start_provider_edge( + backend: EdgeBackend, + *, + mounts: Mapping[str, str] = EDGE_MOUNTS, + bind_host: str = "127.0.0.1", + advertise_host: str | None = None, + forward_timeout: float = 60.0, +) -> RunningEdge: + """Boot an edge server on an OS-assigned port in a daemon thread. + ``advertise_host`` is what api_base URLs name (it differs from the bind + host when the proxy runs in a container and reaches the host machine via + a gateway address like host.docker.internal).""" + server: Final = _EdgeHTTPServer( + (bind_host, 0), backend=backend, mounts=mounts, forward_timeout=forward_timeout + ) + thread: Final = threading.Thread(target=server.serve_forever, name="e2e-provider-edge", daemon=True) + thread.start() + return RunningEdge( + edge=ProviderEdge(port=server.server_address[1], advertise_host=advertise_host or bind_host), + server=server, + ) + + +@functools.lru_cache(maxsize=8) +def _shared_recorder(root: Path) -> BundleRecorder: + prepared = prepare_bundle(root) + if isinstance(prepared, UnsafeBundleDir): + raise ValueError(f"E2E_FIXTURE_DIR {prepared.path} {prepared.reason}") + return prepared + + +@functools.lru_cache(maxsize=8) +def _shared_replay_source(root: Path) -> ReplaySource: + loaded = load_bundle(root) + if isinstance(loaded, UnreadableBundle): + raise ValueError(f"cannot replay from {root}: {loaded.reason}") + return ReplaySource(bundle=loaded) + + +@functools.lru_cache(maxsize=8) +def _shared_edge( + mode: Literal["record", "replay"], + bundle_dir: Path, + bind_host: str, + advertise_host: str, + forward_timeout: float, +) -> ProviderEdge: + backend: Final[EdgeBackend] = ( + RecordEdge(recorder=_shared_recorder(bundle_dir), lock=threading.Lock()) + if mode == "record" + else ReplayEdge(source=_shared_replay_source(bundle_dir)) + ) + return start_provider_edge( + backend, + mounts=EDGE_MOUNTS, + bind_host=bind_host, + advertise_host=advertise_host, + forward_timeout=forward_timeout, + ).edge + + +def replay_leftover_error(*, mode_raw: str, bundle_dir: Path, test_key: str) -> str | None: + """Teardown-time completeness check: in replay mode a passed test with + unconsumed recorded interactions must fail instead of passing against a + recording it no longer matches. Inert in every other mode.""" + if parse_fixture_mode(mode_raw) != "replay": + return None + return _shared_replay_source(bundle_dir).leftover_error(test_key) + + +def provider_edge_api_base( + mount: str, + *, + mode_raw: str, + bundle_dir: Path, + bind_host: str, + advertise_host: str, + forward_timeout: float = 60.0, +) -> str | None: + """The api_base a suite gives an edge-wired deployment: None in live mode + (the deployment keeps its real provider api_base) and the process-wide edge + server's mount URL in record and replay, booting the server on first use.""" + mode: Final = parse_fixture_mode(mode_raw) + match mode: + case InvalidFixtureMode(value=value): + raise ValueError(f"E2E_FIXTURE_MODE={value!r} is not one of {', '.join(FIXTURE_MODES)}") + case "live": + return None + case "record" | "replay": + if mount not in EDGE_MOUNTS: + raise ValueError( + f"unknown provider mount {mount!r}; known mounts: {', '.join(sorted(EDGE_MOUNTS))}" + ) + return _shared_edge(mode, bundle_dir, bind_host, advertise_host, forward_timeout).api_base(mount) + case _: + assert_never(mode) diff --git a/tests/e2e/proxy_client.py b/tests/e2e/proxy_client.py index 3cae337a5ff..6cdd3354bf7 100644 --- a/tests/e2e/proxy_client.py +++ b/tests/e2e/proxy_client.py @@ -65,8 +65,6 @@ from models import ( ) from e2e_config import ( CONTROL_PLANE_BASE_URL, - FIXTURE_DIR, - FIXTURE_MODE_RAW, MASTER_KEY, POLL_INTERVAL, POLL_TIMEOUT, @@ -74,7 +72,6 @@ from e2e_config import ( REQUEST_TIMEOUT, settle_propagation, ) -from fixture_transport import select_transport from transport import HttpTransport, SplitTransport, Transport RowsPredicate = Callable[[list[SpendLogRow]], bool] @@ -547,9 +544,9 @@ def build_proxy_client( pass all three together, since a caller that overrides only the data plane would leave management calls pointed at the env default. - E2E_FIXTURE_MODE wraps (record) or replaces (replay) the transport here, so - every client built from this seam records or replays without changing shape; - unset it stays the plain SplitTransport (see fixture_transport.py).""" + Test-to-proxy traffic always goes over the wire, in every E2E_FIXTURE_MODE: + record and replay scope to the proxy's provider-bound calls via the + provider edge (see provider_edge.py), never to this transport.""" split = SplitTransport( data=HttpTransport( base_url=base_url, @@ -563,12 +560,7 @@ def build_proxy_client( ), ) return ProxyClient( - transport=select_transport( - split, - mode_raw=FIXTURE_MODE_RAW, - bundle_dir=FIXTURE_DIR, - master_key=master_key, - ), + transport=split, poll_timeout=POLL_TIMEOUT, poll_interval=POLL_INTERVAL, ) diff --git a/tests/e2e/quota_management/spend_tracking/test_provider_edge_spend_e2e.py b/tests/e2e/quota_management/spend_tracking/test_provider_edge_spend_e2e.py new file mode 100644 index 00000000000..ced7c819d42 --- /dev/null +++ b/tests/e2e/quota_management/spend_tracking/test_provider_edge_spend_e2e.py @@ -0,0 +1,50 @@ +"""The provider-edge demonstrator: one spend-tracking flow wired through the +record/replay edge (LIT-5745). + +This is the reference for wiring a suite to the edge: register a deployment +whose ``api_base`` comes from ``e2e_config.provider_edge_base``, then exercise +the proxy exactly as a live test would. In live mode the base is None and the +deployment talks to the real provider; in record mode it talks through the +local edge, which forwards to the provider and captures the exchange; in +replay mode the same test drives the REAL proxy and REAL database on the +recorded provider traffic alone, so key auth, routing, and the spend-log +write path are all still under test with zero provider calls. +""" + +import pytest + +from e2e_config import CHEAP_OPENAI_MODEL, provider_edge_base +from lifecycle import ResourceManager +from models import LiteLLMParamsBody +from spend_e2e_client import SpendClient, unique_marker, unwrap + +pytestmark = pytest.mark.e2e + + +@pytest.mark.covers("quota_management.spend_tracking.chat_completions.logs_cost") +def test_edge_wired_chat_writes_nonzero_spend_row( + client: SpendClient, resources: ResourceManager, scoped_key: str +) -> None: + base = provider_edge_base("openai") + model = f"e2e-edge-openai-{unique_marker()}" + model_id = client.proxy.create_model( + model, + LiteLLMParamsBody( + model=f"openai/{CHEAP_OPENAI_MODEL}", + api_key="os.environ/OPENAI_API_KEY", + api_base=None if base is None else f"{base}/v1", + ), + ) + resources.defer(lambda: client.proxy.delete_model(model_id)) + + chat = unwrap( + client.chat(scoped_key, model, f"reply with one word {unique_marker()}", max_tokens=16) + ) + assert chat.id + + rows = client.poll_logs_for_key( + scoped_key, predicate=lambda rs: any((r.spend or 0) > 0 for r in rs) + ) + matching = [row for row in rows if row.request_id == chat.id] + assert matching, f"no SpendLogs row for request_id {chat.id}; saw {len(rows)} row(s)" + assert (matching[0].spend or 0) > 0, f"spend row for {chat.id} has zero spend" diff --git a/tests/e2e/test_fixture_bundle.py b/tests/e2e/test_fixture_bundle.py index fd4cca6451f..b49ab565e39 100644 --- a/tests/e2e/test_fixture_bundle.py +++ b/tests/e2e/test_fixture_bundle.py @@ -1,9 +1,9 @@ -"""Harness coverage for the on-disk fixture bundle format (LIT-5729). +"""Harness coverage for the on-disk fixture bundle format (LIT-5729/LIT-5745). No proxy and no ``e2e`` marker: these pin the bundle CONTRACT - the seven-day freshness gate that names the bundle's age, record mode's wipe safety (never delete a directory that is not a bundle), collision-free per-test slugs, and -lossless Result round-trips - so replay can never silently drift from what +grouped-in-order loading - so replay can never silently drift from what record wrote. """ @@ -12,18 +12,6 @@ from __future__ import annotations from datetime import datetime, timedelta, timezone from pathlib import Path -import pytest -from pydantic import BaseModel - -from e2e_http import ( - NetworkError, - RateLimitedError, - Result, - Success, - UnauthorizedError, - UnknownApiError, - ValidationError, -) from fixture_bundle import ( BUNDLE_FORMAT_VERSION, MANIFEST_FILENAME, @@ -32,28 +20,22 @@ from fixture_bundle import ( FreshBundle, LoadedBundle, Manifest, + RecordedHttpResponse, RecordedRequest, - RecordedResult, StaleBundle, UnreadableBundle, UnsafeBundleDir, check_freshness, format_age, - from_result, interaction_filename, load_bundle, prepare_bundle, slug_for_test, - to_result, ) NOW = datetime(2026, 8, 18, 12, 0, 0, tzinfo=timezone.utc) -class Payload(BaseModel): - value: str - - def write_manifest( root: Path, recorded_at: datetime, *, format_version: int = BUNDLE_FORMAT_VERSION ) -> None: @@ -74,20 +56,8 @@ def plain_request(path: str) -> RecordedRequest: return RecordedRequest(method="post", path=path, headers={}) -class TestResultRoundTrip: - @pytest.mark.parametrize( - "result", - [ - Success(status_code=201, data=Payload(value="ok")), - NetworkError(message="connection refused"), - UnauthorizedError(), - RateLimitedError(retry_after_seconds=7, body="slow down"), - ValidationError(message="bad shape"), - UnknownApiError(status_code=502, body="upstream exploded"), - ], - ) - def test_every_result_kind_survives_disk_and_back(self, result: Result[Payload]) -> None: - assert to_result(from_result(result), Payload) == result +def plain_response() -> RecordedHttpResponse: + return RecordedHttpResponse(status_code=401, headers={}, body_b64="") class TestFreshness: @@ -144,7 +114,7 @@ class TestPrepareBundle: prepared(root).record( test_key="old.py::test_old", request=plain_request("/stale"), - response=RecordedResult(kind="unauthorized"), + response=plain_response(), ) assert any(entry.is_dir() for entry in root.iterdir()) prepared(root) @@ -193,7 +163,7 @@ class TestRecordAndLoad: recorder.record( test_key=key, request=plain_request(path), - response=RecordedResult(kind="unauthorized"), + response=plain_response(), ) loaded = load_bundle(root) assert isinstance(loaded, LoadedBundle) @@ -208,7 +178,7 @@ class TestRecordAndLoad: recorder.record( test_key=key, request=plain_request(f"/{key[-3:]}"), - response=RecordedResult(kind="unauthorized"), + response=plain_response(), ) loaded = load_bundle(root) assert isinstance(loaded, LoadedBundle) diff --git a/tests/e2e/test_fixture_mode.py b/tests/e2e/test_fixture_mode.py new file mode 100644 index 00000000000..109bb9e1b11 --- /dev/null +++ b/tests/e2e/test_fixture_mode.py @@ -0,0 +1,114 @@ +"""Harness coverage for fixture-mode selection and determinism (LIT-5729/LIT-5745). + +No proxy and no ``e2e`` marker. Pins the mode parser, the deterministic +per-test marker sequence a replay run must regenerate, the collection-time +gate (including the stale message that names the bundle's age), and the pytest +report header. The provider-edge record/replay behavior itself is pinned in +test_provider_edge.py. +""" + +from __future__ import annotations + +import hashlib +from datetime import datetime, timedelta, timezone +from pathlib import Path + +import pytest + +from fixture_bundle import BUNDLE_FORMAT_VERSION, MANIFEST_FILENAME, Manifest +from fixture_mode import ( + InvalidFixtureMode, + current_test_key, + deterministic_marker, + fixture_mode_collection_error, + fixture_report_lines, + parse_fixture_mode, +) + +NOW = datetime(2026, 8, 18, 12, 0, 0, tzinfo=timezone.utc) + + +def write_manifest(root: Path, recorded_at: datetime) -> None: + root.mkdir(parents=True, exist_ok=True) + manifest = Manifest( + format_version=BUNDLE_FORMAT_VERSION, recorded_at=recorded_at, harness_version="abc1234" + ) + (root / MANIFEST_FILENAME).write_text(manifest.model_dump_json(), encoding="utf-8") + + +class TestParseFixtureMode: + @pytest.mark.parametrize( + ("raw", "expected"), + [("live", "live"), ("record", "record"), ("replay", "replay"), ("", "live"), (" REPLAY ", "replay")], + ) + def test_known_values_normalize(self, raw: str, expected: str) -> None: + assert parse_fixture_mode(raw) == expected + + def test_unknown_value_is_invalid_with_the_original_spelling(self) -> None: + assert parse_fixture_mode("cached") == InvalidFixtureMode(value="cached") + + +class TestDeterministicMarker: + def test_sequence_is_a_pure_function_of_test_and_ordinal(self) -> None: + """A replay process must regenerate exactly the markers the record + process generated, so the Nth marker of a test is pinned to a pure + function of the node id and N.""" + key = current_test_key() + assert deterministic_marker() == hashlib.sha1(f"{key}#0".encode()).hexdigest()[:12] + assert deterministic_marker() == hashlib.sha1(f"{key}#1".encode()).hexdigest()[:12] + + +class TestCurrentTestKey: + def test_names_this_test_and_strips_the_phase(self) -> None: + key = current_test_key() + assert key.endswith("TestCurrentTestKey::test_names_this_test_and_strips_the_phase") + assert "(call)" not in key + + +class TestCollectionGate: + def test_invalid_mode_names_the_value_and_the_choices(self, tmp_path: Path) -> None: + assert ( + fixture_mode_collection_error("cached", tmp_path, now=NOW) + == "E2E_FIXTURE_MODE='cached' is not one of live, record, replay" + ) + + @pytest.mark.parametrize("mode_raw", ["live", "", "record"]) + def test_live_and_record_never_block_collection(self, mode_raw: str, tmp_path: Path) -> None: + assert fixture_mode_collection_error(mode_raw, tmp_path / "missing", now=NOW) is None + + def test_replay_with_no_bundle_says_how_to_record_one(self, tmp_path: Path) -> None: + reason = fixture_mode_collection_error("replay", tmp_path / "missing", now=NOW) + assert reason is not None + assert f"no {MANIFEST_FILENAME}" in reason + assert "E2E_FIXTURE_MODE=record" in reason + + def test_stale_replay_bundle_fails_naming_its_age(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + write_manifest(root, NOW - timedelta(days=9, hours=5)) + reason = fixture_mode_collection_error("replay", root, now=NOW) + assert reason is not None + assert "age 9d5h exceeds the 7-day limit" in reason + assert "re-record with E2E_FIXTURE_MODE=record" in reason + + def test_fresh_replay_bundle_collects(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + write_manifest(root, NOW - timedelta(days=2)) + assert fixture_mode_collection_error("replay", root, now=NOW) is None + + +class TestReportHeader: + def test_live_mode_prints_nothing(self, tmp_path: Path) -> None: + assert fixture_report_lines("live", tmp_path, now=NOW) == [] + assert fixture_report_lines("", tmp_path, now=NOW) == [] + + def test_record_and_replay_name_the_bundle(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + recorded_at = NOW - timedelta(days=1) + write_manifest(root, recorded_at) + assert fixture_report_lines("record", root, now=NOW) == [ + f"e2e fixture mode: record -> {root}" + ] + replay_lines = fixture_report_lines("replay", root, now=NOW) + assert len(replay_lines) == 1 + assert "replay" in replay_lines[0] + assert recorded_at.isoformat() in replay_lines[0] diff --git a/tests/e2e/test_fixture_transport.py b/tests/e2e/test_fixture_transport.py deleted file mode 100644 index e61088d841c..00000000000 --- a/tests/e2e/test_fixture_transport.py +++ /dev/null @@ -1,676 +0,0 @@ -"""Harness coverage for the record/replay transports (LIT-5729). - -No proxy and no ``e2e`` marker. A fake in-memory ``Transport`` stands in for -the live one (dependency injection, no monkeypatching): recording must pass -every value through unchanged while writing one redacted interaction file per -call, and replay must serve identical values from the bundle alone - the -fake's call log proves nothing reaches the inner transport - failing hard -(``ReplayMiss``) on any content drift, printing the computed canonical key and -the closest recorded key (LIT-5741; the pure canonicalizer is pinned in -test_fixture_canonical.py). The collection-time gate and report header are -pinned here too, including the stale message that names the bundle's age. -""" - -from __future__ import annotations - -import hashlib -import sys -import threading -from concurrent.futures import ThreadPoolExecutor -from dataclasses import dataclass, field -from datetime import datetime, timedelta, timezone -from pathlib import Path -from uuid import uuid4 - -import pytest -from pydantic import BaseModel - -from e2e_http import ( - AuthHeaders, - BinaryStream, - ProbeResult, - Result, - StreamingResponse, - Success, -) -from fixture_bundle import ( - BUNDLE_FORMAT_VERSION, - MANIFEST_FILENAME, - BundleRecorder, - Interaction, - LoadedBundle, - Manifest, - RecordedResult, - load_bundle, - prepare_bundle, - slug_for_test, -) -from fixture_canonical import canonicalize -from fixture_transport import ( - InvalidFixtureMode, - RecordingTransport, - ReplayMiss, - ReplaySource, - ReplayTransport, - current_test_key, - deterministic_marker, - fixture_mode_collection_error, - fixture_report_lines, - parse_fixture_mode, - recorded_request, - replay_leftover_error, - select_transport, -) -from transport import Transport - -NOW = datetime(2026, 8, 18, 12, 0, 0, tzinfo=timezone.utc) - - -class Payload(BaseModel): - value: str - - -class Body(BaseModel): - prompt: str - - -class Query(BaseModel): - q: str - - -class DeployParams(BaseModel): - model: str - api_key: str | None = None - aws_secret_access_key: str | None = None - - -class DeployBody(BaseModel): - model_name: str - litellm_params: DeployParams - - -STREAMING = StreamingResponse( - status_code=200, - body="", - content_type="text/event-stream", - chunks=2, - stream_events=["one", "two"], - stream_done=True, -) -BINARY = BinaryStream(status_code=200, content_type="audio/mpeg", chunk_count=3, total_bytes=42) -PROBE = ProbeResult(status_code=200, body="alive") - - -@dataclass -class FakeTransport: - calls: list[str] = field(default_factory=list) - - def bearer(self, key: str) -> AuthHeaders: - return AuthHeaders(authorization=f"Bearer {key}") - - @property - def master(self) -> AuthHeaders: - return self.bearer("sk-fake-master") - - def _success[R: BaseModel](self, response_type: type[R]) -> Result[R]: - return Success(status_code=200, data=response_type.model_validate({"value": "live"})) - - def post[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - self.calls.append(f"post {path}") - return self._success(response_type) - - def get[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - params: BaseModel, - response_type: type[R], - timeout: float | None = None, - ) -> Result[R]: - self.calls.append(f"get {path}") - return self._success(response_type) - - def delete[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - json: BaseModel, - response_type: type[R], - params: BaseModel | None = None, - ) -> Result[R]: - self.calls.append(f"delete {path}") - return self._success(response_type) - - def patch[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - self.calls.append(f"patch {path}") - return self._success(response_type) - - def put[R: BaseModel]( - self, path: str, *, headers: BaseModel, json: BaseModel, response_type: type[R] - ) -> Result[R]: - self.calls.append(f"put {path}") - return self._success(response_type) - - def stream(self, path: str, *, headers: BaseModel, json: BaseModel) -> StreamingResponse: - self.calls.append(f"stream {path}") - return STREAMING - - def stream_binary( - self, path: str, *, headers: BaseModel, json: BaseModel, chunk_size: int = 8192 - ) -> BinaryStream: - self.calls.append(f"stream_binary {path}") - return BINARY - - def send( - self, - path: str, - *, - headers: BaseModel, - json: BaseModel, - params: BaseModel | None = None, - stream: bool = False, - ) -> StreamingResponse: - self.calls.append(f"send {path}") - return STREAMING - - def probe(self, path: str, *, params: BaseModel) -> ProbeResult: - self.calls.append(f"probe {path}") - return PROBE - - def upload[R: BaseModel]( - self, - path: str, - *, - headers: BaseModel, - form: BaseModel, - filename: str, - content: bytes, - file_content_type: str = "application/jsonl", - file_field: str = "file", - params: BaseModel | None = None, - response_type: type[R], - ) -> Result[R]: - self.calls.append(f"upload {path}") - return self._success(response_type) - - def download(self, path: str, *, headers: BaseModel) -> StreamingResponse: - self.calls.append(f"download {path}") - return STREAMING - - -def make_recorder(root: Path) -> BundleRecorder: - recorder = prepare_bundle(root) - assert isinstance(recorder, BundleRecorder) - return recorder - - -def replay_source(root: Path) -> ReplaySource: - loaded = load_bundle(root) - assert isinstance(loaded, LoadedBundle) - return ReplaySource(bundle=loaded) - - -def this_tests_files(root: Path) -> list[Path]: - slug_dir = root / slug_for_test(current_test_key()) - return sorted(slug_dir.glob("*.json")) if slug_dir.is_dir() else [] - - -def write_manifest(root: Path, recorded_at: datetime) -> None: - root.mkdir(parents=True, exist_ok=True) - manifest = Manifest( - format_version=BUNDLE_FORMAT_VERSION, recorded_at=recorded_at, harness_version="abc1234" - ) - (root / MANIFEST_FILENAME).write_text(manifest.model_dump_json(), encoding="utf-8") - - -class TestParseFixtureMode: - @pytest.mark.parametrize( - ("raw", "expected"), - [("live", "live"), ("record", "record"), ("replay", "replay"), ("", "live"), (" REPLAY ", "replay")], - ) - def test_known_values_normalize(self, raw: str, expected: str) -> None: - assert parse_fixture_mode(raw) == expected - - def test_unknown_value_is_invalid_with_the_original_spelling(self) -> None: - assert parse_fixture_mode("cached") == InvalidFixtureMode(value="cached") - - -class TestDeterministicMarker: - def test_sequence_is_a_pure_function_of_test_and_ordinal(self) -> None: - """A replay process must regenerate exactly the markers the record - process generated, so the Nth marker of a test is pinned to a pure - function of the node id and N.""" - key = current_test_key() - assert deterministic_marker() == hashlib.sha1(f"{key}#0".encode()).hexdigest()[:12] - assert deterministic_marker() == hashlib.sha1(f"{key}#1".encode()).hexdigest()[:12] - - -class TestCurrentTestKey: - def test_names_this_test_and_strips_the_phase(self) -> None: - key = current_test_key() - assert key.endswith("TestCurrentTestKey::test_names_this_test_and_strips_the_phase") - assert "(call)" not in key - - -class TestRecordingTransport: - def test_passes_the_result_through_and_writes_one_file_per_call(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - result = recording.post( - "/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload - ) - assert result == Success(status_code=200, data=Payload(value="live")) - assert fake.calls == ["post /model/new"] - files = this_tests_files(root) - assert [file.name for file in files] == ["0000-post-model-new.json"] - interaction = Interaction.model_validate_json(files[0].read_text(encoding="utf-8")) - assert interaction.request.method == "post" - assert interaction.request.path == "/model/new" - - def test_redacts_auth_header_values_in_the_recorded_request(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - headers = AuthHeaders.model_validate( - {"authorization": "Bearer sk-secret", "x-litellm-api-key": "sk-other"} - ) - recording.post("/key/generate", headers=headers, json=Body(prompt="x"), response_type=Payload) - interaction = Interaction.model_validate_json( - this_tests_files(root)[0].read_text(encoding="utf-8") - ) - assert interaction.request.headers == { - "authorization": "", - "x-litellm-api-key": "", - } - assert "sk-secret" not in this_tests_files(root)[0].read_text(encoding="utf-8") - - def test_redacts_credential_body_fields_in_the_recorded_request(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post( - "/model/new", - headers=fake.master, - json=DeployBody( - model_name="m", - litellm_params=DeployParams(model="openai/gpt", api_key="sk-live-provider-secret-123456"), - ), - response_type=Payload, - ) - raw = this_tests_files(root)[0].read_text(encoding="utf-8") - interaction = Interaction.model_validate_json(raw) - assert "sk-live-provider-secret-123456" not in raw - assert isinstance(interaction.request.body, dict) - params = interaction.request.body["litellm_params"] - assert isinstance(params, dict) - assert params["api_key"] == "" - assert params["aws_secret_access_key"] is None - - def test_upload_records_a_content_digest_not_the_bytes(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.upload( - "/v1/files", - headers=fake.master, - form=Query(q="batch"), - filename="batch.jsonl", - content=b'{"custom_id": "1"}', - response_type=Payload, - ) - interaction = Interaction.model_validate_json( - this_tests_files(root)[0].read_text(encoding="utf-8") - ) - assert interaction.request.file_name == "batch.jsonl" - assert interaction.request.file_bytes == len(b'{"custom_id": "1"}') - assert interaction.request.file_sha256 is not None - assert "custom_id" not in interaction.request.model_dump_json() - - -class TestReplayTransport: - def test_serves_recorded_values_without_touching_the_inner_transport( - self, tmp_path: Path - ) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recorded_post = recording.post( - "/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload - ) - recorded_get = recording.get( - "/v1/models", headers=fake.master, params=Query(q="all"), response_type=Payload - ) - recorded_stream = recording.stream( - "/chat/completions", headers=fake.master, json=Body(prompt="hi") - ) - recorded_probe = recording.probe("/health/liveliness", params=Query(q="1")) - recorded_binary = recording.stream_binary( - "/v1/audio/speech", headers=fake.master, json=Body(prompt="say") - ) - calls_after_record = list(fake.calls) - - replay: Transport = ReplayTransport(source=replay_source(root), master_key="sk-1234") - assert ( - replay.post("/model/new", headers=replay.master, json=Body(prompt="x"), response_type=Payload) - == recorded_post - ) - assert ( - replay.get("/v1/models", headers=replay.master, params=Query(q="all"), response_type=Payload) - == recorded_get - ) - assert ( - replay.stream("/chat/completions", headers=replay.master, json=Body(prompt="hi")) - == recorded_stream - ) - assert replay.probe("/health/liveliness", params=Query(q="1")) == recorded_probe - assert ( - replay.stream_binary("/v1/audio/speech", headers=replay.master, json=Body(prompt="say")) - == recorded_binary - ) - assert fake.calls == calls_after_record - - def test_miss_names_the_computed_key_and_the_closest_recorded_key(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post("/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload) - replay: Transport = ReplayTransport(source=replay_source(root), master_key="sk-1234") - with pytest.raises(ReplayMiss) as excinfo: - replay.get("/v1/models", headers=replay.master, params=Query(q="all"), response_type=Payload) - message = str(excinfo.value) - assert "no recorded interaction matches key get /v1/models #" in message - assert "closest recorded key is post /model/new #" in message - assert "0000-post-model-new.json" in message - assert "re-record with E2E_FIXTURE_MODE=record" in message - - def test_content_drift_on_the_same_route_misses_with_no_live_call(self, tmp_path: Path) -> None: - """The naive verb+path match replayed a stale response for a request - whose content had changed, silently passing; a content key must miss, - print both canonical forms' diff, and never reach the inner transport.""" - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post("/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload) - calls_after_record = list(fake.calls) - replay: Transport = ReplayTransport(source=replay_source(root), master_key="sk-1234") - with pytest.raises(ReplayMiss) as excinfo: - replay.post("/model/new", headers=replay.master, json=Body(prompt="y"), response_type=Payload) - message = str(excinfo.value) - assert "no recorded interaction matches key post /model/new #" in message - assert "closest recorded key is post /model/new #" in message - assert '- "prompt": "x"' in message - assert '+ "prompt": "y"' in message - assert fake.calls == calls_after_record - - def test_exhausted_key_names_the_key(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post("/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload) - replay: Transport = ReplayTransport(source=replay_source(root), master_key="sk-1234") - replay.post("/model/new", headers=replay.master, json=Body(prompt="x"), response_type=Payload) - with pytest.raises( - ReplayMiss, match=r"every recorded interaction for key post /model/new #\w{16} is already consumed" - ): - replay.post("/model/new", headers=replay.master, json=Body(prompt="x"), response_type=Payload) - - def test_replays_out_of_recorded_order_across_distinct_keys(self, tmp_path: Path) -> None: - """Concurrent tests interleave independent calls nondeterministically - (e.g. a burst of parallel chat calls), so replay matches by content, - never by recorded position.""" - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post("/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload) - recording.post("/key/generate", headers=fake.master, json=Body(prompt="k"), response_type=Payload) - source = replay_source(root) - replay: Transport = ReplayTransport(source=source, master_key="sk-1234") - replay.post("/key/generate", headers=replay.master, json=Body(prompt="k"), response_type=Payload) - replay.post("/model/new", headers=replay.master, json=Body(prompt="x"), response_type=Payload) - assert source.leftover_error(current_test_key()) is None - - def test_identical_requests_replay_their_responses_in_recorded_order(self, tmp_path: Path) -> None: - """A poll loop makes the same request repeatedly and asserts on the - progression, so duplicates under one key stay FIFO.""" - root = tmp_path / "bundle" - recorder = make_recorder(root) - recorder.record( - test_key=current_test_key(), - request=recorded_request( - "get", "/v1/models", headers=AuthHeaders(authorization="Bearer sk-x"), params=Query(q="all") - ), - response=RecordedResult(kind="success", status_code=200, data={"value": "first"}), - ) - recorder.record( - test_key=current_test_key(), - request=recorded_request( - "get", "/v1/models", headers=AuthHeaders(authorization="Bearer sk-x"), params=Query(q="all") - ), - response=RecordedResult(kind="success", status_code=200, data={"value": "second"}), - ) - replay: Transport = ReplayTransport(source=replay_source(root), master_key="sk-1234") - first = replay.get("/v1/models", headers=replay.master, params=Query(q="all"), response_type=Payload) - second = replay.get("/v1/models", headers=replay.master, params=Query(q="all"), response_type=Payload) - assert first == Success(status_code=200, data=Payload(value="first")) - assert second == Success(status_code=200, data=Payload(value="second")) - - def test_concurrent_replays_of_one_key_serve_each_recording_exactly_once(self, tmp_path: Path) -> None: - """A burst of parallel identical calls consumes one shared pool: no - response duplicated, none forgotten, nothing left over at teardown. - The tiny switch interval forces thread preemption inside pool setup - and consumption, so a non-atomic pool build or pop fails this test.""" - root = tmp_path / "bundle" - recorder = make_recorder(root) - for ordinal in range(32): - recorder.record( - test_key=current_test_key(), - request=recorded_request( - "get", "/v1/models", headers=AuthHeaders(authorization="Bearer sk-x"), params=Query(q="all") - ), - response=RecordedResult(kind="success", status_code=200, data={"value": f"v{ordinal:02d}"}), - ) - source = replay_source(root) - replay: Transport = ReplayTransport(source=source, master_key="sk-1234") - barrier = threading.Barrier(8) - - def consume_one() -> str: - result = replay.get( - "/v1/models", headers=replay.master, params=Query(q="all"), response_type=Payload - ) - assert isinstance(result, Success) - return result.data.value - - def consume(_: int) -> tuple[str, ...]: - barrier.wait() - return tuple(consume_one() for _call in range(4)) - - previous_interval = sys.getswitchinterval() - sys.setswitchinterval(1e-6) - try: - with ThreadPoolExecutor(max_workers=8) as executor: - served = sorted(value for values in executor.map(consume, range(8)) for value in values) - finally: - sys.setswitchinterval(previous_interval) - assert served == [f"v{ordinal:02d}" for ordinal in range(32)] - assert source.leftover_error(current_test_key()) is None - - -class TestRecordedKeySets: - def test_two_separate_recordings_of_one_flow_produce_identical_key_sets( - self, tmp_path: Path - ) -> None: - """Everything a run randomizes (markers, virtual keys, dates) must - canonicalize out, so separately recorded runs of the same suite agree - on every match key and a bundle recorded elsewhere replays here.""" - - def record_flow(root: Path, run_date: str) -> list[str]: - fake = FakeTransport() - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - marker = deterministic_marker() - recording.post( - "/model/new", - headers=fake.master, - json=DeployBody( - model_name=f"e2e-chat-{marker}", - litellm_params=DeployParams(model="openai/gpt", api_key=f"sk-live-{uuid4().hex}"), - ), - response_type=Payload, - ) - recording.post( - "/chat/completions", - headers=recording.bearer(f"sk-{uuid4().hex}"), - json=Body(prompt=f"Reply with the single word ok. {marker}"), - response_type=Payload, - ) - recording.get( - "/spend/logs", headers=fake.master, params=Query(q=run_date), response_type=Payload - ) - loaded = load_bundle(root) - assert isinstance(loaded, LoadedBundle) - return sorted( - canonicalize(interaction.request).key - for interactions in loaded.interactions.values() - for interaction in interactions - ) - - first_keys = record_flow(tmp_path / "one", "2026-08-18") - second_keys = record_flow(tmp_path / "two", "2026-08-19") - assert first_keys == second_keys - assert len(first_keys) == 3 - - -class TestReplayLeftover: - def test_fully_consumed_recording_leaves_nothing(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post("/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload) - source = replay_source(root) - replay: Transport = ReplayTransport(source=source, master_key="sk-1234") - replay.post("/model/new", headers=replay.master, json=Body(prompt="x"), response_type=Payload) - assert source.leftover_error(current_test_key()) is None - - def test_unconsumed_trailing_interactions_name_the_next_call(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post("/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload) - recording.probe("/health/liveliness", params=Query(q="1")) - source = replay_source(root) - replay: Transport = ReplayTransport(source=source, master_key="sk-1234") - replay.post("/model/new", headers=replay.master, json=Body(prompt="x"), response_type=Payload) - error = source.leftover_error(current_test_key()) - assert error is not None - assert "1 of 2 recorded interactions never consumed" in error - assert "e.g. probe /health/liveliness #" in error - assert "re-record with E2E_FIXTURE_MODE=record" in error - - def test_test_without_recordings_has_no_leftover(self, tmp_path: Path) -> None: - root = tmp_path / "bundle" - make_recorder(root) - assert replay_source(root).leftover_error("suite.py::test_never_recorded") is None - - def test_inert_outside_replay_mode(self, tmp_path: Path) -> None: - missing = tmp_path / "missing" - assert replay_leftover_error(mode_raw="", bundle_dir=missing, test_key="k") is None - assert replay_leftover_error(mode_raw="record", bundle_dir=missing, test_key="k") is None - - def test_replay_mode_reads_the_shared_bundle(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - recording: Transport = RecordingTransport(inner=fake, recorder=make_recorder(root)) - recording.post("/model/new", headers=fake.master, json=Body(prompt="x"), response_type=Payload) - error = replay_leftover_error(mode_raw="replay", bundle_dir=root, test_key=current_test_key()) - assert error is not None - assert "1 of 1 recorded interactions never consumed" in error - - -class TestSelectTransport: - def test_live_returns_the_live_transport_untouched(self, tmp_path: Path) -> None: - fake = FakeTransport() - for mode_raw in ("live", ""): - assert ( - select_transport(fake, mode_raw=mode_raw, bundle_dir=tmp_path / "b", master_key="sk") - is fake - ) - - def test_record_wraps_live_and_starts_a_fresh_bundle(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - write_manifest(root, NOW - timedelta(days=30)) - (root / "old-test-slug").mkdir() - (root / "old-test-slug" / "0000-post-old.json").write_text("{}", encoding="utf-8") - selected = select_transport(fake, mode_raw="record", bundle_dir=root, master_key="sk") - assert isinstance(selected, RecordingTransport) - assert selected.inner is fake - assert {entry.name for entry in root.iterdir()} == {MANIFEST_FILENAME} - - def test_replay_builds_a_transport_from_the_bundle_alone(self, tmp_path: Path) -> None: - fake = FakeTransport() - root = tmp_path / "bundle" - make_recorder(root) - selected = select_transport(fake, mode_raw="replay", bundle_dir=root, master_key="sk-master") - assert isinstance(selected, ReplayTransport) - assert selected.master == AuthHeaders(authorization="Bearer sk-master") - - def test_invalid_mode_raises_naming_the_value(self, tmp_path: Path) -> None: - with pytest.raises(ValueError, match="cached"): - select_transport( - FakeTransport(), mode_raw="cached", bundle_dir=tmp_path / "b", master_key="sk" - ) - - -class TestCollectionGate: - def test_invalid_mode_names_the_value_and_the_choices(self, tmp_path: Path) -> None: - assert ( - fixture_mode_collection_error("cached", tmp_path, now=NOW) - == "E2E_FIXTURE_MODE='cached' is not one of live, record, replay" - ) - - @pytest.mark.parametrize("mode_raw", ["live", "", "record"]) - def test_live_and_record_never_block_collection(self, mode_raw: str, tmp_path: Path) -> None: - assert fixture_mode_collection_error(mode_raw, tmp_path / "missing", now=NOW) is None - - def test_replay_with_no_bundle_says_how_to_record_one(self, tmp_path: Path) -> None: - reason = fixture_mode_collection_error("replay", tmp_path / "missing", now=NOW) - assert reason is not None - assert f"no {MANIFEST_FILENAME}" in reason - assert "E2E_FIXTURE_MODE=record" in reason - - def test_stale_replay_bundle_fails_naming_its_age(self, tmp_path: Path) -> None: - root = tmp_path / "bundle" - write_manifest(root, NOW - timedelta(days=9, hours=5)) - reason = fixture_mode_collection_error("replay", root, now=NOW) - assert reason is not None - assert "age 9d5h exceeds the 7-day limit" in reason - assert "re-record with E2E_FIXTURE_MODE=record" in reason - - def test_fresh_replay_bundle_collects(self, tmp_path: Path) -> None: - root = tmp_path / "bundle" - write_manifest(root, NOW - timedelta(days=2)) - assert fixture_mode_collection_error("replay", root, now=NOW) is None - - -class TestReportHeader: - def test_live_mode_prints_nothing(self, tmp_path: Path) -> None: - assert fixture_report_lines("live", tmp_path, now=NOW) == [] - assert fixture_report_lines("", tmp_path, now=NOW) == [] - - def test_record_and_replay_name_the_bundle(self, tmp_path: Path) -> None: - root = tmp_path / "bundle" - recorded_at = NOW - timedelta(days=1) - write_manifest(root, recorded_at) - assert fixture_report_lines("record", root, now=NOW) == [ - f"e2e fixture mode: record -> {root}" - ] - replay_lines = fixture_report_lines("replay", root, now=NOW) - assert len(replay_lines) == 1 - assert "replay" in replay_lines[0] - assert recorded_at.isoformat() in replay_lines[0] diff --git a/tests/e2e/test_provider_edge.py b/tests/e2e/test_provider_edge.py new file mode 100644 index 00000000000..492eee57aaf --- /dev/null +++ b/tests/e2e/test_provider_edge.py @@ -0,0 +1,492 @@ +"""Harness coverage for the provider-edge record/replay server (LIT-5745). + +No proxy and no ``e2e`` marker. A stdlib http.server stands in for the +provider (dependency injection via the mounts mapping, no monkeypatching): +record mode must forward each edge call to it verbatim, persist one +interaction file, and serve the proxy the same filtered response replay will +serve later; replay mode must serve byte-identical responses from the bundle +alone, with the fake provider's hit log proving nothing leaves the process, +and answer any drifted call with HTTP ``REPLAY_MISS_STATUS`` naming the +computed and closest recorded canonical keys (LIT-5741; the pure canonicalizer +is pinned in test_fixture_canonical.py). Requests are made through +``e2e_http.forward`` so the whole HTTP surface of the edge is exercised; the +pure ``handle_edge_request`` core is pinned socket-free alongside. +""" + +from __future__ import annotations + +import base64 +import json +import threading +from collections.abc import Generator, Mapping +from concurrent.futures import ThreadPoolExecutor +from contextlib import contextmanager +from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer +from pathlib import Path + +import pytest +from pydantic import TypeAdapter + +from e2e_http import RawResponse, forward +from fixture_bundle import ( + BundleRecorder, + Interaction, + LoadedBundle, + RecordedHttpResponse, + RecordedRequest, + load_bundle, + prepare_bundle, + slug_for_test, +) +from fixture_mode import current_test_key +from provider_edge import ( + REPLAY_MISS_STATUS, + EdgeBackend, + ProviderEdge, + RecordEdge, + ReplayEdge, + ReplaySource, + handle_edge_request, + provider_edge_api_base, + replay_leftover_error, + start_provider_edge, +) + +CHAT_PATH = "/openai/v1/chat/completions" +REPLAY_MOUNTS = {"openai": "https://replay.invalid"} +JSON_OBJECT = TypeAdapter(dict[str, object]) + + +def json_object(body: bytes) -> dict[str, object]: + return JSON_OBJECT.validate_json(body) + + +class _FakeProvider(ThreadingHTTPServer): + daemon_threads = True + + def __init__(self, bind: tuple[str, int]) -> None: + super().__init__(bind, _FakeProviderHandler) + self.hits: list[str] = [] + + +class _FakeProviderHandler(BaseHTTPRequestHandler): + protocol_version = "HTTP/1.1" + + def do_POST(self) -> None: + self._respond() + + def do_GET(self) -> None: + self._respond() + + def _respond(self) -> None: + provider = self.server + assert isinstance(provider, _FakeProvider) + length = int(self.headers.get("content-length") or "0") + body = self.rfile.read(length) if length else b"" + provider.hits.append(f"{self.command} {self.path}") + payload = json.dumps( + {"echo": body.decode("utf-8"), "path": self.path, "hit": len(provider.hits)} + ).encode() + self.send_response(200) + self.send_header("content-type", "application/json") + self.send_header("content-length", str(len(payload))) + self.send_header("x-upstream", "fake") + self.send_header("set-cookie", "session=fake-cookie") + self.end_headers() + self.wfile.write(payload) + + def log_message(self, format: str, *args: object) -> None: + """Silence the per-request stderr line BaseHTTPRequestHandler emits.""" + + +@contextmanager +def fake_provider() -> Generator[_FakeProvider]: + server = _FakeProvider(("127.0.0.1", 0)) + thread = threading.Thread(target=server.serve_forever, daemon=True) + thread.start() + try: + yield server + finally: + server.shutdown() + server.server_close() + + +def provider_url(server: _FakeProvider) -> str: + return f"http://127.0.0.1:{server.server_address[1]}" + + +@contextmanager +def running_edge(backend: EdgeBackend, mounts: Mapping[str, str]) -> Generator[ProviderEdge]: + running = start_provider_edge(backend, mounts=mounts, bind_host="127.0.0.1") + try: + yield running.edge + finally: + running.shutdown() + + +def record_backend(root: Path) -> RecordEdge: + recorder = prepare_bundle(root) + assert isinstance(recorder, BundleRecorder) + return RecordEdge(recorder=recorder, lock=threading.Lock()) + + +def replay_source(root: Path) -> ReplaySource: + loaded = load_bundle(root) + assert isinstance(loaded, LoadedBundle) + return ReplaySource(bundle=loaded) + + +def call_edge( + edge: ProviderEdge, + method: str, + path: str, + *, + body: bytes | None = None, + headers: dict[str, str] | None = None, +) -> RawResponse: + outcome = forward( + method, + f"http://{edge.advertise_host}:{edge.port}{path}", + headers=headers or {}, + body=body, + timeout=10.0, + ) + assert isinstance(outcome, RawResponse) + return outcome + + +def this_tests_files(root: Path) -> list[Path]: + slug_dir = root / slug_for_test(current_test_key()) + return sorted(slug_dir.glob("*.json")) if slug_dir.is_dir() else [] + + +def chat_body(prompt: str) -> bytes: + return json.dumps({"model": "gpt", "messages": [{"role": "user", "content": prompt}]}).encode() + + +class TestRecordMode: + def test_forwards_to_the_provider_and_writes_one_interaction_file(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + reply = call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + assert provider.hits == ["POST /v1/chat/completions"] + assert reply.status_code == 200 + served = json_object(reply.body) + assert served["echo"] == chat_body("hi").decode() + files = this_tests_files(root) + assert [file.name for file in files] == ["0000-post-openai-v1-chat-completions.json"] + interaction = Interaction.model_validate_json(files[0].read_text(encoding="utf-8")) + assert interaction.request.method == "post" + assert interaction.request.path == CHAT_PATH + assert interaction.request.body == json_object(chat_body("hi")) + assert interaction.response.status_code == 200 + + def test_never_stores_headers_so_credentials_never_touch_disk(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge( + edge, + "POST", + CHAT_PATH, + body=chat_body("hi"), + headers={"authorization": "Bearer sk-live-provider-secret-abc123"}, + ) + raw = this_tests_files(root)[0].read_text(encoding="utf-8") + assert "sk-live-provider-secret-abc123" not in raw + interaction = Interaction.model_validate_json(raw) + assert interaction.request.headers == {} + + def test_strips_volatile_response_headers_and_serves_the_filtered_copy(self, tmp_path: Path) -> None: + """What record serves the proxy must equal what replay will serve later + (record/replay parity), so the filtered stored copy is served in both.""" + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + reply = call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + assert reply.headers.get("x-upstream") == "fake" + assert "set-cookie" not in reply.headers + interaction = Interaction.model_validate_json( + this_tests_files(root)[0].read_text(encoding="utf-8") + ) + assert interaction.response.headers.get("x-upstream") == "fake" + assert "set-cookie" not in interaction.response.headers + assert "content-length" not in interaction.response.headers + + def test_unreachable_provider_records_and_serves_a_502(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with running_edge(record_backend(root), {"openai": "http://127.0.0.1:9"}) as edge: + reply = call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + assert reply.status_code == 502 + assert b"could not reach the provider" in reply.body + interaction = Interaction.model_validate_json( + this_tests_files(root)[0].read_text(encoding="utf-8") + ) + assert interaction.response.status_code == 502 + + +class TestReplayMode: + def test_serves_recorded_bytes_with_zero_provider_hits(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + recorded = call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + hits_after_record = list(provider.hits) + with running_edge( + ReplayEdge(source=replay_source(root)), {"openai": provider_url(provider)} + ) as edge: + replayed = call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + assert provider.hits == hits_after_record + assert replayed.status_code == recorded.status_code + assert replayed.body == recorded.body + assert replayed.headers.get("x-upstream") == "fake" + + def test_request_identity_ignores_auth_headers(self, tmp_path: Path) -> None: + """The proxy sends different bearer tokens across runs (fresh virtual + keys, rotated provider keys), so headers are no part of the match.""" + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge( + edge, "POST", CHAT_PATH, body=chat_body("hi"), + headers={"authorization": "Bearer sk-first-run"}, + ) + with running_edge(ReplayEdge(source=replay_source(root)), REPLAY_MOUNTS) as edge: + replayed = call_edge( + edge, "POST", CHAT_PATH, body=chat_body("hi"), + headers={"authorization": "Bearer sk-second-run"}, + ) + assert replayed.status_code == 200 + + def test_content_drift_returns_the_miss_status_naming_both_keys(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("x")) + with running_edge(ReplayEdge(source=replay_source(root)), REPLAY_MOUNTS) as edge: + missed = call_edge(edge, "POST", CHAT_PATH, body=chat_body("y")) + assert missed.status_code == REPLAY_MISS_STATUS + message = missed.body.decode() + assert f"no recorded interaction matches key post {CHAT_PATH} #" in message + assert f"closest recorded key is post {CHAT_PATH} #" in message + assert '"content": "x"' in message + assert '"content": "y"' in message + assert "re-record with E2E_FIXTURE_MODE=record" in message + + def test_query_params_are_part_of_the_identity(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge(edge, "GET", "/openai/v1/models?purpose=batch") + assert provider.hits == ["GET /v1/models?purpose=batch"] + with running_edge(ReplayEdge(source=replay_source(root)), REPLAY_MOUNTS) as edge: + missed = call_edge(edge, "GET", "/openai/v1/models?purpose=other") + matched = call_edge(edge, "GET", "/openai/v1/models?purpose=batch") + assert missed.status_code == REPLAY_MISS_STATUS + assert matched.status_code == 200 + + def test_identical_requests_replay_their_responses_in_recorded_order(self, tmp_path: Path) -> None: + """A poll or retry loop repeats the same request and the proxy asserts + on the progression, so duplicates under one key stay FIFO.""" + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + with running_edge(ReplayEdge(source=replay_source(root)), REPLAY_MOUNTS) as edge: + first = json_object(call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")).body) + second = json_object(call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")).body) + assert first["hit"] == 1 + assert second["hit"] == 2 + + def test_exhausted_key_returns_the_miss_status(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + with running_edge(ReplayEdge(source=replay_source(root)), REPLAY_MOUNTS) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + exhausted = call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + assert exhausted.status_code == REPLAY_MISS_STATUS + assert b"already consumed" in exhausted.body + + def test_non_json_bodies_match_by_canonical_digest_without_storing_them(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + opaque = b"custom_id one\ncustom_id two\n" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge(edge, "POST", "/openai/v1/files", body=opaque) + raw = this_tests_files(root)[0].read_text(encoding="utf-8") + interaction = Interaction.model_validate_json(raw) + assert interaction.request.body is None + assert interaction.request.file_sha256 is not None + assert interaction.request.file_bytes == len(opaque) + assert "custom_id" not in interaction.request.model_dump_json() + with running_edge(ReplayEdge(source=replay_source(root)), REPLAY_MOUNTS) as edge: + replayed = call_edge(edge, "POST", "/openai/v1/files", body=opaque) + assert replayed.status_code == 200 + + +class TestReplayLeftover: + def test_partially_consumed_recording_names_the_leftover(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + call_edge(edge, "GET", "/openai/v1/models") + source = replay_source(root) + with running_edge(ReplayEdge(source=source), REPLAY_MOUNTS) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + error = source.leftover_error(current_test_key()) + assert error is not None + assert "1 of 2 recorded interactions never consumed" in error + assert "e.g. get /openai/v1/models #" in error + assert "re-record with E2E_FIXTURE_MODE=record" in error + + def test_fully_consumed_recording_leaves_nothing(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + with fake_provider() as provider: + with running_edge(record_backend(root), {"openai": provider_url(provider)}) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + source = replay_source(root) + with running_edge(ReplayEdge(source=source), REPLAY_MOUNTS) as edge: + call_edge(edge, "POST", CHAT_PATH, body=chat_body("hi")) + assert source.leftover_error(current_test_key()) is None + + def test_test_without_recordings_has_no_leftover(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + assert isinstance(prepare_bundle(root), BundleRecorder) + assert replay_source(root).leftover_error("suite.py::test_never_recorded") is None + + def test_inert_outside_replay_mode(self, tmp_path: Path) -> None: + missing = tmp_path / "missing" + assert replay_leftover_error(mode_raw="", bundle_dir=missing, test_key="k") is None + assert replay_leftover_error(mode_raw="record", bundle_dir=missing, test_key="k") is None + + +class TestConcurrentReplay: + def test_parallel_identical_calls_serve_each_recording_exactly_once(self, tmp_path: Path) -> None: + """The edge server handles requests on concurrent threads and a burst + of parallel identical calls consumes one shared pool: no response + duplicated, none forgotten, nothing left over at teardown.""" + root = tmp_path / "bundle" + recorder = prepare_bundle(root) + assert isinstance(recorder, BundleRecorder) + for ordinal in range(32): + recorder.record( + test_key=current_test_key(), + request=RecordedRequest(method="post", path=CHAT_PATH, headers={}, body={"n": "same"}), + response=RecordedHttpResponse( + status_code=200, + headers={"content-type": "application/json"}, + body_b64=base64.b64encode(json.dumps({"value": f"v{ordinal:02d}"}).encode()).decode(), + ), + ) + source = replay_source(root) + body = json.dumps({"n": "same"}).encode() + barrier = threading.Barrier(8) + with running_edge(ReplayEdge(source=source), REPLAY_MOUNTS) as edge: + + def consume(_: int) -> tuple[str, ...]: + barrier.wait() + return tuple( + str(json_object(call_edge(edge, "POST", CHAT_PATH, body=body).body)["value"]) + for _call in range(4) + ) + + with ThreadPoolExecutor(max_workers=8) as executor: + served = sorted(value for values in executor.map(consume, range(8)) for value in values) + assert served == [f"v{ordinal:02d}" for ordinal in range(32)] + assert source.leftover_error(current_test_key()) is None + + +class TestHandleEdgeRequestPure: + def test_unknown_mount_404s_naming_the_known_mounts(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + assert isinstance(prepare_bundle(root), BundleRecorder) + reply = handle_edge_request( + ReplayEdge(source=replay_source(root)), + {"openai": "https://api.openai.com", "anthropic": "https://api.anthropic.com"}, + "POST", + "/bedrock/model/invoke", + {}, + b"{}", + timeout=1.0, + ) + assert reply.status_code == 404 + assert b"unknown provider mount 'bedrock'" in reply.body + assert b"anthropic, openai" in reply.body + + def test_replay_serves_a_directly_recorded_interaction(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + recorder = prepare_bundle(root) + assert isinstance(recorder, BundleRecorder) + recorder.record( + test_key=current_test_key(), + request=RecordedRequest(method="post", path=CHAT_PATH, headers={}, body={"prompt": "x"}), + response=RecordedHttpResponse( + status_code=201, headers={"x-upstream": "fake"}, body_b64=base64.b64encode(b"ok").decode() + ), + ) + reply = handle_edge_request( + ReplayEdge(source=replay_source(root)), + {"openai": "https://api.openai.com"}, + "POST", + CHAT_PATH, + {"authorization": "Bearer sk-anything"}, + json.dumps({"prompt": "x"}).encode(), + timeout=1.0, + ) + assert reply.status_code == 201 + assert reply.body == b"ok" + assert reply.headers == {"x-upstream": "fake"} + + +class TestApiBaseSeam: + def test_live_mode_returns_none(self, tmp_path: Path) -> None: + for mode_raw in ("live", ""): + assert ( + provider_edge_api_base( + "openai", + mode_raw=mode_raw, + bundle_dir=tmp_path / "bundle", + bind_host="127.0.0.1", + advertise_host="127.0.0.1", + ) + is None + ) + + def test_invalid_mode_raises_naming_the_value(self, tmp_path: Path) -> None: + with pytest.raises(ValueError, match="cached"): + provider_edge_api_base( + "openai", + mode_raw="cached", + bundle_dir=tmp_path / "bundle", + bind_host="127.0.0.1", + advertise_host="127.0.0.1", + ) + + def test_unknown_mount_raises_naming_the_known_mounts(self, tmp_path: Path) -> None: + with pytest.raises(ValueError, match="unknown provider mount 'bedrock'"): + provider_edge_api_base( + "bedrock", + mode_raw="record", + bundle_dir=tmp_path / "bundle", + bind_host="127.0.0.1", + advertise_host="127.0.0.1", + ) + + def test_record_mode_boots_one_shared_edge_and_prepares_the_bundle(self, tmp_path: Path) -> None: + root = tmp_path / "bundle" + first = provider_edge_api_base( + "openai", mode_raw="record", bundle_dir=root, bind_host="127.0.0.1", advertise_host="127.0.0.1" + ) + second = provider_edge_api_base( + "anthropic", mode_raw="record", bundle_dir=root, bind_host="127.0.0.1", advertise_host="127.0.0.1" + ) + assert first is not None and second is not None + assert first.endswith("/openai") + assert second.endswith("/anthropic") + assert first.rsplit("/", 1)[0] == second.rsplit("/", 1)[0] + assert (root / "manifest.json").is_file() From 0de829d3e44094a808e9c1166d12c4d86b29c6b0 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 18:57:35 -0700 Subject: [PATCH 018/220] feat(cli): store the lite login credential in the OS keychain lite login used to write the minted cli-session key in cleartext to ~/.litellm/token.json. The secret material (key plus any JWT) now goes to the OS keychain through the optional keyring package, with the 0600 file kept for non-secret metadata and as the fallback on headless boxes. Legacy plaintext files keep authenticating and are migrated into the keychain, then scrubbed, on first read. A secret still on disk always outranks the keychain entry, so a failed keychain write can never resurrect a stale key. LITELLM_PROXY_API_KEY and --api-key precedence is unchanged, lite logout clears both stores and warns when the keychain will not release the entry, and ~/.litellm is created 0700 (tightened from 0755 where an older CLI left it broader). LITELLM_CLI_DISABLE_KEYRING=1 forces the file fallback. --- basedpyright-code-budget.json | 8 +- .../litellm_proxy_server/cli_token_usage.py | 2 +- litellm/litellm_core_utils/cli_keyring.py | 125 ++++ litellm/litellm_core_utils/cli_token_utils.py | 184 +++++- .../private_json.py | 10 + litellm/proxy/client/README.md | 12 +- litellm/proxy/client/cli/commands/agents.py | 4 +- litellm/proxy/client/cli/commands/auth.py | 148 +++-- .../client/cli/commands/claude_settings.py | 2 +- litellm/proxy/client/cli/commands/config.py | 6 +- litellm/proxy/client/cli/commands/up.py | 22 +- litellm/proxy/client/cli/main.py | 4 +- pyproject.toml | 2 + tests/test_litellm/conftest.py | 65 ++ .../test_cli_token_utils.py | 478 ++++++++++++--- .../proxy/client/cli/test_agents.py | 7 +- .../proxy/client/cli/test_auth_commands.py | 577 +++++++++--------- .../proxy/client/cli/test_claude_settings.py | 15 +- .../proxy/client/cli/test_config_commands.py | 4 +- .../proxy/client/cli/test_up_commands.py | 23 +- type-discipline-budget.json | 8 +- uv.lock | 100 ++- 22 files changed, 1295 insertions(+), 511 deletions(-) create mode 100644 litellm/litellm_core_utils/cli_keyring.py rename litellm/{proxy/client/cli/commands => litellm_core_utils}/private_json.py (64%) diff --git a/basedpyright-code-budget.json b/basedpyright-code-budget.json index 1ce71c5bd2c..59d56a3f63d 100644 --- a/basedpyright-code-budget.json +++ b/basedpyright-code-budget.json @@ -57,7 +57,7 @@ "limit": 5663 }, "reportMissingTypeArgument": { - "limit": 15557 + "limit": 15556 }, "reportMissingTypeStubs": { "limit": 40 @@ -105,13 +105,13 @@ "limit": 109 }, "reportUnknownMemberType": { - "limit": 39043 + "limit": 39042 }, "reportUnknownParameterType": { - "limit": 19887 + "limit": 19886 }, "reportUnknownVariableType": { - "limit": 30574 + "limit": 30571 }, "reportUnnecessaryCast": { "limit": 117 diff --git a/cookbook/litellm_proxy_server/cli_token_usage.py b/cookbook/litellm_proxy_server/cli_token_usage.py index 6306970cdde..e6b3744019c 100644 --- a/cookbook/litellm_proxy_server/cli_token_usage.py +++ b/cookbook/litellm_proxy_server/cli_token_usage.py @@ -60,4 +60,4 @@ if __name__ == "__main__": print("\n💡 Tips:") print("1. Run 'litellm-proxy login' to authenticate first") print("2. Replace 'https://your-proxy.com' with your actual proxy URL") - print("3. The token is stored locally at ~/.litellm/token.json") + print("3. The token is stored in your OS keychain, or in ~/.litellm/token.json when there is none") diff --git a/litellm/litellm_core_utils/cli_keyring.py b/litellm/litellm_core_utils/cli_keyring.py new file mode 100644 index 00000000000..873db64a728 --- /dev/null +++ b/litellm/litellm_core_utils/cli_keyring.py @@ -0,0 +1,125 @@ +""" +CLI Keyring Access + +SDK-level access to the OS keychain (macOS Keychain, Windows Credential Manager, +Linux Secret Service) that holds the credential minted by `lite login`. + +The `keyring` package is optional and imported lazily, so importing this module +never pulls it in. Every failure is returned as a value: a machine with no +keychain, or one whose keychain is locked, must degrade to the token file rather +than break `lite` or the SDK. +""" + +import os +from dataclasses import dataclass +from typing import Final, Protocol, TypeAlias + +KEYRING_SERVICE: Final = "litellm-cli" +KEYRING_ACCOUNT: Final = "credential" +DISABLE_KEYRING_ENV_VAR: Final = "LITELLM_CLI_DISABLE_KEYRING" + +_DISABLED_VALUES: Final = frozenset(("1", "true", "yes", "on")) + + +@dataclass(frozen=True, slots=True) +class SecretFound: + blob: str + + +@dataclass(frozen=True, slots=True) +class SecretMissing: + pass + + +@dataclass(frozen=True, slots=True) +class SecretUnavailable: + pass + + +SecretRead: TypeAlias = SecretFound | SecretMissing | SecretUnavailable + + +class SecretVault(Protocol): + """The single slot holding the CLI credential's secret material.""" + + def read(self) -> SecretRead: ... + + def write(self, blob: str) -> bool: ... + + def erase(self) -> bool: ... + + +class KeyringApi(Protocol): + def get_password(self, service_name: str, username: str) -> str | None: ... + + def set_password(self, service_name: str, username: str, password: str) -> None: ... + + def delete_password(self, service_name: str, username: str) -> None: ... + + +def _keyring_disabled() -> bool: + return os.getenv(DISABLE_KEYRING_ENV_VAR, "").strip().lower() in _DISABLED_VALUES + + +def _import_keyring() -> KeyringApi | None: + try: + import keyring + except ImportError: + return None + return keyring + + +def _keyring_api() -> KeyringApi | None: + return None if _keyring_disabled() else _import_keyring() + + +@dataclass(frozen=True, slots=True) +class KeyringVault: + """The OS keychain, reached through the optional `keyring` package.""" + + def read(self) -> SecretRead: + api: Final = _keyring_api() + if api is None: + return SecretUnavailable() + try: + blob: Final = api.get_password(KEYRING_SERVICE, KEYRING_ACCOUNT) + except Exception: # noqa: BLE001 # backends raise outside keyring.errors; never break the SDK + return SecretUnavailable() + return SecretMissing() if blob is None else SecretFound(blob) + + def write(self, blob: str) -> bool: + api: Final = _keyring_api() + if api is None: + return False + try: + api.set_password(KEYRING_SERVICE, KEYRING_ACCOUNT, blob) + except Exception: # noqa: BLE001 # a keychain that refuses the write falls back to the token file + return False + return True + + def erase(self) -> bool: + if _import_keyring() is None: + return True + if _keyring_disabled(): + # a credential stored before the kill switch was set may still be in the keychain + return False + match self.read(): + case SecretUnavailable(): + return False + case SecretMissing(): + return True + case SecretFound(): + return self._delete() + + def _delete(self) -> bool: + api: Final = _keyring_api() + if api is None: + return False + try: + api.delete_password(KEYRING_SERVICE, KEYRING_ACCOUNT) + except Exception: # noqa: BLE001 # report the failure as a value so `lite logout` can warn + return False + return True + + +SYSTEM_KEYRING: Final[SecretVault] = KeyringVault() diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index a44ce431f4e..9960192180c 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -1,17 +1,68 @@ """ CLI Token Utilities -SDK-level utilities for reading CLI authentication tokens. +SDK-level utilities for reading the credential minted by `lite login`. + +Non-secret metadata lives in ~/.litellm/token.json. The secret material (the +bearer key, plus a JWT when one is issued) lives in the OS keychain when the +machine has one, and in that same 0600 file otherwise. This module hides the +split from callers, and migrates a legacy plaintext file into the keychain the +first time it reads one. + This module has no dependencies on proxy code and can be safely imported at the SDK level. """ -import json -import os +import contextlib import time -from collections.abc import Mapping from pathlib import Path +from types import MappingProxyType from typing import Final +from pydantic import BaseModel, ConfigDict, ValidationError + +from litellm.litellm_core_utils.cli_keyring import ( + SYSTEM_KEYRING, + SecretFound, + SecretMissing, + SecretUnavailable, + SecretVault, +) +from litellm.litellm_core_utils.private_json import ensure_private_dir, write_private_json + + +class CliTokenRecord(BaseModel): + """A stored CLI credential. + + `key is None` means the metadata was found but the secret could not be + produced: the keychain holds nothing for us, or we could not reach it. + """ + + model_config = ConfigDict(frozen=True, extra="allow") + + base_url: str = "" + key: str | None = None + user_id: str = "" + user_email: str = "" + user_role: str = "" + auth_header_name: str = "Authorization" + jwt_token: str = "" + timestamp: float = 0.0 + + +class CliTokenSecret(BaseModel): + """The secret material as stored in the OS keychain. + + `base_url` is duplicated from the metadata file purely as a pairing tag: a + secret minted for one server is never handed to another, even if the + metadata file is edited underneath us. + """ + + model_config = ConfigDict(frozen=True) + + base_url: str + key: str + jwt_token: str = "" + def get_cli_token_file_path() -> str: """Get the path to the CLI token file""" @@ -20,26 +71,39 @@ def get_cli_token_file_path() -> str: return str(config_dir / "token.json") -def load_cli_token() -> dict | None: - """Load CLI token data from file""" - token_file: Final = get_cli_token_file_path() - if not os.path.exists(token_file): +def load_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> CliTokenRecord | None: + """Load the stored CLI credential, or None when this machine has none""" + record: Final = _read_token_file() + if record is None: return None + return _resolve_secret(record, vault) - try: - with open(token_file, "r") as f: - return json.load(f) - except (OSError, json.JSONDecodeError): - return None + +def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRING) -> bool: + """Store a freshly minted credential. Returns whether the keychain took the secret""" + if record.key is None or not vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)): + _write_token_file(record) + return False + _write_token_file(_without_secret(record)) + return True + + +def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> bool: + """Remove the credential from both stores. Returns whether the keychain is now free of it""" + erased: Final = vault.erase() + Path(get_cli_token_file_path()).unlink(missing_ok=True) + return erased def get_litellm_gateway_api_key( expected_base_url: str | None = None, + *, + vault: SecretVault = SYSTEM_KEYRING, ) -> str | None: """ Get the stored CLI API key for use with LiteLLM SDK. - This function reads the token file created by `lite login` + This function reads the credential created by `lite login` and returns the API key for use in Python scripts. Args: @@ -47,6 +111,7 @@ def get_litellm_gateway_api_key( originally issued for this URL. Pass the target server URL to prevent credential leakage when the client is pointed at a different (possibly malicious) server. + vault: Where the secret material is stored. Defaults to the OS keychain. Returns: str: The API key if found (and origin matches), None otherwise @@ -62,25 +127,84 @@ def get_litellm_gateway_api_key( >>> base_url="https://your-proxy.com/v1" >>> ) """ - token_data: Final = load_cli_token() - if not token_data or "key" not in token_data: + record: Final = _read_token_file() + if record is None: return None - if expected_base_url is not None: - stored_url: Final = token_data.get("base_url") - if stored_url != expected_base_url.rstrip("/"): - return None - return token_data["key"] + if expected_base_url is not None and record.base_url != expected_base_url.rstrip("/"): + return None + resolved: Final = _resolve_secret(record, vault) + return None if resolved is None else resolved.key -def is_cli_token_fresh(token_data: Mapping[str, object], buffer_hours: float = 0.1) -> bool: - """Check whether a cached CLI token (as stored in token.json) is still - within its expiration window. Used by `lite auth print-token` to fail - fast, without a network round trip, once the cached token is past - `LITELLM_CLI_JWT_EXPIRATION_HOURS`.""" +def is_cli_token_fresh(token_data: CliTokenRecord, buffer_hours: float = 0.1) -> bool: + """Check whether a cached CLI token is still within its expiration window. + Used by `lite auth print-token` to fail fast, without a network round trip, + once the cached token is past `LITELLM_CLI_JWT_EXPIRATION_HOURS`.""" from litellm.constants import CLI_JWT_EXPIRATION_HOURS - timestamp: Final = token_data.get("timestamp") - if not isinstance(timestamp, (int, float)): - return False - age_hours: Final = (time.time() - timestamp) / 3600 + age_hours: Final = (time.time() - token_data.timestamp) / 3600 return age_hours < (CLI_JWT_EXPIRATION_HOURS - buffer_hours) + + +def _read_token_file() -> CliTokenRecord | None: + try: + raw: Final = Path(get_cli_token_file_path()).read_text() + except OSError: + return None + try: + return CliTokenRecord.model_validate_json(raw) + except ValidationError: + return None + + +def _resolve_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecord | None: + match vault.read(): + case SecretFound(blob=blob): + return _apply_vault_secret(record, blob, vault) + case SecretMissing(): + return _migrate_file_secret(record, vault) + case SecretUnavailable(): + return record + + +def _apply_vault_secret(record: CliTokenRecord, blob: str, vault: SecretVault) -> CliTokenRecord | None: + if record.key is not None: + # a secret still on disk means the last keychain write failed: the file outranks the vault + return _migrate_file_secret(record, vault) + try: + secret: Final = CliTokenSecret.model_validate_json(blob) + except ValidationError: + return _migrate_file_secret(record, vault) + if secret.base_url != record.base_url: + return _migrate_file_secret(record, vault) + _scrub_file_secret(record) + return record.model_copy(update=MappingProxyType({"key": secret.key, "jwt_token": secret.jwt_token})) + + +def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecord | None: + if record.key is None: + return None + if vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)): + _scrub_file_secret(record) + return record + + +def _scrub_file_secret(record: CliTokenRecord) -> None: + if record.key is None and not record.jwt_token: + return + with contextlib.suppress(OSError): + _write_token_file(_without_secret(record)) + + +def _without_secret(record: CliTokenRecord) -> CliTokenRecord: + return record.model_copy(update=MappingProxyType({"key": None, "jwt_token": ""})) + + +def _encode_secret(base_url: str, key: str, jwt_token: str) -> str: + return CliTokenSecret(base_url=base_url, key=key, jwt_token=jwt_token).model_dump_json() + + +def _write_token_file(record: CliTokenRecord) -> None: + path: Final = Path(get_cli_token_file_path()) + ensure_private_dir(path.parent) + write_private_json(str(path), record.model_dump(exclude_none=True)) diff --git a/litellm/proxy/client/cli/commands/private_json.py b/litellm/litellm_core_utils/private_json.py similarity index 64% rename from litellm/proxy/client/cli/commands/private_json.py rename to litellm/litellm_core_utils/private_json.py index 31062e4a799..32bc2e169e2 100644 --- a/litellm/proxy/client/cli/commands/private_json.py +++ b/litellm/litellm_core_utils/private_json.py @@ -1,10 +1,20 @@ import json import os +import stat import tempfile from collections.abc import Mapping from pathlib import Path from typing import Final +PRIVATE_DIR_MODE: Final = 0o700 + + +def ensure_private_dir(directory: Path) -> None: + """Create directory (and parents) owner-only, tightening it if it already exists group/world readable""" + directory.mkdir(mode=PRIVATE_DIR_MODE, parents=True, exist_ok=True) + if stat.S_IMODE(directory.stat().st_mode) & 0o077: + directory.chmod(PRIVATE_DIR_MODE) + def write_private_json(path: str, data: Mapping[str, object]) -> None: """Atomically write JSON to path with owner-only permissions (0600)""" diff --git a/litellm/proxy/client/README.md b/litellm/proxy/client/README.md index 6b28f43ac73..9ece4c2be3d 100644 --- a/litellm/proxy/client/README.md +++ b/litellm/proxy/client/README.md @@ -331,7 +331,7 @@ sequenceDiagram CLI->>Proxy: Poll /sso/cli/poll/login_id with poll_secret header Proxy->>CLI: Return {"status": "ready", "key": "jwt"} - CLI->>CLI: Save key to ~/.litellm/token.json + CLI->>CLI: Save key to the OS keychain (metadata to ~/.litellm/token.json) ``` ### Authentication Commands @@ -352,7 +352,7 @@ The CLI provides these authentication commands: 5. **Callback Processing**: SSO provider redirects back to proxy with state parameter 6. **User Code Verification**: Browser confirms the verification code shown in the CLI 7. **Polling**: CLI polls `/sso/cli/poll/{login_id}` with the polling secret header until the JWT is ready. When `CLI_SSO_CLAIM_MAP` is configured on the proxy, the poll response may include `attribution_metadata` (allowlisted scalar OIDC claims for client attribution). -8. **Token Storage**: CLI saves the authentication token to `~/.litellm/token.json` +8. **Token Storage**: CLI saves the key to the OS keychain and the non-secret session metadata to `~/.litellm/token.json` ### Benefits of This Approach @@ -364,11 +364,11 @@ The CLI provides these authentication commands: ### Token Storage -Authentication tokens are stored in `~/.litellm/token.json` with restricted file permissions (600). The stored token includes: +The key itself goes into the OS keychain (macOS Keychain, Windows Credential Manager, or the Linux Secret Service) under service `litellm-cli`, account `credential`. Only the non-secret session metadata is written to `~/.litellm/token.json`, in a `0700` directory with `0600` file permissions: ```json { - "key": "sk-...", + "base_url": "https://your-proxy.com", "user_id": "cli-user", "user_email": "user@example.com", "user_role": "cli", @@ -377,6 +377,10 @@ Authentication tokens are stored in `~/.litellm/token.json` with restricted file } ``` +Headless boxes and CI runners usually have no keychain. There the key stays in the same `0600` file alongside the metadata, exactly as it did before, and `lite login` tells you which of the two happened. Set `LITELLM_CLI_DISABLE_KEYRING=1` to force the file even where a keychain exists. A `token.json` written by an older `lite` keeps working and is moved into the keychain, and scrubbed from the file, the first time a keychain-capable `lite` reads it. + +`lite logout` clears both stores. If the keychain is locked at that moment it says so, and re-running it once the keychain is unlocked finishes the job. + The stored credential is a short-lived, per-session agent token, not a managed virtual key. It is scoped to the user and team you logged in as and inherits their models and budgets; spend is tracked against the shared team and user budgets rather than a separate per-session cap, so multiple logins or several concurrent agents all draw down the same allowance. It is short-lived by design (default 24h, configurable via `LITELLM_CLI_JWT_EXPIRATION_HOURS`); re-run `lite login` to refresh it and pick up your latest team and user settings. `lite auth print-token` (usable as Claude Code's `apiKeyHelper`) prints it while fresh and fails once it expires -- there is no silent renewal. It is accepted on a default deployment without `EXPERIMENTAL_UI_LOGIN`, does not appear in the Keys UI, and cannot be rotated or revoked mid-session. For a long-lived, rotatable, Keys-UI-visible credential, create a dedicated virtual key in the dashboard and pass it via `--api-key` or `LITELLM_PROXY_API_KEY`. ### Usage diff --git a/litellm/proxy/client/cli/commands/agents.py b/litellm/proxy/client/cli/commands/agents.py index ed2bf2be03d..e05e85ae483 100644 --- a/litellm/proxy/client/cli/commands/agents.py +++ b/litellm/proxy/client/cli/commands/agents.py @@ -8,7 +8,7 @@ from typing import Final import click import requests -from .auth import get_stored_api_key, login +from .auth import context_secret_vault, get_stored_api_key, login ANTHROPIC_BASE_URL_ENV: Final = "ANTHROPIC_BASE_URL" ANTHROPIC_AUTH_TOKEN_ENV: Final = "ANTHROPIC_AUTH_TOKEN" @@ -316,7 +316,7 @@ def resolve_api_key(ctx: click.Context) -> str: click.echo("No LiteLLM credentials found; starting login...") ctx.invoke(login) - api_key = get_stored_api_key(expected_base_url=base_url) + api_key = get_stored_api_key(expected_base_url=base_url, vault=context_secret_vault(ctx)) if not api_key: raise click.ClickException("Login did not produce an API key; cannot start the agent.") return api_key diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index 0a0bcf80ee5..8b9ef5633da 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -1,9 +1,6 @@ -import json -import os import sys import time import webbrowser -from pathlib import Path from typing import Any, Final from urllib.parse import urlencode @@ -11,10 +8,19 @@ import click import requests from rich.console import Console from rich.table import Table -from typing_extensions import NotRequired, TypedDict +from typing_extensions import NotRequired, ReadOnly, TypedDict from litellm.constants import CLI_JWT_EXPIRATION_HOURS -from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh +from litellm.litellm_core_utils.cli_keyring import SYSTEM_KEYRING, SecretVault +from litellm.litellm_core_utils.cli_token_utils import ( + CliTokenRecord, + clear_cli_token, + get_cli_token_file_path, + get_litellm_gateway_api_key, + is_cli_token_fresh, + load_cli_token, + save_cli_token, +) from .claude_settings import ( CLAUDE_SETTINGS_PATH, @@ -22,18 +28,6 @@ from .claude_settings import ( ClaudeSettingsError, write_claude_settings, ) -from .private_json import write_private_json - - -class CliTokenData(TypedDict): - base_url: str - key: str - user_id: str - user_email: str - user_role: str - auth_header_name: str - jwt_token: str - timestamp: float class CliTeam(TypedDict, total=False): @@ -46,6 +40,7 @@ class CliTeam(TypedDict, total=False): class CliContextObj(TypedDict): base_url: str base_url_explicit: NotRequired[bool] + secret_vault: NotRequired[ReadOnly[SecretVault]] class CliPollData(TypedDict, total=False): @@ -76,50 +71,32 @@ class CliAuthResult(TypedDict): team_id: str | None +KEYCHAIN_UNREACHABLE_MESSAGE: Final = ( + "Your credential is stored in your OS keychain, which could not be read. Unlock it and retry, or run 'lite login'." +) + + # Token storage utilities -def get_token_file_path() -> str: - """Get the path to store the authentication token""" - home_dir: Final = Path.home() - config_dir: Final = home_dir / ".litellm" - config_dir.mkdir(exist_ok=True) - return str(config_dir / "token.json") +def context_secret_vault(ctx: click.Context) -> SecretVault: + """Where this invocation reads and writes secret material; injectable through ctx.obj for tests""" + ctx_obj: Final[CliContextObj | None] = ctx.obj + if ctx_obj is None: + return SYSTEM_KEYRING + return ctx_obj.get("secret_vault") or SYSTEM_KEYRING -def save_token(token_data: CliTokenData) -> None: - """Save token data to file""" - write_private_json(get_token_file_path(), token_data) - - -def load_token() -> CliTokenData | None: - """Load token data from file""" - token_file: Final = get_token_file_path() - if not os.path.exists(token_file): - return None - - try: - with open(token_file, "r") as f: - return json.load(f) - except (OSError, json.JSONDecodeError): - return None - - -def clear_token() -> None: - """Clear stored token""" - token_file: Final = get_token_file_path() - if os.path.exists(token_file): - os.remove(token_file) - - -def get_stored_api_key(expected_base_url: str | None = None) -> str | None: - """Get the stored API key from token file. +def get_stored_api_key( + expected_base_url: str | None = None, + *, + vault: SecretVault = SYSTEM_KEYRING, +) -> str | None: + """Get the stored API key. If expected_base_url is provided, the key is only returned when it was originally issued for that URL. This prevents credential leakage when the CLI is pointed at a different (possibly malicious) server. """ - from litellm.litellm_core_utils.cli_token_utils import get_litellm_gateway_api_key - - return get_litellm_gateway_api_key(expected_base_url=expected_base_url) + return get_litellm_gateway_api_key(expected_base_url=expected_base_url, vault=vault) # Team selection utilities @@ -689,23 +666,27 @@ def login(ctx: click.Context, config_claude: bool): api_key: Final = auth_result["api_key"] user_id: Final = auth_result["user_id"] - # Save token data. base_url is stored so we can verify origin - # before reusing the key on a subsequent CLI invocation. - save_token( - { - "base_url": base_url.rstrip("/"), - "key": api_key, - "user_id": user_id or "cli-user", - "user_email": "unknown", - "user_role": "cli", - "auth_header_name": "Authorization", - "jwt_token": "", - "timestamp": time.time(), - } + # base_url is stored so we can verify origin before reusing the + # key on a subsequent CLI invocation. + record: Final = CliTokenRecord( + base_url=base_url.rstrip("/"), + key=api_key, + user_id=user_id or "cli-user", + user_email="unknown", + user_role="cli", + auth_header_name="Authorization", + jwt_token="", + timestamp=time.time(), ) + in_keychain: Final = save_cli_token(record, vault=context_secret_vault(ctx)) click.echo("\nLogin successful!") click.echo(f"JWT Token: {api_key[:20]}...") + click.echo( + "Credential stored in your OS keychain." + if in_keychain + else f"No OS keychain available; credential stored in {get_cli_token_file_path()} (owner-only)." + ) click.echo("You can now use the CLI without specifying --api-key") if config_claude: @@ -736,10 +717,14 @@ def login(ctx: click.Context, config_claude: bool): @click.command(name="logout") -def logout(): +@click.pass_context +def logout(ctx: click.Context): """Logout and clear stored authentication""" - clear_token() - click.echo("Logged out successfully. Authentication token cleared.") + if clear_cli_token(vault=context_secret_vault(ctx)): + click.echo("Logged out successfully. Authentication token cleared.") + return + click.echo("Logged out. The local token file is gone, but the OS keychain entry could not be removed.") + click.echo("Unlock your keychain and run 'lite logout' again to clear it.") @click.command(name="print-token") @@ -753,7 +738,7 @@ def print_token(ctx: click.Context): expires after `LITELLM_CLI_JWT_EXPIRATION_HOURS` (default 24h); once expired, run `lite login` again. """ - token_data: Final = load_token() + token_data: Final = load_cli_token(vault=context_secret_vault(ctx)) if not token_data: click.echo("Not authenticated. Run 'lite login'.", err=True) sys.exit(1) @@ -765,7 +750,7 @@ def print_token(ctx: click.Context): ctx_obj: Final[CliContextObj] = ctx.obj if ctx_obj.get("base_url_explicit"): base_url: Final = ctx_obj["base_url"] - if token_data.get("base_url") != base_url.rstrip("/"): + if token_data.base_url != base_url.rstrip("/"): click.echo("Not authenticated for this server. Run 'lite login'.", err=True) sys.exit(1) @@ -773,33 +758,36 @@ def print_token(ctx: click.Context): click.echo("Token expired. Run 'lite login' again.", err=True) sys.exit(1) - api_key: Final = token_data.get("key") + api_key: Final = token_data.key if not api_key: - click.echo("No token available. Run 'lite login'.", err=True) + click.echo(KEYCHAIN_UNREACHABLE_MESSAGE, err=True) sys.exit(1) click.echo(api_key) @click.command(name="whoami") -def whoami(): +@click.pass_context +def whoami(ctx: click.Context): """Show current authentication status""" - token_data: Final = load_token() + token_data: Final = load_cli_token(vault=context_secret_vault(ctx)) if not token_data: click.echo("Not authenticated. Run 'lite login' to authenticate.") return click.echo("Authenticated") - click.echo(f"User Email: {token_data.get('user_email', 'Unknown')}") - click.echo(f"User ID: {token_data.get('user_id', 'Unknown')}") - click.echo(f"User Role: {token_data.get('user_role', 'Unknown')}") + click.echo(f"User Email: {token_data.user_email or 'Unknown'}") + click.echo(f"User ID: {token_data.user_id or 'Unknown'}") + click.echo(f"User Role: {token_data.user_role or 'Unknown'}") # Check if token is still valid (basic timestamp check) - timestamp: Final = token_data.get("timestamp", 0) - age_hours: Final = (time.time() - timestamp) / 3600 + age_hours: Final = (time.time() - token_data.timestamp) / 3600 click.echo(f"Token age: {age_hours:.1f} hours") + if token_data.key is None: + click.echo(KEYCHAIN_UNREACHABLE_MESSAGE) + if age_hours > CLI_JWT_EXPIRATION_HOURS: click.echo(f"Warning: Token is more than {CLI_JWT_EXPIRATION_HOURS} hours old and may have expired.") diff --git a/litellm/proxy/client/cli/commands/claude_settings.py b/litellm/proxy/client/cli/commands/claude_settings.py index e9a6a25a064..e18e5b1b7ee 100644 --- a/litellm/proxy/client/cli/commands/claude_settings.py +++ b/litellm/proxy/client/cli/commands/claude_settings.py @@ -15,7 +15,7 @@ from typing import Final from pydantic import JsonValue, TypeAdapter, ValidationError -from .private_json import write_private_json +from litellm.litellm_core_utils.private_json import write_private_json ENV_KEY: Final = "env" API_KEY_HELPER_KEY: Final = "apiKeyHelper" diff --git a/litellm/proxy/client/cli/commands/config.py b/litellm/proxy/client/cli/commands/config.py index 19dd407ba19..2715a0a9a38 100644 --- a/litellm/proxy/client/cli/commands/config.py +++ b/litellm/proxy/client/cli/commands/config.py @@ -10,7 +10,7 @@ from urllib.parse import urlparse import click from pydantic import TypeAdapter -from .private_json import write_private_json +from litellm.litellm_core_utils.private_json import ensure_private_dir, write_private_json HIDDEN_COMMANDS_KEY: Final = "hidden_commands" @@ -42,7 +42,9 @@ def load_config() -> Mapping[str, str]: def save_config(config: Mapping[str, str]) -> None: """Save CLI config to file""" - write_private_json(get_config_file_path(), config) + config_file: Final = Path(get_config_file_path()) + ensure_private_dir(config_file.parent) + write_private_json(str(config_file), config) def get_config_value(key: str) -> str | None: diff --git a/litellm/proxy/client/cli/commands/up.py b/litellm/proxy/client/cli/commands/up.py index dd266b4afa1..a0fd4af8f72 100644 --- a/litellm/proxy/client/cli/commands/up.py +++ b/litellm/proxy/client/cli/commands/up.py @@ -14,10 +14,12 @@ from typing import IO, Final import click from pydantic import JsonValue, TypeAdapter, ValidationError -from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh +from litellm.litellm_core_utils.cli_keyring import SecretVault +from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh, load_cli_token +from litellm.litellm_core_utils.private_json import ensure_private_dir from .agents import AgentRunError, resolve_api_key, verify_proxy_key -from .auth import load_token, login +from .auth import context_secret_vault, login from .claude_settings import ( BACKUP_PATH, CLAUDE_SETTINGS_PATH, @@ -66,7 +68,7 @@ def secure_create(path: Path) -> Iterator[IO[str]]: def write_backup(record: BackupRecord, backup_path: Path | None = None) -> None: path: Final = backup_path if backup_path is not None else BACKUP_PATH - path.parent.mkdir(exist_ok=True) + ensure_private_dir(path.parent) with secure_create(path) as f: json.dump({"existed": record.existed, "content": record.content}, f, indent=2) @@ -103,10 +105,17 @@ def restore_claude_settings(settings_path: Path | None = None, backup_path: Path return record +def _has_fresh_login(base_url: str, vault: SecretVault) -> bool: + token_data: Final = load_cli_token(vault=vault) + if token_data is None or token_data.key is None or token_data.base_url != base_url: + return False + return is_cli_token_fresh(token_data) + + def _ensure_fresh_login(ctx: click.Context) -> None: base_url: Final = ctx.obj["base_url"].rstrip("/") - token_data = load_token() - if token_data and token_data.get("base_url") == base_url and is_cli_token_fresh(token_data): + vault: Final = context_secret_vault(ctx) + if _has_fresh_login(base_url, vault): return if not sys.stdin.isatty(): @@ -117,8 +126,7 @@ def _ensure_fresh_login(ctx: click.Context) -> None: click.echo("No fresh LiteLLM login found for this proxy; starting login...") ctx.invoke(login) - token_data = load_token() - if not token_data or token_data.get("base_url") != base_url or not is_cli_token_fresh(token_data): + if not _has_fresh_login(base_url, vault): raise UpError("Login did not produce a usable token; cannot start `lite up`.") diff --git a/litellm/proxy/client/cli/main.py b/litellm/proxy/client/cli/main.py index 3a289736c66..664bf5a216c 100644 --- a/litellm/proxy/client/cli/main.py +++ b/litellm/proxy/client/cli/main.py @@ -9,7 +9,7 @@ from litellm._version import version as litellm_version from litellm.proxy.client.health import HealthManagementClient from .commands.agents import agent_commands -from .commands.auth import auth_group, get_stored_api_key, login, logout, whoami +from .commands.auth import auth_group, context_secret_vault, get_stored_api_key, login, logout, whoami from .commands.autoroute.commands import autoroute_group from .commands.chat import chat from .commands.config import config_commands, get_config_value, hidden_command_names @@ -94,7 +94,7 @@ def cli(ctx: click.Context, show_version: bool, base_url: str | None, api_key: s # If no API key provided via flag or environment variable, try to load from saved token. # Pass base_url so we only use the stored key when it was issued for this server. if api_key is None: - api_key = get_stored_api_key(expected_base_url=base_url) + api_key = get_stored_api_key(expected_base_url=base_url, vault=context_secret_vault(ctx)) ctx.obj["base_url"] = base_url ctx.obj["api_key"] = api_key diff --git a/pyproject.toml b/pyproject.toml index ffbc96eefb9..32921e14d31 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -85,6 +85,7 @@ cli = [ "pyyaml>=6.0.3,<7.0", "requests>=2.32.0,<3.0", "InquirerPy>=0.3.4,<1.0", + "keyring>=25.6.0,<26.0", ] extra_proxy = [ "prisma>=0.11.0,<1.0", @@ -166,6 +167,7 @@ litellm-proxy = "litellm.proxy.client.cli:cli" dev = [ "diff-cover==9.7.2", "basedpyright==1.39.7", + "keyring==25.7.0", "pytest==9.0.3", "pytest-mock==3.15.1", "pytest-asyncio==1.3.0", diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py index c0644c88291..ce0fd197538 100644 --- a/tests/test_litellm/conftest.py +++ b/tests/test_litellm/conftest.py @@ -22,6 +22,12 @@ import litellm from litellm import router as litellm_router_module from litellm import utils as litellm_utils_module from litellm._logging import ALL_LOGGERS +from litellm.litellm_core_utils.cli_keyring import ( + SecretFound, + SecretMissing, + SecretRead, + SecretUnavailable, +) from litellm.litellm_core_utils.prompt_templates import ( image_handling as image_handling_module, ) @@ -106,6 +112,65 @@ def isolate_host_proxy_base_url(monkeypatch): monkeypatch.delenv("PROXY_BASE_URL", raising=False) +@pytest.fixture(scope="function", autouse=True) +def isolate_host_os_keychain(monkeypatch): + """Keep any code path that resolves a CLI credential out of the developer's real OS keychain. + + Tests that exercise keychain behaviour inject their own vault instead. + """ + monkeypatch.setenv("LITELLM_CLI_DISABLE_KEYRING", "1") + + +class FakeSecretVault: + """In-memory stand-in for the OS keychain, injected wherever CLI credential storage is exercised. + + `available=False` models a keychain that is locked or has no backend, `writable=False` one that + refuses to store, and `erasable=False` one that will not release what it already holds. + """ + + def __init__( + self, + blob: str | None = None, + *, + available: bool = True, + writable: bool = True, + erasable: bool = True, + ) -> None: + self.blob: str | None = blob + self.available: bool = available + self.writable: bool = writable + self.erasable: bool = erasable + self.reads: int = 0 + self.writes: list[str] = [] + self.erases: int = 0 + + def read(self) -> SecretRead: + self.reads += 1 + if not self.available: + return SecretUnavailable() + return SecretMissing() if self.blob is None else SecretFound(self.blob) + + def write(self, blob: str) -> bool: + self.writes.append(blob) + if not (self.available and self.writable): + return False + self.blob = blob + return True + + def erase(self) -> bool: + self.erases += 1 + if not (self.available and self.erasable): + return False + self.blob = None + return True + + +@pytest.fixture +def secret_vault_factory(): + """Build FakeSecretVault instances; see its docstring for the failure modes it can model.""" + return FakeSecretVault + + def _run_coroutine_if_needed(result): if not asyncio.iscoroutine(result): return diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 27fc5eb4bd0..56ab6bcbfe0 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -1,89 +1,429 @@ -""" -Unit tests for CLI token utilities -""" - import json -import os -import tempfile -from pathlib import Path -from unittest.mock import mock_open, patch +import stat +import sys +import time import pytest -from litellm.litellm_core_utils.cli_token_utils import get_litellm_gateway_api_key +from litellm.constants import CLI_JWT_EXPIRATION_HOURS +from litellm.litellm_core_utils.cli_keyring import ( + DISABLE_KEYRING_ENV_VAR, + KEYRING_ACCOUNT, + KEYRING_SERVICE, + KeyringVault, + SecretFound, + SecretMissing, + SecretUnavailable, +) +from litellm.litellm_core_utils.cli_token_utils import ( + CliTokenRecord, + clear_cli_token, + get_cli_token_file_path, + get_litellm_gateway_api_key, + is_cli_token_fresh, + load_cli_token, + save_cli_token, +) + +SERVER = "https://proxy.example.com" +OTHER_SERVER = "https://other-proxy.example.com" -class TestCLITokenUtils: - """Test CLI token utility functions""" +@pytest.fixture +def isolated_home(monkeypatch, tmp_path): + monkeypatch.setenv("HOME", str(tmp_path)) + monkeypatch.setenv("USERPROFILE", str(tmp_path)) + return tmp_path - def test_get_litellm_gateway_api_key_success(self): - """Test getting CLI API key when token file exists and is valid""" - token_data = { - "key": "sk-test-cli-key-123", - "user_id": "test-user", - "user_email": "test@example.com", - "timestamp": 1234567890, - } - with ( - patch("os.path.exists", return_value=True), - patch("builtins.open", mock_open(read_data=json.dumps(token_data))), - patch( - "litellm.litellm_core_utils.cli_token_utils.get_cli_token_file_path", - return_value="/test/.litellm/token.json", - ), - ): +def _token_file(home): + return home / ".litellm" / "token.json" - result = get_litellm_gateway_api_key() - assert result == "sk-test-cli-key-123" +def _write_legacy_file(home, **overrides): + payload = { + "base_url": SERVER, + "key": "sk-legacy", + "user_id": "u-1", + "user_email": "user@example.com", + "user_role": "cli", + "timestamp": time.time(), + **overrides, + } + path = _token_file(home) + path.parent.mkdir(exist_ok=True) + path.write_text(json.dumps(payload)) + path.chmod(0o600) + return path - def test_get_litellm_gateway_api_key_no_file(self): - """Test getting CLI API key when token file doesn't exist""" - with ( - patch("os.path.exists", return_value=False), - patch( - "litellm.litellm_core_utils.cli_token_utils.get_cli_token_file_path", - return_value="/test/.litellm/token.json", - ), - ): - result = get_litellm_gateway_api_key() +def _write_metadata_only_file(home): + """What a post-migration token.json looks like: everything except the secret material.""" + path = _token_file(home) + path.parent.mkdir(exist_ok=True) + path.write_text(json.dumps({"base_url": SERVER, "user_id": "u-1", "timestamp": time.time()})) + path.chmod(0o600) + return path - assert result is None - def test_get_litellm_gateway_api_key_invalid_json(self): - """Test getting CLI API key when token file has invalid JSON""" - with ( - patch("os.path.exists", return_value=True), - patch("builtins.open", mock_open(read_data="invalid json")), - patch( - "litellm.litellm_core_utils.cli_token_utils.get_cli_token_file_path", - return_value="/test/.litellm/token.json", - ), - ): +def _blob(base_url=SERVER, key="sk-vault", jwt_token=""): + return json.dumps({"base_url": base_url, "key": key, "jwt_token": jwt_token}) - result = get_litellm_gateway_api_key() - assert result is None +class TestGetCliTokenFilePath: + def test_points_at_the_home_config_file(self, isolated_home): + assert get_cli_token_file_path() == str(isolated_home / ".litellm" / "token.json") - def test_get_litellm_gateway_api_key_no_key_field(self): - """Test getting CLI API key when token file exists but has no key field""" - token_data = { - "user_id": "test-user", - "user_email": "test@example.com", - # Missing 'key' field - } + def test_does_not_create_the_directory(self, isolated_home): + """Merely asking for the path must not leave a directory behind, so an SDK import that + never logs in cannot create a ~/.litellm on someone's machine.""" + get_cli_token_file_path() - with ( - patch("os.path.exists", return_value=True), - patch("builtins.open", mock_open(read_data=json.dumps(token_data))), - patch( - "litellm.litellm_core_utils.cli_token_utils.get_cli_token_file_path", - return_value="/test/.litellm/token.json", - ), - ): + assert not (isolated_home / ".litellm").exists() - result = get_litellm_gateway_api_key() - assert result is None +class TestLoadCliToken: + def test_no_token_file_never_touches_the_keychain(self, isolated_home, secret_vault_factory): + """The SDK calls this on machines that never ran `lite login`; it must not prompt for + keychain access there.""" + vault = secret_vault_factory(blob=_blob()) + + assert load_cli_token(vault=vault) is None + assert vault.reads == 0 + + def test_secret_comes_from_the_vault_when_the_file_holds_only_metadata(self, isolated_home, secret_vault_factory): + _write_metadata_only_file(isolated_home) + vault = secret_vault_factory(blob=_blob(key="sk-from-keychain")) + + record = load_cli_token(vault=vault) + + assert record.key == "sk-from-keychain" + assert "sk-from-keychain" not in _token_file(isolated_home).read_text() + + def test_jwt_token_round_trips_through_the_vault(self, isolated_home, secret_vault_factory): + _write_metadata_only_file(isolated_home) + vault = secret_vault_factory(blob=_blob(key="sk-a", jwt_token="jwt-a")) + + record = load_cli_token(vault=vault) + + assert (record.key, record.jwt_token) == ("sk-a", "jwt-a") + + def test_legacy_plaintext_file_still_authenticates_and_is_migrated(self, isolated_home, secret_vault_factory): + """A token.json written by an older `lite` keeps working, and reading it moves the secret + into the keychain and scrubs it from disk.""" + path = _write_legacy_file(isolated_home) + vault = secret_vault_factory() + + record = load_cli_token(vault=vault) + + assert record.key == "sk-legacy" + assert json.loads(vault.blob)["key"] == "sk-legacy" + on_disk = json.loads(path.read_text()) + assert "key" not in on_disk + assert on_disk["user_email"] == "user@example.com" + assert stat.S_IMODE(path.stat().st_mode) == 0o600 + + def test_legacy_file_survives_a_vault_that_refuses_to_store(self, isolated_home, secret_vault_factory): + """Scrubbing the only copy of the secret after a failed keychain write would log the user + out for good.""" + path = _write_legacy_file(isolated_home) + before = path.read_text() + + record = load_cli_token(vault=secret_vault_factory(writable=False)) + + assert record.key == "sk-legacy" + assert path.read_text() == before + + def test_a_secret_left_on_disk_outranks_a_stale_keychain_entry(self, isolated_home, secret_vault_factory): + """A failed keychain write leaves the fresh secret on disk while the vault still holds the + previous one; the next read must serve the file's secret and move it into the vault, never + resurrect the stale key or scrub the only copy of the fresh one.""" + path = _write_legacy_file(isolated_home, key="sk-fresh") + vault = secret_vault_factory(blob=_blob(key="sk-stale")) + + record = load_cli_token(vault=vault) + + assert record.key == "sk-fresh" + assert json.loads(vault.blob)["key"] == "sk-fresh" + assert "key" not in json.loads(path.read_text()) + + def test_a_disk_secret_survives_when_the_stale_vault_refuses_the_rewrite( + self, isolated_home, secret_vault_factory + ): + path = _write_legacy_file(isolated_home, key="sk-fresh") + before = path.read_text() + + record = load_cli_token(vault=secret_vault_factory(blob=_blob(key="sk-stale"), writable=False)) + + assert record.key == "sk-fresh" + assert path.read_text() == before + + def test_legacy_file_survives_an_unreachable_vault_without_write_attempts( + self, isolated_home, secret_vault_factory + ): + path = _write_legacy_file(isolated_home) + before = path.read_text() + vault = secret_vault_factory(available=False) + + record = load_cli_token(vault=vault) + + assert record.key == "sk-legacy" + assert vault.writes == [] + assert path.read_text() == before + + def test_metadata_only_file_with_an_empty_vault_is_not_a_login(self, isolated_home, secret_vault_factory): + _write_metadata_only_file(isolated_home) + + assert load_cli_token(vault=secret_vault_factory()) is None + + def test_metadata_only_file_with_an_unreachable_vault_reports_a_missing_secret( + self, isolated_home, secret_vault_factory + ): + """The caller needs to tell "never logged in" apart from "locked keychain", so the record + comes back with no key rather than as None.""" + _write_metadata_only_file(isolated_home) + + record = load_cli_token(vault=secret_vault_factory(available=False)) + + assert record.key is None + assert record.user_id == "u-1" + + def test_a_secret_minted_for_another_server_is_never_handed_out(self, isolated_home, secret_vault_factory): + _write_metadata_only_file(isolated_home) + + assert load_cli_token(vault=secret_vault_factory(blob=_blob(base_url=OTHER_SERVER))) is None + + def test_a_secret_minted_for_another_server_loses_to_the_file(self, isolated_home, secret_vault_factory): + _write_legacy_file(isolated_home) + vault = secret_vault_factory(blob=_blob(base_url=OTHER_SERVER, key="sk-elsewhere")) + + record = load_cli_token(vault=vault) + + assert record.key == "sk-legacy" + assert json.loads(vault.blob)["key"] == "sk-legacy" + + def test_unreadable_vault_blob_falls_back_to_the_file_secret(self, isolated_home, secret_vault_factory): + _write_legacy_file(isolated_home) + + record = load_cli_token(vault=secret_vault_factory(blob="not json at all {{{")) + + assert record.key == "sk-legacy" + + def test_corrupt_token_file_is_not_a_login(self, isolated_home, secret_vault_factory): + _token_file(isolated_home).parent.mkdir() + _token_file(isolated_home).write_text("not json at all {{{") + + assert load_cli_token(vault=secret_vault_factory(blob=_blob())) is None + + +class TestGetLitellmGatewayApiKey: + def test_returns_the_vault_secret_when_the_origin_matches(self, isolated_home, secret_vault_factory): + _write_metadata_only_file(isolated_home) + + key = get_litellm_gateway_api_key(expected_base_url=SERVER, vault=secret_vault_factory(blob=_blob())) + + assert key == "sk-vault" + + def test_trailing_slash_on_the_expected_url_is_normalised(self, isolated_home, secret_vault_factory): + _write_metadata_only_file(isolated_home) + + key = get_litellm_gateway_api_key(expected_base_url=SERVER + "/", vault=secret_vault_factory(blob=_blob())) + + assert key == "sk-vault" + + def test_origin_mismatch_returns_nothing_without_reading_the_keychain(self, isolated_home, secret_vault_factory): + """Pointing the SDK at a different server must fail before the keychain is even consulted, + so a hostile base_url cannot provoke an unlock prompt.""" + _write_legacy_file(isolated_home) + vault = secret_vault_factory(blob=_blob()) + + assert get_litellm_gateway_api_key(expected_base_url=OTHER_SERVER, vault=vault) is None + assert vault.reads == 0 + + def test_no_token_file_returns_nothing(self, isolated_home, secret_vault_factory): + assert get_litellm_gateway_api_key(vault=secret_vault_factory(blob=_blob())) is None + + +class TestSaveCliToken: + def test_secret_goes_to_the_keychain_and_never_to_the_file(self, isolated_home, secret_vault_factory): + vault = secret_vault_factory() + + stored = save_cli_token( + CliTokenRecord(base_url=SERVER, key="sk-new", user_id="u-1", timestamp=time.time()), + vault=vault, + ) + + assert stored is True + assert "sk-new" not in _token_file(isolated_home).read_text() + assert json.loads(vault.blob)["key"] == "sk-new" + assert load_cli_token(vault=vault).key == "sk-new" + + def test_falls_back_to_the_owner_only_file_when_there_is_no_keychain(self, isolated_home, secret_vault_factory): + stored = save_cli_token( + CliTokenRecord(base_url=SERVER, key="sk-new", timestamp=time.time()), + vault=secret_vault_factory(available=False), + ) + + path = _token_file(isolated_home) + assert stored is False + assert json.loads(path.read_text())["key"] == "sk-new" + assert stat.S_IMODE(path.stat().st_mode) == 0o600 + assert list(path.parent.glob(".tmp-*")) == [] + + def test_creates_the_config_directory_owner_only(self, isolated_home, secret_vault_factory): + """A 0755 ~/.litellm lets any local process list, and in the fallback case read, the + credential's directory.""" + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=secret_vault_factory()) + + assert stat.S_IMODE((isolated_home / ".litellm").stat().st_mode) == 0o700 + + def test_tightens_a_directory_left_group_readable_by_an_older_cli(self, isolated_home, secret_vault_factory): + config_dir = isolated_home / ".litellm" + config_dir.mkdir(mode=0o755) + + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=secret_vault_factory()) + + assert stat.S_IMODE(config_dir.stat().st_mode) == 0o700 + + def test_a_failed_write_leaves_the_previous_credential_intact(self, isolated_home, secret_vault_factory, monkeypatch): + path = _write_legacy_file(isolated_home) + before = path.read_text() + + def _explode(*args, **kwargs): + raise TypeError("not serialisable") + + monkeypatch.setattr("litellm.litellm_core_utils.private_json.json.dump", _explode) + + with pytest.raises(TypeError): + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=secret_vault_factory(available=False)) + + assert path.read_text() == before + assert list(path.parent.glob(".tmp-*")) == [] + + +class TestClearCliToken: + def test_removes_the_credential_from_both_stores(self, isolated_home, secret_vault_factory): + vault = secret_vault_factory() + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=vault) + + assert clear_cli_token(vault=vault) is True + assert vault.blob is None + assert not _token_file(isolated_home).exists() + assert load_cli_token(vault=vault) is None + + def test_reports_a_keychain_that_will_not_release_the_secret(self, isolated_home, secret_vault_factory): + _write_legacy_file(isolated_home) + vault = secret_vault_factory(blob=_blob(), erasable=False) + + assert clear_cli_token(vault=vault) is False + assert not _token_file(isolated_home).exists() + + def test_is_safe_when_nothing_was_ever_stored(self, isolated_home, secret_vault_factory): + assert clear_cli_token(vault=secret_vault_factory()) is True + + +class TestIsCliTokenFresh: + def test_a_just_issued_token_is_fresh(self): + assert is_cli_token_fresh(CliTokenRecord(timestamp=time.time())) is True + + def test_a_token_past_its_expiry_is_stale(self): + stale = CliTokenRecord(timestamp=time.time() - (CLI_JWT_EXPIRATION_HOURS + 1) * 3600) + + assert is_cli_token_fresh(stale) is False + + def test_the_buffer_retires_a_token_just_before_it_expires(self): + almost = CliTokenRecord(timestamp=time.time() - (CLI_JWT_EXPIRATION_HOURS * 3600 - 60)) + + assert is_cli_token_fresh(almost, buffer_hours=0.1) is False + + +class _FakeKeyringModule: + def __init__(self, stored=None, *, get_error=None, set_error=None, delete_error=None): + self.stored = stored + self.get_error = get_error + self.set_error = set_error + self.delete_error = delete_error + self.calls = [] + + def get_password(self, service_name, username): + self.calls.append(("get", service_name, username)) + if self.get_error is not None: + raise self.get_error + return self.stored + + def set_password(self, service_name, username, password): + self.calls.append(("set", service_name, username)) + if self.set_error is not None: + raise self.set_error + self.stored = password + + def delete_password(self, service_name, username): + self.calls.append(("delete", service_name, username)) + if self.delete_error is not None: + raise self.delete_error + self.stored = None + + +@pytest.fixture +def install_fake_keyring(monkeypatch): + def _install(fake): + monkeypatch.delenv(DISABLE_KEYRING_ENV_VAR, raising=False) + monkeypatch.setitem(sys.modules, "keyring", fake) + return fake + + return _install + + +class TestKeyringVault: + def test_round_trips_through_the_installed_keyring(self, install_fake_keyring): + fake = install_fake_keyring(_FakeKeyringModule()) + vault = KeyringVault() + + assert vault.write("blob-1") is True + assert vault.read() == SecretFound("blob-1") + assert vault.erase() is True + assert vault.read() == SecretMissing() + assert {call[1:] for call in fake.calls} == {(KEYRING_SERVICE, KEYRING_ACCOUNT)} + + def test_the_kill_switch_reports_no_keychain(self, monkeypatch): + """`LITELLM_CLI_DISABLE_KEYRING` has to work without importing keyring, because keyring + caches its backend on first use and cannot be reconfigured later. Erase still fails: a + credential stored before the switch was set may be in the keychain, and with reads + disabled `lite logout` cannot verify it is gone, so it must warn instead.""" + monkeypatch.setenv(DISABLE_KEYRING_ENV_VAR, "1") + vault = KeyringVault() + + assert vault.read() == SecretUnavailable() + assert vault.write("blob-1") is False + assert vault.erase() is False + + def test_an_uninstalled_keyring_library_degrades_to_the_file(self, monkeypatch): + """keyring is an optional extra, so the SDK must survive its absence rather than raise on + the hot path.""" + monkeypatch.delenv(DISABLE_KEYRING_ENV_VAR, raising=False) + monkeypatch.setitem(sys.modules, "keyring", None) + vault = KeyringVault() + + assert vault.read() == SecretUnavailable() + assert vault.write("blob-1") is False + assert vault.erase() is True + + def test_a_locked_keychain_is_reported_not_raised(self, install_fake_keyring): + install_fake_keyring(_FakeKeyringModule(get_error=RuntimeError("keyring is locked"))) + + assert KeyringVault().read() == SecretUnavailable() + + def test_a_refused_write_is_reported_not_raised(self, install_fake_keyring): + install_fake_keyring(_FakeKeyringModule(set_error=RuntimeError("no backend"))) + + assert KeyringVault().write("blob-1") is False + + def test_a_refused_delete_is_reported_so_logout_can_warn(self, install_fake_keyring): + install_fake_keyring(_FakeKeyringModule(stored="blob-1", delete_error=RuntimeError("locked"))) + + assert KeyringVault().erase() is False + + def test_erasing_a_locked_keychain_is_a_failure(self, install_fake_keyring): + install_fake_keyring(_FakeKeyringModule(get_error=RuntimeError("locked"))) + + assert KeyringVault().erase() is False diff --git a/tests/test_litellm/proxy/client/cli/test_agents.py b/tests/test_litellm/proxy/client/cli/test_agents.py index a23c573047f..c2858c84c6d 100644 --- a/tests/test_litellm/proxy/client/cli/test_agents.py +++ b/tests/test_litellm/proxy/client/cli/test_agents.py @@ -672,8 +672,9 @@ class TestAgentCommands: assert "LITELLM_PROXY_API_KEY" in result.output mock_run.assert_not_called() - def test_interactive_without_key_logs_in_then_launches(self): + def test_interactive_without_key_logs_in_then_launches(self, secret_vault_factory): captured = {} + vault = secret_vault_factory() @click.command() def fake_login(): @@ -695,12 +696,12 @@ class TestAgentCommands: result = self.runner.invoke( _agent_command("claude"), [], - obj={"base_url": "http://localhost:4000", "api_key": None}, + obj={"base_url": "http://localhost:4000", "api_key": None, "secret_vault": vault}, ) assert result.exit_code == 0, result.output assert captured["api_key"] == "sk-after-login" - mock_get.assert_called_once_with(expected_base_url="http://localhost:4000") + mock_get.assert_called_once_with(expected_base_url="http://localhost:4000", vault=vault) def test_child_exit_code_reaches_the_shell(self): with patch(f"{AGENTS_MODULE}.run_agent", side_effect=SystemExit(42)): diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 59048067674..e93f05cb4aa 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -4,7 +4,7 @@ import stat import sys import time from pathlib import Path -from unittest.mock import Mock, mock_open, patch +from unittest.mock import Mock, patch sys.path.insert(0, os.path.abspath("../../..")) # Adds the parent directory to the system path @@ -13,21 +13,38 @@ import pytest from click.testing import CliRunner from litellm.constants import CLI_JWT_EXPIRATION_HOURS +from litellm.litellm_core_utils.cli_token_utils import CliTokenRecord, save_cli_token from litellm.proxy.client.cli import cli from litellm.proxy.client.cli.commands.auth import ( - clear_token, + KEYCHAIN_UNREACHABLE_MESSAGE, get_stored_api_key, - get_token_file_path, - load_token, login, logout, print_token, - save_token, whoami, ) from litellm.proxy.client.cli.commands.claude_settings import SettingsFileOwner +@pytest.fixture +def isolated_home(monkeypatch, tmp_path): + monkeypatch.setenv("HOME", str(tmp_path)) + monkeypatch.setenv("USERPROFILE", str(tmp_path)) + monkeypatch.delenv("LITELLM_PROXY_URL", raising=False) + monkeypatch.delenv("LITELLM_PROXY_API_KEY", raising=False) + return tmp_path + + +def _write_home_json(home: Path, filename: str, payload: dict[str, object]) -> None: + litellm_dir = home / ".litellm" + litellm_dir.mkdir(exist_ok=True) + (litellm_dir / filename).write_text(json.dumps(payload)) + + +def _secret_blob(base_url: str, key: str) -> str: + return json.dumps({"base_url": base_url, "key": key, "jwt_token": ""}) + + def _mock_cli_sso_start_response( login_id: str = "cli-session-uuid-456", poll_secret: str = "poll-secret", @@ -176,200 +193,50 @@ class TestStartCliSsoFlowErrors: assert "https://unreachable.example.com/sso/cli/start" in message -class TestTokenUtilities: - """Test token file utility functions""" +class TestStoredApiKeyLookup: + """`get_stored_api_key` is what every other `lite` subcommand authenticates with, so the + keychain split and the origin check both have to be invisible to it.""" - def test_get_token_file_path(self): - """Test getting token file path""" - with ( - patch("pathlib.Path.home") as mock_home, - patch("pathlib.Path.mkdir") as mock_mkdir, - ): - mock_home.return_value = Path("/home/user") + def test_returns_the_secret_the_keychain_holds(self, isolated_home, secret_vault_factory): + _write_home_json(isolated_home, "token.json", {"base_url": "https://real-proxy.com", "user_id": "u-1"}) + vault = secret_vault_factory(blob=_secret_blob("https://real-proxy.com", "sk-from-keychain")) - result = get_token_file_path() + assert get_stored_api_key(vault=vault) == "sk-from-keychain" - assert result == "/home/user/.litellm/token.json" - mock_mkdir.assert_called_once_with(exist_ok=True) + def test_returns_a_legacy_plaintext_key(self, isolated_home, secret_vault_factory): + _write_home_json(isolated_home, "token.json", {"base_url": "https://real-proxy.com", "key": "sk-legacy"}) - def test_get_token_file_path_creates_directory(self): - """Test that get_token_file_path creates the config directory""" - with ( - patch("pathlib.Path.home") as mock_home, - patch("pathlib.Path.mkdir") as mock_mkdir, - ): - mock_home.return_value = Path("/home/user") + assert get_stored_api_key(vault=secret_vault_factory()) == "sk-legacy" - get_token_file_path() + def test_no_token_at_all_returns_nothing(self, isolated_home, secret_vault_factory): + assert get_stored_api_key(vault=secret_vault_factory()) is None - mock_mkdir.assert_called_once_with(exist_ok=True) + def test_metadata_without_a_secret_returns_nothing(self, isolated_home, secret_vault_factory): + _write_home_json(isolated_home, "token.json", {"base_url": "https://real-proxy.com", "user_id": "u-1"}) - def test_save_token(self, tmp_path): - """Test saving token data to file""" - token_data = { - "key": "test-key", - "user_id": "test-user", - "timestamp": 1234567890, - } - token_file = tmp_path / "token.json" + assert get_stored_api_key(vault=secret_vault_factory()) is None - with patch("litellm.proxy.client.cli.commands.auth.get_token_file_path") as mock_path: - mock_path.return_value = str(token_file) + def test_matching_base_url_returns_the_key(self, isolated_home, secret_vault_factory): + _write_home_json(isolated_home, "token.json", {"base_url": "https://real-proxy.com", "key": "sk-prod"}) - save_token(token_data) + assert get_stored_api_key("https://real-proxy.com", vault=secret_vault_factory()) == "sk-prod" - assert json.loads(token_file.read_text()) == token_data - assert stat.S_IMODE(token_file.stat().st_mode) == 0o600 + def test_trailing_slash_on_the_expected_url_is_normalised(self, isolated_home, secret_vault_factory): + _write_home_json(isolated_home, "token.json", {"base_url": "https://real-proxy.com", "key": "sk-prod"}) - def test_load_token_success(self): - """Test loading token data from file successfully""" - token_data = { - "key": "test-key", - "user_id": "test-user", - "timestamp": 1234567890, - } + assert get_stored_api_key("https://real-proxy.com/", vault=secret_vault_factory()) == "sk-prod" - with ( - patch("builtins.open", mock_open(read_data=json.dumps(token_data))), - patch("litellm.proxy.client.cli.commands.auth.get_token_file_path") as mock_path, - patch("os.path.exists", return_value=True), - ): - mock_path.return_value = "/test/path/token.json" + def test_mismatched_base_url_withholds_the_key(self, isolated_home, secret_vault_factory): + _write_home_json(isolated_home, "token.json", {"base_url": "https://real-proxy.com", "key": "sk-prod"}) - result = load_token() + assert get_stored_api_key("https://evil.com", vault=secret_vault_factory()) is None - assert result == token_data + def test_old_tokens_without_a_base_url_are_rejected_when_an_origin_is_expected( + self, isolated_home, secret_vault_factory + ): + _write_home_json(isolated_home, "token.json", {"key": "sk-old-token"}) - def test_load_token_file_not_exists(self): - """Test loading token when file doesn't exist""" - with ( - patch("litellm.proxy.client.cli.commands.auth.get_token_file_path") as mock_path, - patch("os.path.exists", return_value=False), - ): - mock_path.return_value = "/test/path/token.json" - - result = load_token() - - assert result is None - - def test_load_token_json_decode_error(self): - """Test loading token with invalid JSON""" - with ( - patch("builtins.open", mock_open(read_data="invalid json")), - patch("litellm.proxy.client.cli.commands.auth.get_token_file_path") as mock_path, - patch("os.path.exists", return_value=True), - ): - mock_path.return_value = "/test/path/token.json" - - result = load_token() - - assert result is None - - def test_load_token_io_error(self): - """Test loading token with IO error""" - with ( - patch("builtins.open", side_effect=OSError("Permission denied")), - patch("litellm.proxy.client.cli.commands.auth.get_token_file_path") as mock_path, - patch("os.path.exists", return_value=True), - ): - mock_path.return_value = "/test/path/token.json" - - result = load_token() - - assert result is None - - def test_clear_token_file_exists(self): - """Test clearing token when file exists""" - with ( - patch("litellm.proxy.client.cli.commands.auth.get_token_file_path") as mock_path, - patch("os.path.exists", return_value=True), - patch("os.remove") as mock_remove, - ): - mock_path.return_value = "/test/path/token.json" - - clear_token() - - mock_remove.assert_called_once_with("/test/path/token.json") - - def test_clear_token_file_not_exists(self): - """Test clearing token when file doesn't exist""" - with ( - patch("litellm.proxy.client.cli.commands.auth.get_token_file_path") as mock_path, - patch("os.path.exists", return_value=False), - patch("os.remove") as mock_remove, - ): - mock_path.return_value = "/test/path/token.json" - - clear_token() - - mock_remove.assert_not_called() - - def test_get_stored_api_key_success(self): - """Test getting stored API key successfully""" - token_data = {"key": "test-api-key-123", "user_id": "test-user"} - - with patch( - "litellm.litellm_core_utils.cli_token_utils.load_cli_token", - return_value=token_data, - ): - result = get_stored_api_key() - assert result == "test-api-key-123" - - def test_get_stored_api_key_no_token(self): - """Test getting stored API key when no token exists""" - with patch( - "litellm.litellm_core_utils.cli_token_utils.load_cli_token", - return_value=None, - ): - result = get_stored_api_key() - assert result is None - - def test_get_stored_api_key_no_key_field(self): - """Test getting stored API key when token has no key field""" - token_data = {"user_id": "test-user"} - - with patch( - "litellm.litellm_core_utils.cli_token_utils.load_cli_token", - return_value=token_data, - ): - result = get_stored_api_key() - assert result is None - - def test_get_stored_api_key_base_url_match(self): - """Stored key is returned when expected_base_url matches stored origin""" - token_data = {"key": "sk-prod", "base_url": "https://real-proxy.com"} - with patch( - "litellm.litellm_core_utils.cli_token_utils.load_cli_token", - return_value=token_data, - ): - assert get_stored_api_key(expected_base_url="https://real-proxy.com") == "sk-prod" - - def test_get_stored_api_key_base_url_match_trailing_slash(self): - """Trailing slash on expected_base_url is normalised before comparison""" - token_data = {"key": "sk-prod", "base_url": "https://real-proxy.com"} - with patch( - "litellm.litellm_core_utils.cli_token_utils.load_cli_token", - return_value=token_data, - ): - assert get_stored_api_key(expected_base_url="https://real-proxy.com/") == "sk-prod" - - def test_get_stored_api_key_base_url_mismatch(self): - """Stored key is NOT returned when expected_base_url differs from stored origin""" - token_data = {"key": "sk-prod", "base_url": "https://real-proxy.com"} - with patch( - "litellm.litellm_core_utils.cli_token_utils.load_cli_token", - return_value=token_data, - ): - assert get_stored_api_key(expected_base_url="https://evil.com") is None - - def test_get_stored_api_key_old_token_no_base_url(self): - """Old tokens without a base_url field are rejected when origin check is requested""" - token_data = {"key": "sk-old-token"} - with patch( - "litellm.litellm_core_utils.cli_token_utils.load_cli_token", - return_value=token_data, - ): - assert get_stored_api_key(expected_base_url="https://real-proxy.com") is None + assert get_stored_api_key("https://real-proxy.com", vault=secret_vault_factory()) is None class TestLoginCommand: @@ -402,7 +269,7 @@ class TestLoginCommand: return_value=_mock_cli_sso_start_response(login_id="cli-test-uuid-123"), ) as mock_post, patch("requests.get", return_value=mock_response) as mock_get, - patch("litellm.proxy.client.cli.commands.auth.save_token") as mock_save, + patch("litellm.proxy.client.cli.commands.auth.save_cli_token") as mock_save, patch("litellm.proxy.client.cli.interface.show_commands") as mock_show_commands, ): result = self.runner.invoke(login, obj=mock_context.obj) @@ -424,8 +291,8 @@ class TestLoginCommand: # Verify JWT was saved mock_save.assert_called_once() saved_data = mock_save.call_args[0][0] - assert saved_data["key"] == "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.test.jwt" - assert saved_data["user_id"] == "test-user-123" + assert saved_data.key == "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.test.jwt" + assert saved_data.user_id == "test-user-123" # Verify commands were shown mock_show_commands.assert_called_once() @@ -557,7 +424,7 @@ class TestLogoutCommand: def test_logout_success(self): """Test successful logout""" - with patch("litellm.proxy.client.cli.commands.auth.clear_token") as mock_clear: + with patch("litellm.proxy.client.cli.commands.auth.clear_cli_token") as mock_clear: result = self.runner.invoke(logout) assert result.exit_code == 0 @@ -574,14 +441,15 @@ class TestWhoamiCommand: def test_whoami_authenticated(self): """Test whoami when user is authenticated""" - token_data = { - "user_email": "test@example.com", - "user_id": "test-user-123", - "user_role": "admin", - "timestamp": time.time() - 3600, # 1 hour ago - } + token_data = CliTokenRecord( + user_email="test@example.com", + user_id="test-user-123", + user_role="admin", + key="sk-live", + timestamp=time.time() - 3600, + ) - with patch("litellm.proxy.client.cli.commands.auth.load_token", return_value=token_data): + with patch("litellm.proxy.client.cli.commands.auth.load_cli_token", return_value=token_data): result = self.runner.invoke(whoami) assert result.exit_code == 0 @@ -593,7 +461,7 @@ class TestWhoamiCommand: def test_whoami_not_authenticated(self): """Test whoami when user is not authenticated""" - with patch("litellm.proxy.client.cli.commands.auth.load_token", return_value=None): + with patch("litellm.proxy.client.cli.commands.auth.load_cli_token", return_value=None): result = self.runner.invoke(whoami) assert result.exit_code == 0 @@ -602,14 +470,15 @@ class TestWhoamiCommand: def test_whoami_old_token(self): """Test whoami with old token showing warning""" - token_data = { - "user_email": "test@example.com", - "user_id": "test-user-123", - "user_role": "admin", - "timestamp": time.time() - (25 * 3600), # 25 hours ago - } + token_data = CliTokenRecord( + user_email="test@example.com", + user_id="test-user-123", + user_role="admin", + key="sk-live", + timestamp=time.time() - (25 * 3600), + ) - with patch("litellm.proxy.client.cli.commands.auth.load_token", return_value=token_data): + with patch("litellm.proxy.client.cli.commands.auth.load_cli_token", return_value=token_data): result = self.runner.invoke(whoami) assert result.exit_code == 0 @@ -618,12 +487,9 @@ class TestWhoamiCommand: def test_whoami_missing_fields(self): """Test whoami with token missing some fields""" - token_data = { - "timestamp": time.time() - 3600 - # Missing user_email, user_id, user_role - } + token_data = CliTokenRecord(key="sk-live", timestamp=time.time() - 3600) - with patch("litellm.proxy.client.cli.commands.auth.load_token", return_value=token_data): + with patch("litellm.proxy.client.cli.commands.auth.load_cli_token", return_value=token_data): result = self.runner.invoke(whoami) assert result.exit_code == 0 @@ -632,16 +498,16 @@ class TestWhoamiCommand: def test_whoami_no_timestamp(self): """Test whoami with token missing timestamp""" - token_data = { - "user_email": "test@example.com", - "user_id": "test-user-123", - "user_role": "admin", - # Missing timestamp - } + token_data = CliTokenRecord( + user_email="test@example.com", + user_id="test-user-123", + user_role="admin", + key="sk-live", + ) with ( patch( - "litellm.proxy.client.cli.commands.auth.load_token", + "litellm.proxy.client.cli.commands.auth.load_cli_token", return_value=token_data, ), patch("time.time", return_value=1000), @@ -701,7 +567,7 @@ class TestCLIKeyRegenerationFlow: return_value=_mock_cli_sso_start_response(login_id="cli-session-uuid-456"), ), patch("requests.get", side_effect=[mock_first_response, mock_second_response]) as mock_get, - patch("litellm.proxy.client.cli.commands.auth.save_token") as mock_save, + patch("litellm.proxy.client.cli.commands.auth.save_cli_token") as mock_save, patch("litellm.proxy.client.cli.interface.show_commands") as mock_show_commands, patch("click.prompt", return_value="2"), ): # User selects index 2 @@ -734,8 +600,8 @@ class TestCLIKeyRegenerationFlow: # Verify JWT was saved mock_save.assert_called_once() saved_data = mock_save.call_args[0][0] - assert saved_data["key"] == "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.team-beta.jwt" - assert saved_data["user_id"] == "test-user-456" + assert saved_data.key == "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.team-beta.jwt" + assert saved_data.user_id == "test-user-456" mock_show_commands.assert_called_once() @@ -762,7 +628,7 @@ class TestCLIKeyRegenerationFlow: return_value=_mock_cli_sso_start_response(login_id="cli-session-uuid-solo"), ), patch("requests.get", return_value=mock_response), - patch("litellm.proxy.client.cli.commands.auth.save_token") as mock_save, + patch("litellm.proxy.client.cli.commands.auth.save_cli_token") as mock_save, patch("litellm.proxy.client.cli.interface.show_commands"), ): result = self.runner.invoke(login, obj=mock_context.obj) @@ -780,8 +646,8 @@ class TestCLIKeyRegenerationFlow: # Verify JWT was saved mock_save.assert_called_once() saved_data = mock_save.call_args[0][0] - assert saved_data["key"] == "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.no-team.jwt" - assert saved_data["user_id"] == "test-user-solo" + assert saved_data.key == "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.no-team.jwt" + assert saved_data.user_id == "test-user-solo" class TestPrintTokenCommand: @@ -810,7 +676,7 @@ class TestPrintTokenCommand: self.runner = CliRunner() def test_no_stored_token_fails_cleanly(self): - with patch("litellm.proxy.client.cli.commands.auth.load_token", return_value=None): + with patch("litellm.proxy.client.cli.commands.auth.load_cli_token", return_value=None): result = self.runner.invoke(print_token, obj={}) assert result.exit_code != 0 @@ -822,12 +688,12 @@ class TestPrintTokenCommand: one). Must use token.json's own base_url, not a hardcoded default.""" with ( patch( - "litellm.proxy.client.cli.commands.auth.load_token", - return_value={ - "base_url": "https://litellm-proxy.corp.com", - "key": "sk-prod-fresh", - "timestamp": time.time(), - }, + "litellm.proxy.client.cli.commands.auth.load_cli_token", + return_value=CliTokenRecord( + base_url="https://litellm-proxy.corp.com", + key="sk-prod-fresh", + timestamp=time.time(), + ), ), patch("requests.post") as mock_post, ): @@ -844,12 +710,12 @@ class TestPrintTokenCommand: token minted for proxy A must not reach a helper invocation aimed at proxy B, even though the token itself is otherwise fresh.""" with patch( - "litellm.proxy.client.cli.commands.auth.load_token", - return_value={ - "base_url": "https://other-server.com", - "key": "sk-should-not-print", - "timestamp": time.time(), - }, + "litellm.proxy.client.cli.commands.auth.load_cli_token", + return_value=CliTokenRecord( + base_url="https://other-server.com", + key="sk-should-not-print", + timestamp=time.time(), + ), ): result = self.runner.invoke( print_token, @@ -863,12 +729,12 @@ class TestPrintTokenCommand: """`lite up`'s own bound invocation shape: --base-url matching the token's origin must succeed exactly like the bare/legacy invocation does.""" with patch( - "litellm.proxy.client.cli.commands.auth.load_token", - return_value={ - "base_url": "http://localhost:4000", - "key": "sk-matches", - "timestamp": time.time(), - }, + "litellm.proxy.client.cli.commands.auth.load_cli_token", + return_value=CliTokenRecord( + base_url="http://localhost:4000", + key="sk-matches", + timestamp=time.time(), + ), ): result = self.runner.invoke( print_token, @@ -884,12 +750,12 @@ class TestPrintTokenCommand: frequently).""" with ( patch( - "litellm.proxy.client.cli.commands.auth.load_token", - return_value={ - "base_url": "http://localhost:4000", - "key": "sk-cached-fresh", - "timestamp": time.time(), - }, + "litellm.proxy.client.cli.commands.auth.load_cli_token", + return_value=CliTokenRecord( + base_url="http://localhost:4000", + key="sk-cached-fresh", + timestamp=time.time(), + ), ), patch("requests.post") as mock_post, ): @@ -908,12 +774,12 @@ class TestPrintTokenCommand: with ( patch( - "litellm.proxy.client.cli.commands.auth.load_token", - return_value={ - "base_url": "http://localhost:4000", - "key": "sk-stale-key", - "timestamp": old_timestamp, - }, + "litellm.proxy.client.cli.commands.auth.load_cli_token", + return_value=CliTokenRecord( + base_url="http://localhost:4000", + key="sk-stale-key", + timestamp=old_timestamp, + ), ), patch("requests.post") as mock_post, ): @@ -925,25 +791,11 @@ class TestPrintTokenCommand: mock_post.assert_not_called() -def _write_home_json(home: Path, filename: str, payload: dict[str, object]) -> None: - litellm_dir = home / ".litellm" - litellm_dir.mkdir(exist_ok=True) - (litellm_dir / filename).write_text(json.dumps(payload)) - - class TestPrintTokenWithConfigFile: """A config-file base_url is a drop-in replacement for exporting LITELLM_PROXY_URL, so print-token must treat it as an explicit server choice: a token minted for a different proxy is never handed out.""" - @pytest.fixture - def isolated_home(self, monkeypatch, tmp_path): - monkeypatch.setenv("HOME", str(tmp_path)) - monkeypatch.setenv("USERPROFILE", str(tmp_path)) - monkeypatch.delenv("LITELLM_PROXY_URL", raising=False) - monkeypatch.delenv("LITELLM_PROXY_API_KEY", raising=False) - return tmp_path - def test_config_base_url_mismatch_fails_closed(self, isolated_home): _write_home_json( isolated_home, @@ -1001,37 +853,196 @@ class TestPrintTokenWithConfigFile: assert result.stdout.strip() == "sk-issued-for-a" -class TestSaveTokenPrivateWrite: - """token.json holds the real API key: it must never be world-readable at any - instant, and a failed write must not destroy the previously stored token.""" +class TestFileFallbackStorage: + """On a headless box with no keychain the token file is still the only store, so it has to + stay owner-only and survive a failed write.""" - @pytest.fixture - def isolated_home(self, monkeypatch, tmp_path): - monkeypatch.setenv("HOME", str(tmp_path)) - monkeypatch.setenv("USERPROFILE", str(tmp_path)) - monkeypatch.delenv("LITELLM_PROXY_URL", raising=False) - monkeypatch.delenv("LITELLM_PROXY_API_KEY", raising=False) - return tmp_path - - def test_save_token_owner_only_permissions_and_no_temp_leftovers(self, isolated_home): - save_token({"key": "sk-secret", "user_id": "u-1", "timestamp": 1234567890}) + def test_owner_only_file_and_directory_with_no_temp_leftovers(self, isolated_home, secret_vault_factory): + save_cli_token( + CliTokenRecord(base_url="https://proxy.example.com", key="sk-secret", user_id="u-1", timestamp=1234567890), + vault=secret_vault_factory(available=False), + ) token_file = isolated_home / ".litellm" / "token.json" - assert json.loads(token_file.read_text()) == {"key": "sk-secret", "user_id": "u-1", "timestamp": 1234567890} + assert json.loads(token_file.read_text())["key"] == "sk-secret" assert stat.S_IMODE(token_file.stat().st_mode) == 0o600 + assert stat.S_IMODE(token_file.parent.stat().st_mode) == 0o700 assert list(token_file.parent.glob(".tmp-*")) == [] - def test_save_token_failure_mid_write_preserves_existing_token(self, isolated_home): + def test_a_failed_write_preserves_the_existing_token(self, isolated_home, secret_vault_factory, monkeypatch): _write_home_json(isolated_home, "token.json", {"key": "sk-original", "timestamp": 1234567890}) token_file = isolated_home / ".litellm" / "token.json" + def _explode(*args, **kwargs): + raise TypeError("not serialisable") + + monkeypatch.setattr("litellm.litellm_core_utils.private_json.json.dump", _explode) + with pytest.raises(TypeError): - save_token({"key": object()}) + save_cli_token(CliTokenRecord(key="sk-new"), vault=secret_vault_factory(available=False)) assert json.loads(token_file.read_text()) == {"key": "sk-original", "timestamp": 1234567890} assert list(token_file.parent.glob(".tmp-*")) == [] +class TestKeychainBackedCommands: + """End-to-end through the `lite` commands: the secret lives in the keychain, the file keeps + only metadata, and every command still reads and writes through that split.""" + + def setup_method(self): + self.runner = CliRunner() + + def _login(self, vault, base_url="https://test.example.com"): + poll_response = Mock() + poll_response.status_code = 200 + poll_response.json.return_value = { + "status": "ready", + "key": "sk-minted", + "user_id": "test-user-123", + "team_id": "team-1", + "teams": ["team-1"], + } + with ( + patch("webbrowser.open"), + patch("requests.post", return_value=_mock_cli_sso_start_response()), + patch("requests.get", return_value=poll_response), + patch("litellm.proxy.client.cli.interface.show_commands"), + ): + return self.runner.invoke(login, obj={"base_url": base_url, "secret_vault": vault}) + + def test_login_puts_the_secret_in_the_keychain_and_not_in_the_file(self, isolated_home, secret_vault_factory): + vault = secret_vault_factory() + + result = self._login(vault) + + token_file = isolated_home / ".litellm" / "token.json" + assert result.exit_code == 0 + assert "Credential stored in your OS keychain." in result.output + assert json.loads(vault.blob)["key"] == "sk-minted" + assert "sk-minted" not in token_file.read_text() + assert json.loads(token_file.read_text())["user_id"] == "test-user-123" + + def test_login_without_a_keychain_says_where_the_credential_went(self, isolated_home, secret_vault_factory): + result = self._login(secret_vault_factory(available=False)) + + token_file = isolated_home / ".litellm" / "token.json" + assert result.exit_code == 0 + assert "No OS keychain available" in result.output + assert str(token_file) in result.output + assert json.loads(token_file.read_text())["key"] == "sk-minted" + + def test_whoami_and_print_token_read_through_the_keychain(self, isolated_home, secret_vault_factory): + vault = secret_vault_factory() + self._login(vault) + obj = {"base_url": "https://test.example.com", "secret_vault": vault} + + whoami_result = self.runner.invoke(whoami, obj=obj) + print_result = self.runner.invoke(print_token, obj=obj) + + assert "Authenticated" in whoami_result.output + assert "test-user-123" in whoami_result.output + assert print_result.exit_code == 0 + assert print_result.stdout.strip() == "sk-minted" + + def test_logout_clears_the_keychain_as_well_as_the_file(self, isolated_home, secret_vault_factory): + vault = secret_vault_factory() + self._login(vault) + + result = self.runner.invoke(logout, obj={"base_url": "https://test.example.com", "secret_vault": vault}) + + assert result.exit_code == 0 + assert "Logged out successfully" in result.output + assert vault.blob is None + assert not (isolated_home / ".litellm" / "token.json").exists() + + def test_logout_warns_when_the_keychain_will_not_release_the_secret(self, isolated_home, secret_vault_factory): + """Silently reporting success would leave a live credential in the keychain.""" + vault = secret_vault_factory(erasable=False) + self._login(vault) + + result = self.runner.invoke(logout, obj={"base_url": "https://test.example.com", "secret_vault": vault}) + + assert result.exit_code == 0 + assert "could not be removed" in result.output + assert not (isolated_home / ".litellm" / "token.json").exists() + + def test_print_token_explains_a_locked_keychain_instead_of_printing_nothing( + self, isolated_home, secret_vault_factory + ): + _write_home_json( + isolated_home, + "token.json", + {"base_url": "https://test.example.com", "user_id": "u-1", "timestamp": time.time()}, + ) + obj = {"base_url": "https://test.example.com", "secret_vault": secret_vault_factory(available=False)} + + result = self.runner.invoke(print_token, obj=obj) + + assert result.exit_code == 1 + assert KEYCHAIN_UNREACHABLE_MESSAGE in result.output + + def test_whoami_flags_a_locked_keychain(self, isolated_home, secret_vault_factory): + _write_home_json( + isolated_home, + "token.json", + {"base_url": "https://test.example.com", "user_id": "u-1", "timestamp": time.time()}, + ) + obj = {"base_url": "https://test.example.com", "secret_vault": secret_vault_factory(available=False)} + + result = self.runner.invoke(whoami, obj=obj) + + assert "Authenticated" in result.output + assert KEYCHAIN_UNREACHABLE_MESSAGE in result.output + + +class TestApiKeyPrecedence: + """`LITELLM_PROXY_API_KEY` and `--api-key` outrank the stored credential; moving the secret + into the keychain must not disturb that order.""" + + def _resolved_key(self, args, obj=None): + with patch("litellm.proxy.client.cli.main.print_version") as mock_print_version: + result = CliRunner().invoke(cli, [*args, "version"], obj=obj) + assert result.exit_code == 0, result.output + return mock_print_version.call_args[0][1] + + def test_the_stored_credential_is_the_fallback(self, isolated_home): + _write_home_json( + isolated_home, + "token.json", + {"base_url": "http://localhost:4000", "key": "sk-stored", "timestamp": time.time()}, + ) + + assert self._resolved_key([]) == "sk-stored" + + def test_the_stored_credential_is_read_through_the_injected_keychain(self, isolated_home, secret_vault_factory): + """The vault handed to the CLI through ctx.obj must be the one the group callback reads, + so a keychain-held secret resolves without ever touching the host OS keychain.""" + _write_home_json(isolated_home, "token.json", {"base_url": "http://localhost:4000", "timestamp": time.time()}) + vault = secret_vault_factory(_secret_blob("http://localhost:4000", "sk-keychain")) + + assert self._resolved_key([], obj={"secret_vault": vault}) == "sk-keychain" + + def test_env_var_beats_the_stored_credential(self, isolated_home, monkeypatch): + _write_home_json( + isolated_home, + "token.json", + {"base_url": "http://localhost:4000", "key": "sk-stored", "timestamp": time.time()}, + ) + monkeypatch.setenv("LITELLM_PROXY_API_KEY", "sk-from-env") + + assert self._resolved_key([]) == "sk-from-env" + + def test_explicit_api_key_beats_both(self, isolated_home, monkeypatch): + _write_home_json( + isolated_home, + "token.json", + {"base_url": "http://localhost:4000", "key": "sk-stored", "timestamp": time.time()}, + ) + monkeypatch.setenv("LITELLM_PROXY_API_KEY", "sk-from-env") + + assert self._resolved_key(["--api-key", "sk-explicit"]) == "sk-explicit" + + class TestLoginConfigClaude: """`lite login --config-claude` wiring into ~/.claude/settings.json""" @@ -1054,7 +1065,7 @@ class TestLoginConfigClaude: patch("webbrowser.open"), patch("requests.post", return_value=_mock_cli_sso_start_response()), patch("requests.get", return_value=poll_response), - patch("litellm.proxy.client.cli.commands.auth.save_token"), + patch("litellm.proxy.client.cli.commands.auth.save_cli_token"), patch("litellm.proxy.client.cli.interface.show_commands"), patch("litellm.proxy.client.cli.commands.auth.CLAUDE_SETTINGS_PATH", settings_path), patch( diff --git a/tests/test_litellm/proxy/client/cli/test_claude_settings.py b/tests/test_litellm/proxy/client/cli/test_claude_settings.py index bc9744eb410..9010fb4c022 100644 --- a/tests/test_litellm/proxy/client/cli/test_claude_settings.py +++ b/tests/test_litellm/proxy/client/cli/test_claude_settings.py @@ -7,6 +7,7 @@ from unittest.mock import patch import pytest from click.testing import CliRunner +from litellm.litellm_core_utils.cli_token_utils import CliTokenRecord from litellm.proxy.client.cli import cli from litellm.proxy.client.cli.commands.claude_settings import ( AUTOROUTE_BACKUP_PATH, @@ -181,18 +182,18 @@ class TestApiKeyHelperIsActuallyInvocable: assert result.exit_code != 2 def test_the_generated_command_reaches_print_token(self): - with patch(f"{AUTH_MODULE}.load_token", return_value=None): + with patch(f"{AUTH_MODULE}.load_cli_token", return_value=None): result = CliRunner().invoke(cli, self._helper_args("http://localhost:4000")) assert "Not authenticated" in result.output def test_the_generated_command_carries_the_base_url_through(self): - stale = { - "base_url": "http://other-proxy.example.com", - "key": "sk-stale", - "timestamp": time.time(), - } - with patch(f"{AUTH_MODULE}.load_token", return_value=stale): + stale = CliTokenRecord( + base_url="http://other-proxy.example.com", + key="sk-stale", + timestamp=time.time(), + ) + with patch(f"{AUTH_MODULE}.load_cli_token", return_value=stale): result = CliRunner().invoke(cli, self._helper_args("http://localhost:4000")) assert "Not authenticated for this server" in result.output diff --git a/tests/test_litellm/proxy/client/cli/test_config_commands.py b/tests/test_litellm/proxy/client/cli/test_config_commands.py index d81ee6bd2b1..6f3f4e4b268 100644 --- a/tests/test_litellm/proxy/client/cli/test_config_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_config_commands.py @@ -18,7 +18,7 @@ from litellm.proxy.client.cli.commands.config import ( load_config, save_config, ) -from litellm.proxy.client.cli.commands.private_json import write_private_json +from litellm.litellm_core_utils.private_json import write_private_json from litellm.proxy.client.cli.interface import show_commands @@ -355,7 +355,7 @@ class TestWritePrivateJson: def _interrupt(*args: object, **kwargs: object) -> None: raise KeyboardInterrupt() - monkeypatch.setattr("litellm.proxy.client.cli.commands.private_json.json.dump", _interrupt) + monkeypatch.setattr("litellm.litellm_core_utils.private_json.json.dump", _interrupt) target = tmp_path / "config.json" with pytest.raises(KeyboardInterrupt): diff --git a/tests/test_litellm/proxy/client/cli/test_up_commands.py b/tests/test_litellm/proxy/client/cli/test_up_commands.py index 51de0dcf11d..aebf441f777 100644 --- a/tests/test_litellm/proxy/client/cli/test_up_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_up_commands.py @@ -8,6 +8,7 @@ import click import pytest from click.testing import CliRunner +from litellm.litellm_core_utils.cli_token_utils import CliTokenRecord from litellm.proxy.client.cli.commands import up as up_module from litellm.proxy.client.cli.commands.agents import AgentRunError from litellm.proxy.client.cli.commands.claude_settings import ClaudeSettingsError @@ -220,13 +221,17 @@ def _make_ctx(base_url): return click.Context(click.Command("test"), obj={"base_url": base_url}) +def _token(key, base_url): + return CliTokenRecord(key=key, base_url=base_url) + + class TestEnsureFreshLogin: """A token that is fresh but was issued for a *different* proxy must not be trusted: without this check, a user logged into proxy A who runs `up --base-url proxy-b` would silently get an apiKeyHelper wired up around proxy A's real token, which print-token would then hand to proxy B.""" def test_reuses_a_fresh_token_issued_for_the_same_proxy(self, monkeypatch): - monkeypatch.setattr(up_module, "load_token", lambda: {"key": "sk-a", "base_url": "http://proxy-a:4000"}) + monkeypatch.setattr(up_module, "load_cli_token", lambda **_: _token("sk-a", "http://proxy-a:4000")) monkeypatch.setattr(up_module, "is_cli_token_fresh", lambda token_data: True) login_calls = [] monkeypatch.setattr(up_module, "login", lambda ctx: login_calls.append(ctx)) @@ -239,11 +244,11 @@ class TestEnsureFreshLogin: monkeypatch.setattr(up_module.sys.stdin, "isatty", lambda: True) tokens = iter( [ - {"key": "sk-a", "base_url": "http://proxy-a:4000"}, - {"key": "sk-b", "base_url": "http://proxy-b:4000"}, + _token("sk-a", "http://proxy-a:4000"), + _token("sk-b", "http://proxy-b:4000"), ] ) - monkeypatch.setattr(up_module, "load_token", lambda: next(tokens)) + monkeypatch.setattr(up_module, "load_cli_token", lambda **_: next(tokens)) monkeypatch.setattr(up_module, "is_cli_token_fresh", lambda token_data: True) login_calls = [] @@ -259,7 +264,7 @@ class TestEnsureFreshLogin: def test_fails_cleanly_non_interactively_when_only_a_different_proxys_token_is_cached(self, monkeypatch): monkeypatch.setattr(up_module.sys.stdin, "isatty", lambda: False) - monkeypatch.setattr(up_module, "load_token", lambda: {"key": "sk-a", "base_url": "http://proxy-a:4000"}) + monkeypatch.setattr(up_module, "load_cli_token", lambda **_: _token("sk-a", "http://proxy-a:4000")) monkeypatch.setattr(up_module, "is_cli_token_fresh", lambda token_data: True) with pytest.raises(UpError, match="lite login"): @@ -276,7 +281,7 @@ class TestUpCommand: backup_path.write_text(json.dumps(existing_backup)) with ( - patch(f"{UP_MODULE}.load_token", return_value={"key": "sk-fresh", "base_url": "http://localhost:4000"}), + patch(f"{UP_MODULE}.load_cli_token", return_value=_token("sk-fresh", "http://localhost:4000")), patch(f"{UP_MODULE}.is_cli_token_fresh", return_value=True), patch(f"{UP_MODULE}.resolve_api_key", return_value="sk-fresh"), patch(f"{UP_MODULE}.verify_proxy_key"), @@ -293,7 +298,7 @@ class TestUpCommand: _patch_paths(monkeypatch, tmp_path) monkeypatch.setattr(sys.stdin, "isatty", lambda: False) - with patch(f"{UP_MODULE}.load_token", return_value=None): + with patch(f"{UP_MODULE}.load_cli_token", return_value=None): result = self.runner.invoke(up, obj={"base_url": "http://localhost:4000"}) assert result.exit_code != 0 @@ -303,7 +308,7 @@ class TestUpCommand: _patch_paths(monkeypatch, tmp_path) with ( - patch(f"{UP_MODULE}.load_token", return_value={"key": "sk-fresh", "base_url": "http://localhost:4000"}), + patch(f"{UP_MODULE}.load_cli_token", return_value=_token("sk-fresh", "http://localhost:4000")), patch(f"{UP_MODULE}.is_cli_token_fresh", return_value=True), patch(f"{UP_MODULE}.resolve_api_key", return_value="sk-fresh"), patch( @@ -329,7 +334,7 @@ class TestUpCommand: return True with ( - patch(f"{UP_MODULE}.load_token", return_value={"key": "sk-fresh", "base_url": "http://localhost:4000"}), + patch(f"{UP_MODULE}.load_cli_token", return_value=_token("sk-fresh", "http://localhost:4000")), patch(f"{UP_MODULE}.is_cli_token_fresh", return_value=True), patch(f"{UP_MODULE}.resolve_api_key", return_value="sk-fresh"), patch(f"{UP_MODULE}.verify_proxy_key"), diff --git a/type-discipline-budget.json b/type-discipline-budget.json index 31726acfbaa..a5c5a9f135b 100644 --- a/type-discipline-budget.json +++ b/type-discipline-budget.json @@ -1,9 +1,9 @@ { "LIT001": { - "limit": 22806 + "limit": 22805 }, "LIT002": { - "limit": 26878 + "limit": 26877 }, "LIT003": { "limit": 269 @@ -27,12 +27,12 @@ "limit": 0 }, "LIT010": { - "limit": 16695 + "limit": 16693 }, "LIT011": { "limit": 5588 }, "LIT012": { - "limit": 4519 + "limit": 4511 } } diff --git a/uv.lock b/uv.lock index d9e2fb94667..53bb0cb8f82 100644 --- a/uv.lock +++ b/uv.lock @@ -10,7 +10,7 @@ resolution-markers = [ ] [options] -exclude-newer = "2026-08-16T00:41:08.185444Z" +exclude-newer = "2026-08-17T01:06:38.502388Z" exclude-newer-span = "P3D" [manifest] @@ -710,6 +710,15 @@ wheels = [ { url = "https://files.pythonhosted.org/packages/a0/59/76ab57e3fe74484f48a53f8e337171b4a2349e506eabe136d7e01d059086/backports_asyncio_runner-1.2.0-py3-none-any.whl", hash = "sha256:0da0a936a8aeb554eccb426dc55af3ba63bcdc69fa1a600b5bb305413a4477b5", size = 12313, upload-time = "2025-07-02T02:27:14.263Z" }, ] +[[package]] +name = "backports-tarfile" +version = "1.2.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/86/72/cd9b395f25e290e633655a100af28cb253e4393396264a98bd5f5951d50f/backports_tarfile-1.2.0.tar.gz", hash = "sha256:d75e02c268746e1b8144c278978b6e98e85de6ad16f8e4b0844a154557eca991", size = 86406, upload-time = "2024-05-28T17:01:54.731Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/b9/fa/123043af240e49752f1c4bd24da5053b6bd00cad78c2be53c0d1e8b975bc/backports.tarfile-1.2.0-py3-none-any.whl", hash = "sha256:77e284d754527b01fb1e6fa8a1afe577858ebe4e9dad8919e34c862cb399bc34", size = 30181, upload-time = "2024-05-28T17:01:53.112Z" }, +] + [[package]] name = "basedpyright" version = "1.39.7" @@ -3538,6 +3547,51 @@ wheels = [ { url = "https://files.pythonhosted.org/packages/04/96/92447566d16df59b2a776c0fb82dbc4d9e07cd95062562af01e408583fc4/itsdangerous-2.2.0-py3-none-any.whl", hash = "sha256:c6242fc49e35958c8b15141343aa660db5fc54d4f13a1db01a3f5891b98700ef", size = 16234, upload-time = "2024-04-16T21:28:14.499Z" }, ] +[[package]] +name = "jaraco-classes" +version = "3.4.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "more-itertools" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/06/c0/ed4a27bc5571b99e3cff68f8a9fa5b56ff7df1c2251cc715a652ddd26402/jaraco.classes-3.4.0.tar.gz", hash = "sha256:47a024b51d0239c0dd8c8540c6c7f484be3b8fcf0b2d85c13825780d3b3f3acd", size = 11780, upload-time = "2024-03-31T07:27:36.643Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/7f/66/b15ce62552d84bbfcec9a4873ab79d993a1dd4edb922cbfccae192bd5b5f/jaraco.classes-3.4.0-py3-none-any.whl", hash = "sha256:f662826b6bed8cace05e7ff873ce0f9283b5c924470fe664fff1c2f00f581790", size = 6777, upload-time = "2024-03-31T07:27:34.792Z" }, +] + +[[package]] +name = "jaraco-context" +version = "6.1.2" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "backports-tarfile", marker = "python_full_version < '3.12'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/af/50/4763cd07e722bb6285316d390a164bc7e479db9d90daa769f22578f698b4/jaraco_context-6.1.2.tar.gz", hash = "sha256:f1a6c9d391e661cc5b8d39861ff077a7dc24dc23833ccee564b234b81c82dfe3", size = 16801, upload-time = "2026-03-20T22:13:33.922Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/f2/58/bc8954bda5fcda97bd7c19be11b85f91973d67a706ed4a3aec33e7de22db/jaraco_context-6.1.2-py3-none-any.whl", hash = "sha256:bf8150b79a2d5d91ae48629d8b427a8f7ba0e1097dd6202a9059f29a36379535", size = 7871, upload-time = "2026-03-20T22:13:32.808Z" }, +] + +[[package]] +name = "jaraco-functools" +version = "4.6.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "more-itertools" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/6c/1f/c23395957d41ccf27c4e535c3d334c4051e5395b3752057ba4cbaec35c56/jaraco_functools-4.6.0.tar.gz", hash = "sha256:880c577ec9720b3a052d5bc611fb9f2269b3d87902ef42440df443b88e443280", size = 20837, upload-time = "2026-07-14T01:28:02.544Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/02/36/ecc85bc96c273dc8a11273ed4782272975e6338d4a3e9228621175edf0e3/jaraco_functools-4.6.0-py3-none-any.whl", hash = "sha256:99e3dc0060c5cbe8fcd1cdb36258e2a65ca40f1566b2033b12abb1bb44dd3c30", size = 11677, upload-time = "2026-07-14T01:28:01.59Z" }, +] + +[[package]] +name = "jeepney" +version = "0.9.0" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/7b/6f/357efd7602486741aa73ffc0617fb310a29b588ed0fd69c2399acbb85b0c/jeepney-0.9.0.tar.gz", hash = "sha256:cf0e9e845622b81e4a28df94c40345400256ec608d0e55bb8a3feaa9163f5732", size = 106758, upload-time = "2025-02-27T18:51:01.684Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/b2/a3/e137168c9c44d18eff0376253da9f1e9234d0239e0ee230d2fee6cea8e55/jeepney-0.9.0-py3-none-any.whl", hash = "sha256:97e5714520c16fc0a45695e5365a2e11b81ea79bba796e26f9f1d178cb182683", size = 49010, upload-time = "2025-02-27T18:51:00.104Z" }, +] + [[package]] name = "jinja2" version = "3.1.6" @@ -3764,6 +3818,24 @@ wheels = [ { url = "https://files.pythonhosted.org/packages/41/45/1a4ed80516f02155c51f51e8cedb3c1902296743db0bbc66608a0db2814f/jsonschema_specifications-2025.9.1-py3-none-any.whl", hash = "sha256:98802fee3a11ee76ecaca44429fda8a41bff98b00a0f2838151b113f210cc6fe", size = 18437, upload-time = "2025-09-08T01:34:57.871Z" }, ] +[[package]] +name = "keyring" +version = "25.7.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "importlib-metadata", marker = "python_full_version < '3.12'" }, + { name = "jaraco-classes" }, + { name = "jaraco-context" }, + { name = "jaraco-functools" }, + { name = "jeepney", marker = "sys_platform == 'linux'" }, + { name = "pywin32-ctypes", marker = "sys_platform == 'win32'" }, + { name = "secretstorage", marker = "sys_platform == 'linux'" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/43/4b/674af6ef2f97d56f0ab5153bf0bfa28ccb6c3ed4d1babf4305449668807b/keyring-25.7.0.tar.gz", hash = "sha256:fe01bd85eb3f8fb3dd0405defdeac9a5b4f6f0439edbb3149577f244a2e8245b", size = 63516, upload-time = "2025-11-16T16:26:09.482Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/81/db/e655086b7f3a705df045bf0933bdd9c2f79bb3c97bfef1384598bb79a217/keyring-25.7.0-py3-none-any.whl", hash = "sha256:be4a0b195f149690c166e850609a477c532ddbfbaed96a404d4e43f8d5e2689f", size = 39160, upload-time = "2025-11-16T16:26:08.402Z" }, +] + [[package]] name = "kiwisolver" version = "1.5.0" @@ -4222,6 +4294,7 @@ caching = [ ] cli = [ { name = "inquirerpy" }, + { name = "keyring" }, { name = "pyyaml" }, { name = "requests" }, { name = "rich" }, @@ -4352,6 +4425,7 @@ dev = [ { name = "diff-cover" }, { name = "fakeredis" }, { name = "fastapi-offline" }, + { name = "keyring" }, { name = "langfuse" }, { name = "openapi-core" }, { name = "opentelemetry-api" }, @@ -4446,6 +4520,7 @@ requires-dist = [ { name = "inquirerpy", marker = "extra == 'proxy'", specifier = ">=0.3.4,<1.0" }, { name = "jinja2", specifier = ">=3.1.6,<4.0" }, { name = "jsonschema", specifier = ">=4.0.0,<5.0" }, + { name = "keyring", marker = "extra == 'cli'", specifier = ">=25.6.0,<26.0" }, { name = "langfuse", marker = "extra == 'proxy-runtime'", specifier = ">=2.59.7,<3.0" }, { name = "litellm-enterprise", marker = "extra == 'proxy'", editable = "enterprise" }, { name = "litellm-proxy-extras", marker = "extra == 'proxy'", editable = "litellm-proxy-extras" }, @@ -4532,6 +4607,7 @@ dev = [ { name = "diff-cover", specifier = "==9.7.2" }, { name = "fakeredis", specifier = "==2.34.1" }, { name = "fastapi-offline", specifier = "==1.7.6" }, + { name = "keyring", specifier = "==25.7.0" }, { name = "langfuse", specifier = "==2.59.7" }, { name = "openapi-core", specifier = "==0.22.0" }, { name = "opentelemetry-api", specifier = "==1.28.0" }, @@ -7790,6 +7866,15 @@ wheels = [ { url = "https://files.pythonhosted.org/packages/c0/d2/21af5c535501a7233e734b8af901574572da66fcc254cb35d0609c9080dd/pywin32-311-cp314-cp314-win_arm64.whl", hash = "sha256:a508e2d9025764a8270f93111a970e1d0fbfc33f4153b388bb649b7eec4f9b42", size = 8932540, upload-time = "2025-07-14T20:13:36.379Z" }, ] +[[package]] +name = "pywin32-ctypes" +version = "0.2.3" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/85/9f/01a1a99704853cb63f253eea009390c88e7131c67e66a0a02099a8c917cb/pywin32-ctypes-0.2.3.tar.gz", hash = "sha256:d162dc04946d704503b2edc4d55f3dba5c1d539ead017afa00142c38b9885755", size = 29471, upload-time = "2024-08-14T10:15:34.626Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/de/3d/8161f7711c017e01ac9f008dfddd9410dff3674334c233bde66e7ba65bbf/pywin32_ctypes-0.2.3-py3-none-any.whl", hash = "sha256:8a1513379d709975552d202d942d9837758905c8d01eb82b8bcc30918929e7b8", size = 30756, upload-time = "2024-08-14T10:15:33.187Z" }, +] + [[package]] name = "pyyaml" version = "6.0.3" @@ -8632,6 +8717,19 @@ wheels = [ { url = "https://files.pythonhosted.org/packages/07/39/338d9219c4e87f3e708f18857ecd24d22a0c3094752393319553096b98af/scipy-1.17.1-cp314-cp314t-win_arm64.whl", hash = "sha256:200e1050faffacc162be6a486a984a0497866ec54149a01270adc8a59b7c7d21", size = 25489165, upload-time = "2026-02-23T00:22:29.563Z" }, ] +[[package]] +name = "secretstorage" +version = "3.5.0" +source = { registry = "https://pypi.org/simple" } +dependencies = [ + { name = "cryptography" }, + { name = "jeepney" }, +] +sdist = { url = "https://files.pythonhosted.org/packages/1c/03/e834bcd866f2f8a49a85eaff47340affa3bfa391ee9912a952a1faa68c7b/secretstorage-3.5.0.tar.gz", hash = "sha256:f04b8e4689cbce351744d5537bf6b1329c6fc68f91fa666f60a380edddcd11be", size = 19884, upload-time = "2025-11-23T19:02:53.191Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/b7/46/f5af3402b579fd5e11573ce652019a67074317e18c1935cc0b4ba9b35552/secretstorage-3.5.0-py3-none-any.whl", hash = "sha256:0ce65888c0725fcb2c5bc0fdb8e5438eece02c523557ea40ce0703c266248137", size = 15554, upload-time = "2025-11-23T19:02:51.545Z" }, +] + [[package]] name = "semantic-router" version = "0.1.15" From bd322ed8a7eb6968b2af8bebce28eb9f19251fd3 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 19:11:01 -0700 Subject: [PATCH 019/220] refactor(cli): state the credential-store precedence rules as contracts Drop the inline notes on keychain erasure and disk-vs-vault precedence in favour of docstrings on the two functions that own those rules, and remove a stale section header and a field note that the code already says plainly. --- litellm/litellm_core_utils/cli_keyring.py | 6 +++++- litellm/litellm_core_utils/cli_token_utils.py | 6 +++++- litellm/proxy/client/cli/commands/auth.py | 3 --- 3 files changed, 10 insertions(+), 5 deletions(-) diff --git a/litellm/litellm_core_utils/cli_keyring.py b/litellm/litellm_core_utils/cli_keyring.py index 873db64a728..fcbf5ada55a 100644 --- a/litellm/litellm_core_utils/cli_keyring.py +++ b/litellm/litellm_core_utils/cli_keyring.py @@ -98,10 +98,14 @@ class KeyringVault: return True def erase(self) -> bool: + """Whether the keychain is guaranteed to hold no credential afterwards. + + An uninstalled `keyring` package can never have stored one. A kill switch set after + a credential was stored leaves that entry out of reach, so erasure cannot be promised. + """ if _import_keyring() is None: return True if _keyring_disabled(): - # a credential stored before the kill switch was set may still be in the keychain return False match self.read(): case SecretUnavailable(): diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 9960192180c..dd263ccd412 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -168,8 +168,12 @@ def _resolve_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecor def _apply_vault_secret(record: CliTokenRecord, blob: str, vault: SecretVault) -> CliTokenRecord | None: + """Resolve the credential when both stores hold one. + + A secret still on disk is the fresher of the two, because it is only left there when the + keychain write that should have removed it failed, so it outranks the vault entry. + """ if record.key is not None: - # a secret still on disk means the last keychain write failed: the file outranks the vault return _migrate_file_secret(record, vault) try: secret: Final = CliTokenSecret.model_validate_json(blob) diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index 8b9ef5633da..c2b5b6a620f 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -76,7 +76,6 @@ KEYCHAIN_UNREACHABLE_MESSAGE: Final = ( ) -# Token storage utilities def context_secret_vault(ctx: click.Context) -> SecretVault: """Where this invocation reads and writes secret material; injectable through ctx.obj for tests""" ctx_obj: Final[CliContextObj | None] = ctx.obj @@ -666,8 +665,6 @@ def login(ctx: click.Context, config_claude: bool): api_key: Final = auth_result["api_key"] user_id: Final = auth_result["user_id"] - # base_url is stored so we can verify origin before reusing the - # key on a subsequent CLI invocation. record: Final = CliTokenRecord( base_url=base_url.rstrip("/"), key=api_key, From 01add582982a253c2fff3466b608c1d0409ed1ae Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 19:45:38 -0700 Subject: [PATCH 020/220] fix(cli): name why a login fell back to the token file lite ships with every install of litellm, but the keyring package it needs for keychain storage only ships with the cli extra. Such a user on a Mac was told 'No OS keychain available' about a machine that plainly has one, with nothing pointing at the missing package. The vault now reports which of the three unusable states it is in, so login can point at the install, name the kill switch, or report a genuinely absent keychain. --- litellm/litellm_core_utils/cli_keyring.py | 56 +++++++++++++------ litellm/litellm_core_utils/cli_token_utils.py | 26 +++++---- litellm/proxy/client/README.md | 2 +- litellm/proxy/client/cli/commands/auth.py | 41 +++++++++++--- tests/test_litellm/conftest.py | 18 ++++-- .../test_cli_token_utils.py | 23 ++++---- .../proxy/client/cli/test_auth_commands.py | 28 ++++++++++ 7 files changed, 141 insertions(+), 53 deletions(-) diff --git a/litellm/litellm_core_utils/cli_keyring.py b/litellm/litellm_core_utils/cli_keyring.py index fcbf5ada55a..b19b3d3bc83 100644 --- a/litellm/litellm_core_utils/cli_keyring.py +++ b/litellm/litellm_core_utils/cli_keyring.py @@ -5,9 +5,9 @@ SDK-level access to the OS keychain (macOS Keychain, Windows Credential Manager, Linux Secret Service) that holds the credential minted by `lite login`. The `keyring` package is optional and imported lazily, so importing this module -never pulls it in. Every failure is returned as a value: a machine with no -keychain, or one whose keychain is locked, must degrade to the token file rather -than break `lite` or the SDK. +never pulls it in. Every failure is returned as a value, naming which of the +three ways the keychain can be out of reach applies, so callers can degrade to +the token file and tell the user what to do about it. """ import os @@ -32,11 +32,28 @@ class SecretMissing: @dataclass(frozen=True, slots=True) -class SecretUnavailable: +class SecretStored: pass -SecretRead: TypeAlias = SecretFound | SecretMissing | SecretUnavailable +@dataclass(frozen=True, slots=True) +class KeyringNotInstalled: + pass + + +@dataclass(frozen=True, slots=True) +class KeyringDisabled: + pass + + +@dataclass(frozen=True, slots=True) +class KeyringUnreachable: + pass + + +KeyringUnusable: TypeAlias = KeyringNotInstalled | KeyringDisabled | KeyringUnreachable +SecretRead: TypeAlias = SecretFound | SecretMissing | KeyringUnusable +SecretWrite: TypeAlias = SecretStored | KeyringUnusable class SecretVault(Protocol): @@ -44,7 +61,7 @@ class SecretVault(Protocol): def read(self) -> SecretRead: ... - def write(self, blob: str) -> bool: ... + def write(self, blob: str) -> SecretWrite: ... def erase(self) -> bool: ... @@ -69,8 +86,11 @@ def _import_keyring() -> KeyringApi | None: return keyring -def _keyring_api() -> KeyringApi | None: - return None if _keyring_disabled() else _import_keyring() +def _keyring_api() -> KeyringApi | KeyringNotInstalled | KeyringDisabled: + if _keyring_disabled(): + return KeyringDisabled() + api: Final = _import_keyring() + return KeyringNotInstalled() if api is None else api @dataclass(frozen=True, slots=True) @@ -79,23 +99,23 @@ class KeyringVault: def read(self) -> SecretRead: api: Final = _keyring_api() - if api is None: - return SecretUnavailable() + if isinstance(api, (KeyringNotInstalled, KeyringDisabled)): + return api try: blob: Final = api.get_password(KEYRING_SERVICE, KEYRING_ACCOUNT) except Exception: # noqa: BLE001 # backends raise outside keyring.errors; never break the SDK - return SecretUnavailable() + return KeyringUnreachable() return SecretMissing() if blob is None else SecretFound(blob) - def write(self, blob: str) -> bool: + def write(self, blob: str) -> SecretWrite: api: Final = _keyring_api() - if api is None: - return False + if isinstance(api, (KeyringNotInstalled, KeyringDisabled)): + return api try: api.set_password(KEYRING_SERVICE, KEYRING_ACCOUNT, blob) except Exception: # noqa: BLE001 # a keychain that refuses the write falls back to the token file - return False - return True + return KeyringUnreachable() + return SecretStored() def erase(self) -> bool: """Whether the keychain is guaranteed to hold no credential afterwards. @@ -108,7 +128,7 @@ class KeyringVault: if _keyring_disabled(): return False match self.read(): - case SecretUnavailable(): + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable(): return False case SecretMissing(): return True @@ -117,7 +137,7 @@ class KeyringVault: def _delete(self) -> bool: api: Final = _keyring_api() - if api is None: + if isinstance(api, (KeyringNotInstalled, KeyringDisabled)): return False try: api.delete_password(KEYRING_SERVICE, KEYRING_ACCOUNT) diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index dd263ccd412..40822e5b335 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -22,10 +22,14 @@ from pydantic import BaseModel, ConfigDict, ValidationError from litellm.litellm_core_utils.cli_keyring import ( SYSTEM_KEYRING, + KeyringDisabled, + KeyringNotInstalled, + KeyringUnreachable, SecretFound, SecretMissing, - SecretUnavailable, + SecretStored, SecretVault, + SecretWrite, ) from litellm.litellm_core_utils.private_json import ensure_private_dir, write_private_json @@ -79,13 +83,15 @@ def load_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> CliTokenRecord | N return _resolve_secret(record, vault) -def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRING) -> bool: - """Store a freshly minted credential. Returns whether the keychain took the secret""" - if record.key is None or not vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)): - _write_token_file(record) - return False - _write_token_file(_without_secret(record)) - return True +def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRING) -> SecretWrite: + """Store a freshly minted credential. Reports whether the keychain took the secret, and why not""" + outcome: Final = ( + SecretStored() + if record.key is None + else vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)) + ) + _write_token_file(_without_secret(record) if isinstance(outcome, SecretStored) else record) + return outcome def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> bool: @@ -163,7 +169,7 @@ def _resolve_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecor return _apply_vault_secret(record, blob, vault) case SecretMissing(): return _migrate_file_secret(record, vault) - case SecretUnavailable(): + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable(): return record @@ -188,7 +194,7 @@ def _apply_vault_secret(record: CliTokenRecord, blob: str, vault: SecretVault) - def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecord | None: if record.key is None: return None - if vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)): + if isinstance(vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)), SecretStored): _scrub_file_secret(record) return record diff --git a/litellm/proxy/client/README.md b/litellm/proxy/client/README.md index 9ece4c2be3d..d46c7174b4c 100644 --- a/litellm/proxy/client/README.md +++ b/litellm/proxy/client/README.md @@ -377,7 +377,7 @@ The key itself goes into the OS keychain (macOS Keychain, Windows Credential Man } ``` -Headless boxes and CI runners usually have no keychain. There the key stays in the same `0600` file alongside the metadata, exactly as it did before, and `lite login` tells you which of the two happened. Set `LITELLM_CLI_DISABLE_KEYRING=1` to force the file even where a keychain exists. A `token.json` written by an older `lite` keeps working and is moved into the keychain, and scrubbed from the file, the first time a keychain-capable `lite` reads it. +Keychain storage needs the `keyring` package, which ships with `pip install 'litellm[cli]'`. Headless boxes and CI runners usually have no keychain either. In all of those cases the key stays in the same `0600` file alongside the metadata, exactly as it did before, and `lite login` names which one applies: the package is missing, the machine has no keychain, or you set `LITELLM_CLI_DISABLE_KEYRING=1` to force the file even where a keychain exists. A `token.json` written by an older `lite` keeps working and is moved into the keychain, and scrubbed from the file, the first time a keychain-capable `lite` reads it. `lite logout` clears both stores. If the keychain is locked at that moment it says so, and re-running it once the keychain is unlocked finishes the job. diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index c2b5b6a620f..03906f9b6df 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -11,7 +11,16 @@ from rich.table import Table from typing_extensions import NotRequired, ReadOnly, TypedDict from litellm.constants import CLI_JWT_EXPIRATION_HOURS -from litellm.litellm_core_utils.cli_keyring import SYSTEM_KEYRING, SecretVault +from litellm.litellm_core_utils.cli_keyring import ( + DISABLE_KEYRING_ENV_VAR, + SYSTEM_KEYRING, + KeyringDisabled, + KeyringNotInstalled, + KeyringUnreachable, + SecretStored, + SecretVault, + SecretWrite, +) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, clear_cli_token, @@ -72,9 +81,29 @@ class CliAuthResult(TypedDict): KEYCHAIN_UNREACHABLE_MESSAGE: Final = ( - "Your credential is stored in your OS keychain, which could not be read. Unlock it and retry, or run 'lite login'." + "Your credential is stored in your OS keychain, which could not be read. Unlock it and retry, " + "install the keyring package with: pip install 'litellm[cli]', or run 'lite login'." ) +KEYRING_INSTALL_HINT: Final = "pip install 'litellm[cli]'" + + +def storage_notice(outcome: SecretWrite) -> str: + """Tell the user where the credential ended up, and how to get keychain storage if it did not.""" + path: Final = get_cli_token_file_path() + match outcome: + case SecretStored(): + return "Credential stored in your OS keychain." + case KeyringNotInstalled(): + return ( + f"Credential stored in {path} (owner-only). " + f"For OS keychain storage, install the keyring package with: {KEYRING_INSTALL_HINT}" + ) + case KeyringDisabled(): + return f"Keychain storage is off ({DISABLE_KEYRING_ENV_VAR}). Credential stored in {path} (owner-only)." + case KeyringUnreachable(): + return f"No OS keychain available. Credential stored in {path} (owner-only)." + def context_secret_vault(ctx: click.Context) -> SecretVault: """Where this invocation reads and writes secret material; injectable through ctx.obj for tests""" @@ -675,15 +704,11 @@ def login(ctx: click.Context, config_claude: bool): jwt_token="", timestamp=time.time(), ) - in_keychain: Final = save_cli_token(record, vault=context_secret_vault(ctx)) + stored: Final = save_cli_token(record, vault=context_secret_vault(ctx)) click.echo("\nLogin successful!") click.echo(f"JWT Token: {api_key[:20]}...") - click.echo( - "Credential stored in your OS keychain." - if in_keychain - else f"No OS keychain available; credential stored in {get_cli_token_file_path()} (owner-only)." - ) + click.echo(storage_notice(stored)) click.echo("You can now use the CLI without specifying --api-key") if config_claude: diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py index ce0fd197538..a716ec0e0aa 100644 --- a/tests/test_litellm/conftest.py +++ b/tests/test_litellm/conftest.py @@ -23,10 +23,13 @@ from litellm import router as litellm_router_module from litellm import utils as litellm_utils_module from litellm._logging import ALL_LOGGERS from litellm.litellm_core_utils.cli_keyring import ( + KeyringUnreachable, + KeyringUnusable, SecretFound, SecretMissing, SecretRead, - SecretUnavailable, + SecretStored, + SecretWrite, ) from litellm.litellm_core_utils.prompt_templates import ( image_handling as image_handling_module, @@ -125,7 +128,8 @@ class FakeSecretVault: """In-memory stand-in for the OS keychain, injected wherever CLI credential storage is exercised. `available=False` models a keychain that is locked or has no backend, `writable=False` one that - refuses to store, and `erasable=False` one that will not release what it already holds. + refuses to store, `erasable=False` one that will not release what it already holds, and `failure` + picks which unusable state those report. """ def __init__( @@ -135,11 +139,13 @@ class FakeSecretVault: available: bool = True, writable: bool = True, erasable: bool = True, + failure: KeyringUnusable = KeyringUnreachable(), ) -> None: self.blob: str | None = blob self.available: bool = available self.writable: bool = writable self.erasable: bool = erasable + self.failure: KeyringUnusable = failure self.reads: int = 0 self.writes: list[str] = [] self.erases: int = 0 @@ -147,15 +153,15 @@ class FakeSecretVault: def read(self) -> SecretRead: self.reads += 1 if not self.available: - return SecretUnavailable() + return self.failure return SecretMissing() if self.blob is None else SecretFound(self.blob) - def write(self, blob: str) -> bool: + def write(self, blob: str) -> SecretWrite: self.writes.append(blob) if not (self.available and self.writable): - return False + return self.failure self.blob = blob - return True + return SecretStored() def erase(self) -> bool: self.erases += 1 diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 56ab6bcbfe0..961d986e8f5 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -13,7 +13,10 @@ from litellm.litellm_core_utils.cli_keyring import ( KeyringVault, SecretFound, SecretMissing, - SecretUnavailable, + KeyringDisabled, + KeyringNotInstalled, + KeyringUnreachable, + SecretStored, ) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, @@ -253,7 +256,7 @@ class TestSaveCliToken: vault=vault, ) - assert stored is True + assert stored == SecretStored() assert "sk-new" not in _token_file(isolated_home).read_text() assert json.loads(vault.blob)["key"] == "sk-new" assert load_cli_token(vault=vault).key == "sk-new" @@ -265,7 +268,7 @@ class TestSaveCliToken: ) path = _token_file(isolated_home) - assert stored is False + assert stored == KeyringUnreachable() assert json.loads(path.read_text())["key"] == "sk-new" assert stat.S_IMODE(path.stat().st_mode) == 0o600 assert list(path.parent.glob(".tmp-*")) == [] @@ -379,7 +382,7 @@ class TestKeyringVault: fake = install_fake_keyring(_FakeKeyringModule()) vault = KeyringVault() - assert vault.write("blob-1") is True + assert vault.write("blob-1") == SecretStored() assert vault.read() == SecretFound("blob-1") assert vault.erase() is True assert vault.read() == SecretMissing() @@ -393,8 +396,8 @@ class TestKeyringVault: monkeypatch.setenv(DISABLE_KEYRING_ENV_VAR, "1") vault = KeyringVault() - assert vault.read() == SecretUnavailable() - assert vault.write("blob-1") is False + assert vault.read() == KeyringDisabled() + assert vault.write("blob-1") == KeyringDisabled() assert vault.erase() is False def test_an_uninstalled_keyring_library_degrades_to_the_file(self, monkeypatch): @@ -404,19 +407,19 @@ class TestKeyringVault: monkeypatch.setitem(sys.modules, "keyring", None) vault = KeyringVault() - assert vault.read() == SecretUnavailable() - assert vault.write("blob-1") is False + assert vault.read() == KeyringNotInstalled() + assert vault.write("blob-1") == KeyringNotInstalled() assert vault.erase() is True def test_a_locked_keychain_is_reported_not_raised(self, install_fake_keyring): install_fake_keyring(_FakeKeyringModule(get_error=RuntimeError("keyring is locked"))) - assert KeyringVault().read() == SecretUnavailable() + assert KeyringVault().read() == KeyringUnreachable() def test_a_refused_write_is_reported_not_raised(self, install_fake_keyring): install_fake_keyring(_FakeKeyringModule(set_error=RuntimeError("no backend"))) - assert KeyringVault().write("blob-1") is False + assert KeyringVault().write("blob-1") == KeyringUnreachable() def test_a_refused_delete_is_reported_so_logout_can_warn(self, install_fake_keyring): install_fake_keyring(_FakeKeyringModule(stored="blob-1", delete_error=RuntimeError("locked"))) diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index e93f05cb4aa..1f2d48f6547 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -13,6 +13,11 @@ import pytest from click.testing import CliRunner from litellm.constants import CLI_JWT_EXPIRATION_HOURS +from litellm.litellm_core_utils.cli_keyring import ( + DISABLE_KEYRING_ENV_VAR, + KeyringDisabled, + KeyringNotInstalled, +) from litellm.litellm_core_utils.cli_token_utils import CliTokenRecord, save_cli_token from litellm.proxy.client.cli import cli from litellm.proxy.client.cli.commands.auth import ( @@ -931,6 +936,29 @@ class TestKeychainBackedCommands: assert str(token_file) in result.output assert json.loads(token_file.read_text())["key"] == "sk-minted" + def test_login_points_a_user_missing_the_keyring_package_at_the_install( + self, isolated_home, secret_vault_factory + ): + """`lite` ships with every install, the keyring package only with the cli extra. Telling + that user their machine has no keychain sends them looking for a problem they do not have.""" + result = self._login(secret_vault_factory(available=False, failure=KeyringNotInstalled())) + + token_file = isolated_home / ".litellm" / "token.json" + assert result.exit_code == 0 + assert "pip install 'litellm[cli]'" in result.output + assert "No OS keychain available" not in result.output + assert json.loads(token_file.read_text())["key"] == "sk-minted" + + def test_login_names_the_kill_switch_instead_of_blaming_the_machine( + self, isolated_home, secret_vault_factory + ): + result = self._login(secret_vault_factory(available=False, failure=KeyringDisabled())) + + assert result.exit_code == 0 + assert DISABLE_KEYRING_ENV_VAR in result.output + assert "No OS keychain available" not in result.output + assert json.loads((isolated_home / ".litellm" / "token.json").read_text())["key"] == "sk-minted" + def test_whoami_and_print_token_read_through_the_keychain(self, isolated_home, secret_vault_factory): vault = secret_vault_factory() self._login(vault) From 1750893a6905f45d7fa9cc65167856ae6c182208 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 20:26:41 -0700 Subject: [PATCH 021/220] fix(cli): never report success while a credential is still readable Migration moved the secret into the keychain and then suppressed any OSError from rewriting token.json, so a file that could not be rewritten kept the credential in cleartext while every command reported success. That file is now removed instead: signing in again costs one command, a stranded live credential costs the credential `lite logout` also reported a clean logout whenever the keyring package was missing, on the reasoning that an install without it could never have stored anything. The entry belongs to the OS, so a keychain-backed login survives a logout run from a venv without the cli extra. erase() now reports which keychain state applies, and logout warns with the advice that fixes each one, staying quiet for file-backed logins whose token file still carries its own secret Also pins the migration path's tightening of a world-readable legacy token.json, and moves the logout tests off patch() onto the injected vault --- litellm/litellm_core_utils/cli_keyring.py | 41 ++++++---- litellm/litellm_core_utils/cli_token_utils.py | 49 ++++++++++-- litellm/proxy/client/cli/commands/auth.py | 31 +++++--- tests/test_litellm/conftest.py | 13 +++- .../test_cli_token_utils.py | 77 ++++++++++++++++--- .../proxy/client/cli/test_auth_commands.py | 66 ++++++++++++++-- 6 files changed, 226 insertions(+), 51 deletions(-) diff --git a/litellm/litellm_core_utils/cli_keyring.py b/litellm/litellm_core_utils/cli_keyring.py index b19b3d3bc83..0497991c2a0 100644 --- a/litellm/litellm_core_utils/cli_keyring.py +++ b/litellm/litellm_core_utils/cli_keyring.py @@ -36,6 +36,16 @@ class SecretStored: pass +@dataclass(frozen=True, slots=True) +class SecretErased: + pass + + +@dataclass(frozen=True, slots=True) +class SecretStranded: + pass + + @dataclass(frozen=True, slots=True) class KeyringNotInstalled: pass @@ -54,6 +64,7 @@ class KeyringUnreachable: KeyringUnusable: TypeAlias = KeyringNotInstalled | KeyringDisabled | KeyringUnreachable SecretRead: TypeAlias = SecretFound | SecretMissing | KeyringUnusable SecretWrite: TypeAlias = SecretStored | KeyringUnusable +SecretErase: TypeAlias = SecretErased | SecretStranded | KeyringUnusable class SecretVault(Protocol): @@ -63,7 +74,7 @@ class SecretVault(Protocol): def write(self, blob: str) -> SecretWrite: ... - def erase(self) -> bool: ... + def erase(self) -> SecretErase: ... class KeyringApi(Protocol): @@ -117,33 +128,31 @@ class KeyringVault: return KeyringUnreachable() return SecretStored() - def erase(self) -> bool: - """Whether the keychain is guaranteed to hold no credential afterwards. + def erase(self) -> SecretErase: + """Remove our entry, reporting whether the keychain is guaranteed to be free of it. - An uninstalled `keyring` package can never have stored one. A kill switch set after - a credential was stored leaves that entry out of reach, so erasure cannot be promised. + A keychain out of reach is never an erasure: the entry belongs to the OS, not to this + install, so it outlives an uninstalled `keyring` package and a kill switch set after login. + Those cases are reported apart from a confirmed entry that would not delete, because only + the caller knows whether this machine ever put a secret in a keychain. """ - if _import_keyring() is None: - return True - if _keyring_disabled(): - return False match self.read(): - case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable(): - return False + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable() as unusable: + return unusable case SecretMissing(): - return True + return SecretErased() case SecretFound(): return self._delete() - def _delete(self) -> bool: + def _delete(self) -> SecretErase: api: Final = _keyring_api() if isinstance(api, (KeyringNotInstalled, KeyringDisabled)): - return False + return api try: api.delete_password(KEYRING_SERVICE, KEYRING_ACCOUNT) except Exception: # noqa: BLE001 # report the failure as a value so `lite logout` can warn - return False - return True + return SecretStranded() + return SecretErased() SYSTEM_KEYRING: Final[SecretVault] = KeyringVault() diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 40822e5b335..cd69f9470a3 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -25,9 +25,12 @@ from litellm.litellm_core_utils.cli_keyring import ( KeyringDisabled, KeyringNotInstalled, KeyringUnreachable, + SecretErase, + SecretErased, SecretFound, SecretMissing, SecretStored, + SecretStranded, SecretVault, SecretWrite, ) @@ -84,7 +87,7 @@ def load_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> CliTokenRecord | N def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRING) -> SecretWrite: - """Store a freshly minted credential. Reports whether the keychain took the secret, and why not""" + """Store a freshly minted credential. Reports where its secret material ended up, and why""" outcome: Final = ( SecretStored() if record.key is None @@ -94,11 +97,33 @@ def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRIN return outcome -def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> bool: - """Remove the credential from both stores. Returns whether the keychain is now free of it""" - erased: Final = vault.erase() +def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretErase: + """Remove the credential from both stores. Reports whether the keychain is now free of it""" + outcome: Final = vault.erase() + settled: Final = _nothing_left_behind(outcome) Path(get_cli_token_file_path()).unlink(missing_ok=True) - return erased + return SecretErased() if settled else outcome + + +def _nothing_left_behind(outcome: SecretErase) -> bool: + """Whether the keychain can be trusted to hold no credential of ours once the file is gone""" + match outcome: + case SecretErased(): + return True + case SecretStranded(): + return False + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable(): + return not _secret_lives_in_keychain() + + +def _secret_lives_in_keychain() -> bool: + """Whether the token file is the metadata half of a pair whose secret half went to a keychain. + + A file that still carries its own secret rules one out, which keeps `lite logout` quiet on the + machines that never had a keychain to begin with. + """ + record: Final = _read_token_file() + return record is not None and record.key is None and not record.jwt_token def get_litellm_gateway_api_key( @@ -200,10 +225,22 @@ def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliToken def _scrub_file_secret(record: CliTokenRecord) -> None: + """Leave no secret material in the token file once the vault holds it. + + A file that cannot be rewritten without the secret is removed instead. Signing in again costs + the user one command; a live credential left behind in cleartext costs them the credential. + """ if record.key is None and not record.jwt_token: return - with contextlib.suppress(OSError): + try: _write_token_file(_without_secret(record)) + except OSError: + _discard_token_file() + + +def _discard_token_file() -> None: + with contextlib.suppress(OSError): + Path(get_cli_token_file_path()).unlink(missing_ok=True) def _without_secret(record: CliTokenRecord) -> CliTokenRecord: diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index 03906f9b6df..4cf18435473 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -17,7 +17,9 @@ from litellm.litellm_core_utils.cli_keyring import ( KeyringDisabled, KeyringNotInstalled, KeyringUnreachable, + SecretErased, SecretStored, + SecretStranded, SecretVault, SecretWrite, ) @@ -80,12 +82,16 @@ class CliAuthResult(TypedDict): team_id: str | None -KEYCHAIN_UNREACHABLE_MESSAGE: Final = ( - "Your credential is stored in your OS keychain, which could not be read. Unlock it and retry, " - "install the keyring package with: pip install 'litellm[cli]', or run 'lite login'." +KEYRING_INSTALL_HINT: Final = "pip install 'litellm[cli]'" + +STRANDED_CREDENTIAL_MESSAGE: Final = ( + "Logged out locally, but your credential is still in the OS keychain and could not be removed." ) -KEYRING_INSTALL_HINT: Final = "pip install 'litellm[cli]'" +KEYCHAIN_UNREACHABLE_MESSAGE: Final = ( + "Your credential is stored in your OS keychain, which could not be read. Unlock it, or install " + f"the keyring package with: {KEYRING_INSTALL_HINT}. Run 'lite login' to start over." +) def storage_notice(outcome: SecretWrite) -> str: @@ -742,11 +748,18 @@ def login(ctx: click.Context, config_claude: bool): @click.pass_context def logout(ctx: click.Context): """Logout and clear stored authentication""" - if clear_cli_token(vault=context_secret_vault(ctx)): - click.echo("Logged out successfully. Authentication token cleared.") - return - click.echo("Logged out. The local token file is gone, but the OS keychain entry could not be removed.") - click.echo("Unlock your keychain and run 'lite logout' again to clear it.") + match clear_cli_token(vault=context_secret_vault(ctx)): + case SecretErased(): + click.echo("Logged out successfully. Authentication token cleared.") + case KeyringNotInstalled(): + click.echo(STRANDED_CREDENTIAL_MESSAGE) + click.echo(f"Install the keyring package with: {KEYRING_INSTALL_HINT}, then run 'lite logout' again.") + case KeyringDisabled(): + click.echo(STRANDED_CREDENTIAL_MESSAGE) + click.echo(f"Unset {DISABLE_KEYRING_ENV_VAR} and run 'lite logout' again to clear it.") + case SecretStranded() | KeyringUnreachable(): + click.echo(STRANDED_CREDENTIAL_MESSAGE) + click.echo("Unlock your keychain and run 'lite logout' again to clear it.") @click.command(name="print-token") diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py index a716ec0e0aa..b42355fa045 100644 --- a/tests/test_litellm/conftest.py +++ b/tests/test_litellm/conftest.py @@ -25,10 +25,13 @@ from litellm._logging import ALL_LOGGERS from litellm.litellm_core_utils.cli_keyring import ( KeyringUnreachable, KeyringUnusable, + SecretErase, + SecretErased, SecretFound, SecretMissing, SecretRead, SecretStored, + SecretStranded, SecretWrite, ) from litellm.litellm_core_utils.prompt_templates import ( @@ -163,12 +166,14 @@ class FakeSecretVault: self.blob = blob return SecretStored() - def erase(self) -> bool: + def erase(self) -> SecretErase: self.erases += 1 - if not (self.available and self.erasable): - return False + if not self.available: + return self.failure + if not self.erasable: + return SecretStranded() if self.blob is not None else SecretErased() self.blob = None - return True + return SecretErased() @pytest.fixture diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 961d986e8f5..925b440dfb7 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -16,7 +16,9 @@ from litellm.litellm_core_utils.cli_keyring import ( KeyringDisabled, KeyringNotInstalled, KeyringUnreachable, + SecretErased, SecretStored, + SecretStranded, ) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, @@ -126,6 +128,16 @@ class TestLoadCliToken: assert on_disk["user_email"] == "user@example.com" assert stat.S_IMODE(path.stat().st_mode) == 0o600 + def test_migration_tightens_a_world_readable_legacy_file(self, isolated_home, secret_vault_factory): + """An older `lite`, a loose umask, or a restored backup can leave token.json readable by + every account on the box. Migrating it must not preserve those permissions.""" + path = _write_legacy_file(isolated_home) + path.chmod(0o644) + + load_cli_token(vault=secret_vault_factory()) + + assert stat.S_IMODE(path.stat().st_mode) == 0o600 + def test_legacy_file_survives_a_vault_that_refuses_to_store(self, isolated_home, secret_vault_factory): """Scrubbing the only copy of the secret after a failed keychain write would log the user out for good.""" @@ -304,12 +316,35 @@ class TestSaveCliToken: assert list(path.parent.glob(".tmp-*")) == [] +class TestScrubFailure: + """A keychain that took the secret while the file kept it is the worst of both stores: the + credential is live, it is in cleartext on disk, and every command reports success.""" + + def test_a_file_that_cannot_be_rewritten_is_removed_instead( + self, isolated_home, secret_vault_factory, monkeypatch + ): + path = _write_legacy_file(isolated_home) + vault = secret_vault_factory() + + def _explode(*args, **kwargs): + raise OSError("no space left on device") + + monkeypatch.setattr("litellm.litellm_core_utils.private_json.json.dump", _explode) + + record = load_cli_token(vault=vault) + + assert record.key == "sk-legacy" + assert json.loads(vault.blob)["key"] == "sk-legacy" + assert not path.exists() + assert list(path.parent.glob(".tmp-*")) == [] + + class TestClearCliToken: def test_removes_the_credential_from_both_stores(self, isolated_home, secret_vault_factory): vault = secret_vault_factory() save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=vault) - assert clear_cli_token(vault=vault) is True + assert clear_cli_token(vault=vault) == SecretErased() assert vault.blob is None assert not _token_file(isolated_home).exists() assert load_cli_token(vault=vault) is None @@ -318,11 +353,32 @@ class TestClearCliToken: _write_legacy_file(isolated_home) vault = secret_vault_factory(blob=_blob(), erasable=False) - assert clear_cli_token(vault=vault) is False + assert clear_cli_token(vault=vault) == SecretStranded() + assert not _token_file(isolated_home).exists() + + def test_logout_from_an_install_without_keyring_does_not_claim_the_keychain_is_clear( + self, isolated_home, secret_vault_factory + ): + """Log in where `litellm[cli]` is installed and the secret goes to the OS keychain; log out + from a venv without it and the entry survives, because it belongs to the OS rather than to + the package. Reporting a clean logout there leaves a live credential the user thinks is gone.""" + _write_metadata_only_file(isolated_home) + vault = secret_vault_factory(available=False, failure=KeyringNotInstalled()) + + assert clear_cli_token(vault=vault) == KeyringNotInstalled() + assert not _token_file(isolated_home).exists() + + def test_a_file_backed_login_logs_out_quietly_without_keyring(self, isolated_home, secret_vault_factory): + """The complement: a user who never had a keychain keeps their whole credential in the file, + so removing it is a complete logout and must not warn about an entry that cannot exist.""" + _write_legacy_file(isolated_home) + vault = secret_vault_factory(available=False, failure=KeyringNotInstalled()) + + assert clear_cli_token(vault=vault) == SecretErased() assert not _token_file(isolated_home).exists() def test_is_safe_when_nothing_was_ever_stored(self, isolated_home, secret_vault_factory): - assert clear_cli_token(vault=secret_vault_factory()) is True + assert clear_cli_token(vault=secret_vault_factory()) == SecretErased() class TestIsCliTokenFresh: @@ -384,7 +440,7 @@ class TestKeyringVault: assert vault.write("blob-1") == SecretStored() assert vault.read() == SecretFound("blob-1") - assert vault.erase() is True + assert vault.erase() == SecretErased() assert vault.read() == SecretMissing() assert {call[1:] for call in fake.calls} == {(KEYRING_SERVICE, KEYRING_ACCOUNT)} @@ -392,24 +448,25 @@ class TestKeyringVault: """`LITELLM_CLI_DISABLE_KEYRING` has to work without importing keyring, because keyring caches its backend on first use and cannot be reconfigured later. Erase still fails: a credential stored before the switch was set may be in the keychain, and with reads - disabled `lite logout` cannot verify it is gone, so it must warn instead.""" + disabled `lite logout` cannot verify it is gone, so it must say so instead.""" monkeypatch.setenv(DISABLE_KEYRING_ENV_VAR, "1") vault = KeyringVault() assert vault.read() == KeyringDisabled() assert vault.write("blob-1") == KeyringDisabled() - assert vault.erase() is False + assert vault.erase() == KeyringDisabled() def test_an_uninstalled_keyring_library_degrades_to_the_file(self, monkeypatch): """keyring is an optional extra, so the SDK must survive its absence rather than raise on - the hot path.""" + the hot path. Erase cannot succeed: the entry belongs to the OS and outlives the package, + so an install without it is not evidence that the keychain is empty.""" monkeypatch.delenv(DISABLE_KEYRING_ENV_VAR, raising=False) monkeypatch.setitem(sys.modules, "keyring", None) vault = KeyringVault() assert vault.read() == KeyringNotInstalled() assert vault.write("blob-1") == KeyringNotInstalled() - assert vault.erase() is True + assert vault.erase() == KeyringNotInstalled() def test_a_locked_keychain_is_reported_not_raised(self, install_fake_keyring): install_fake_keyring(_FakeKeyringModule(get_error=RuntimeError("keyring is locked"))) @@ -424,9 +481,9 @@ class TestKeyringVault: def test_a_refused_delete_is_reported_so_logout_can_warn(self, install_fake_keyring): install_fake_keyring(_FakeKeyringModule(stored="blob-1", delete_error=RuntimeError("locked"))) - assert KeyringVault().erase() is False + assert KeyringVault().erase() == SecretStranded() def test_erasing_a_locked_keychain_is_a_failure(self, install_fake_keyring): install_fake_keyring(_FakeKeyringModule(get_error=RuntimeError("locked"))) - assert KeyringVault().erase() is False + assert KeyringVault().erase() == KeyringUnreachable() diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 1f2d48f6547..33bd8307c21 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -46,6 +46,12 @@ def _write_home_json(home: Path, filename: str, payload: dict[str, object]) -> N (litellm_dir / filename).write_text(json.dumps(payload)) +def _write_token_file(home: Path, *, key: str | None) -> None: + """A stored login: `key=None` is the metadata half of a keychain-backed pair, a key is a file-backed one.""" + payload: dict[str, object] = {"base_url": "https://test.example.com", "user_id": "u-1", "timestamp": time.time()} + _write_home_json(home, "token.json", payload if key is None else {**payload, "key": key}) + + def _secret_blob(base_url: str, key: str) -> str: return json.dumps({"base_url": base_url, "key": key, "jwt_token": ""}) @@ -427,14 +433,62 @@ class TestLogoutCommand: """Setup for each test""" self.runner = CliRunner() - def test_logout_success(self): + def test_logout_success(self, isolated_home, secret_vault_factory): """Test successful logout""" - with patch("litellm.proxy.client.cli.commands.auth.clear_cli_token") as mock_clear: - result = self.runner.invoke(logout) + vault = secret_vault_factory(blob=_secret_blob("https://test.example.com", "sk-stored")) + _write_token_file(isolated_home, key=None) - assert result.exit_code == 0 - assert "Logged out successfully" in result.output - mock_clear.assert_called_once() + result = self.runner.invoke(logout, obj={"secret_vault": vault}) + + assert result.exit_code == 0 + assert "Logged out successfully" in result.output + assert vault.blob is None + assert not (isolated_home / ".litellm" / "token.json").exists() + + def test_logout_without_the_keyring_package_does_not_claim_the_keychain_is_clear( + self, isolated_home, secret_vault_factory + ): + """Logging out from an install without the cli extra cannot touch an entry a keychain-backed + login left behind, so it must point at the package rather than report a clean logout.""" + _write_token_file(isolated_home, key=None) + + result = self.runner.invoke( + logout, obj={"secret_vault": secret_vault_factory(available=False, failure=KeyringNotInstalled())} + ) + + assert result.exit_code == 0 + assert "Logged out successfully" not in result.output + assert "still in the OS keychain" in result.output + assert "pip install 'litellm[cli]'" in result.output + + def test_logout_warns_when_the_keychain_refuses_to_release_the_entry( + self, isolated_home, secret_vault_factory + ): + """A locked keychain leaves a live credential behind that the user believes is gone.""" + vault = secret_vault_factory( + blob=_secret_blob("https://test.example.com", "sk-stored"), erasable=False + ) + _write_token_file(isolated_home, key=None) + + result = self.runner.invoke(logout, obj={"secret_vault": vault}) + + assert result.exit_code == 0 + assert "Logged out successfully" not in result.output + assert "still in the OS keychain" in result.output + assert "Unlock your keychain" in result.output + + def test_logout_from_a_file_only_login_stays_quiet(self, isolated_home, secret_vault_factory): + """The credential never went to a keychain, so removing the file is the whole logout and + warning about a keychain entry would send the user chasing one that cannot exist.""" + _write_token_file(isolated_home, key="sk-in-file") + + result = self.runner.invoke( + logout, obj={"secret_vault": secret_vault_factory(available=False, failure=KeyringNotInstalled())} + ) + + assert result.exit_code == 0 + assert "Logged out successfully" in result.output + assert "still in the OS keychain" not in result.output class TestWhoamiCommand: From 424e74ba9ffc9f33757d3c68f90d0fef2cda4079 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 20:31:43 -0700 Subject: [PATCH 022/220] fix(cli): roll the keychain write back when the plaintext copy cannot be removed Removing the file when it could not be rewritten covered a full disk, but not a ~/.litellm that permits neither the rewrite nor the delete, which is what a `sudo lite login` leaves behind. There the secret was copied into the keychain and kept in cleartext on disk, so migration widened exposure instead of narrowing it Migration now only keeps the vault copy if the file's copy is gone. When it is not, the write is rolled back and the user is left exactly as they were, logged in with one copy of the credential --- litellm/litellm_core_utils/cli_token_utils.py | 26 +++++++++++++------ .../test_cli_token_utils.py | 20 ++++++++++++++ 2 files changed, 38 insertions(+), 8 deletions(-) diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index cd69f9470a3..53289770c1a 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -12,7 +12,6 @@ first time it reads one. This module has no dependencies on proxy code and can be safely imported at the SDK level. """ -import contextlib import time from pathlib import Path from types import MappingProxyType @@ -217,30 +216,41 @@ def _apply_vault_secret(record: CliTokenRecord, blob: str, vault: SecretVault) - def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecord | None: + """Move a file-held secret into the vault, but only if the file's copy can be taken away. + + Migrating without scrubbing would leave the credential live in two stores instead of one, so a + file that will not give its copy up rolls the vault write back rather than widening exposure. + """ if record.key is None: return None - if isinstance(vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)), SecretStored): - _scrub_file_secret(record) + if not isinstance(vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)), SecretStored): + return record + if not _scrub_file_secret(record): + vault.erase() return record -def _scrub_file_secret(record: CliTokenRecord) -> None: +def _scrub_file_secret(record: CliTokenRecord) -> bool: """Leave no secret material in the token file once the vault holds it. A file that cannot be rewritten without the secret is removed instead. Signing in again costs the user one command; a live credential left behind in cleartext costs them the credential. """ if record.key is None and not record.jwt_token: - return + return True try: _write_token_file(_without_secret(record)) except OSError: - _discard_token_file() + return _discard_token_file() + return True -def _discard_token_file() -> None: - with contextlib.suppress(OSError): +def _discard_token_file() -> bool: + try: Path(get_cli_token_file_path()).unlink(missing_ok=True) + except OSError: + return False + return True def _without_secret(record: CliTokenRecord) -> CliTokenRecord: diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 925b440dfb7..cbc93bbde71 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -1,4 +1,5 @@ import json +import os import stat import sys import time @@ -320,6 +321,25 @@ class TestScrubFailure: """A keychain that took the secret while the file kept it is the worst of both stores: the credential is live, it is in cleartext on disk, and every command reports success.""" + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") + def test_a_file_that_will_not_give_its_copy_up_rolls_the_vault_write_back( + self, isolated_home, secret_vault_factory + ): + """Handing the keychain a copy without taking the file's away leaves the credential live in + two stores instead of one. A directory that permits neither the rewrite nor the delete, a + root-owned ~/.litellm left behind by a `sudo lite login`, must widen nothing.""" + path = _write_legacy_file(isolated_home) + vault = secret_vault_factory() + path.parent.chmod(0o500) + try: + record = load_cli_token(vault=vault) + finally: + path.parent.chmod(0o700) + + assert record.key == "sk-legacy" + assert json.loads(path.read_text())["key"] == "sk-legacy" + assert vault.blob is None + def test_a_file_that_cannot_be_rewritten_is_removed_instead( self, isolated_home, secret_vault_factory, monkeypatch ): From 3c73a39877fa97db3dff4e22dc685bcb77fa4041 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 21:10:59 -0700 Subject: [PATCH 023/220] fix(cli): verify every credential store transition before reporting it done A keyring backend can accept a write and keep nothing. That is exactly what `keyring --disable` and PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring select, and it raises nothing to distinguish itself, so `lite login` was handing the credential to a black hole, scrubbing its own copy from token.json, and printing a success message over a login that no longer worked. Reading the value back is the only way to tell that backend apart from a keychain that really stored the secret. The same rule closes the rest of the gaps. A credential the token file will not record is taken back out of the keychain instead of being left live on a machine with no record of it, and is reported rather than raised. The migration stages its scrubbed file before the keychain is handed anything, so a directory that will not accept the rewrite stops the move rather than leaving the secret in two places. Logout no longer reads a key in the file as proof that the keychain is clear, which was never sound across two separate runs, and only draws that conclusion when the `keyring` package is missing outright, where nothing could have reached a keychain at all. --- litellm/litellm_core_utils/cli_keyring.py | 26 +++- litellm/litellm_core_utils/cli_token_utils.py | 104 +++++++++----- litellm/litellm_core_utils/private_json.py | 32 ++++- litellm/proxy/client/cli/commands/auth.py | 63 +++++++-- .../test_cli_token_utils.py | 131 ++++++++++++++++-- .../proxy/client/cli/test_auth_commands.py | 62 ++++++++- 6 files changed, 356 insertions(+), 62 deletions(-) diff --git a/litellm/litellm_core_utils/cli_keyring.py b/litellm/litellm_core_utils/cli_keyring.py index 0497991c2a0..15282fc522c 100644 --- a/litellm/litellm_core_utils/cli_keyring.py +++ b/litellm/litellm_core_utils/cli_keyring.py @@ -6,8 +6,12 @@ Linux Secret Service) that holds the credential minted by `lite login`. The `keyring` package is optional and imported lazily, so importing this module never pulls it in. Every failure is returned as a value, naming which of the -three ways the keychain can be out of reach applies, so callers can degrade to -the token file and tell the user what to do about it. +ways the keychain can be out of reach applies, so callers can degrade to the +token file and tell the user what to do about it. + +A write is only reported as stored once it has been read back, because keyring's +null backend, which `keyring --disable` and headless CI images both select, +accepts every write and keeps nothing. """ import os @@ -61,7 +65,12 @@ class KeyringUnreachable: pass -KeyringUnusable: TypeAlias = KeyringNotInstalled | KeyringDisabled | KeyringUnreachable +@dataclass(frozen=True, slots=True) +class KeyringDiscardsWrites: + pass + + +KeyringUnusable: TypeAlias = KeyringNotInstalled | KeyringDisabled | KeyringUnreachable | KeyringDiscardsWrites SecretRead: TypeAlias = SecretFound | SecretMissing | KeyringUnusable SecretWrite: TypeAlias = SecretStored | KeyringUnusable SecretErase: TypeAlias = SecretErased | SecretStranded | KeyringUnusable @@ -119,6 +128,13 @@ class KeyringVault: return SecretMissing() if blob is None else SecretFound(blob) def write(self, blob: str) -> SecretWrite: + """Store the secret, reporting stored only once the keychain hands the same bytes back. + + A backend that accepts writes and keeps nothing, which is exactly what `keyring --disable` + and `PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring` select, raises nothing to + distinguish itself. Reading the value back is the only way to tell it apart from a keychain + that really stored the credential, and the caller is about to drop its own copy on our word. + """ api: Final = _keyring_api() if isinstance(api, (KeyringNotInstalled, KeyringDisabled)): return api @@ -126,7 +142,7 @@ class KeyringVault: api.set_password(KEYRING_SERVICE, KEYRING_ACCOUNT, blob) except Exception: # noqa: BLE001 # a keychain that refuses the write falls back to the token file return KeyringUnreachable() - return SecretStored() + return SecretStored() if self.read() == SecretFound(blob) else KeyringDiscardsWrites() def erase(self) -> SecretErase: """Remove our entry, reporting whether the keychain is guaranteed to be free of it. @@ -137,7 +153,7 @@ class KeyringVault: the caller knows whether this machine ever put a secret in a keychain. """ match self.read(): - case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable() as unusable: + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable() | KeyringDiscardsWrites() as unusable: return unusable case SecretMissing(): return SecretErased() diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 53289770c1a..af1b918fc2a 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -13,15 +13,17 @@ This module has no dependencies on proxy code and can be safely imported at the """ import time +from dataclasses import dataclass from pathlib import Path from types import MappingProxyType -from typing import Final +from typing import Final, TypeAlias from pydantic import BaseModel, ConfigDict, ValidationError from litellm.litellm_core_utils.cli_keyring import ( SYSTEM_KEYRING, KeyringDisabled, + KeyringDiscardsWrites, KeyringNotInstalled, KeyringUnreachable, SecretErase, @@ -33,7 +35,23 @@ from litellm.litellm_core_utils.cli_keyring import ( SecretVault, SecretWrite, ) -from litellm.litellm_core_utils.private_json import ensure_private_dir, write_private_json +from litellm.litellm_core_utils.private_json import ( + commit_staged_json, + discard_staged_json, + ensure_private_dir, + stage_private_json, + write_private_json, +) + + +@dataclass(frozen=True, slots=True) +class CredentialNotSaved: + """The credential was minted but no store would keep it, so this machine has none.""" + + detail: str + + +SecretSave: TypeAlias = SecretWrite | CredentialNotSaved class CliTokenRecord(BaseModel): @@ -85,14 +103,24 @@ def load_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> CliTokenRecord | N return _resolve_secret(record, vault) -def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRING) -> SecretWrite: - """Store a freshly minted credential. Reports where its secret material ended up, and why""" +def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRING) -> SecretSave: + """Store a freshly minted credential. Reports where its secret material ended up, and why. + + The token file is what makes a keychain-backed credential findable again, so a file that will + not be written takes the keychain copy down with it rather than leaving a live credential + stored under a machine that has no record of it. + """ outcome: Final = ( SecretStored() if record.key is None else vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)) ) - _write_token_file(_without_secret(record) if isinstance(outcome, SecretStored) else record) + try: + _write_token_file(_without_secret(record) if isinstance(outcome, SecretStored) else record) + except OSError as error: + if record.key is not None and isinstance(outcome, SecretStored): + vault.erase() + return CredentialNotSaved(str(error)) return outcome @@ -105,24 +133,30 @@ def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretErase: def _nothing_left_behind(outcome: SecretErase) -> bool: - """Whether the keychain can be trusted to hold no credential of ours once the file is gone""" + """Whether the keychain can be trusted to hold no credential of ours once the file is gone. + + A keychain that exists but is out of reach right now is never trusted, whatever the token file + looks like: the login that stored a secret there and the logout that cannot remove it are + separate runs, free to differ in whether the keychain was usable at the time. + """ match outcome: case SecretErased(): return True - case SecretStranded(): + case SecretStranded() | KeyringDisabled() | KeyringUnreachable() | KeyringDiscardsWrites(): return False - case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable(): - return not _secret_lives_in_keychain() + case KeyringNotInstalled(): + return _file_holds_its_own_secret() -def _secret_lives_in_keychain() -> bool: - """Whether the token file is the metadata half of a pair whose secret half went to a keychain. +def _file_holds_its_own_secret() -> bool: + """Whether the stored login keeps its secret in the token file, ruling out a keychain entry. - A file that still carries its own secret rules one out, which keeps `lite logout` quiet on the - machines that never had a keychain to begin with. + Sound only against a missing `keyring` package, the one way to lose the keychain that had to + hold at storage time too, since nothing here can reach a keychain without it. A file whose + secret half is absent went to a keychain by definition, and so rules nothing out. """ record: Final = _read_token_file() - return record is not None and record.key is None and not record.jwt_token + return record is not None and record.key is not None def get_litellm_gateway_api_key( @@ -179,7 +213,7 @@ def is_cli_token_fresh(token_data: CliTokenRecord, buffer_hours: float = 0.1) -> def _read_token_file() -> CliTokenRecord | None: try: raw: Final = Path(get_cli_token_file_path()).read_text() - except OSError: + except (OSError, ValueError): return None try: return CliTokenRecord.model_validate_json(raw) @@ -193,7 +227,7 @@ def _resolve_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecor return _apply_vault_secret(record, blob, vault) case SecretMissing(): return _migrate_file_secret(record, vault) - case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable(): + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable() | KeyringDiscardsWrites(): return record @@ -216,38 +250,46 @@ def _apply_vault_secret(record: CliTokenRecord, blob: str, vault: SecretVault) - def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecord | None: - """Move a file-held secret into the vault, but only if the file's copy can be taken away. + """Move a file-held secret into the vault, but only once the file's copy can be taken away. - Migrating without scrubbing would leave the credential live in two stores instead of one, so a - file that will not give its copy up rolls the vault write back rather than widening exposure. + The scrubbed file is staged first so a directory that will not accept it stops the migration + before the keychain is handed anything. Copying the credential into a second store and only + then discovering the first one cannot be cleaned would widen exposure instead of narrowing it, + which is the opposite of what moving it into the keychain is for. """ if record.key is None: return None - if not isinstance(vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)), SecretStored): + staged: Final = _stage_scrubbed_file(record) + if staged is None: return record - if not _scrub_file_secret(record): + if not isinstance(vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)), SecretStored): + discard_staged_json(staged) + return record + if not _commit_scrubbed_file(staged): vault.erase() return record def _scrub_file_secret(record: CliTokenRecord) -> bool: - """Leave no secret material in the token file once the vault holds it. - - A file that cannot be rewritten without the secret is removed instead. Signing in again costs - the user one command; a live credential left behind in cleartext costs them the credential. - """ + """Leave no secret material in the token file once the vault holds it""" if record.key is None and not record.jwt_token: return True + staged: Final = _stage_scrubbed_file(record) + return staged is not None and _commit_scrubbed_file(staged) + + +def _stage_scrubbed_file(record: CliTokenRecord) -> str | None: + path: Final = Path(get_cli_token_file_path()) try: - _write_token_file(_without_secret(record)) + ensure_private_dir(path.parent) + return stage_private_json(str(path), _without_secret(record).model_dump(exclude_none=True)) except OSError: - return _discard_token_file() - return True + return None -def _discard_token_file() -> bool: +def _commit_scrubbed_file(staged: str) -> bool: try: - Path(get_cli_token_file_path()).unlink(missing_ok=True) + commit_staged_json(staged, get_cli_token_file_path()) except OSError: return False return True diff --git a/litellm/litellm_core_utils/private_json.py b/litellm/litellm_core_utils/private_json.py index 32bc2e169e2..fbeb74aab5a 100644 --- a/litellm/litellm_core_utils/private_json.py +++ b/litellm/litellm_core_utils/private_json.py @@ -16,8 +16,12 @@ def ensure_private_dir(directory: Path) -> None: directory.chmod(PRIVATE_DIR_MODE) -def write_private_json(path: str, data: Mapping[str, object]) -> None: - """Atomically write JSON to path with owner-only permissions (0600)""" +def stage_private_json(path: str, data: Mapping[str, object]) -> str: + """Write JSON to a private temp file beside `path`, ready for `commit_staged_json`. + + Staging is the half that can fail on a read-only or full directory, so callers with something + to lose can find that out before they act on the assumption that the rewrite will land. + """ parent: Final = Path(path).parent parent.mkdir(parents=True, exist_ok=True) fd, tmp_path = tempfile.mkstemp(dir=str(parent), prefix=".tmp-", suffix=".json") @@ -26,6 +30,26 @@ def write_private_json(path: str, data: Mapping[str, object]) -> None: json.dump(data, f, indent=2) f.flush() os.fsync(f.fileno()) - os.replace(tmp_path, path) - finally: + except BaseException: Path(tmp_path).unlink(missing_ok=True) + raise + return tmp_path + + +def commit_staged_json(staged: str, path: str) -> None: + """Move a staged file into place, replacing whatever is there in one step""" + try: + os.replace(staged, path) + except OSError: + Path(staged).unlink(missing_ok=True) + raise + + +def discard_staged_json(staged: str) -> None: + """Throw a staged file away when the change it was part of is abandoned""" + Path(staged).unlink(missing_ok=True) + + +def write_private_json(path: str, data: Mapping[str, object]) -> None: + """Atomically write JSON to path with owner-only permissions (0600)""" + commit_staged_json(stage_private_json(path, data), path) diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index 4cf18435473..eba9994f7ec 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -15,16 +15,20 @@ from litellm.litellm_core_utils.cli_keyring import ( DISABLE_KEYRING_ENV_VAR, SYSTEM_KEYRING, KeyringDisabled, + KeyringDiscardsWrites, KeyringNotInstalled, KeyringUnreachable, SecretErased, + SecretFound, + SecretMissing, SecretStored, SecretStranded, SecretVault, - SecretWrite, ) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, + CredentialNotSaved, + SecretSave, clear_cli_token, get_cli_token_file_path, get_litellm_gateway_api_key, @@ -84,17 +88,19 @@ class CliAuthResult(TypedDict): KEYRING_INSTALL_HINT: Final = "pip install 'litellm[cli]'" +KEYRING_ENABLE_HINT: Final = "keyring --enable (or unset PYTHON_KEYRING_BACKEND)" + STRANDED_CREDENTIAL_MESSAGE: Final = ( "Logged out locally, but your credential is still in the OS keychain and could not be removed." ) -KEYCHAIN_UNREACHABLE_MESSAGE: Final = ( - "Your credential is stored in your OS keychain, which could not be read. Unlock it, or install " - f"the keyring package with: {KEYRING_INSTALL_HINT}. Run 'lite login' to start over." +UNCHECKED_KEYCHAIN_MESSAGE: Final = ( + "Logged out locally, but your OS keychain could not be checked, so a credential stored there by " + "an earlier login may still be usable." ) -def storage_notice(outcome: SecretWrite) -> str: +def storage_notice(outcome: SecretSave) -> str: """Tell the user where the credential ended up, and how to get keychain storage if it did not.""" path: Final = get_cli_token_file_path() match outcome: @@ -109,6 +115,38 @@ def storage_notice(outcome: SecretWrite) -> str: return f"Keychain storage is off ({DISABLE_KEYRING_ENV_VAR}). Credential stored in {path} (owner-only)." case KeyringUnreachable(): return f"No OS keychain available. Credential stored in {path} (owner-only)." + case KeyringDiscardsWrites(): + return ( + f"Your keyring backend keeps nothing it is given, so the credential was stored in {path} " + f"(owner-only) instead. For OS keychain storage, run: {KEYRING_ENABLE_HINT}" + ) + case CredentialNotSaved(detail=detail): + return ( + f"Signed in, but the credential could not be saved to {path}: {detail}. " + "Nothing was kept, so run 'lite login' again once that path is writable." + ) + + +def keychain_unreadable_notice(vault: SecretVault) -> str: + """Explain why the secret half of a stored login cannot be produced, and what fixes it""" + match vault.read(): + case KeyringNotInstalled(): + return ( + "Your credential is in your OS keychain, which this install cannot read without the " + f"keyring package. Install it with: {KEYRING_INSTALL_HINT}, or run 'lite login' to start over." + ) + case KeyringDisabled(): + return ( + f"Your credential is in your OS keychain, which {DISABLE_KEYRING_ENV_VAR} is blocking. " + "Unset it, or run 'lite login' to start over." + ) + case KeyringUnreachable() | KeyringDiscardsWrites(): + return ( + "Your credential is in your OS keychain, which could not be read. Unlock it, or run " + "'lite login' to start over." + ) + case SecretFound() | SecretMissing(): + return "Your credential could not be read from your OS keychain. Run 'lite login' to start over." def context_secret_vault(ctx: click.Context) -> SecretVault: @@ -715,6 +753,8 @@ def login(ctx: click.Context, config_claude: bool): click.echo("\nLogin successful!") click.echo(f"JWT Token: {api_key[:20]}...") click.echo(storage_notice(stored)) + if isinstance(stored, CredentialNotSaved): + return click.echo("You can now use the CLI without specifying --api-key") if config_claude: @@ -751,14 +791,17 @@ def logout(ctx: click.Context): match clear_cli_token(vault=context_secret_vault(ctx)): case SecretErased(): click.echo("Logged out successfully. Authentication token cleared.") + case SecretStranded(): + click.echo(STRANDED_CREDENTIAL_MESSAGE) + click.echo("Unlock your keychain and run 'lite logout' again to clear it.") case KeyringNotInstalled(): click.echo(STRANDED_CREDENTIAL_MESSAGE) click.echo(f"Install the keyring package with: {KEYRING_INSTALL_HINT}, then run 'lite logout' again.") case KeyringDisabled(): - click.echo(STRANDED_CREDENTIAL_MESSAGE) + click.echo(UNCHECKED_KEYCHAIN_MESSAGE) click.echo(f"Unset {DISABLE_KEYRING_ENV_VAR} and run 'lite logout' again to clear it.") - case SecretStranded() | KeyringUnreachable(): - click.echo(STRANDED_CREDENTIAL_MESSAGE) + case KeyringUnreachable() | KeyringDiscardsWrites(): + click.echo(UNCHECKED_KEYCHAIN_MESSAGE) click.echo("Unlock your keychain and run 'lite logout' again to clear it.") @@ -795,7 +838,7 @@ def print_token(ctx: click.Context): api_key: Final = token_data.key if not api_key: - click.echo(KEYCHAIN_UNREACHABLE_MESSAGE, err=True) + click.echo(keychain_unreadable_notice(context_secret_vault(ctx)), err=True) sys.exit(1) click.echo(api_key) @@ -821,7 +864,7 @@ def whoami(ctx: click.Context): click.echo(f"Token age: {age_hours:.1f} hours") if token_data.key is None: - click.echo(KEYCHAIN_UNREACHABLE_MESSAGE) + click.echo(keychain_unreadable_notice(context_secret_vault(ctx))) if age_hours > CLI_JWT_EXPIRATION_HOURS: click.echo(f"Warning: Token is more than {CLI_JWT_EXPIRATION_HOURS} hours old and may have expired.") diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index cbc93bbde71..e0f5f99dc1b 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -15,6 +15,7 @@ from litellm.litellm_core_utils.cli_keyring import ( SecretFound, SecretMissing, KeyringDisabled, + KeyringDiscardsWrites, KeyringNotInstalled, KeyringUnreachable, SecretErased, @@ -23,6 +24,7 @@ from litellm.litellm_core_utils.cli_keyring import ( ) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, + CredentialNotSaved, clear_cli_token, get_cli_token_file_path, get_litellm_gateway_api_key, @@ -225,6 +227,15 @@ class TestLoadCliToken: assert record.key == "sk-legacy" + def test_a_token_file_that_is_not_text_is_not_a_login(self, isolated_home, secret_vault_factory): + """A truncated write or a half-synced backup can leave bytes that are not UTF-8 at all. + Reading them must fail the way an absent file does, not crash every `lite` command.""" + path = _token_file(isolated_home) + path.parent.mkdir(parents=True, exist_ok=True) + path.write_bytes(b"\xff\xfe not utf-8 at all") + + assert load_cli_token(vault=secret_vault_factory()) is None + def test_corrupt_token_file_is_not_a_login(self, isolated_home, secret_vault_factory): _token_file(isolated_home).parent.mkdir() _token_file(isolated_home).write_text("not json at all {{{") @@ -301,6 +312,40 @@ class TestSaveCliToken: assert stat.S_IMODE(config_dir.stat().st_mode) == 0o700 + def test_a_credential_no_store_would_keep_is_reported_rather_than_raised( + self, isolated_home, secret_vault_factory, monkeypatch + ): + """`lite login` catches whatever escapes here and calls it an authentication failure, which + is the one thing that did not happen: the proxy minted a real credential. Saying so lets the + user act on the actual problem instead of retrying a sign-in that already worked.""" + + def _explode(*args, **kwargs): + raise OSError("read-only file system") + + monkeypatch.setattr("litellm.litellm_core_utils.private_json.json.dump", _explode) + + outcome = save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=secret_vault_factory()) + + assert isinstance(outcome, CredentialNotSaved) + assert "read-only file system" in outcome.detail + + def test_a_credential_the_file_will_not_record_is_taken_back_out_of_the_keychain( + self, isolated_home, secret_vault_factory, monkeypatch + ): + """The token file is what makes a keychain entry findable again. Leaving the secret in the + keychain with nothing pointing at it strands a live credential under a machine that has no + idea it is there, and no `lite logout` would ever go looking for it.""" + vault = secret_vault_factory() + + def _explode(*args, **kwargs): + raise OSError("read-only file system") + + monkeypatch.setattr("litellm.litellm_core_utils.private_json.json.dump", _explode) + + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=vault) + + assert vault.blob is None + def test_a_failed_write_leaves_the_previous_credential_intact(self, isolated_home, secret_vault_factory, monkeypatch): path = _write_legacy_file(isolated_home) before = path.read_text() @@ -340,9 +385,11 @@ class TestScrubFailure: assert json.loads(path.read_text())["key"] == "sk-legacy" assert vault.blob is None - def test_a_file_that_cannot_be_rewritten_is_removed_instead( + def test_a_full_disk_stops_the_migration_before_the_keychain_is_handed_anything( self, isolated_home, secret_vault_factory, monkeypatch ): + """The scrubbed file is staged first precisely so this is knowable in advance. A disk that + cannot take the rewrite leaves the credential where it already was, in one store.""" path = _write_legacy_file(isolated_home) vault = secret_vault_factory() @@ -354,8 +401,8 @@ class TestScrubFailure: record = load_cli_token(vault=vault) assert record.key == "sk-legacy" - assert json.loads(vault.blob)["key"] == "sk-legacy" - assert not path.exists() + assert vault.blob is None + assert json.loads(path.read_text())["key"] == "sk-legacy" assert list(path.parent.glob(".tmp-*")) == [] @@ -376,12 +423,39 @@ class TestClearCliToken: assert clear_cli_token(vault=vault) == SecretStranded() assert not _token_file(isolated_home).exists() + @pytest.mark.parametrize( + "failure", [KeyringDisabled(), KeyringUnreachable(), KeyringDiscardsWrites()] + ) + def test_a_secret_in_the_file_is_no_evidence_about_a_keychain_that_exists( + self, isolated_home, secret_vault_factory, failure + ): + """Store a secret in the keychain, sign in again while the keychain is unusable so the new + secret lands in the file, then log out while it is still unusable. The file now carries its + own secret and the first login's entry is still there, so reading the file as proof of a + clean keychain reports a logout that did not happen.""" + _write_legacy_file(isolated_home) + vault = secret_vault_factory(available=False, failure=failure) + + assert clear_cli_token(vault=vault) == failure + assert not _token_file(isolated_home).exists() + + def test_a_second_logout_still_reports_the_keychain_it_could_not_clear( + self, isolated_home, secret_vault_factory + ): + """The first logout deletes the file and tells the user to run it again once the keychain is + reachable. If the second run reads that missing file as proof of a clean keychain, the advice + turns into the very false all-clear it was issued to prevent.""" + _write_metadata_only_file(isolated_home) + vault = secret_vault_factory(available=False, failure=KeyringUnreachable()) + + assert clear_cli_token(vault=vault) == KeyringUnreachable() + assert clear_cli_token(vault=vault) == KeyringUnreachable() + def test_logout_from_an_install_without_keyring_does_not_claim_the_keychain_is_clear( self, isolated_home, secret_vault_factory ): - """Log in where `litellm[cli]` is installed and the secret goes to the OS keychain; log out - from a venv without it and the entry survives, because it belongs to the OS rather than to - the package. Reporting a clean logout there leaves a live credential the user thinks is gone.""" + """A file holding only metadata put its secret in a keychain by definition. Losing the + package that reaches it does not take the entry with it, so this cannot report success.""" _write_metadata_only_file(isolated_home) vault = secret_vault_factory(available=False, failure=KeyringNotInstalled()) @@ -389,8 +463,9 @@ class TestClearCliToken: assert not _token_file(isolated_home).exists() def test_a_file_backed_login_logs_out_quietly_without_keyring(self, isolated_home, secret_vault_factory): - """The complement: a user who never had a keychain keeps their whole credential in the file, - so removing it is a complete logout and must not warn about an entry that cannot exist.""" + """The complement, and the one inference the file does support: nothing here can reach a + keychain without the package, so an install that lacks it and a file that still holds its + own secret between them account for the whole credential.""" _write_legacy_file(isolated_home) vault = secret_vault_factory(available=False, failure=KeyringNotInstalled()) @@ -417,11 +492,12 @@ class TestIsCliTokenFresh: class _FakeKeyringModule: - def __init__(self, stored=None, *, get_error=None, set_error=None, delete_error=None): + def __init__(self, stored=None, *, get_error=None, set_error=None, delete_error=None, discard=False): self.stored = stored self.get_error = get_error self.set_error = set_error self.delete_error = delete_error + self.discard = discard self.calls = [] def get_password(self, service_name, username): @@ -434,6 +510,8 @@ class _FakeKeyringModule: self.calls.append(("set", service_name, username)) if self.set_error is not None: raise self.set_error + if self.discard: + return self.stored = password def delete_password(self, service_name, username): @@ -503,6 +581,41 @@ class TestKeyringVault: assert KeyringVault().erase() == SecretStranded() + def test_a_backend_that_keeps_nothing_is_not_a_successful_write(self, install_fake_keyring): + """keyring's null backend accepts every write, stores nothing, and raises nothing to say so. + Taking its silence for success is how a credential gets deleted: the caller drops its own + copy on our word. Only reading the value back tells the two apart.""" + fake = install_fake_keyring(_FakeKeyringModule(discard=True)) + + assert KeyringVault().write("blob-1") == KeyringDiscardsWrites() + assert fake.stored is None + + def test_the_real_null_backend_is_rejected(self, monkeypatch): + """Pinned against the actual library rather than the double above, because the whole risk is + that upstream's no-op write looks exactly like a successful one.""" + keyring = pytest.importorskip("keyring") + null_backend = pytest.importorskip("keyring.backends.null") + monkeypatch.delenv(DISABLE_KEYRING_ENV_VAR, raising=False) + previous = keyring.get_keyring() + keyring.set_keyring(null_backend.Keyring()) + try: + assert KeyringVault().write("blob-1") == KeyringDiscardsWrites() + finally: + keyring.set_keyring(previous) + + def test_a_credential_survives_a_backend_that_keeps_nothing( + self, isolated_home, install_fake_keyring + ): + """The end of the same story: the credential must still be usable afterwards. Reporting the + discard is only worth anything if the token file then keeps the copy the keychain refused.""" + install_fake_keyring(_FakeKeyringModule(discard=True)) + + outcome = save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-only-copy")) + + assert outcome == KeyringDiscardsWrites() + assert json.loads(_token_file(isolated_home).read_text())["key"] == "sk-only-copy" + assert load_cli_token().key == "sk-only-copy" + def test_erasing_a_locked_keychain_is_a_failure(self, install_fake_keyring): install_fake_keyring(_FakeKeyringModule(get_error=RuntimeError("locked"))) diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 33bd8307c21..8e1551c0720 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -16,12 +16,13 @@ from litellm.constants import CLI_JWT_EXPIRATION_HOURS from litellm.litellm_core_utils.cli_keyring import ( DISABLE_KEYRING_ENV_VAR, KeyringDisabled, + KeyringDiscardsWrites, KeyringNotInstalled, ) from litellm.litellm_core_utils.cli_token_utils import CliTokenRecord, save_cli_token from litellm.proxy.client.cli import cli from litellm.proxy.client.cli.commands.auth import ( - KEYCHAIN_UNREACHABLE_MESSAGE, + DISABLE_KEYRING_ENV_VAR, get_stored_api_key, login, logout, @@ -461,6 +462,20 @@ class TestLogoutCommand: assert "still in the OS keychain" in result.output assert "pip install 'litellm[cli]'" in result.output + def test_logout_does_not_call_an_unusable_keychain_clean(self, isolated_home, secret_vault_factory): + """A keychain-backed login, then a login that fell back to the file because the keychain had + become unusable, leaves the first entry live. The file's own secret says nothing about it, + so a clean bill of health here is the one answer that cannot be justified.""" + _write_token_file(isolated_home, key="sk-in-file") + vault = secret_vault_factory(available=False, failure=KeyringDisabled()) + + result = self.runner.invoke(logout, obj={"secret_vault": vault}) + + assert result.exit_code == 0 + assert "Logged out successfully" not in result.output + assert "could not be checked" in result.output + assert DISABLE_KEYRING_ENV_VAR in result.output + def test_logout_warns_when_the_keychain_refuses_to_release_the_entry( self, isolated_home, secret_vault_factory ): @@ -1003,6 +1018,19 @@ class TestKeychainBackedCommands: assert "No OS keychain available" not in result.output assert json.loads(token_file.read_text())["key"] == "sk-minted" + def test_login_keeps_the_credential_when_the_backend_keeps_nothing( + self, isolated_home, secret_vault_factory + ): + """A backend that accepts writes and stores nothing must not be reported as keychain + storage, because the file is then told to drop the only remaining copy.""" + result = self._login(secret_vault_factory(available=False, failure=KeyringDiscardsWrites())) + + token_file = isolated_home / ".litellm" / "token.json" + assert result.exit_code == 0 + assert "Credential stored in your OS keychain." not in result.output + assert "keyring --enable" in result.output + assert json.loads(token_file.read_text())["key"] == "sk-minted" + def test_login_names_the_kill_switch_instead_of_blaming_the_machine( self, isolated_home, secret_vault_factory ): @@ -1061,7 +1089,8 @@ class TestKeychainBackedCommands: result = self.runner.invoke(print_token, obj=obj) assert result.exit_code == 1 - assert KEYCHAIN_UNREACHABLE_MESSAGE in result.output + assert "could not be read" in result.output + assert "lite login" in result.output def test_whoami_flags_a_locked_keychain(self, isolated_home, secret_vault_factory): _write_home_json( @@ -1074,7 +1103,34 @@ class TestKeychainBackedCommands: result = self.runner.invoke(whoami, obj=obj) assert "Authenticated" in result.output - assert KEYCHAIN_UNREACHABLE_MESSAGE in result.output + assert "could not be read" in result.output + + def test_whoami_names_the_kill_switch_rather_than_a_missing_package( + self, isolated_home, secret_vault_factory + ): + """Every unreachable keychain used to be described as a locked one needing the keyring + package installed. Someone who set the kill switch has the package and an unlocked keychain, + so that advice sends them to fix two things that were never wrong.""" + _write_token_file(isolated_home, key=None) + vault = secret_vault_factory(available=False, failure=KeyringDisabled()) + + result = self.runner.invoke(whoami, obj={"base_url": "https://test.example.com", "secret_vault": vault}) + + assert DISABLE_KEYRING_ENV_VAR in result.output + assert "pip install" not in result.output + + def test_print_token_points_an_install_without_keyring_at_the_package( + self, isolated_home, secret_vault_factory + ): + _write_token_file(isolated_home, key=None) + vault = secret_vault_factory(available=False, failure=KeyringNotInstalled()) + obj = {"base_url": "https://test.example.com", "secret_vault": vault} + + result = self.runner.invoke(print_token, obj=obj) + + assert result.exit_code == 1 + assert "pip install 'litellm[cli]'" in result.output + assert DISABLE_KEYRING_ENV_VAR not in result.output class TestApiKeyPrecedence: From 4829bb3a151c4b4e58580313784862eb966921e1 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 21:25:06 -0700 Subject: [PATCH 024/220] fix(prompt_management): don't route requests without prompt_id to prompt managers that can't run them UI-injected empty vector_store_ids/tags/guardrails on a DB model tripped the dynamic-param check, and the prompt-management fallback then handed the request to the first registered prompt manager (e.g. a saved dotprompt), whose sync path raised "prompt_id is required" as a 500 on every /chat/completions call. Empty dynamic params no longer count as a trigger, the fallback skips managers whose should_run_prompt_management declines a None prompt_id, and the sync base path returns the request unchanged for a None prompt_id like the async path. --- .../integrations/prompt_management_base.py | 2 +- litellm/litellm_core_utils/litellm_logging.py | 30 ++++++- .../test_litellm_logging.py | 78 +++++++++++++++++++ 3 files changed, 105 insertions(+), 5 deletions(-) diff --git a/litellm/integrations/prompt_management_base.py b/litellm/integrations/prompt_management_base.py index d16afa92ec2..81c01599e77 100644 --- a/litellm/integrations/prompt_management_base.py +++ b/litellm/integrations/prompt_management_base.py @@ -165,7 +165,7 @@ class PromptManagementBase(ABC): ignore_prompt_manager_optional_params: bool | None = False, ) -> tuple[str, list[AllMessageValues], dict]: if prompt_id is None: - raise ValueError("prompt_id is required for Prompt Management Base class") + return model, messages, non_default_params if not self.should_run_prompt_management( prompt_id=prompt_id, prompt_spec=prompt_spec, diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py index 91a312b4f45..9b7707eabe1 100644 --- a/litellm/litellm_core_utils/litellm_logging.py +++ b/litellm/litellm_core_utils/litellm_logging.py @@ -832,8 +832,8 @@ class Logging(LiteLLMLoggingBaseClass): eg. AnthropicCacheControlHook and BedrockKnowledgeBaseHook both don't require a `prompt_id` to be passed in, they are triggered by dynamic params """ - for param in non_default_params: - if param in DynamicPromptManagementParamLiteral.list_all_params(): + for param in DynamicPromptManagementParamLiteral.list_all_params(): + if non_default_params.get(param): return True ############################################################################# @@ -966,6 +966,23 @@ class Logging(LiteLLMLoggingBaseClass): return None + @staticmethod + def _prompt_manager_runs_without_prompt_id( + logger: CustomLogger, + prompt_spec: PromptSpec | None, + dynamic_callback_params: StandardCallbackDynamicParams | None, + ) -> bool: + if not isinstance(logger, CustomPromptManagement): + return False + try: + return logger.should_run_prompt_management( + prompt_id=None, + prompt_spec=prompt_spec, + dynamic_callback_params=dynamic_callback_params or StandardCallbackDynamicParams(), + ) + except Exception: + return False + def get_custom_logger_for_prompt_management( self, model: str, @@ -1016,8 +1033,13 @@ class Logging(LiteLLMLoggingBaseClass): callback_type=CustomPromptManagement ) - if prompt_management_loggers: - logger: Final = prompt_management_loggers[0] + for logger in prompt_management_loggers: + if prompt_id is None and not self._prompt_manager_runs_without_prompt_id( + logger=logger, + prompt_spec=prompt_spec, + dynamic_callback_params=dynamic_callback_params, + ): + continue self.model_call_details["prompt_integration"] = logger.__class__.__name__ return logger diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py index c2d73ea467d..05aa65034fb 100644 --- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py +++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py @@ -5105,3 +5105,81 @@ def test_set_cost_breakdown_stores_vertex_location(): cost_for_built_in_tools_cost_usd_dollar=0.0, ) assert no_location.cost_breakdown.get("vertex_location") is None + + +def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_path, monkeypatch): + """ + Regression for UI-injected `vector_store_ids: []` and always-on non-empty `vector_store_ids` + with a registered prompt manager (e.g. dotprompt): requests without a prompt_id 500'd with + "prompt_id is required for Prompt Management Base class" instead of completing normally. + """ + from litellm.integrations.dotprompt.dotprompt_manager import DotpromptManager + from litellm.integrations.vector_store_integrations.vector_store_pre_call_hook import ( + VectorStorePreCallHook, + ) + from litellm.types.vector_stores import LiteLLM_ManagedVectorStore + from litellm.vector_stores.vector_store_registry import VectorStoreRegistry + + (tmp_path / "stem.prompt").write_text("---\nmodel: gemini-2.5-flash\n---\nyou are a stem tutor\n") + dotprompt_manager = DotpromptManager(prompt_directory=str(tmp_path)) + litellm.logging_callback_manager.add_litellm_callback(dotprompt_manager) + monkeypatch.setattr( + litellm, + "vector_store_registry", + VectorStoreRegistry( + vector_stores=[LiteLLM_ManagedVectorStore(vector_store_id="vs_123", custom_llm_provider="openai")] + ), + ) + + messages = [{"role": "user", "content": "hi"}] + try: + assert not logging_obj.should_run_prompt_management_hooks( + prompt_id=None, non_default_params={"vector_store_ids": []} + ) + + assert logging_obj.get_chat_completion_prompt( + model="gemini-2.5-flash", + messages=messages, + non_default_params={"vector_store_ids": []}, + prompt_variables=None, + prompt_id=None, + ) == ("gemini-2.5-flash", messages, {"vector_store_ids": []}) + + assert dotprompt_manager.get_chat_completion_prompt( + model="gemini-2.5-flash", + messages=messages, + non_default_params={}, + prompt_id=None, + prompt_variables=None, + dynamic_callback_params={}, + ) == ("gemini-2.5-flash", messages, {}) + + assert logging_obj.should_run_prompt_management_hooks( + prompt_id=None, non_default_params={"vector_store_ids": ["vs_123"]} + ) + assert isinstance( + logging_obj.get_custom_logger_for_prompt_management( + model="gemini-2.5-flash", + non_default_params={"vector_store_ids": ["vs_123"]}, + prompt_id=None, + dynamic_callback_params={}, + ), + VectorStorePreCallHook, + ) + + assert isinstance( + logging_obj.get_custom_logger_for_prompt_management( + model="gemini-2.5-flash", + non_default_params={}, + prompt_id="stem", + dynamic_callback_params={}, + ), + DotpromptManager, + ) + finally: + litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, dotprompt_manager) + litellm.logging_callback_manager.remove_callback_from_list_by_object( + litellm._async_success_callback, dotprompt_manager + ) + for hook in [cb for cb in litellm.callbacks if isinstance(cb, VectorStorePreCallHook)]: + litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, hook) From b780b1e23c160ee8b8eb3effe471c280547fa10f Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 21:38:15 -0700 Subject: [PATCH 025/220] fix(arize): decline prompt management runs without a prompt_id Arize Phoenix claimed it could run without a prompt_id while its compiler requires one, so the no-prompt_id fallback could select it and fail instead of reaching the vector-store hook. It now declines like the other managers. --- .../arize/arize_phoenix_prompt_manager.py | 6 +++--- .../litellm_core_utils/test_litellm_logging.py | 16 ++++++++++++---- 2 files changed, 15 insertions(+), 7 deletions(-) diff --git a/litellm/integrations/arize/arize_phoenix_prompt_manager.py b/litellm/integrations/arize/arize_phoenix_prompt_manager.py index fa178a02752..71f4902bbe5 100644 --- a/litellm/integrations/arize/arize_phoenix_prompt_manager.py +++ b/litellm/integrations/arize/arize_phoenix_prompt_manager.py @@ -359,10 +359,10 @@ class ArizePhoenixPromptManager(CustomPromptManagement): """ Determine if prompt management should run based on the prompt_id. - For Arize Phoenix, we always return True and handle the prompt loading - in the _compile_prompt_helper method. + Arize Phoenix needs a prompt_id to compile, so it declines requests without one; + prompt loading itself happens in the _compile_prompt_helper method. """ - return True + return prompt_id is not None def _compile_prompt_helper( self, diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py index 05aa65034fb..db64340f895 100644 --- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py +++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py @@ -5113,6 +5113,7 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa with a registered prompt manager (e.g. dotprompt): requests without a prompt_id 500'd with "prompt_id is required for Prompt Management Base class" instead of completing normally. """ + from litellm.integrations.arize.arize_phoenix_prompt_manager import ArizePhoenixPromptManager from litellm.integrations.dotprompt.dotprompt_manager import DotpromptManager from litellm.integrations.vector_store_integrations.vector_store_pre_call_hook import ( VectorStorePreCallHook, @@ -5122,7 +5123,9 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa (tmp_path / "stem.prompt").write_text("---\nmodel: gemini-2.5-flash\n---\nyou are a stem tutor\n") dotprompt_manager = DotpromptManager(prompt_directory=str(tmp_path)) + arize_manager = ArizePhoenixPromptManager(api_key="fake-key", api_base="http://127.0.0.1:9") litellm.logging_callback_manager.add_litellm_callback(dotprompt_manager) + litellm.logging_callback_manager.add_litellm_callback(arize_manager) monkeypatch.setattr( litellm, "vector_store_registry", @@ -5154,6 +5157,10 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa dynamic_callback_params={}, ) == ("gemini-2.5-flash", messages, {}) + assert not arize_manager.should_run_prompt_management( + prompt_id=None, prompt_spec=None, dynamic_callback_params={} + ) + assert logging_obj.should_run_prompt_management_hooks( prompt_id=None, non_default_params={"vector_store_ids": ["vs_123"]} ) @@ -5177,9 +5184,10 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa DotpromptManager, ) finally: - litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, dotprompt_manager) - litellm.logging_callback_manager.remove_callback_from_list_by_object( - litellm._async_success_callback, dotprompt_manager - ) + for manager in (dotprompt_manager, arize_manager): + litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, manager) + litellm.logging_callback_manager.remove_callback_from_list_by_object( + litellm._async_success_callback, manager + ) for hook in [cb for cb in litellm.callbacks if isinstance(cb, VectorStorePreCallHook)]: litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, hook) From 5c7604d7facd84df1767a8c2b65a3fb0cfb3873d Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 19 Aug 2026 23:05:30 -0700 Subject: [PATCH 026/220] test(prompt_management): cover _prompt_manager_runs_without_prompt_id directly --- .../test_litellm_logging.py | 28 +++++++++++++------ 1 file changed, 20 insertions(+), 8 deletions(-) diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py index db64340f895..82de634b488 100644 --- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py +++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py @@ -5115,6 +5115,7 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa """ from litellm.integrations.arize.arize_phoenix_prompt_manager import ArizePhoenixPromptManager from litellm.integrations.dotprompt.dotprompt_manager import DotpromptManager + from litellm.integrations.vector_store_integrations.base_vector_store import BaseVectorStore from litellm.integrations.vector_store_integrations.vector_store_pre_call_hook import ( VectorStorePreCallHook, ) @@ -5164,14 +5165,25 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa assert logging_obj.should_run_prompt_management_hooks( prompt_id=None, non_default_params={"vector_store_ids": ["vs_123"]} ) - assert isinstance( - logging_obj.get_custom_logger_for_prompt_management( - model="gemini-2.5-flash", - non_default_params={"vector_store_ids": ["vs_123"]}, - prompt_id=None, - dynamic_callback_params={}, - ), - VectorStorePreCallHook, + selected_logger = logging_obj.get_custom_logger_for_prompt_management( + model="gemini-2.5-flash", + non_default_params={"vector_store_ids": ["vs_123"]}, + prompt_id=None, + dynamic_callback_params={}, + ) + assert isinstance(selected_logger, VectorStorePreCallHook) + + assert logging_obj._prompt_manager_runs_without_prompt_id( + logger=BaseVectorStore(), prompt_spec=None, dynamic_callback_params=None + ) + assert not logging_obj._prompt_manager_runs_without_prompt_id( + logger=selected_logger, prompt_spec=None, dynamic_callback_params=None + ) + assert not logging_obj._prompt_manager_runs_without_prompt_id( + logger=dotprompt_manager, prompt_spec=None, dynamic_callback_params=None + ) + assert not logging_obj._prompt_manager_runs_without_prompt_id( + logger=arize_manager, prompt_spec=None, dynamic_callback_params=None ) assert isinstance( From 26f62377451369bd1b96d2fbab11205c1018575c Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 01:39:43 -0700 Subject: [PATCH 027/220] build: skip deleted files in the changed-file ruff format check `make lint` hands every path in the diff against the base branch to `ruff format --check`, including the ones the branch deleted, so any branch that moves or removes a file under `litellm/` fails the gate with "No such file or directory" instead of a formatting complaint. test-linting.yml already filters those out with `--diff-filter=ACMR`, so the Makefile was the half that drifted. Match it. --- Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Makefile b/Makefile index 5e5f7c80027..ab6eba880bf 100644 --- a/Makefile +++ b/Makefile @@ -146,7 +146,7 @@ lint-install: # only the litellm Python files changed vs the base are checked, so a pre-existing # format issue elsewhere doesn't block an unrelated commit. lint-format-check-changed: $(LINT_DEP_INSTALL) $(LINT_DEP_BASE) - @files=$$(git diff --name-only origin/litellm_internal_staging...HEAD -- 'litellm/**/*.py' | grep -v '^litellm/enterprise/' || true); \ + @files=$$(git diff --name-only --diff-filter=ACMR origin/litellm_internal_staging...HEAD -- 'litellm/**/*.py' | grep -v '^litellm/enterprise/' || true); \ if [ -z "$$files" ]; then \ echo "No changed litellm Python files to format-check."; \ else \ From ba637553f8cded70ddab429c429c9c030f29dcc5 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 01:39:43 -0700 Subject: [PATCH 028/220] fix(cli): keep lite login and logout honest when the keychain will not answer Three ways the credential commands could mislead or hang. `lite logout` on a machine that never logged in warned that a credential may be stranded in a keychain it could not check, and told the user to install keyring to go clear it. There was nothing there. A missing token file is now read as the evidence it is, because logout keeps a secret-free file behind whenever the keychain is left unconfirmed, so a later run can tell a machine with a credential it cannot reach apart from one that never had a login. That holds on the LITELLM_CLI_DISABLE_KEYRING path too. `KeyringDiscardsWrites` was handled on the read and erase paths, which cannot produce it: the null backend returns None from `get_password` rather than raising, so only a write ever detects it. It now lives on `SecretWrite` alone and the unreachable arms are gone. `keyring.set_password` blocks forever under a HOME with no usable login keychain, which is what containers, CI images, `sudo -H`, and service accounts run with, and reads answer normally there so nothing cheaper tells them apart. `lite login` never touched a keychain before this, so a sign-in that simply never returns would be a new way for it to fail. Writes are pre-flighted with a throwaway value on a bounded wait, and a keychain that stays silent falls back to the token file. The real credential is never the thing handed to a call that might land long after we stopped waiting. Saving also stages the token file before the keychain is given anything, since the file is the half a read-only or full directory refuses. A save that cannot land now leaves both stores as it found them, which matters most when the login it failed to replace still works. --- litellm/litellm_core_utils/cli_keyring.py | 53 +++++++- litellm/litellm_core_utils/cli_token_utils.py | 88 +++++++----- litellm/proxy/client/cli/commands/auth.py | 7 +- pyproject.toml | 2 +- .../test_cli_token_utils.py | 125 +++++++++++++++++- .../proxy/client/cli/test_auth_commands.py | 2 +- 6 files changed, 227 insertions(+), 50 deletions(-) diff --git a/litellm/litellm_core_utils/cli_keyring.py b/litellm/litellm_core_utils/cli_keyring.py index 15282fc522c..8da3e5226d4 100644 --- a/litellm/litellm_core_utils/cli_keyring.py +++ b/litellm/litellm_core_utils/cli_keyring.py @@ -11,18 +11,24 @@ token file and tell the user what to do about it. A write is only reported as stored once it has been read back, because keyring's null backend, which `keyring --disable` and headless CI images both select, -accepts every write and keeps nothing. +accepts every write and keeps nothing. Writes are also pre-flighted with a +throwaway value, because a keychain can answer neither way and block forever. """ import os +import threading +from contextlib import suppress from dataclasses import dataclass from typing import Final, Protocol, TypeAlias KEYRING_SERVICE: Final = "litellm-cli" KEYRING_ACCOUNT: Final = "credential" +KEYRING_PREFLIGHT_ACCOUNT: Final = "credential-preflight" DISABLE_KEYRING_ENV_VAR: Final = "LITELLM_CLI_DISABLE_KEYRING" _DISABLED_VALUES: Final = frozenset(("1", "true", "yes", "on")) +_PREFLIGHT_VALUE: Final = "preflight" +_PREFLIGHT_TIMEOUT_SECONDS: Final = 5.0 @dataclass(frozen=True, slots=True) @@ -70,9 +76,9 @@ class KeyringDiscardsWrites: pass -KeyringUnusable: TypeAlias = KeyringNotInstalled | KeyringDisabled | KeyringUnreachable | KeyringDiscardsWrites +KeyringUnusable: TypeAlias = KeyringNotInstalled | KeyringDisabled | KeyringUnreachable SecretRead: TypeAlias = SecretFound | SecretMissing | KeyringUnusable -SecretWrite: TypeAlias = SecretStored | KeyringUnusable +SecretWrite: TypeAlias = SecretStored | KeyringUnusable | KeyringDiscardsWrites SecretErase: TypeAlias = SecretErased | SecretStranded | KeyringUnusable @@ -113,10 +119,43 @@ def _keyring_api() -> KeyringApi | KeyringNotInstalled | KeyringDisabled: return KeyringNotInstalled() if api is None else api +def _answers_a_write(api: KeyringApi, timeout_seconds: float) -> bool: + """Whether the keychain answers a write at all, asked with a value worth nothing. + + macOS derives the login keychain from `$HOME`, and `set_password` against a HOME with no usable + one blocks forever with no timeout of its own. Containers, CI images, `sudo -H`, and service + accounts all run there, and reads answer normally, so nothing cheaper tells them apart. Asking + with a throwaway value keeps a keychain that never answers from taking `lite login` down with + it, and keeps the real credential out of a store that might accept it long after we gave up. + A keychain that refuses the probe outright still answered it, so only silence counts against it. + """ + answered: Final = threading.Event() + + def ask() -> None: + with suppress(Exception): + api.set_password(KEYRING_SERVICE, KEYRING_PREFLIGHT_ACCOUNT, _PREFLIGHT_VALUE) + answered.set() + + threading.Thread(target=ask, daemon=True, name="litellm-cli-keyring-preflight").start() + return answered.wait(timeout_seconds) + + +def _forget_the_preflight(api: KeyringApi) -> None: + """Take the throwaway probe back out. + + A backend that kept nothing has nothing to remove, and the probe is worth nothing either way, + so a keychain that refuses to give it up costs the caller nothing. + """ + with suppress(Exception): + api.delete_password(KEYRING_SERVICE, KEYRING_PREFLIGHT_ACCOUNT) + + @dataclass(frozen=True, slots=True) class KeyringVault: """The OS keychain, reached through the optional `keyring` package.""" + preflight_timeout_seconds: float = _PREFLIGHT_TIMEOUT_SECONDS + def read(self) -> SecretRead: api: Final = _keyring_api() if isinstance(api, (KeyringNotInstalled, KeyringDisabled)): @@ -134,10 +173,16 @@ class KeyringVault: and `PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring` select, raises nothing to distinguish itself. Reading the value back is the only way to tell it apart from a keychain that really stored the credential, and the caller is about to drop its own copy on our word. + + The keychain is pre-flighted first, because one that blocks rather than answering would + otherwise hang `lite login` outright. """ api: Final = _keyring_api() if isinstance(api, (KeyringNotInstalled, KeyringDisabled)): return api + if not _answers_a_write(api, self.preflight_timeout_seconds): + return KeyringUnreachable() + _forget_the_preflight(api) try: api.set_password(KEYRING_SERVICE, KEYRING_ACCOUNT, blob) except Exception: # noqa: BLE001 # a keychain that refuses the write falls back to the token file @@ -153,7 +198,7 @@ class KeyringVault: the caller knows whether this machine ever put a secret in a keychain. """ match self.read(): - case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable() | KeyringDiscardsWrites() as unusable: + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable() as unusable: return unusable case SecretMissing(): return SecretErased() diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index af1b918fc2a..825967e6866 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -23,7 +23,6 @@ from pydantic import BaseModel, ConfigDict, ValidationError from litellm.litellm_core_utils.cli_keyring import ( SYSTEM_KEYRING, KeyringDisabled, - KeyringDiscardsWrites, KeyringNotInstalled, KeyringUnreachable, SecretErase, @@ -53,6 +52,8 @@ class CredentialNotSaved: SecretSave: TypeAlias = SecretWrite | CredentialNotSaved +_UNREPLACEABLE_FILE: Final = "the staged file could not replace the one already there" + class CliTokenRecord(BaseModel): """A stored CLI credential. @@ -106,57 +107,71 @@ def load_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> CliTokenRecord | N def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRING) -> SecretSave: """Store a freshly minted credential. Reports where its secret material ended up, and why. - The token file is what makes a keychain-backed credential findable again, so a file that will - not be written takes the keychain copy down with it rather than leaving a live credential - stored under a machine that has no record of it. + The token file is what makes a keychain-backed credential findable again, and it is also the + half that a read-only or full directory refuses, so it is staged before the keychain is handed + anything. A save that cannot land then leaves both stores exactly as it found them, which + matters most when the login it failed to replace is still perfectly good. """ + staged: Final = _stage_token_file(_without_secret(record)) + if isinstance(staged, CredentialNotSaved): + return staged outcome: Final = ( SecretStored() if record.key is None else vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)) ) + if isinstance(outcome, SecretStored): + return outcome if _commit_token_file(staged) else CredentialNotSaved(_UNREPLACEABLE_FILE) + discard_staged_json(staged) + return _keep_the_secret_in_the_file(record, outcome) + + +def _keep_the_secret_in_the_file(record: CliTokenRecord, outcome: SecretWrite) -> SecretSave: + """Fall back to the owner-only file, which is all that is left when no keychain took the secret""" try: - _write_token_file(_without_secret(record) if isinstance(outcome, SecretStored) else record) + _write_token_file(record) except OSError as error: - if record.key is not None and isinstance(outcome, SecretStored): - vault.erase() return CredentialNotSaved(str(error)) return outcome def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretErase: - """Remove the credential from both stores. Reports whether the keychain is now free of it""" + """Remove the credential from both stores. Reports whether the keychain is now free of it. + + A file that holds no secret of its own is kept when the keychain will not confirm the entry is + gone, because it is the only remaining record that something is still in there to remove. That + is what lets a later run tell a machine with a credential it cannot reach apart from one that + never had a login at all. Anything still holding a secret is removed either way. + """ outcome: Final = vault.erase() - settled: Final = _nothing_left_behind(outcome) - Path(get_cli_token_file_path()).unlink(missing_ok=True) + record: Final = _read_token_file() + settled: Final = _nothing_left_behind(outcome, record) + if settled or record is None or record.key is not None: + Path(get_cli_token_file_path()).unlink(missing_ok=True) return SecretErased() if settled else outcome -def _nothing_left_behind(outcome: SecretErase) -> bool: +def _nothing_left_behind(outcome: SecretErase, record: CliTokenRecord | None) -> bool: """Whether the keychain can be trusted to hold no credential of ours once the file is gone. - A keychain that exists but is out of reach right now is never trusted, whatever the token file - looks like: the login that stored a secret there and the logout that cannot remove it are - separate runs, free to differ in whether the keychain was usable at the time. + A machine with no token file has no stored login to end, and `clear_cli_token` keeps one behind + whenever the keychain is left unconfirmed, so a missing file is real evidence rather than the + absence of it. Past that, a keychain that exists but is out of reach right now is + never trusted, whatever the file looks like: the login that stored a secret there and the + logout that cannot remove it are separate runs, free to differ in whether the keychain was + usable at the time. The exception is a missing `keyring` package, which had to be missing when + the credential was stored too, so a file still holding its own secret proves no keychain was + ever involved. `SecretStranded` is the keychain answering for itself and outranks the file. """ match outcome: case SecretErased(): return True - case SecretStranded() | KeyringDisabled() | KeyringUnreachable() | KeyringDiscardsWrites(): + case SecretStranded(): return False + case KeyringDisabled() | KeyringUnreachable(): + return record is None case KeyringNotInstalled(): - return _file_holds_its_own_secret() - - -def _file_holds_its_own_secret() -> bool: - """Whether the stored login keeps its secret in the token file, ruling out a keychain entry. - - Sound only against a missing `keyring` package, the one way to lose the keychain that had to - hold at storage time too, since nothing here can reach a keychain without it. A file whose - secret half is absent went to a keychain by definition, and so rules nothing out. - """ - record: Final = _read_token_file() - return record is not None and record.key is not None + return record is None or record.key is not None def get_litellm_gateway_api_key( @@ -227,7 +242,7 @@ def _resolve_secret(record: CliTokenRecord, vault: SecretVault) -> CliTokenRecor return _apply_vault_secret(record, blob, vault) case SecretMissing(): return _migrate_file_secret(record, vault) - case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable() | KeyringDiscardsWrites(): + case KeyringNotInstalled() | KeyringDisabled() | KeyringUnreachable(): return record @@ -265,7 +280,7 @@ def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliToken if not isinstance(vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)), SecretStored): discard_staged_json(staged) return record - if not _commit_scrubbed_file(staged): + if not _commit_token_file(staged): vault.erase() return record @@ -275,19 +290,24 @@ def _scrub_file_secret(record: CliTokenRecord) -> bool: if record.key is None and not record.jwt_token: return True staged: Final = _stage_scrubbed_file(record) - return staged is not None and _commit_scrubbed_file(staged) + return staged is not None and _commit_token_file(staged) def _stage_scrubbed_file(record: CliTokenRecord) -> str | None: + staged: Final = _stage_token_file(_without_secret(record)) + return None if isinstance(staged, CredentialNotSaved) else staged + + +def _stage_token_file(record: CliTokenRecord) -> str | CredentialNotSaved: path: Final = Path(get_cli_token_file_path()) try: ensure_private_dir(path.parent) - return stage_private_json(str(path), _without_secret(record).model_dump(exclude_none=True)) - except OSError: - return None + return stage_private_json(str(path), record.model_dump(exclude_none=True)) + except OSError as error: + return CredentialNotSaved(str(error)) -def _commit_scrubbed_file(staged: str) -> bool: +def _commit_token_file(staged: str) -> bool: try: commit_staged_json(staged, get_cli_token_file_path()) except OSError: diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index eba9994f7ec..d89641d2366 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -123,7 +123,8 @@ def storage_notice(outcome: SecretSave) -> str: case CredentialNotSaved(detail=detail): return ( f"Signed in, but the credential could not be saved to {path}: {detail}. " - "Nothing was kept, so run 'lite login' again once that path is writable." + "Any login you already had is untouched. Run 'lite login' again once that path is " + "writable, or 'lite logout' to clear whatever is stored now." ) @@ -140,7 +141,7 @@ def keychain_unreadable_notice(vault: SecretVault) -> str: f"Your credential is in your OS keychain, which {DISABLE_KEYRING_ENV_VAR} is blocking. " "Unset it, or run 'lite login' to start over." ) - case KeyringUnreachable() | KeyringDiscardsWrites(): + case KeyringUnreachable(): return ( "Your credential is in your OS keychain, which could not be read. Unlock it, or run " "'lite login' to start over." @@ -800,7 +801,7 @@ def logout(ctx: click.Context): case KeyringDisabled(): click.echo(UNCHECKED_KEYCHAIN_MESSAGE) click.echo(f"Unset {DISABLE_KEYRING_ENV_VAR} and run 'lite logout' again to clear it.") - case KeyringUnreachable() | KeyringDiscardsWrites(): + case KeyringUnreachable(): click.echo(UNCHECKED_KEYCHAIN_MESSAGE) click.echo("Unlock your keychain and run 'lite logout' again to clear it.") diff --git a/pyproject.toml b/pyproject.toml index 32921e14d31..64adb9cd595 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -79,7 +79,7 @@ proxy = [ ] # Thin client install for the `lite` CLI on developer laptops. The CLI's heavy # imports (fastapi, cryptography, ...) are all guarded, so it runs on the base -# SDK plus just these four; none of the server runtime in `proxy` is pulled in. +# SDK plus just these five; none of the server runtime in `proxy` is pulled in. cli = [ "rich>=13.9.4,<14.0", "pyyaml>=6.0.3,<7.0", diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index e0f5f99dc1b..69ce47b25b3 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -2,6 +2,7 @@ import json import os import stat import sys +import threading import time import pytest @@ -10,6 +11,7 @@ from litellm.constants import CLI_JWT_EXPIRATION_HOURS from litellm.litellm_core_utils.cli_keyring import ( DISABLE_KEYRING_ENV_VAR, KEYRING_ACCOUNT, + KEYRING_PREFLIGHT_ACCOUNT, KEYRING_SERVICE, KeyringVault, SecretFound, @@ -329,12 +331,12 @@ class TestSaveCliToken: assert isinstance(outcome, CredentialNotSaved) assert "read-only file system" in outcome.detail - def test_a_credential_the_file_will_not_record_is_taken_back_out_of_the_keychain( + def test_a_file_that_will_not_be_written_stops_the_save_before_the_keychain_is_touched( self, isolated_home, secret_vault_factory, monkeypatch ): - """The token file is what makes a keychain entry findable again. Leaving the secret in the - keychain with nothing pointing at it strands a live credential under a machine that has no - idea it is there, and no `lite logout` would ever go looking for it.""" + """The token file is what makes a keychain entry findable again, so it is staged first. + Handing the keychain a secret and only then finding out that nothing will point at it + would strand a live credential under a machine with no idea it is there.""" vault = secret_vault_factory() def _explode(*args, **kwargs): @@ -346,6 +348,26 @@ class TestSaveCliToken: assert vault.blob is None + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") + def test_a_login_that_cannot_be_saved_leaves_the_working_one_alone( + self, isolated_home, secret_vault_factory + ): + """Signing in again on a machine whose ~/.litellm has gone read-only must not cost the user + the credential they already had. Overwriting the keychain and then failing to record it, or + undoing that write afterwards, would take a login that still works out from under them.""" + _write_legacy_file(isolated_home, key=None) + vault = secret_vault_factory(blob=_blob(key="sk-in-use")) + path = _token_file(isolated_home) + path.parent.chmod(0o500) + try: + outcome = save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=vault) + finally: + path.parent.chmod(0o700) + + assert isinstance(outcome, CredentialNotSaved) + assert json.loads(vault.blob)["key"] == "sk-in-use" + assert load_cli_token(vault=vault).key == "sk-in-use" + def test_a_failed_write_leaves_the_previous_credential_intact(self, isolated_home, secret_vault_factory, monkeypatch): path = _write_legacy_file(isolated_home) before = path.read_text() @@ -460,8 +482,44 @@ class TestClearCliToken: vault = secret_vault_factory(available=False, failure=KeyringNotInstalled()) assert clear_cli_token(vault=vault) == KeyringNotInstalled() + assert json.loads(_token_file(isolated_home).read_text()).get("key") is None + + def test_a_logout_that_cannot_clear_the_keychain_keeps_the_record_that_it_has_to( + self, isolated_home, secret_vault_factory + ): + """The file left behind holds no secret. It is what a later run reads to tell a machine with + a credential it cannot reach apart from one that never had a login, which is the difference + between warning the user and inventing a credential for them to worry about.""" + _write_metadata_only_file(isolated_home) + vault = secret_vault_factory(available=False, failure=KeyringUnreachable()) + + clear_cli_token(vault=vault) + + assert json.loads(_token_file(isolated_home).read_text()).get("key") is None + + def test_a_logout_that_cannot_clear_the_keychain_still_takes_the_file_secret_away( + self, isolated_home, secret_vault_factory + ): + """Keeping a record of the unreachable keychain must never mean keeping the cleartext copy + the user just asked to be rid of.""" + _write_legacy_file(isolated_home) + vault = secret_vault_factory(available=False, failure=KeyringUnreachable()) + + clear_cli_token(vault=vault) + assert not _token_file(isolated_home).exists() + @pytest.mark.parametrize("failure", [KeyringNotInstalled(), KeyringDisabled(), KeyringUnreachable()]) + def test_logging_out_of_a_machine_that_never_logged_in_invents_nothing_to_warn_about( + self, isolated_home, secret_vault_factory, failure + ): + """`lite logout` with no token file has nothing to end. Warning that a credential may be + stranded in a keychain it cannot check sends the user after something that was never there, + and `pip install keyring` will not make it appear.""" + vault = secret_vault_factory(available=False, failure=failure) + + assert clear_cli_token(vault=vault) == SecretErased() + def test_a_file_backed_login_logs_out_quietly_without_keyring(self, isolated_home, secret_vault_factory): """The complement, and the one inference the file does support: nothing here can reach a keychain without the package, so an install that lacks it and a file that still holds its @@ -510,7 +568,7 @@ class _FakeKeyringModule: self.calls.append(("set", service_name, username)) if self.set_error is not None: raise self.set_error - if self.discard: + if self.discard or username != KEYRING_ACCOUNT: return self.stored = password @@ -518,7 +576,22 @@ class _FakeKeyringModule: self.calls.append(("delete", service_name, username)) if self.delete_error is not None: raise self.delete_error - self.stored = None + if username == KEYRING_ACCOUNT: + self.stored = None + + +class _NeverAnsweringKeyringModule(_FakeKeyringModule): + """A keychain whose writes block instead of returning, the way macOS does under a HOME that + has no usable login keychain.""" + + def __init__(self): + super().__init__() + self.blocked = threading.Event() + + def set_password(self, service_name, username, password): + self.calls.append(("set", service_name, username)) + self.blocked.set() + threading.Event().wait() @pytest.fixture @@ -540,7 +613,8 @@ class TestKeyringVault: assert vault.read() == SecretFound("blob-1") assert vault.erase() == SecretErased() assert vault.read() == SecretMissing() - assert {call[1:] for call in fake.calls} == {(KEYRING_SERVICE, KEYRING_ACCOUNT)} + assert {call[1] for call in fake.calls} == {KEYRING_SERVICE} + assert {call[2] for call in fake.calls} == {KEYRING_ACCOUNT, KEYRING_PREFLIGHT_ACCOUNT} def test_the_kill_switch_reports_no_keychain(self, monkeypatch): """`LITELLM_CLI_DISABLE_KEYRING` has to work without importing keyring, because keyring @@ -590,6 +664,43 @@ class TestKeyringVault: assert KeyringVault().write("blob-1") == KeyringDiscardsWrites() assert fake.stored is None + def test_a_keychain_that_never_answers_does_not_hang_the_login(self, install_fake_keyring): + """macOS derives the login keychain from `$HOME`, and `set_password` under a HOME with no + usable one blocks forever with no timeout of its own. Containers, CI images, `sudo -H`, and + service accounts all run there, and `lite login` never touched a keychain before this, so a + sign-in that simply never returns would be a new way for it to fail.""" + fake = install_fake_keyring(_NeverAnsweringKeyringModule()) + vault = KeyringVault(preflight_timeout_seconds=0.2) + + started = time.monotonic() + outcome = vault.write("blob-1") + + assert outcome == KeyringUnreachable() + assert time.monotonic() - started < 5 + assert fake.blocked.is_set() + + def test_a_keychain_that_never_answers_is_never_handed_the_credential(self, install_fake_keyring): + """Giving up on the write is only safe if the secret was never the thing being written. A + blocked call can still land later, and a keychain copy nobody waited for would sit beside + the file copy the user was told about.""" + fake = install_fake_keyring(_NeverAnsweringKeyringModule()) + + KeyringVault(preflight_timeout_seconds=0.2).write("blob-1") + + assert [call[2] for call in fake.calls] == [KEYRING_PREFLIGHT_ACCOUNT] + + def test_a_login_survives_a_keychain_that_never_answers(self, isolated_home, install_fake_keyring): + """The end of the same story: the credential still has to be usable afterwards.""" + install_fake_keyring(_NeverAnsweringKeyringModule()) + + outcome = save_cli_token( + CliTokenRecord(base_url=SERVER, key="sk-only-copy"), + vault=KeyringVault(preflight_timeout_seconds=0.2), + ) + + assert outcome == KeyringUnreachable() + assert json.loads(_token_file(isolated_home).read_text())["key"] == "sk-only-copy" + def test_the_real_null_backend_is_rejected(self, monkeypatch): """Pinned against the actual library rather than the double above, because the whole risk is that upstream's no-op write looks exactly like a successful one.""" diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 8e1551c0720..e491dbd5aca 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -1074,7 +1074,7 @@ class TestKeychainBackedCommands: assert result.exit_code == 0 assert "could not be removed" in result.output - assert not (isolated_home / ".litellm" / "token.json").exists() + assert json.loads((isolated_home / ".litellm" / "token.json").read_text()).get("key") is None def test_print_token_explains_a_locked_keychain_instead_of_printing_nothing( self, isolated_home, secret_vault_factory From c7da91d47f95aa9aa992264c1979748b916de796 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 01:47:50 -0700 Subject: [PATCH 029/220] fix(cli): say so when the keychain took a credential the file cannot name Staging the token file can succeed and the replacement still fail afterwards, and that is the one save path where the keychain has already taken the new secret. It was reported as a save that kept nothing, which sends the user looking for a credential that is sitting in their keychain, and it claimed the previous login was untouched when the one keychain slot had just been written over. Give that path its own outcome and its own notice. The new secret stays where it is: the entry it replaced went the moment it landed, so no rollback brings that back, and removing the new one too would turn a login this machine may still be able to use into no login at all. The remaining `CredentialNotSaved` paths all leave both stores untouched, so the reassurance they carry is now true wherever it is printed. --- litellm/litellm_core_utils/cli_token_utils.py | 23 ++++++++++--- litellm/proxy/client/cli/commands/auth.py | 9 +++++- .../test_cli_token_utils.py | 32 +++++++++++++++++++ 3 files changed, 59 insertions(+), 5 deletions(-) diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 825967e6866..15b1390f337 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -45,14 +45,25 @@ from litellm.litellm_core_utils.private_json import ( @dataclass(frozen=True, slots=True) class CredentialNotSaved: - """The credential was minted but no store would keep it, so this machine has none.""" + """The credential was minted but no store would keep it, so this machine has none. + + Nothing was touched on the way to this, so a login that already worked still does. + """ detail: str -SecretSave: TypeAlias = SecretWrite | CredentialNotSaved +@dataclass(frozen=True, slots=True) +class CredentialNotRecorded: + """The keychain took the credential, but the file that names it could not be replaced. -_UNREPLACEABLE_FILE: Final = "the staged file could not replace the one already there" + The keychain holds one entry, so the secret that was there is already gone and no rollback + brings it back. Removing the new one as well would only turn a login this machine may still + be able to use into no login at all, so it stays, and the user is told what is where. + """ + + +SecretSave: TypeAlias = SecretWrite | CredentialNotSaved | CredentialNotRecorded class CliTokenRecord(BaseModel): @@ -111,6 +122,10 @@ def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRIN half that a read-only or full directory refuses, so it is staged before the keychain is handed anything. A save that cannot land then leaves both stores exactly as it found them, which matters most when the login it failed to replace is still perfectly good. + + Staging can still succeed and the replacement fail afterwards. That is the one case where the + keychain has already taken the new secret, and it reports itself as such rather than claiming + the previous login survived. """ staged: Final = _stage_token_file(_without_secret(record)) if isinstance(staged, CredentialNotSaved): @@ -121,7 +136,7 @@ def save_cli_token(record: CliTokenRecord, *, vault: SecretVault = SYSTEM_KEYRIN else vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)) ) if isinstance(outcome, SecretStored): - return outcome if _commit_token_file(staged) else CredentialNotSaved(_UNREPLACEABLE_FILE) + return outcome if _commit_token_file(staged) else CredentialNotRecorded() discard_staged_json(staged) return _keep_the_secret_in_the_file(record, outcome) diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index d89641d2366..eb956cd536c 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -27,6 +27,7 @@ from litellm.litellm_core_utils.cli_keyring import ( ) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, + CredentialNotRecorded, CredentialNotSaved, SecretSave, clear_cli_token, @@ -126,6 +127,12 @@ def storage_notice(outcome: SecretSave) -> str: "Any login you already had is untouched. Run 'lite login' again once that path is " "writable, or 'lite logout' to clear whatever is stored now." ) + case CredentialNotRecorded(): + return ( + f"Signed in, and the credential is in your OS keychain, but {path} could not be " + "replaced, so this machine may still be using your previous login. Run 'lite login' " + "again once that path is writable, or 'lite logout' to clear both." + ) def keychain_unreadable_notice(vault: SecretVault) -> str: @@ -754,7 +761,7 @@ def login(ctx: click.Context, config_claude: bool): click.echo("\nLogin successful!") click.echo(f"JWT Token: {api_key[:20]}...") click.echo(storage_notice(stored)) - if isinstance(stored, CredentialNotSaved): + if isinstance(stored, (CredentialNotSaved, CredentialNotRecorded)): return click.echo("You can now use the CLI without specifying --api-key") diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 69ce47b25b3..0cf1e55d363 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -26,6 +26,7 @@ from litellm.litellm_core_utils.cli_keyring import ( ) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, + CredentialNotRecorded, CredentialNotSaved, clear_cli_token, get_cli_token_file_path, @@ -368,6 +369,37 @@ class TestSaveCliToken: assert json.loads(vault.blob)["key"] == "sk-in-use" assert load_cli_token(vault=vault).key == "sk-in-use" + def test_a_keychain_write_the_file_cannot_be_pointed_at_is_reported_as_that( + self, isolated_home, secret_vault_factory + ): + """Staging the file can succeed and the replacement still fail, and that is the one path + where the keychain already took the new secret. Reporting it as a save that kept nothing + would send the user looking for a credential that is sitting in their keychain.""" + vault = secret_vault_factory() + path = _token_file(isolated_home) + path.parent.mkdir(parents=True, exist_ok=True) + path.mkdir() + + outcome = save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=vault) + + assert isinstance(outcome, CredentialNotRecorded) + assert json.loads(vault.blob)["key"] == "sk-new" + + def test_the_credential_the_file_cannot_name_is_left_in_the_keychain( + self, isolated_home, secret_vault_factory + ): + """The keychain holds one entry, so the secret that was there went the moment this one + landed. Taking the new one back out would turn a login this machine may still be able to + use into no login at all, and it cannot restore the old one either way.""" + vault = secret_vault_factory(blob=_blob(key="sk-in-use")) + path = _token_file(isolated_home) + path.parent.mkdir(parents=True, exist_ok=True) + path.mkdir() + + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-new"), vault=vault) + + assert vault.blob is not None + def test_a_failed_write_leaves_the_previous_credential_intact(self, isolated_home, secret_vault_factory, monkeypatch): path = _write_legacy_file(isolated_home) before = path.read_text() From bcb6a6eaab2f5080fb905014fc60f413df84fad3 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 01:50:01 -0700 Subject: [PATCH 030/220] test(e2e): pin prompt-cache, service-tier, and cost-header billing Seven live e2e tests covering cost-tracking regressions that currently ship unnoticed: cache-write tokens billed at the cache-creation rate (#34046), per-component cost_breakdown on the spend row (#31686), cache reads billed at the cache-read discount on streamed calls (#34812), cache tokens surviving the anthropic-messages to Responses bridge (#34957), priority-tier rates applied to input, output and reasoning (#35923, #35925), the per-component response cost headers summing to the total (#36965), and cost injected into the final usage frame of an /openai passthrough stream (#36503). Every test registers its own deployment with a distinct custom rate per component, so a component billed at the wrong rate cannot pass. The shared helpers in cost_rows.py encode the one thing the two surfaces disagree on: the spend row's input_cost is gross of cache while the response's cost-input header is net of it. --- .../coverage_registry/quota_management.yaml | 7 + tests/e2e/models.py | 25 +- .../spend_tracking/cost_rows.py | 204 ++++++++++++++ .../test_cache_cost_accounting_e2e.py | 263 ++++++++++++++++++ .../spend_tracking/test_cost_headers_e2e.py | 136 +++++++++ .../test_passthrough_stream_cost_e2e.py | 68 +++++ .../test_service_tier_pricing_e2e.py | 116 ++++++++ 7 files changed, 817 insertions(+), 2 deletions(-) create mode 100644 tests/e2e/quota_management/spend_tracking/cost_rows.py create mode 100644 tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py create mode 100644 tests/e2e/quota_management/spend_tracking/test_cost_headers_e2e.py create mode 100644 tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py create mode 100644 tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py diff --git a/tests/e2e/coverage_registry/quota_management.yaml b/tests/e2e/coverage_registry/quota_management.yaml index 2dfa7adddea..5438ed8534a 100644 --- a/tests/e2e/coverage_registry/quota_management.yaml +++ b/tests/e2e/coverage_registry/quota_management.yaml @@ -44,3 +44,10 @@ - {id: quota_management.spend_tracking.failure.writes_failure_row, module: quota_management, tier: P1, behavior: spend_tracking, variant: failure, assertions: [writes_failure_row], exercised_on: [chat_completions], source: "proxy/spend_tracking/spend_log_error_logger.py", rationale: "A failed call writes a failure-status spend row"} - {id: quota_management.spend_tracking.spend_calculate.returns_cost, module: quota_management, tier: P2, behavior: spend_tracking, variant: spend_calculate, assertions: [returns_cost], exercised_on: [spend_calculate], source: "proxy/spend_tracking/spend_management_endpoints.py", rationale: "/spend/calculate prices a hypothetical request at nonzero cost"} - {id: quota_management.spend_tracking.pagination.keeps_total, module: quota_management, tier: P2, behavior: spend_tracking, variant: pagination, assertions: [keeps_total], exercised_on: [chat_completions], source: "proxy/spend_tracking/spend_management_endpoints.py", rationale: "Spend-logs v2 pagination caps page size without losing the total"} +- {id: quota_management.spend_tracking.cache_write.bills_cache_creation_rate, module: quota_management, tier: P1, behavior: spend_tracking, variant: cache_write, assertions: [bills_cache_creation_rate], exercised_on: [chat_completions], source: "litellm_core_utils/llm_cost_calc/utils.py", rationale: "OpenAI cache-write tokens land on the spend row as cache-creation tokens billed at the cache-creation rate, not silently at the input rate (#34046)"} +- {id: quota_management.spend_tracking.cost_breakdown.reports_component_costs, module: quota_management, tier: P1, behavior: spend_tracking, variant: cost_breakdown, assertions: [reports_component_costs], exercised_on: [chat_completions], source: "proxy/spend_tracking/spend_tracking_utils.py", rationale: "The spend row's metadata.cost_breakdown itemizes cache-read, cache-creation, output, and reasoning costs at the deployment's own rates and they sum to the row's spend (#31686)"} +- {id: quota_management.spend_tracking.stream_cache_read.bills_cache_read_rate, module: quota_management, tier: P1, behavior: spend_tracking, variant: stream_cache_read, assertions: [bills_cache_read_rate], exercised_on: [chat_completions], source: "litellm_core_utils/streaming_chunk_builder_utils.py", rationale: "A streamed call's reassembled usage keeps the cached-token detail so cache reads bill at the cache-read discount, not full input price (#34812)"} +- {id: quota_management.spend_tracking.messages_bridge.keeps_cache_tokens, module: quota_management, tier: P1, behavior: spend_tracking, variant: messages_bridge, assertions: [keeps_cache_tokens], exercised_on: [messages], source: "llms/anthropic/experimental_pass_through/responses_adapters/handler.py", rationale: "A /v1/messages request served by a Responses-only OpenAI model keeps its cache-read tokens and their discounted billing across the bridge (#34957)"} +- {id: quota_management.spend_tracking.service_tier.bills_tier_rates, module: quota_management, tier: P1, behavior: spend_tracking, variant: service_tier, assertions: [bills_tier_rates], exercised_on: [chat_completions], source: "cost_calculator.py", rationale: "A priority service_tier call bills input, output, and reasoning at the deployment's *_priority rates and records the tier on the row (#35923, #35925)"} +- {id: quota_management.spend_tracking.cost_headers.additive_components, module: quota_management, tier: P1, behavior: spend_tracking, variant: cost_headers, assertions: [additive_components], exercised_on: [chat_completions], source: "proxy/common_request_processing.py", rationale: "The x-litellm-response-cost-* component headers sum to the total, input covers only fresh tokens, and reasoning stays a subset of output (#36965)"} +- {id: quota_management.spend_tracking.passthrough_stream.injects_usage_cost, module: quota_management, tier: P1, behavior: spend_tracking, variant: passthrough_stream, assertions: [injects_usage_cost], exercised_on: [openai_passthrough], source: "proxy/pass_through_endpoints/streaming_handler.py", rationale: "With include_cost_in_streaming_usage on, the /openai passthrough's final streaming usage frame carries the proxy-computed cost (#36503)"} diff --git a/tests/e2e/models.py b/tests/e2e/models.py index 619f4dcacfe..fe93b13e0a4 100644 --- a/tests/e2e/models.py +++ b/tests/e2e/models.py @@ -216,10 +216,19 @@ class McpChatTool(BaseModel): allowed_tools: list[str] | None = None +class StreamOptions(BaseModel): + """OpenAI `stream_options`: `include_usage` asks for a final usage-only SSE + frame, which is where the proxy's `include_cost_in_streaming_usage` setting + injects `usage.cost`.""" + + include_usage: bool = True + + class ChatBody(BaseModel): model: str messages: list[ChatMessage] stream: bool = False + stream_options: StreamOptions | None = None max_tokens: int | None = None max_completion_tokens: int | None = None temperature: float | None = None @@ -322,6 +331,9 @@ class CompletionTokensDetails(BaseModel): class Usage(BaseModel): + """`cost` exists only on streaming usage frames from a proxy running with + `include_cost_in_streaming_usage: true`; providers never send it.""" + prompt_tokens: int | None = None completion_tokens: int | None = None total_tokens: int | None = None @@ -329,6 +341,7 @@ class Usage(BaseModel): cache_creation_input_tokens: int | None = None prompt_tokens_details: PromptTokensDetails | None = None completion_tokens_details: CompletionTokensDetails | None = None + cost: float | None = None class ChatResponse(BaseModel): @@ -449,9 +462,11 @@ class AnthropicMessagesResponse(BaseModel): for triage.""" model_config = ConfigDict(extra="allow") + id: str | None = None model: str | None = None content: list[AnthropicContentBlock] | None = None choices: list[ChatChoice] | None = None + usage: Usage | None = None class CountTokensResponse(BaseModel): @@ -716,8 +731,10 @@ class FineTuningJobsResponse(BaseModel): class LiteLLMParamsBody(BaseModel): """POST /model/new litellm_params: `model` is the only required field; `api_key` et al may be an `os.environ/FOO` reference the proxy resolves at call time. - `input_cost_per_token`/`output_cost_per_token` register a per-deployment custom - pricing override; left None (and dropped from the body) the deployment keeps the + The `*_cost_per_token` / `*_token_cost` fields register a per-deployment custom + pricing override (the cache and `_priority` rates only apply when both base + rates are set, which is what makes the proxy register the deployment's full + pricing entry); left None (and dropped from the body) the deployment keeps the backend's canonical rate.""" model: str @@ -744,6 +761,10 @@ class LiteLLMParamsBody(BaseModel): aws_external_id: str | None = None input_cost_per_token: float | None = None output_cost_per_token: float | None = None + cache_read_input_token_cost: float | None = None + cache_creation_input_token_cost: float | None = None + input_cost_per_token_priority: float | None = None + output_cost_per_token_priority: float | None = None extra_headers: dict[str, str] | None = None use_in_pass_through: bool | None = None complexity_router_config: dict[str, object] | None = None diff --git a/tests/e2e/quota_management/spend_tracking/cost_rows.py b/tests/e2e/quota_management/spend_tracking/cost_rows.py new file mode 100644 index 00000000000..87af54fe83f --- /dev/null +++ b/tests/e2e/quota_management/spend_tracking/cost_rows.py @@ -0,0 +1,204 @@ +"""Cost-accounting helpers for the spend-tracking suite: the /spend/logs row shape +that carries the per-component cost breakdown, a poll that waits for it, and the +builders the cache-pricing tests share. + +The shared SpendLogRow deliberately stays thin (most tests only read totals), so +the component-cost tests model the metadata they assert on here instead: +`metadata.cost_breakdown` (input/output/cache-read/cache-creation/reasoning costs +plus the service-tier pricing basis) and `metadata.additional_usage_values` (the +cache token counts the biller derived from the provider's usage). + +Determinism strategy: every test registers its own deployment with explicit custom +rates for each component it asserts on (`register_priced_model`), so expected cost +is exactly tokens-on-the-row times configured rate, immune to provider price +changes. The rates are chosen ~100x above canonical and distinct from one another, +so a component billed at the wrong rate can never accidentally match. + +OpenAI prompt caching is implicit and keyed on the exact token prefix, with a +1024-token minimum. `cacheable_prefix` builds a prefix whose first word is the +run's unique marker: unique marker = the whole prefix is novel (a fresh cache +write), same marker + different question = a cache read that still misses the +proxy's own response cache. How long the prefix has to be before the provider +actually reports a read varies by model, so callers pass `words` to suit theirs. + +Two facts about the recorded bill that the assertions here encode, because the +two surfaces disagree on purpose. On the spend row, `input_cost` is gross: it +already contains the cache-read and cache-creation costs, so the row's total is +input + output + tool-usage and the fresh-token cost is input minus the two cache +components. In the response headers, `x-litellm-response-cost-input` is net of +cache, which is what makes the component headers sum to the total. +""" + +import time +from collections.abc import Callable + +from pydantic import BaseModel, RootModel + +from e2e_config import unique_marker +from e2e_http import Success +from lifecycle import ResourceManager +from models import LiteLLMParamsBody, SpendLogsParams +from proxy_client import ProxyClient + + +class CostBreakdownRow(BaseModel): + input_cost: float | None = None + output_cost: float | None = None + cache_read_cost: float | None = None + cache_creation_cost: float | None = None + reasoning_cost: float | None = None + tool_usage_cost: float | None = None + total_cost: float | None = None + service_tier: str | None = None + + +class AdditionalUsageValues(BaseModel): + cache_read_input_tokens: int | None = None + cache_creation_input_tokens: int | None = None + + +class CostRowMetadata(BaseModel): + cost_breakdown: CostBreakdownRow | None = None + additional_usage_values: AdditionalUsageValues | None = None + + +class CostRow(BaseModel): + request_id: str | None = None + spend: float | None = None + prompt_tokens: int | None = None + completion_tokens: int | None = None + metadata: CostRowMetadata | None = None + + @property + def breakdown(self) -> CostBreakdownRow: + assert self.metadata and self.metadata.cost_breakdown, ( + f"spend row {self.request_id} landed without a cost breakdown" + ) + return self.metadata.cost_breakdown + + @property + def cache_read_tokens(self) -> int: + if self.metadata and self.metadata.additional_usage_values: + return self.metadata.additional_usage_values.cache_read_input_tokens or 0 + return 0 + + @property + def cache_creation_tokens(self) -> int: + if self.metadata and self.metadata.additional_usage_values: + return self.metadata.additional_usage_values.cache_creation_input_tokens or 0 + return 0 + + +class CostRows(RootModel[list[CostRow]]): + pass + + +def approx_equal(actual: float, expected: float) -> bool: + """Within 1% or 1e-9 absolute - spend math, not exact float identity.""" + return abs(actual - expected) <= max(1e-9, abs(expected) * 1e-2) + + +def assert_total_is_sum_of_components(row: CostRow) -> None: + """The row's total is input + output + tool usage. The cache components are + already inside the gross input cost, so adding them again would double-bill.""" + breakdown = row.breakdown + components = sum( + cost or 0.0 + for cost in (breakdown.input_cost, breakdown.output_cost, breakdown.tool_usage_cost) + ) + assert breakdown.total_cost is not None and approx_equal(breakdown.total_cost, components), ( + f"total_cost {breakdown.total_cost} != input + output + tool usage ({components}): {breakdown}" + ) + assert row.spend is not None and approx_equal(row.spend, breakdown.total_cost), ( + f"row spend {row.spend} != breakdown total {breakdown.total_cost}" + ) + + +def assert_fresh_tokens_billed_at(row: CostRow, input_rate: float) -> None: + """Strip the cache components out of the gross input cost and what is left must + be the freshly-read tokens at the deployment's input rate.""" + breakdown = row.breakdown + fresh_tokens = (row.prompt_tokens or 0) - row.cache_read_tokens - row.cache_creation_tokens + fresh_cost = ( + (breakdown.input_cost or 0.0) + - (breakdown.cache_read_cost or 0.0) + - (breakdown.cache_creation_cost or 0.0) + ) + assert breakdown.input_cost is not None and approx_equal(fresh_cost, fresh_tokens * input_rate), ( + f"input_cost {breakdown.input_cost} less cache read {breakdown.cache_read_cost} and " + f"cache creation {breakdown.cache_creation_cost} leaves {fresh_cost}, not " + f"{fresh_tokens} fresh tokens * {input_rate} (prompt {row.prompt_tokens}, " + f"cache read {row.cache_read_tokens}, cache creation {row.cache_creation_tokens}); " + "cached tokens are being billed at the input rate" + ) + + +def poll_cost_row(proxy: ProxyClient, request_id: str) -> CostRow | None: + """Poll /spend/logs for the call's row until it lands with a cost breakdown + (rows flush ~60s behind the call via proxy_batch_write_at); None on timeout.""" + deadline = time.monotonic() + proxy.poll_timeout + while time.monotonic() < deadline: + result = proxy.transport.get( + "/spend/logs", + headers=proxy.transport.master, + params=SpendLogsParams(request_id=request_id), + response_type=CostRows, + ) + match result: + case Success(data=data): + rows = data.root + case _: + rows = [] + for row in rows: + if row.metadata and row.metadata.cost_breakdown: + return row + time.sleep(proxy.poll_interval) + return None + + +def poll_cost_row_where( + proxy: ProxyClient, api_key: str, predicate: Callable[[CostRow], bool] +) -> CostRow | None: + """Poll the key's own /spend/logs until one of its rows carries a cost breakdown + the predicate accepts; None on timeout. For calls whose response id is not the + id the bill is filed under, which is how a user finds the row in the UI anyway.""" + deadline = time.monotonic() + proxy.poll_timeout + while time.monotonic() < deadline: + result = proxy.transport.get( + "/spend/logs", + headers=proxy.transport.master, + params=SpendLogsParams(api_key=api_key), + response_type=CostRows, + ) + match result: + case Success(data=data): + rows = data.root + case _: + rows = [] + for row in rows: + if row.metadata and row.metadata.cost_breakdown and predicate(row): + return row + time.sleep(proxy.poll_interval) + return None + + +def register_priced_model( + proxy: ProxyClient, + resources: ResourceManager, + name_prefix: str, + litellm_params: LiteLLMParamsBody, +) -> str: + """Register a deployment with explicit custom rates (deleted on teardown) and + return its unique model name.""" + model_name = f"{name_prefix}-{unique_marker()}" + model_id = proxy.create_model(model_name, litellm_params) + resources.defer(lambda: proxy.delete_model(model_id)) + return model_name + + +def cacheable_prefix(marker: str, *, words: int = 1200) -> str: + """A prompt prefix above OpenAI's 1024-token caching minimum whose identity is + fully determined by `marker` (it is the first word, and prefix caching matches + from token zero). Raise `words` for models that only report a cache read on a + substantially longer prefix.""" + return " ".join(marker if i == 0 else f"token{i:04d}" for i in range(words)) diff --git a/tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py b/tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py new file mode 100644 index 00000000000..ca5985fcda6 --- /dev/null +++ b/tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py @@ -0,0 +1,263 @@ +"""Live e2e: prompt-cache token accounting bills each cache component at its own rate. + +Four regressions the gateway has shipped fixes for, pinned against real OpenAI +prompt caching (implicit, keyed on the token prefix). Every test registers its own +deployment with distinct custom rates for input / output / cache-read / +cache-creation, so the expected bill is exactly the row's token counts times the +configured rates and a component billed at the wrong rate can never pass: + +- cache writes: gpt-5.6's cache-write tokens must land on the spend row as + cache-creation tokens billed at the cache-creation rate, not silently at the + input rate (#34046) +- breakdown components: the row's metadata.cost_breakdown must itemize cache-read, + cache-creation, and reasoning costs, with reasoning a subset of output (#31686) +- streaming: a streamed call's reassembled usage must keep the cached-token detail + so cache reads bill at the cache-read discount, not full input price (#34812) +- /v1/messages bridge: a request served by a Responses-only OpenAI model crosses + the anthropic-messages -> Responses adapter and must keep its cache-read tokens + and their discounted billing (#34957) + +Each test drives the model that actually reports the component it bills, which is +not the same model throughout. gpt-5.6-luna reports cache-write tokens on every +call over the caching minimum and never reports a cache read, so it is the one +model that can prove cache-write billing and the one model that can never prove +cache-read billing. gpt-5.5 is the reverse: it reports cached tokens on the second +call and no cache writes at all. gpt-5.3-codex is Responses-only, which is what +forces the /v1/messages bridge, and it starts reporting cache reads once the +prefix is a few thousand tokens rather than one. + +OpenAI caching is best-effort, so each test retries with a fresh prefix (new +marker = brand-new cache identity) up to three times before failing; the prime and +measured calls share the prefix but differ in the trailing question, which defeats +the proxy's own response cache without touching the provider's prefix cache. +""" + +import pytest + +from cost_rows import ( + CostRow, + approx_equal, + assert_fresh_tokens_billed_at, + assert_total_is_sum_of_components, + cacheable_prefix, + poll_cost_row, + poll_cost_row_where, + register_priced_model, +) +from e2e_config import unique_marker +from e2e_http import unwrap +from lifecycle import ResourceManager +from models import AnthropicMessagesBody, ChatBody, ChatMessage, LiteLLMParamsBody +from pydantic import BaseModel +from spend_e2e_client import SpendClient + +pytestmark = pytest.mark.e2e + +CACHE_WRITE_BACKEND = "openai/gpt-5.6-luna" +CACHE_READ_BACKEND = "openai/gpt-5.5" +BRIDGE_BACKEND = "openai/gpt-5.3-codex" +BRIDGE_PREFIX_WORDS = 3000 +OPENAI_API_KEY = "os.environ/OPENAI_API_KEY" +CACHE_ATTEMPTS = 3 + +INPUT_RATE = 4e-05 +OUTPUT_RATE = 8e-05 +CACHE_READ_RATE = 1e-05 +CACHE_WRITE_RATE = 5e-05 + +PRIME_QUESTION = "Reply with the single word ready." +REASONING_QUESTION = "Compute 47*83 - 19*7 step by step, then reply with just the final number." + + +class _StreamChunk(BaseModel): + id: str | None = None + + +def _cache_priced_params(backend: str) -> LiteLLMParamsBody: + return LiteLLMParamsBody( + model=backend, + api_key=OPENAI_API_KEY, + input_cost_per_token=INPUT_RATE, + output_cost_per_token=OUTPUT_RATE, + cache_read_input_token_cost=CACHE_READ_RATE, + cache_creation_input_token_cost=CACHE_WRITE_RATE, + ) + + +def _chat_body(model: str, content: str, *, stream: bool = False) -> ChatBody: + return ChatBody( + model=model, + messages=[ChatMessage(role="user", content=content)], + stream=stream, + max_completion_tokens=4000, + ) + + +def _require_row(client: SpendClient, request_id: str) -> CostRow: + row = poll_cost_row(client.proxy, request_id) + assert row is not None, f"no spend row with a cost breakdown landed for {request_id}" + return row + + +def _assert_cache_read_billed(row: CostRow) -> None: + assert row.breakdown.cache_read_cost is not None and approx_equal( + row.breakdown.cache_read_cost, row.cache_read_tokens * CACHE_READ_RATE + ), ( + f"cache_read_cost {row.breakdown.cache_read_cost} != " + f"{row.cache_read_tokens} cached tokens * {CACHE_READ_RATE}" + ) + assert_fresh_tokens_billed_at(row, INPUT_RATE) + assert_total_is_sum_of_components(row) + + +class TestCacheCostAccounting: + @pytest.mark.covers("quota_management.spend_tracking.cache_write.bills_cache_creation_rate") + def test_cache_write_tokens_billed_at_cache_creation_rate( + self, client: SpendClient, resources: ResourceManager, scoped_key: str + ) -> None: + model = register_priced_model( + client.proxy, resources, "cache-write-priced", _cache_priced_params(CACHE_WRITE_BACKEND) + ) + + for _ in range(CACHE_ATTEMPTS): + prompt = f"{cacheable_prefix(unique_marker())}\n{PRIME_QUESTION}" + chat = unwrap(client.proxy.chat(scoped_key, _chat_body(model, prompt))) + assert chat.id, f"chat response carried no id: {chat}" + row = _require_row(client, chat.id) + if row.cache_creation_tokens > 0: + break + else: + pytest.fail( + f"OpenAI reported no cache-write tokens across {CACHE_ATTEMPTS} fresh " + "~2k-token prompts; the cache-write billing path was never exercised" + ) + + assert row.breakdown.cache_creation_cost is not None and approx_equal( + row.breakdown.cache_creation_cost, row.cache_creation_tokens * CACHE_WRITE_RATE + ), ( + f"cache_creation_cost {row.breakdown.cache_creation_cost} != " + f"{row.cache_creation_tokens} cache-write tokens * {CACHE_WRITE_RATE}" + ) + assert_fresh_tokens_billed_at(row, INPUT_RATE) + assert_total_is_sum_of_components(row) + + @pytest.mark.covers("quota_management.spend_tracking.cost_breakdown.reports_component_costs") + def test_cost_breakdown_reports_component_costs( + self, client: SpendClient, resources: ResourceManager, scoped_key: str + ) -> None: + model = register_priced_model( + client.proxy, resources, "breakdown-priced", _cache_priced_params(CACHE_READ_BACKEND) + ) + + for _ in range(CACHE_ATTEMPTS): + prefix = cacheable_prefix(unique_marker()) + unwrap(client.proxy.chat(scoped_key, _chat_body(model, f"{prefix}\n{PRIME_QUESTION}"))) + chat = unwrap( + client.proxy.chat(scoped_key, _chat_body(model, f"{prefix}\n{REASONING_QUESTION}")) + ) + assert chat.id, f"chat response carried no id: {chat}" + row = _require_row(client, chat.id) + if row.cache_read_tokens > 0: + break + else: + pytest.fail( + f"no cache read landed across {CACHE_ATTEMPTS} prime+read rounds; " + "the component-cost breakdown was never exercised with cached input" + ) + + usage = chat.usage + assert usage is not None and usage.completion_tokens_details is not None, ( + f"no completion token details on the measured call: {chat}" + ) + reasoning_tokens = usage.completion_tokens_details.reasoning_tokens or 0 + assert reasoning_tokens > 0, f"the reasoning question produced no reasoning tokens: {usage}" + + breakdown = row.breakdown + assert breakdown.output_cost is not None and approx_equal( + breakdown.output_cost, (row.completion_tokens or 0) * OUTPUT_RATE + ), ( + f"output_cost {breakdown.output_cost} != " + f"{row.completion_tokens} completion tokens * {OUTPUT_RATE}" + ) + assert breakdown.reasoning_cost is not None and approx_equal( + breakdown.reasoning_cost, reasoning_tokens * OUTPUT_RATE + ), ( + f"reasoning_cost {breakdown.reasoning_cost} != " + f"{reasoning_tokens} reasoning tokens * {OUTPUT_RATE}" + ) + assert breakdown.reasoning_cost <= (breakdown.output_cost or 0.0) * 1.01, ( + f"reasoning_cost {breakdown.reasoning_cost} exceeds output_cost " + f"{breakdown.output_cost}; reasoning must be a subset of output" + ) + _assert_cache_read_billed(row) + + @pytest.mark.covers("quota_management.spend_tracking.stream_cache_read.bills_cache_read_rate") + def test_streaming_cache_read_billed_at_cache_read_rate( + self, client: SpendClient, resources: ResourceManager, scoped_key: str + ) -> None: + model = register_priced_model( + client.proxy, resources, "stream-cache-priced", _cache_priced_params(CACHE_READ_BACKEND) + ) + + for _ in range(CACHE_ATTEMPTS): + prefix = cacheable_prefix(unique_marker()) + unwrap(client.proxy.chat(scoped_key, _chat_body(model, f"{prefix}\n{PRIME_QUESTION}"))) + result = client.proxy.chat_stream( + scoped_key, + _chat_body(model, f"{prefix}\nReply with the single word cached.", stream=True), + ) + assert result.ok and result.stream_events, ( + f"streamed chat failed (status {result.status_code}): {result.body[:300]}" + ) + stream_id = _StreamChunk.model_validate_json(result.stream_events[0]).id + assert stream_id, f"first stream chunk carried no id: {result.stream_events[0][:200]}" + row = _require_row(client, stream_id) + if row.cache_read_tokens > 0: + break + else: + pytest.fail( + f"no cache read landed across {CACHE_ATTEMPTS} prime+stream rounds; " + "streaming cache-read billing was never exercised" + ) + + _assert_cache_read_billed(row) + + @pytest.mark.covers("quota_management.spend_tracking.messages_bridge.keeps_cache_tokens") + def test_messages_bridge_keeps_cache_tokens( + self, client: SpendClient, resources: ResourceManager, scoped_key: str + ) -> None: + model = register_priced_model( + client.proxy, resources, "bridge-cache-priced", _cache_priced_params(BRIDGE_BACKEND) + ) + + def bridge_call(content: str) -> int: + response = unwrap( + client.proxy.messages( + scoped_key, + AnthropicMessagesBody( + model=model, + messages=[ChatMessage(role="user", content=content)], + max_tokens=4000, + ), + ) + ) + assert response.usage is not None, f"bridged response carried no usage: {response}" + return response.usage.cache_read_input_tokens or 0 + + for _ in range(CACHE_ATTEMPTS): + prefix = cacheable_prefix(unique_marker(), words=BRIDGE_PREFIX_WORDS) + bridge_call(f"{prefix}\n{PRIME_QUESTION}") + if bridge_call(f"{prefix}\nReply with the single word bridged.") > 0: + break + else: + pytest.fail( + f"no cache read survived {CACHE_ATTEMPTS} bridged prime+read rounds; " + "cache tokens are not surviving the anthropic-messages -> Responses bridge" + ) + + row = poll_cost_row_where(client.proxy, scoped_key, lambda r: r.cache_read_tokens > 0) + assert row is not None, ( + "the bridged call reported cached tokens but no spend row for the key " + "recorded any; the cache tokens were dropped on the way to the bill" + ) + _assert_cache_read_billed(row) diff --git a/tests/e2e/quota_management/spend_tracking/test_cost_headers_e2e.py b/tests/e2e/quota_management/spend_tracking/test_cost_headers_e2e.py new file mode 100644 index 00000000000..203be611905 --- /dev/null +++ b/tests/e2e/quota_management/spend_tracking/test_cost_headers_e2e.py @@ -0,0 +1,136 @@ +"""Live e2e: the per-component x-litellm-response-cost-* headers keep their contract. + +Pins the header contract shipped in #36965: alongside the x-litellm-response-cost +total, every response carries the component costs (input, output, cache-read, +cache-creation, reasoning, tool-usage), where input covers only fresh tokens (the +cache components are subtracted out) so the components sum to the total, and +reasoning stays a subset of output. + +The deployment carries distinct custom rates per component, a prime call fills the +provider's prefix cache, and the measured call re-reads it, so the cache-read +header is exercised with a real nonzero value instead of passing vacuously. The +backend is gpt-5.5 because it reports cached tokens on the second call; the +gpt-5.6 line reports cache writes and never a read, which would leave the +cache-read header at zero forever. The raw-transport send is used because the +typed chat client validates bodies and drops headers. OpenAI caching is +best-effort, so the prime+measure round retries with a fresh prefix before +failing. +""" + +import pytest + +from cost_rows import approx_equal, cacheable_prefix, register_priced_model +from e2e_config import unique_marker +from e2e_http import StreamingResponse +from lifecycle import ResourceManager +from models import ChatBody, ChatMessage, ChatResponse, LiteLLMParamsBody +from spend_e2e_client import SpendClient + +pytestmark = pytest.mark.e2e + +BACKEND = "openai/gpt-5.5" +OPENAI_API_KEY = "os.environ/OPENAI_API_KEY" +CACHE_ATTEMPTS = 3 + +INPUT_RATE = 4e-05 +OUTPUT_RATE = 8e-05 +CACHE_READ_RATE = 1e-05 +CACHE_WRITE_RATE = 5e-05 + +COMPONENT_HEADERS = ( + "x-litellm-response-cost-input", + "x-litellm-response-cost-cache-read", + "x-litellm-response-cost-cache-creation", + "x-litellm-response-cost-output", + "x-litellm-response-cost-tool-usage", +) + + +def _header_cost(response: StreamingResponse, name: str) -> float: + value = response.headers.get(name) + return float(value) if value not in (None, "", "None") else 0.0 + + +class TestCostHeaders: + @pytest.mark.covers("quota_management.spend_tracking.cost_headers.additive_components") + def test_component_cost_headers_sum_to_total( + self, client: SpendClient, resources: ResourceManager, scoped_key: str + ) -> None: + model = register_priced_model( + client.proxy, + resources, + "header-priced", + LiteLLMParamsBody( + model=BACKEND, + api_key=OPENAI_API_KEY, + input_cost_per_token=INPUT_RATE, + output_cost_per_token=OUTPUT_RATE, + cache_read_input_token_cost=CACHE_READ_RATE, + cache_creation_input_token_cost=CACHE_WRITE_RATE, + ), + ) + + def priced_call(content: str) -> StreamingResponse: + response = client.proxy.transport.send( + "/chat/completions", + headers=client.proxy.transport.bearer(scoped_key), + json=ChatBody( + model=model, + messages=[ChatMessage(role="user", content=content)], + max_completion_tokens=4000, + ), + ) + assert response.ok, f"chat failed (status {response.status_code}): {response.body[:300]}" + return response + + for _ in range(CACHE_ATTEMPTS): + prefix = cacheable_prefix(unique_marker()) + priced_call(f"{prefix}\nReply with the single word ready.") + measured = priced_call(f"{prefix}\nReply with the single word measured.") + if _header_cost(measured, "x-litellm-response-cost-cache-read") > 0: + break + else: + pytest.fail( + f"no cache read landed across {CACHE_ATTEMPTS} prime+measure rounds; " + "the cache-read cost header was never exercised with a nonzero value" + ) + + total = measured.response_cost + assert total is not None and total > 0, ( + f"x-litellm-response-cost missing or zero: {measured.headers}" + ) + component_sum = sum(_header_cost(measured, name) for name in COMPONENT_HEADERS) + assert approx_equal(component_sum, total), ( + f"component headers sum to {component_sum}, not the total {total}: " + f"{ {name: measured.headers.get(name) for name in COMPONENT_HEADERS} }" + ) + + reasoning = _header_cost(measured, "x-litellm-response-cost-reasoning") + output = _header_cost(measured, "x-litellm-response-cost-output") + assert reasoning <= output * 1.01, ( + f"reasoning header {reasoning} exceeds output header {output}; " + "reasoning must be a subset of output" + ) + + usage = ChatResponse.model_validate_json(measured.body).usage + assert usage is not None, f"measured response carried no usage: {measured.body[:300]}" + cached_tokens = ( + usage.prompt_tokens_details.cached_tokens or 0 if usage.prompt_tokens_details else 0 + ) + cache_creation_tokens = usage.cache_creation_input_tokens or 0 + assert cached_tokens > 0, f"cache-read header nonzero but usage shows no cached tokens: {usage}" + assert approx_equal( + _header_cost(measured, "x-litellm-response-cost-cache-read"), + cached_tokens * CACHE_READ_RATE, + ), ( + f"cache-read header {measured.headers.get('x-litellm-response-cost-cache-read')} != " + f"{cached_tokens} cached tokens * {CACHE_READ_RATE}" + ) + fresh_tokens = (usage.prompt_tokens or 0) - cached_tokens - cache_creation_tokens + assert approx_equal( + _header_cost(measured, "x-litellm-response-cost-input"), fresh_tokens * INPUT_RATE + ), ( + f"input header {measured.headers.get('x-litellm-response-cost-input')} != " + f"{fresh_tokens} fresh tokens * {INPUT_RATE}; the input component is not " + "subtracting the cache components" + ) diff --git a/tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py b/tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py new file mode 100644 index 00000000000..4c3a2a4509f --- /dev/null +++ b/tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py @@ -0,0 +1,68 @@ +"""Live e2e: the /openai passthrough injects usage.cost into streaming usage frames. + +Pins #36503: with the proxy running `include_cost_in_streaming_usage: true`, a +streamed call through the provider passthrough surface must carry the computed +cost inside the final usage-only SSE frame, the same contract the native +/chat/completions stream has. Providers never send `cost` themselves, so a +nonzero value proves the proxy computed and injected it on the passthrough path. + +The row-side spend accounting for passthrough calls is covered elsewhere; this +test pins only the in-stream cost surface, which clients read without ever +touching /spend/logs. +""" + +import pytest +from pydantic import BaseModel + +from e2e_config import unique_marker +from models import ChatBody, ChatMessage, StreamOptions, Usage +from spend_e2e_client import SpendClient + +pytestmark = pytest.mark.e2e + +OPENAI_MODEL = "gpt-5.6-luna" + + +class _StreamFrame(BaseModel): + usage: Usage | None = None + + +class TestPassthroughStreamCost: + @pytest.mark.covers("quota_management.spend_tracking.passthrough_stream.injects_usage_cost") + def test_passthrough_stream_final_usage_frame_carries_cost( + self, client: SpendClient, scoped_key: str + ) -> None: + result = client.proxy.transport.send( + "/openai/v1/chat/completions", + headers=client.proxy.transport.bearer(scoped_key), + json=ChatBody( + model=OPENAI_MODEL, + messages=[ + ChatMessage( + role="user", + content=f"{unique_marker()} Reply with the single word passthrough.", + ) + ], + stream=True, + stream_options=StreamOptions(), + ), + stream=True, + ) + assert result.ok and result.stream_events, ( + f"passthrough stream failed (status {result.status_code}): {result.body[:300]}" + ) + + usage_frames = [ + frame.usage + for frame in (_StreamFrame.model_validate_json(event) for event in result.stream_events) + if frame.usage is not None + ] + assert usage_frames, ( + f"no usage frame in the passthrough stream despite stream_options.include_usage; " + f"last event: {result.stream_events[-1][:300]}" + ) + + final_usage = usage_frames[-1] + assert final_usage.cost is not None and final_usage.cost > 0, ( + f"final passthrough usage frame carries no injected cost: {final_usage}" + ) diff --git a/tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py b/tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py new file mode 100644 index 00000000000..171c849fb4c --- /dev/null +++ b/tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py @@ -0,0 +1,116 @@ +"""Live e2e: a service_tier request bills every component at the tier's own rates. + +Pins the tier-billing fixes (#35923, #35925): a priority-tier call must price +input and output at the deployment's `*_priority` rates, including the reasoning +tokens inside output (the shipped bug billed reasoning at the default-tier rate), +and the spend row must record the tier the bill was computed on. + +The deployment carries custom base AND priority rates, each distinct, so a bill +computed from the wrong tier (or a mix) cannot match the expected numbers. The +prompt is a fresh unique marker per run, keeping cached tokens out of the math. +The response's own `service_tier` echo is asserted first: if OpenAI ever declined +priority processing and served the default tier, the test fails there instead of +producing a vacuous rate comparison. +""" + +import pytest + +from cost_rows import ( + approx_equal, + assert_fresh_tokens_billed_at, + assert_total_is_sum_of_components, + poll_cost_row, + register_priced_model, +) +from e2e_config import unique_marker +from e2e_http import unwrap +from lifecycle import ResourceManager +from models import ChatBody, ChatMessage, LiteLLMParamsBody +from spend_e2e_client import SpendClient + +pytestmark = pytest.mark.e2e + +BACKEND = "openai/gpt-5.6-luna" +OPENAI_API_KEY = "os.environ/OPENAI_API_KEY" + +INPUT_RATE = 4e-05 +OUTPUT_RATE = 8e-05 +PRIORITY_INPUT_RATE = 6e-05 +PRIORITY_OUTPUT_RATE = 1.6e-04 + + +class TestServiceTierPricing: + @pytest.mark.covers("quota_management.spend_tracking.service_tier.bills_tier_rates") + def test_priority_tier_bills_priority_rates( + self, client: SpendClient, resources: ResourceManager, scoped_key: str + ) -> None: + model = register_priced_model( + client.proxy, + resources, + "tier-priced", + LiteLLMParamsBody( + model=BACKEND, + api_key=OPENAI_API_KEY, + input_cost_per_token=INPUT_RATE, + output_cost_per_token=OUTPUT_RATE, + input_cost_per_token_priority=PRIORITY_INPUT_RATE, + output_cost_per_token_priority=PRIORITY_OUTPUT_RATE, + ), + ) + + chat = unwrap( + client.proxy.chat( + scoped_key, + ChatBody( + model=model, + messages=[ + ChatMessage( + role="user", + content=( + f"{unique_marker()} Compute 47*83 - 19*7 step by step, " + "then reply with just the final number." + ), + ) + ], + max_completion_tokens=4000, + service_tier="priority", + ), + ) + ) + assert chat.service_tier == "priority", ( + f"OpenAI served tier {chat.service_tier!r} instead of priority; " + "tier billing was never exercised" + ) + assert chat.id, f"chat response carried no id: {chat}" + + row = poll_cost_row(client.proxy, chat.id) + assert row is not None, f"no spend row with a cost breakdown landed for {chat.id}" + breakdown = row.breakdown + + assert breakdown.service_tier == "priority", ( + f"the bill records pricing basis {breakdown.service_tier!r}, not priority" + ) + + assert_fresh_tokens_billed_at(row, PRIORITY_INPUT_RATE) + assert breakdown.output_cost is not None and approx_equal( + breakdown.output_cost, (row.completion_tokens or 0) * PRIORITY_OUTPUT_RATE + ), ( + f"output_cost {breakdown.output_cost} != {row.completion_tokens} tokens * priority rate " + f"{PRIORITY_OUTPUT_RATE} (base rate would give {(row.completion_tokens or 0) * OUTPUT_RATE})" + ) + + usage = chat.usage + assert usage is not None and usage.completion_tokens_details is not None, ( + f"no completion token details on the priority call: {chat}" + ) + reasoning_tokens = usage.completion_tokens_details.reasoning_tokens or 0 + assert reasoning_tokens > 0, f"the reasoning question produced no reasoning tokens: {usage}" + assert breakdown.reasoning_cost is not None and approx_equal( + breakdown.reasoning_cost, reasoning_tokens * PRIORITY_OUTPUT_RATE + ), ( + f"reasoning_cost {breakdown.reasoning_cost} != {reasoning_tokens} reasoning tokens * " + f"priority rate {PRIORITY_OUTPUT_RATE} (the default-tier rate would give " + f"{reasoning_tokens * OUTPUT_RATE})" + ) + + assert_total_is_sum_of_components(row) From 021a09b1560d9bd79c5abc102d919aa9e18c867b Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:00:29 -0700 Subject: [PATCH 031/220] fix(passthrough): resolve vertex live credentials from db model deployments The /vertex_ai/live WebSocket passthrough only ever looked at default_vertex_config and the DEFAULT_VERTEXAI_* env vars, so a proxy whose Vertex credentials live in the DB as a model entry with use_in_pass_through had nothing to authenticate with. The upgrade still succeeded and the socket then closed with a bare 1000 on the first client frame, which gave the client no way to tell a misconfiguration from a normal end of session. Credentials now also resolve from the router deployments flagged use_in_pass_through, preferring the one matching the requested model, and a failure to mint an access token closes 1011 with a reason naming both ways to configure it. Upstream closes other than a plain 1000 are relayed to the client with their code and reason, so Google's own errors reach the caller. The setup frame's model is rewritten to the full projects/.../publishers/google/models resource path, which is what Vertex expects and what lets a bare model id or a gateway alias work over this route. --- litellm/constants.py | 3 + .../llm_passthrough_endpoints.py | 152 ++++++++++--- .../pass_through_endpoints.py | 75 ++++++- .../passthrough_endpoint_router.py | 65 +++++- .../test_llm_pass_through_endpoints.py | 101 +++++++++ .../test_pass_through_endpoints.py | 206 ++++++++++++++++++ .../test_passthrough_endpoint_router.py | 141 ++++++++++++ 7 files changed, 701 insertions(+), 42 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index facfc6f7c19..3ce2cfd35a5 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -244,6 +244,9 @@ AIOHTTP_NEEDS_CLEANUP_CLOSED: Final = (3, 13, 0) <= sys.version_info < ( _max_size_env: Final = os.getenv("REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES") REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES: Final = int(_max_size_env) if _max_size_env is not None else None +# RFC 6455 caps the close frame payload at 125 bytes, 2 of which carry the status code +WEBSOCKET_CLOSE_REASON_MAX_BYTES: Final = 123 + # SSL/TLS cipher configuration for faster handshakes # Strategy: Strongly prefer fast modern ciphers, but allow fallback to commonly supported ones # This balances performance with broad compatibility diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index c5ab7f1fc63..050ad0fd627 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -9,8 +9,9 @@ Use litellm with Anthropic SDK, Vertex AI SDK, Cohere SDK, etc. import json import os import re +from collections.abc import Callable from types import MappingProxyType -from typing import Annotated, Any, Final, cast +from typing import TYPE_CHECKING, Annotated, Any, Final, cast import httpx from fastapi import APIRouter, Depends, HTTPException, Request, Response, WebSocket @@ -55,11 +56,15 @@ from litellm.types.passthrough_endpoints.pass_through_endpoints import ( LITELLM_PASS_THROUGH_CUSTOM_BODY_STATE_KEY, LITELLM_PASS_THROUGH_RAW_BODY_STATE_KEY, ) +from litellm.types.passthrough_endpoints.vertex_ai import VertexPassThroughCredentials from litellm.types.utils import LlmProviders from litellm.utils import ProviderConfigManager from .passthrough_endpoint_router import PassthroughEndpointRouter +if TYPE_CHECKING: + from litellm.router import Router + vertex_llm_base: Final = VertexBase() router: Final = APIRouter() openai_passthrough_router: Final = APIRouter() @@ -2373,6 +2378,89 @@ async def cursor_proxy_route( return received_value +VERTEX_LIVE_UNCONFIGURED_CLOSE_REASON: Final = ( + "Vertex AI auth failed: set a use_in_pass_through vertex model, default_vertex_config, or DEFAULT_VERTEXAI_* env" +) + +VERTEX_PUBLISHER_MODEL_PREFIX: Final = "publishers/google/models/" + + +def _get_llm_router() -> "Router | None": + from litellm.proxy.proxy_server import llm_router + + return llm_router + + +def _resolve_vertex_live_credentials( + vertex_project: str | None, + vertex_location: str | None, + model: str | None, +) -> VertexPassThroughCredentials | None: + """ + Resolution order: credentials registered for the requested project/location, then any DB model entry + flagged ``use_in_pass_through``, then ``default_vertex_config`` and the ``DEFAULT_VERTEXAI_*`` env vars + """ + keyed: Final = passthrough_endpoint_router.get_vertex_credentials( + project_id=vertex_project, + location=vertex_location, + ) + if keyed is not None and keyed.vertex_project is not None: + return keyed + from_deployments: Final = passthrough_endpoint_router.get_vertex_credentials_from_router_deployments(model=model) + if from_deployments is not None: + return from_deployments + if keyed is not None: + return keyed + passthrough_endpoint_router.set_default_vertex_config() + return passthrough_endpoint_router.get_vertex_credentials( + project_id=vertex_project, + location=vertex_location, + ) + + +def _build_vertex_live_setup_model_rewriter( + vertex_project: str | None, + vertex_location: str | None, + llm_router: "Router | None", +) -> Callable[[str], str] | None: + """ + Rewrite the ``setup`` frame's model into the full Vertex resource path the Live API requires. + + Clients address the gateway the way they address LiteLLM (bare id or model alias); Vertex reads anything + that is not a ``projects/...`` path as a project name and closes the socket + """ + if vertex_project is None or vertex_location is None: + return None + + def rewrite(setup_model: str) -> str: + if setup_model.startswith("projects/"): + return setup_model + aliased: Final = _resolve_alias_to_upstream_model(setup_model, llm_router) + return ( + f"projects/{vertex_project}/locations/{vertex_location}/" + f"{VERTEX_PUBLISHER_MODEL_PREFIX}{aliased.removeprefix(VERTEX_PUBLISHER_MODEL_PREFIX)}" + ) + + return rewrite + + +def _resolve_alias_to_upstream_model(setup_model: str, llm_router: "Router | None") -> str: + if llm_router is None: + return setup_model + upstream: Final = next( + ( + deployment["litellm_params"]["model"] + for deployment in (llm_router.get_model_list() or ()) + if deployment.get("model_name") == setup_model + ), + None, + ) + if upstream is None: + return setup_model + _, provider, _, _ = litellm.get_llm_provider(model=upstream) + return upstream.removeprefix(f"{provider}/") + + async def vertex_ai_live_websocket_passthrough( websocket: WebSocket, model: str | None = None, @@ -2396,51 +2484,40 @@ async def vertex_ai_live_websocket_passthrough( await websocket.accept() incoming_headers: Final = dict(websocket.headers) - vertex_credentials_config = passthrough_endpoint_router.get_vertex_credentials( - project_id=vertex_project, - location=vertex_location, + vertex_credentials_config: Final = _resolve_vertex_live_credentials( + vertex_project=vertex_project, + vertex_location=vertex_location, + model=model, ) - if vertex_credentials_config is None: - # Attempt to load defaults from environment/config if not already initialised - passthrough_endpoint_router.set_default_vertex_config() - vertex_credentials_config = passthrough_endpoint_router.get_vertex_credentials( - project_id=vertex_project, - location=vertex_location, - ) - - resolved_project = vertex_project - resolved_location: str | None = vertex_location - credentials_value: str | None = None - - if vertex_credentials_config is not None: - resolved_project = resolved_project or vertex_credentials_config.vertex_project - temp_location: Final = resolved_location or vertex_credentials_config.vertex_location - # Ensure resolved_location is a string - if isinstance(temp_location, dict) or temp_location is not None: - resolved_location = str(temp_location) - else: - resolved_location = None - credentials_value = ( - str(vertex_credentials_config.vertex_credentials) - if vertex_credentials_config.vertex_credentials is not None - else None - ) + configured_project: Final = vertex_project or ( + vertex_credentials_config.vertex_project if vertex_credentials_config is not None else None + ) + configured_location: Final = vertex_location or ( + vertex_credentials_config.vertex_location if vertex_credentials_config is not None else None + ) + credentials_value: Final = ( + str(vertex_credentials_config.vertex_credentials) + if vertex_credentials_config is not None and vertex_credentials_config.vertex_credentials is not None + else None + ) try: - resolved_location = resolved_location or (vertex_llm_base.get_default_vertex_location()) - if model: - resolved_location = vertex_llm_base.get_vertex_region( - vertex_region=resolved_location, + resolved_location: Final = ( + vertex_llm_base.get_vertex_region( + vertex_region=configured_location or vertex_llm_base.get_default_vertex_location(), model=model, ) + if model + else configured_location or vertex_llm_base.get_default_vertex_location() + ) ( access_token, resolved_project, ) = await vertex_llm_base._ensure_access_token_async( credentials=credentials_value, - project_id=resolved_project, + project_id=configured_project, custom_llm_provider="vertex_ai_beta", ) except Exception as e: @@ -2453,7 +2530,7 @@ async def vertex_ai_live_websocket_passthrough( request_data={}, ) if websocket.client_state != WebSocketState.DISCONNECTED: - await websocket.close(code=1011, reason="Vertex AI authentication failed") + await websocket.close(code=1011, reason=VERTEX_LIVE_UNCONFIGURED_CLOSE_REASON) return host_location: Final = resolved_location or vertex_llm_base.get_default_vertex_location() @@ -2485,6 +2562,11 @@ async def vertex_ai_live_websocket_passthrough( forward_headers=False, endpoint="/vertex_ai/live", accept_websocket=False, + setup_model_rewriter=_build_vertex_live_setup_model_rewriter( + vertex_project=resolved_project, + vertex_location=resolved_location, + llm_router=_get_llm_router(), + ), ) diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py index 0df0aaa1bcd..d68ef8019d2 100644 --- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py @@ -5,7 +5,7 @@ import json import posixpath import traceback from base64 import b64encode -from collections.abc import AsyncGenerator, Callable, Mapping +from collections.abc import AsyncGenerator, Callable, Iterable, Mapping from datetime import datetime from itertools import groupby from typing import Any, Final, TypedDict, cast @@ -32,11 +32,15 @@ from websockets.exceptions import ( ConnectionClosedOK, InvalidStatus, ) +from websockets.frames import Close import litellm from litellm._logging import verbose_proxy_logger from litellm._uuid import uuid -from litellm.constants import MAXIMUM_TRACEBACK_LINES_TO_LOG +from litellm.constants import ( + MAXIMUM_TRACEBACK_LINES_TO_LOG, + WEBSOCKET_CLOSE_REASON_MAX_BYTES, +) from litellm.integrations.custom_guardrail import CustomGuardrail from litellm.integrations.custom_logger import CustomLogger from litellm.litellm_core_utils.core_helpers import ( @@ -1890,6 +1894,52 @@ def create_websocket_passthrough_route( return websocket_endpoint_func +def _rewrite_vertex_live_setup_model(text_data: str, setup_model_rewriter: Callable[[str], str] | None) -> str: + """ + Rewrite the model of a Vertex AI Live ``setup`` frame, leaving every other frame byte-identical + """ + if setup_model_rewriter is None: + return text_data + try: + message: Final = json.loads(text_data) + except json.JSONDecodeError: + return text_data + if not isinstance(message, dict): + return text_data + setup: Final = message.get("setup") + if not isinstance(setup, dict): + return text_data + setup_model: Final = setup.get("model") + if not isinstance(setup_model, str): + return text_data + rewritten_model: Final = setup_model_rewriter(setup_model) + if rewritten_model == setup_model: + return text_data + return json.dumps({**message, "setup": {**setup, "model": rewritten_model}}) # mutable-ok: one-shot json payload + + +def _truncated_close_reason(reason: str) -> str: + """ + Fit a close reason inside the byte budget a WebSocket close frame allows, without splitting a character + """ + encoded: Final = reason.encode("utf-8") + if len(encoded) <= WEBSOCKET_CLOSE_REASON_MAX_BYTES: + return reason + return encoded[:WEBSOCKET_CLOSE_REASON_MAX_BYTES].decode("utf-8", errors="ignore") + + +def _upstream_close_to_relay(task_results: Iterable[object]) -> Close | None: + """ + The upstream close worth telling the client about: anything other than a plain, reasonless normal close + """ + upstream_close: Final = next((result for result in task_results if isinstance(result, Close)), None) + if upstream_close is None: + return None + if upstream_close.code == 1000 and upstream_close.reason == "": + return None + return upstream_close + + async def websocket_passthrough_request( websocket: WebSocket, target: str, @@ -1899,6 +1949,7 @@ async def websocket_passthrough_request( endpoint: str | None = None, cost_per_request: float | None = None, accept_websocket: bool = True, + setup_model_rewriter: Callable[[str], str] | None = None, ): """ WebSocket passthrough request handler. @@ -1911,6 +1962,7 @@ async def websocket_passthrough_request( forward_headers: Whether to forward incoming headers 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 """ from litellm.litellm_core_utils.litellm_logging import Logging from litellm.proxy.proxy_server import proxy_logging_obj @@ -2100,7 +2152,7 @@ async def websocket_passthrough_request( ) # Not a JSON message or doesn't contain setup data - await upstream_ws.send(text_data) + await upstream_ws.send(_rewrite_vertex_live_setup_model(text_data, setup_model_rewriter)) elif bytes_data is not None: await upstream_ws.send(bytes_data) except asyncio.CancelledError: @@ -2111,8 +2163,8 @@ async def websocket_passthrough_request( ) await upstream_ws.close() - async def forward_upstream_to_client() -> None: - """Forward messages from upstream to client WebSocket""" + async def forward_upstream_to_client() -> Close | None: + """Forward messages from upstream to client WebSocket, returning the upstream's close frame""" try: # Wait for the first response from upstream raw_response = await upstream_ws.recv(decode=False) @@ -2177,6 +2229,7 @@ async def websocket_passthrough_request( except (ConnectionClosedOK, ConnectionClosedError) as e: verbose_proxy_logger.debug("Upstream WebSocket connection closed: %s", e) + return e.rcvd except asyncio.CancelledError: verbose_proxy_logger.debug("asyncio.CancelledError in forward_upstream_to_client") raise @@ -2209,6 +2262,13 @@ async def websocket_passthrough_request( if exception is not None: raise exception + upstream_close: Final = _upstream_close_to_relay(task.result() for task in done) + if upstream_close is not None and websocket.application_state != WebSocketState.DISCONNECTED: + await websocket.close( + code=upstream_close.code, + reason=_truncated_close_reason(upstream_close.reason), + ) + end_time: Final = datetime.now() # Update passthrough logging payload with response data @@ -2325,7 +2385,10 @@ async def websocket_passthrough_request( if websocket.client_state != WebSocketState.DISCONNECTED: await websocket.close(code=1011, reason="WebSocket passthrough error") finally: - if websocket.client_state != WebSocketState.DISCONNECTED: + if ( + websocket.client_state != WebSocketState.DISCONNECTED + and websocket.application_state != WebSocketState.DISCONNECTED + ): await websocket.close() diff --git a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py index 1d2b4504d61..e7887e33ca6 100644 --- a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py +++ b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py @@ -10,7 +10,7 @@ from litellm.litellm_core_utils.credential_accessor import CredentialAccessor from litellm.secret_managers.main import get_secret_str from litellm.types.llms.vertex_ai import VERTEX_CREDENTIALS_TYPES from litellm.types.passthrough_endpoints.vertex_ai import VertexPassThroughCredentials -from litellm.types.router import LiteLLMParamsTypedDict +from litellm.types.router import DeploymentTypedDict, LiteLLMParamsTypedDict if TYPE_CHECKING: from litellm.router import Router @@ -120,6 +120,69 @@ class PassthroughEndpointRouter: return None return provider + def get_vertex_credentials_from_router_deployments(self, model: str | None) -> VertexPassThroughCredentials | None: + """ + Resolve vertex pass-through credentials from the live router deployments flagged ``use_in_pass_through``. + + ``deployment_key_to_vertex_credentials`` is only reachable when the caller names a project and location, + which WebSocket clients never do, so DB-stored deployments need this lookup to be usable at all. + """ + llm_router: Final = self.llm_router_getter() + if llm_router is None: + return None + resolved: Final = tuple( + (deployment, credentials) + for deployment in (llm_router.get_model_list() or ()) + if (credentials := self._resolve_vertex_deployment_credentials(deployment["litellm_params"])) is not None + ) + if len(resolved) == 0: + return None + return next( + ( + credentials + for deployment, credentials in resolved + if model is not None and self._deployment_matches_model(deployment, model) + ), + resolved[0][1], + ) + + def _resolve_vertex_deployment_credentials( + self, litellm_params: LiteLLMParamsTypedDict + ) -> VertexPassThroughCredentials | None: + if litellm_params.get("use_in_pass_through") is not True: + return None + if self._get_deployment_provider(litellm_params) != "vertex_ai": + return None + credential_name: Final = litellm_params.get("litellm_credential_name") + credential_values: Final = ( + CredentialAccessor.get_credential_values(credential_name) if credential_name is not None else None + ) + vertex_project: Final = _get_str_value(credential_values, "vertex_project") or litellm_params.get( + "vertex_project" + ) + vertex_location: Final = _get_str_value(credential_values, "vertex_location") or litellm_params.get( + "vertex_location" + ) + vertex_credentials: Final = _get_str_value(credential_values, "vertex_credentials") or litellm_params.get( + "vertex_credentials" + ) + if vertex_project is None or vertex_location is None: + return None + return VertexPassThroughCredentials( + vertex_project=vertex_project, + vertex_location=vertex_location, + vertex_credentials=vertex_credentials, + ) + + @staticmethod + def _deployment_matches_model(deployment: DeploymentTypedDict, model: str) -> bool: + upstream_model: Final = deployment["litellm_params"].get("model") + return model in ( + deployment.get("model_name"), + upstream_model, + upstream_model.split("/", 1)[-1] if upstream_model is not None else None, + ) + def _get_vertex_env_vars(self) -> VertexPassThroughCredentials: """ Helper to get vertex pass through config from environment variables diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index a9454854948..5ed974c7a47 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -3887,3 +3887,104 @@ class TestComprehendMedicalProxyRoute: user_api_key_dict=Mock(), ) assert exc_info.value.status_code == 400 + + +class TestVertexAILiveWebsocketPassthrough: + def _websocket(self): + from starlette.websockets import WebSocketState + + websocket = MagicMock() + websocket.accept = AsyncMock() + websocket.close = AsyncMock() + websocket.headers = {} + websocket.client_state = WebSocketState.CONNECTED + return websocket + + def _clear_vertex_env(self, monkeypatch): + monkeypatch.delenv("DEFAULT_VERTEXAI_PROJECT", raising=False) + monkeypatch.delenv("DEFAULT_VERTEXAI_LOCATION", raising=False) + monkeypatch.delenv("DEFAULT_GOOGLE_APPLICATION_CREDENTIALS", raising=False) + + @pytest.mark.asyncio + async def test_uses_db_deployment_credentials_without_query_params(self, monkeypatch): + from litellm.proxy.pass_through_endpoints import ( + llm_passthrough_endpoints as passthrough_module, + ) + + llm_router = litellm.Router( + model_list=[ + { + "model_name": "gemini-live", + "litellm_params": { + "model": "vertex_ai/gemini-live-2.5-flash", + "use_in_pass_through": True, + "vertex_project": "proj-db", + "vertex_location": "global", + "vertex_credentials": '{"type": "service_account"}', + }, + } + ] + ) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", llm_router) + monkeypatch.setattr( + passthrough_module.passthrough_endpoint_router, "default_vertex_config", None + ) + self._clear_vertex_env(monkeypatch) + websocket = self._websocket() + ensure_token = AsyncMock(return_value=("token-abc", "proj-db")) + ws_passthrough = AsyncMock() + + with ( + patch.object(passthrough_module.vertex_llm_base, "_ensure_access_token_async", ensure_token), + patch.object(passthrough_module, "websocket_passthrough_request", ws_passthrough), + ): + await passthrough_module.vertex_ai_live_websocket_passthrough( + websocket=websocket, + user_api_key_dict=UserAPIKeyAuth(), + ) + + ensure_token.assert_awaited_once_with( + credentials='{"type": "service_account"}', + project_id="proj-db", + custom_llm_provider="vertex_ai_beta", + ) + passthrough_kwargs = ws_passthrough.await_args.kwargs + assert passthrough_kwargs["target"] == ( + "wss://aiplatform.googleapis.com/ws/google.cloud.aiplatform.v1.LlmBidiService/BidiGenerateContent" + ) + assert passthrough_kwargs["custom_headers"]["Authorization"] == "Bearer token-abc" + rewriter = passthrough_kwargs["setup_model_rewriter"] + assert rewriter("gemini-live") == ( + "projects/proj-db/locations/global/publishers/google/models/gemini-live-2.5-flash" + ) + websocket.close.assert_not_awaited() + + @pytest.mark.asyncio + async def test_credential_failure_close_names_configuration_options(self, monkeypatch): + from litellm.proxy.pass_through_endpoints import ( + llm_passthrough_endpoints as passthrough_module, + ) + + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None) + monkeypatch.setattr( + passthrough_module.passthrough_endpoint_router, "default_vertex_config", None + ) + self._clear_vertex_env(monkeypatch) + websocket = self._websocket() + ensure_token = AsyncMock(side_effect=Exception("Unable to find your credentials")) + + with ( + patch.object(passthrough_module.vertex_llm_base, "_ensure_access_token_async", ensure_token), + patch("litellm.proxy.proxy_server.proxy_logging_obj") as mock_proxy_logging, + ): + mock_proxy_logging.post_call_failure_hook = AsyncMock() + await passthrough_module.vertex_ai_live_websocket_passthrough( + websocket=websocket, + user_api_key_dict=UserAPIKeyAuth(), + ) + + close_kwargs = websocket.close.await_args.kwargs + assert close_kwargs["code"] == 1011 + assert "use_in_pass_through" in close_kwargs["reason"] + assert "default_vertex_config" in close_kwargs["reason"] + assert len(close_kwargs["reason"].encode("utf-8")) <= 123 diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index b1b934b0949..844aa099541 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -4994,6 +4994,212 @@ async def test_websocket_passthrough_forwards_non_ascii_first_frame(): assert all(call.kwargs.get("code") != 1011 for call in websocket.close.await_args_list) +class ClosingUpstreamWebSocket: + def __init__(self, close_exc: Exception): + self._close_exc = close_exc + self.close = AsyncMock() + self.send = AsyncMock() + + async def recv(self, decode: bool = True): + raise self._close_exc + + def __aiter__(self): + return self + + async def __anext__(self): + raise StopAsyncIteration + + +class RecordingUpstreamWebSocket: + def __init__(self): + self.close = AsyncMock() + self.send = AsyncMock() + + async def recv(self, decode: bool = True): + await asyncio.Event().wait() + + def __aiter__(self): + return self + + async def __anext__(self): + raise StopAsyncIteration + + +def _client_websocket(receive): + from starlette.websockets import WebSocketState + + websocket = MagicMock() + websocket.accept = AsyncMock() + websocket.send_text = AsyncMock() + websocket.send_bytes = AsyncMock() + websocket.receive = receive + websocket.headers = {} + websocket.client_state = WebSocketState.CONNECTED + websocket.application_state = WebSocketState.CONNECTED + + def _mark_closed(*args, **kwargs): + websocket.application_state = WebSocketState.DISCONNECTED + + websocket.close = AsyncMock(side_effect=_mark_closed) + return websocket + + +@contextmanager +def _patched_websocket_passthrough_environment(upstream_ws): + with ( + patch("litellm.proxy.proxy_server.proxy_logging_obj") as mock_proxy_logging, + patch( + "litellm.proxy.pass_through_endpoints.pass_through_endpoints.connect", + return_value=FakeUpstreamConnect(upstream_ws), + ), + patch( + "litellm.proxy.pass_through_endpoints.pass_through_endpoints.GLOBAL_LOGGING_WORKER" + ) as mock_worker, + ): + 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_worker.ensure_initialized_and_enqueue = MagicMock( + side_effect=lambda async_coroutine: async_coroutine.close() + ) + yield + + +async def _pending_receive(): + await asyncio.Event().wait() + + +@pytest.mark.asyncio +async def test_websocket_passthrough_relays_upstream_policy_close_to_client(): + from websockets.exceptions import ConnectionClosedError + from websockets.frames import Close + + upstream_reason = "Publisher Model `projects/p/locations/global/publishers/google/models/nope` was not found" + upstream_ws = ClosingUpstreamWebSocket( + ConnectionClosedError( + rcvd=Close(1008, upstream_reason), + sent=Close(1008, ""), + rcvd_then_sent=True, + ) + ) + websocket = _client_websocket(_pending_receive) + + with _patched_websocket_passthrough_environment(upstream_ws): + await websocket_passthrough_request( + websocket=websocket, + target="wss://aiplatform.googleapis.com/ws/google.cloud.aiplatform.v1.LlmBidiService/BidiGenerateContent", + custom_headers={"Authorization": "Bearer token"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/vertex_ai/live", + accept_websocket=False, + ) + + websocket.close.assert_awaited_once_with(code=1008, reason=upstream_reason) + + +@pytest.mark.asyncio +async def test_websocket_passthrough_keeps_normal_upstream_close_normal(): + from websockets.exceptions import ConnectionClosedOK + from websockets.frames import Close + + upstream_ws = ClosingUpstreamWebSocket( + ConnectionClosedOK(rcvd=Close(1000, ""), sent=Close(1000, ""), rcvd_then_sent=True) + ) + websocket = _client_websocket(_pending_receive) + + with _patched_websocket_passthrough_environment(upstream_ws): + await websocket_passthrough_request( + websocket=websocket, + target="wss://aiplatform.googleapis.com/ws/google.cloud.aiplatform.v1.LlmBidiService/BidiGenerateContent", + custom_headers={"Authorization": "Bearer token"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/vertex_ai/live", + accept_websocket=False, + ) + + websocket.close.assert_awaited_once_with() + + +async def _run_setup_rewrite_passthrough(setup_model: str, llm_router) -> str: + from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import ( + _build_vertex_live_setup_model_rewriter, + ) + + upstream_ws = RecordingUpstreamWebSocket() + setup_frame = json.dumps({"setup": {"model": setup_model, "generationConfig": {"responseModalities": ["TEXT"]}}}) + websocket = _client_websocket( + AsyncMock( + side_effect=[ + {"type": "websocket.receive", "text": setup_frame}, + {"type": "websocket.disconnect"}, + ] + ) + ) + + with _patched_websocket_passthrough_environment(upstream_ws): + await websocket_passthrough_request( + websocket=websocket, + target="wss://aiplatform.googleapis.com/ws/google.cloud.aiplatform.v1.LlmBidiService/BidiGenerateContent", + custom_headers={"Authorization": "Bearer token"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/vertex_ai/live", + accept_websocket=False, + setup_model_rewriter=_build_vertex_live_setup_model_rewriter( + vertex_project="proj-db", + vertex_location="global", + llm_router=llm_router, + ), + ) + + upstream_ws.send.assert_awaited_once() + return upstream_ws.send.await_args.args[0] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "setup_model", + ["gemini-live-2.5-flash", "publishers/google/models/gemini-live-2.5-flash"], +) +async def test_websocket_passthrough_rewrites_setup_model_to_full_resource(setup_model): + sent_frame = await _run_setup_rewrite_passthrough(setup_model, llm_router=None) + + sent_setup = json.loads(sent_frame)["setup"] + assert sent_setup["model"] == "projects/proj-db/locations/global/publishers/google/models/gemini-live-2.5-flash" + assert sent_setup["generationConfig"] == {"responseModalities": ["TEXT"]} + + +@pytest.mark.asyncio +async def test_websocket_passthrough_rewrites_gateway_alias_setup_model(): + llm_router = litellm.Router( + model_list=[ + { + "model_name": "gemini-live", + "litellm_params": {"model": "vertex_ai/gemini-live-2.5-flash"}, + } + ] + ) + + sent_frame = await _run_setup_rewrite_passthrough("gemini-live", llm_router=llm_router) + + sent_setup = json.loads(sent_frame)["setup"] + assert sent_setup["model"] == "projects/proj-db/locations/global/publishers/google/models/gemini-live-2.5-flash" + + +@pytest.mark.asyncio +async def test_websocket_passthrough_leaves_full_resource_setup_model_untouched(): + full_resource = "projects/other/locations/us-central1/publishers/google/models/gemini-2.0-flash-live-preview-04-09" + + sent_frame = await _run_setup_rewrite_passthrough(full_resource, llm_router=None) + + assert json.loads(sent_frame)["setup"]["model"] == full_resource + assert sent_frame == json.dumps( + {"setup": {"model": full_resource, "generationConfig": {"responseModalities": ["TEXT"]}}} + ) + + def _passthrough_kwargs_for_reservation( user_api_key_dict: UserAPIKeyAuth, parsed_body: Optional[dict] = None ) -> dict: diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py index d7266ecd9ed..7816178471f 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py @@ -172,3 +172,144 @@ def test_returns_none_when_no_router_and_no_env(): passthrough_router = _passthrough_router(None) assert passthrough_router.get_credentials(custom_llm_provider="openai", region_name=None) is None + + +def _vertex_credential(name: str, values: dict) -> CredentialItem: + return CredentialItem(credential_name=name, credential_values=values, credential_info={}) + + +def _vertex_deployment(model_name: str, model: str, **litellm_params) -> dict: + return { + "model_name": model_name, + "litellm_params": {"model": model, "use_in_pass_through": True, **litellm_params}, + } + + +def test_vertex_deployment_resolves_via_named_credential(): + CredentialAccessor.upsert_credentials( + [ + _vertex_credential( + "cred_gcp", + { + "vertex_project": "proj-db", + "vertex_location": "global", + "vertex_credentials": '{"type": "service_account"}', + }, + ) + ] + ) + llm_router = litellm.Router( + model_list=[ + _vertex_deployment( + "gemini-live", "vertex_ai/gemini-live-2.5-flash", litellm_credential_name="cred_gcp" + ) + ] + ) + passthrough_router = _passthrough_router(llm_router) + + resolved = passthrough_router.get_vertex_credentials_from_router_deployments(model=None) + + assert resolved is not None + assert resolved.vertex_project == "proj-db" + assert resolved.vertex_location == "global" + assert resolved.vertex_credentials == '{"type": "service_account"}' + + +def test_vertex_deployment_resolves_from_inline_litellm_params(): + llm_router = litellm.Router( + model_list=[ + _vertex_deployment( + "gemini-live", + "vertex_ai/gemini-live-2.5-flash", + vertex_project="proj-inline", + vertex_location="us-east4", + vertex_credentials='{"type": "service_account", "project_id": "proj-inline"}', + ) + ] + ) + passthrough_router = _passthrough_router(llm_router) + + resolved = passthrough_router.get_vertex_credentials_from_router_deployments(model=None) + + assert resolved is not None + assert resolved.vertex_project == "proj-inline" + assert resolved.vertex_location == "us-east4" + assert resolved.vertex_credentials == '{"type": "service_account", "project_id": "proj-inline"}' + + +def _two_vertex_deployments_router() -> litellm.Router: + return litellm.Router( + model_list=[ + _vertex_deployment( + "gemini-flash", + "vertex_ai/gemini-2.5-flash", + vertex_project="proj-first", + vertex_location="us-central1", + ), + _vertex_deployment( + "gemini-live", + "vertex_ai/gemini-live-2.5-flash", + vertex_project="proj-live", + vertex_location="global", + ), + ] + ) + + +def test_vertex_model_hint_prefers_matching_deployment(): + passthrough_router = _passthrough_router(_two_vertex_deployments_router()) + + by_alias = passthrough_router.get_vertex_credentials_from_router_deployments(model="gemini-live") + by_upstream_id = passthrough_router.get_vertex_credentials_from_router_deployments( + model="gemini-live-2.5-flash" + ) + + assert by_alias is not None and by_alias.vertex_project == "proj-live" + assert by_upstream_id is not None and by_upstream_id.vertex_project == "proj-live" + + +def test_vertex_unmatched_hint_falls_back_to_first_flagged_deployment(): + passthrough_router = _passthrough_router(_two_vertex_deployments_router()) + + unmatched = passthrough_router.get_vertex_credentials_from_router_deployments(model="unknown-model") + no_hint = passthrough_router.get_vertex_credentials_from_router_deployments(model=None) + + assert unmatched is not None and unmatched.vertex_project == "proj-first" + assert no_hint is not None and no_hint.vertex_project == "proj-first" + + +def test_no_flagged_vertex_deployment_returns_none(): + llm_router = litellm.Router( + model_list=[ + { + "model_name": "gemini-live", + "litellm_params": { + "model": "vertex_ai/gemini-live-2.5-flash", + "vertex_project": "proj-unflagged", + "vertex_location": "global", + }, + }, + _flagged_deployment("openai/gpt-4o", api_key="sk-flagged"), + ] + ) + passthrough_router = _passthrough_router(llm_router) + + assert passthrough_router.get_vertex_credentials_from_router_deployments(model=None) is None + assert _passthrough_router(None).get_vertex_credentials_from_router_deployments(model=None) is None + + +def test_vertex_deployment_with_deleted_credential_is_skipped(monkeypatch): + CredentialAccessor.upsert_credentials( + [_vertex_credential("cred_gone", {"vertex_project": "proj-db", "vertex_location": "global"})] + ) + llm_router = litellm.Router( + model_list=[ + _vertex_deployment( + "gemini-live", "vertex_ai/gemini-live-2.5-flash", litellm_credential_name="cred_gone" + ) + ] + ) + passthrough_router = _passthrough_router(llm_router) + monkeypatch.setattr(litellm, "credential_list", []) + + assert passthrough_router.get_vertex_credentials_from_router_deployments(model=None) is None From 9271133bebedd2ece0fe23535fd783d69cae2547 Mon Sep 17 00:00:00 2001 From: Mateo Date: Thu, 20 Aug 2026 02:02:00 -0700 Subject: [PATCH 032/220] fix(realtime): bound Vertex credential resolution and make realtime failures loud A /v1/realtime connection to a Vertex AI Live model accepted the WebSocket upgrade and then went silent: a stalled Google OAuth token fetch blocked the handler before any session event, and the eventual failure closed the socket with a bare 1011 and no error event, so callers saw an open socket, no frames, and no reason. Bound the pre-session token fetch with REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS (20s default) and, on any realtime failure, send an OpenAI-style error event before closing with a reason that names the failure. Close reasons are truncated by bytes, not characters, since an over-long reason makes the close frame itself fail. --- litellm/constants.py | 3 + litellm/litellm_core_utils/realtime_errors.py | 31 +++++ litellm/llms/custom_httpx/llm_http_handler.py | 14 ++- litellm/proxy/proxy_server.py | 21 +++- litellm/realtime_api/main.py | 35 +++++- litellm/types/realtime.py | 12 +- .../test_realtime_errors.py | 47 ++++++++ .../custom_httpx/test_llm_http_handler.py | 68 +++++++++++ .../test_realtime_webrtc_endpoints.py | 108 ++++++++++++++++++ tests/test_litellm/realtime_api/test_main.py | 63 ++++++++++ 10 files changed, 392 insertions(+), 10 deletions(-) create mode 100644 litellm/litellm_core_utils/realtime_errors.py create mode 100644 tests/test_litellm/litellm_core_utils/test_realtime_errors.py diff --git a/litellm/constants.py b/litellm/constants.py index facfc6f7c19..eff82bc268b 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -243,6 +243,9 @@ AIOHTTP_NEEDS_CLEANUP_CLOSED: Final = (3, 13, 0) <= sys.version_info < ( # https://github.com/openai/openai-agents-python/blob/cf1b933660e44fd37b4350c41febab8221801409/src/agents/realtime/openai_realtime.py#L235 _max_size_env: Final = os.getenv("REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES") REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES: Final = int(_max_size_env) if _max_size_env is not None else None +REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS: Final = float( + os.getenv("REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS", "20.0") +) # SSL/TLS cipher configuration for faster handshakes # Strategy: Strongly prefer fast modern ciphers, but allow fallback to commonly supported ones diff --git a/litellm/litellm_core_utils/realtime_errors.py b/litellm/litellm_core_utils/realtime_errors.py new file mode 100644 index 00000000000..e1b957f4325 --- /dev/null +++ b/litellm/litellm_core_utils/realtime_errors.py @@ -0,0 +1,31 @@ +"""Loud-failure helpers for the realtime WebSocket paths. + +A realtime caller that only gets a bare close frame has nothing to act on, so +every failure surfaces as an OpenAI-style ``error`` event plus a close frame +whose reason names the failure. Close reasons are capped at +``WEBSOCKET_CLOSE_REASON_MAX_BYTES``: RFC 6455 control frames carry at most 125 +bytes, two of which hold the status code, and a longer reason makes the close +frame itself fail, which is how a loud failure turns back into a silent one. +""" + +import json +from typing import Final + +from litellm.types.realtime import RealtimeErrorDetail, RealtimeErrorEvent + +WEBSOCKET_CLOSE_REASON_MAX_BYTES: Final = 123 + + +def realtime_error_event(message: str, error_type: str) -> str: + detail: Final[RealtimeErrorDetail] = {"type": error_type, "message": message} + event: Final[RealtimeErrorEvent] = {"type": "error", "error": detail} + return json.dumps(event) + + +def websocket_close_reason(message: str, fallback: str) -> str: + encoded: Final = message.encode("utf-8") + if not encoded: + return fallback + if len(encoded) <= WEBSOCKET_CLOSE_REASON_MAX_BYTES: + return message + return encoded[:WEBSOCKET_CLOSE_REASON_MAX_BYTES].decode("utf-8", errors="ignore") diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index cc522aed1ee..9a950d7f920 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -20,6 +20,7 @@ from litellm._logging import _redact_string, verbose_logger from litellm.anthropic_beta_headers_manager import update_headers_with_filtered_beta from litellm.constants import REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES from litellm.litellm_core_utils.asyncify import run_async_function +from litellm.litellm_core_utils.realtime_errors import realtime_error_event, websocket_close_reason from litellm.litellm_core_utils.realtime_streaming import RealTimeStreaming from litellm.litellm_core_utils.url_utils import encode_url_path_segment from litellm.llms.base_llm.anthropic_messages.transformation import ( @@ -5976,8 +5977,19 @@ class BaseLLMHTTPHandler: await websocket.close(code=e.status_code, reason=_redact_string(str(e))) except Exception as e: verbose_logger.exception("Error connecting to backend: %s", e) + redacted_error: Final = _redact_string(str(e)) try: - await websocket.close(code=1011, reason=_redact_string(f"Internal server error: {e}")) + await websocket.send_text(realtime_error_event(redacted_error, error_type="server_error")) + except Exception: # noqa: BLE001 # best-effort notice: a dead client socket must not skip the close below + verbose_logger.debug("Could not send realtime error event to client; closing anyway") + try: + await websocket.close( + code=1011, + reason=websocket_close_reason( + _redact_string(f"Internal server error: {e}"), + fallback="Internal server error", + ), + ) except RuntimeError as close_error: if "already completed" in str(close_error) or "websocket.close" in str(close_error): # The WebSocket is already closed or the response is completed, so we can ignore this error diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index af082f04706..4a342174277 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -222,7 +222,7 @@ from functools import lru_cache import litellm import litellm._redis from litellm import Router -from litellm._logging import verbose_proxy_logger, verbose_router_logger +from litellm._logging import _redact_string, verbose_proxy_logger, verbose_router_logger from litellm.caching.caching import DualCache, RedisCache from litellm.caching.redis_cluster_cache import RedisClusterCache from litellm.constants import ( @@ -259,6 +259,10 @@ from litellm.litellm_core_utils.core_helpers import ( ) from litellm.litellm_core_utils.credential_accessor import CredentialAccessor from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.litellm_core_utils.realtime_errors import ( + realtime_error_event, + websocket_close_reason, +) from litellm.litellm_core_utils.sensitive_data_masker import ( SensitiveDataMasker, mask_sensitive_keys, @@ -10993,9 +10997,20 @@ async def realtime_websocket_endpoint( except websockets.exceptions.InvalidStatusCode as e: verbose_proxy_logger.exception("Invalid status code") await websocket.close(code=e.status_code, reason="Invalid status code") - except Exception: + except Exception as e: verbose_proxy_logger.exception("Internal server error") - await websocket.close(code=1011, reason="Internal server error") + redacted_error: Final = _redact_string(str(e)) + try: + await websocket.send_text(realtime_error_event(redacted_error, error_type="server_error")) + except Exception: # noqa: BLE001 # best-effort notice: a dead client socket must not skip the close below + verbose_proxy_logger.debug("Could not send realtime error event to client; closing anyway") + try: + await websocket.close( + code=1011, + reason=websocket_close_reason(redacted_error, fallback="Internal server error"), + ) + except Exception: # noqa: BLE001 # the lower layer may have closed the socket already; closing twice is not an error + verbose_proxy_logger.debug("Could not close realtime client websocket; it is already gone") ###################################################################### diff --git a/litellm/realtime_api/main.py b/litellm/realtime_api/main.py index d5195659b1c..8fde7cb75c5 100644 --- a/litellm/realtime_api/main.py +++ b/litellm/realtime_api/main.py @@ -1,15 +1,21 @@ """Abstraction function for OpenAI's realtime API""" +import asyncio import os from typing import Any, Final, cast import litellm -from litellm.constants import REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES, request_timeout +from litellm.constants import ( + REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS, + REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES, + request_timeout, +) from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler from litellm.llms.xai.common_utils import XAIModelInfo from litellm.secret_managers.main import get_secret_str +from litellm.types.llms.vertex_ai import VERTEX_CREDENTIALS_TYPES from litellm.types.realtime import ( RealtimeClientSecretRequest, RealtimeExpiresAfter, @@ -281,6 +287,27 @@ async def arealtime_calls( ) +async def _resolve_vertex_access_token_bounded( + credentials: VERTEX_CREDENTIALS_TYPES | None, + project_id: str | None, +) -> tuple[str, str]: + try: + return await asyncio.wait_for( + vertex_llm_base._ensure_access_token_async( + credentials=credentials, + project_id=project_id, + custom_llm_provider="vertex_ai", + ), + timeout=REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS, + ) + except asyncio.TimeoutError as e: + raise ValueError( + "Vertex AI realtime: timed out fetching Google OAuth access token after " + f"{REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS}s; check network egress from the proxy " + "to the OAuth token endpoint (oauth2.googleapis.com)" + ) from e + + @wrapper_client async def _arealtime( model: str, @@ -478,10 +505,9 @@ async def _arealtime( ( access_token, resolved_project, - ) = await vertex_llm_base._ensure_access_token_async( + ) = await _resolve_vertex_access_token_bounded( credentials=vertex_credentials, project_id=vertex_project, - custom_llm_provider="vertex_ai", ) vertex_realtime_config: Final = VertexAIRealtimeConfig( @@ -559,10 +585,9 @@ async def _realtime_health_check( ( access_token, resolved_project, - ) = await vertex_llm_base._ensure_access_token_async( + ) = await _resolve_vertex_access_token_bounded( credentials=VertexBase.safe_get_vertex_ai_credentials(vertex_model_params), project_id=VertexBase.safe_get_vertex_ai_project(vertex_model_params), - custom_llm_provider="vertex_ai", ) vertex_realtime_config: Final = VertexAIRealtimeConfig( access_token=access_token, diff --git a/litellm/types/realtime.py b/litellm/types/realtime.py index 15238f7e13f..cbd7a8b7ecb 100644 --- a/litellm/types/realtime.py +++ b/litellm/types/realtime.py @@ -1,7 +1,7 @@ from typing import Any, Literal from pydantic import BaseModel -from typing_extensions import TypedDict +from typing_extensions import ReadOnly, TypedDict from .llms.openai import ( OpenAIRealtimeEvents, @@ -152,3 +152,13 @@ class RealtimeTranscriptionSessionResponse(BaseModel): model_config = {"extra": "allow"} client_secret: dict[str, Any] | None = None + + +class RealtimeErrorDetail(TypedDict): + type: ReadOnly[str] + message: ReadOnly[str] + + +class RealtimeErrorEvent(TypedDict): + type: ReadOnly[Literal["error"]] + error: ReadOnly[RealtimeErrorDetail] diff --git a/tests/test_litellm/litellm_core_utils/test_realtime_errors.py b/tests/test_litellm/litellm_core_utils/test_realtime_errors.py new file mode 100644 index 00000000000..263d1654f65 --- /dev/null +++ b/tests/test_litellm/litellm_core_utils/test_realtime_errors.py @@ -0,0 +1,47 @@ +import json +import os +import sys + +sys.path.insert(0, os.path.abspath("../../..")) + +from litellm.litellm_core_utils.realtime_errors import ( + WEBSOCKET_CLOSE_REASON_MAX_BYTES, + realtime_error_event, + websocket_close_reason, +) + + +def test_realtime_error_event_shape(): + event = json.loads(realtime_error_event("token refresh failed", error_type="server_error")) + + assert event == { + "type": "error", + "error": {"type": "server_error", "message": "token refresh failed"}, + } + + +def test_websocket_close_reason_keeps_short_messages_intact(): + assert websocket_close_reason("boom", fallback="Internal server error") == "boom" + + +def test_websocket_close_reason_falls_back_on_empty_message(): + assert websocket_close_reason("", fallback="Internal server error") == "Internal server error" + + +def test_websocket_close_reason_truncates_long_ascii_message(): + reason = websocket_close_reason("x" * 500, fallback="Internal server error") + + assert len(reason.encode("utf-8")) <= WEBSOCKET_CLOSE_REASON_MAX_BYTES + assert reason == "x" * WEBSOCKET_CLOSE_REASON_MAX_BYTES + + +def test_websocket_close_reason_truncates_multibyte_message_by_bytes(): + """A close frame carries at most 123 bytes of reason, not 123 characters: + truncating by characters lets a multibyte message overflow the control + frame, which makes the close itself fail and leaves the caller with a bare + abnormal closure and no reason at all.""" + reason = websocket_close_reason("あ" * 200, fallback="Internal server error") + + assert len(reason.encode("utf-8")) <= WEBSOCKET_CLOSE_REASON_MAX_BYTES + assert reason == "あ" * (WEBSOCKET_CLOSE_REASON_MAX_BYTES // 3) + assert "�" not in reason diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py index 9e9242137e6..c568b82ebba 100644 --- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py +++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py @@ -1738,6 +1738,74 @@ async def test_realtime_backend_open_does_not_retry_auth_failure(rejection): assert fake.attempts == 1 +class _FakeClientWebSocket: + def __init__(self, send_error=None): + self.events = [] + self._send_error = send_error + + async def send_text(self, payload): + if self._send_error is not None: + raise self._send_error + self.events.append(("send_text", payload)) + + async def close(self, code=None, reason=None): + self.events.append(("close", (code, reason))) + + +async def _run_async_realtime_with_backend_failure(client_ws): + import websockets.exceptions # noqa: F401 # binds the submodule so async_realtime's except clause resolves, as in the proxy process + + handler = BaseLLMHTTPHandler() + provider_config = Mock() + provider_config.get_complete_url.return_value = "wss://backend.example/live" + provider_config.validate_environment.return_value = {} + + with patch.object( + handler, + "_open_realtime_backend_ws", + AsyncMock(side_effect=Exception("vertex token refresh exploded")), + ): + await handler.async_realtime( + model="gemini-live-2.5-flash", + websocket=client_ws, + logging_obj=Mock(), + provider_config=provider_config, + headers={}, + ) + + +@pytest.mark.asyncio +async def test_async_realtime_generic_failure_sends_error_event_then_reasoned_close(): + """Regression for the realtime accept-then-silence hang: a generic backend + failure used to close the client socket without any error event, so callers + only saw a bare 1011. The client must receive an OpenAI-style error event + before the reasoned close.""" + client_ws = _FakeClientWebSocket() + + await _run_async_realtime_with_backend_failure(client_ws) + + assert [name for name, _ in client_ws.events] == ["send_text", "close"] + + error_event = json.loads(client_ws.events[0][1]) + assert error_event["type"] == "error" + assert error_event["error"]["type"] == "server_error" + assert "vertex token refresh exploded" in error_event["error"]["message"] + + assert client_ws.events[1][1] == (1011, "Internal server error: vertex token refresh exploded") + + +@pytest.mark.asyncio +async def test_async_realtime_error_event_send_failure_still_closes(): + """A client socket that already dropped must not turn the loud-failure path + into a new exception: the error-event send may fail, but the reasoned close + must still be attempted.""" + client_ws = _FakeClientWebSocket(send_error=RuntimeError("client already disconnected")) + + await _run_async_realtime_with_backend_failure(client_ws) + + assert client_ws.events == [("close", (1011, "Internal server error: vertex token refresh exploded"))] + + class _JSONBodyAudioTranscriptionConfig(BaseAudioTranscriptionConfig): def get_supported_openai_params(self, model): return [] diff --git a/tests/test_litellm/proxy/realtime_endpoints/test_realtime_webrtc_endpoints.py b/tests/test_litellm/proxy/realtime_endpoints/test_realtime_webrtc_endpoints.py index e0e51e7b966..9840de8bcb1 100644 --- a/tests/test_litellm/proxy/realtime_endpoints/test_realtime_webrtc_endpoints.py +++ b/tests/test_litellm/proxy/realtime_endpoints/test_realtime_webrtc_endpoints.py @@ -818,6 +818,114 @@ async def test_realtime_transcription_websocket_default_model_checks_team_scope( assert "not allowed to access model" in close_kwargs["reason"] +@pytest.mark.asyncio +async def test_realtime_websocket_phase2_failure_sends_error_event_and_reasoned_close(): + """Regression for the realtime accept-then-silence hang: a phase-2 failure + (routing / upstream credential resolution) used to close 1011 with the bare + reason "Internal server error" and no error event, leaving the client with + no clue what happened. The client must get an OpenAI-style error event and + a close reason naming the failure.""" + from litellm.proxy import proxy_server + + events = [] + + websocket = MagicMock() + websocket.headers = {} + websocket.scope = {"headers": []} + websocket.accept = AsyncMock() + websocket.send_text = AsyncMock(side_effect=lambda payload: events.append(("send_text", payload))) + websocket.close = AsyncMock(side_effect=lambda **kwargs: events.append(("close", kwargs))) + + mock_processor = MagicMock() + mock_processor.common_processing_pre_call_logic = AsyncMock( + return_value=({"model": "gpt-4o-realtime-preview"}, MagicMock()) + ) + + with ( + patch( + "litellm.proxy.proxy_server.can_key_call_resolved_model", + new=AsyncMock(return_value=None), + ), + patch( + "litellm.proxy.proxy_server.ProxyBaseLLMRequestProcessing", + return_value=mock_processor, + ), + patch( + "litellm.proxy.proxy_server.route_request", + new=AsyncMock(side_effect=RuntimeError("vertex token refresh exploded")), + ), + ): + await proxy_server.realtime_websocket_endpoint( + websocket=websocket, + model="gpt-4o-realtime-preview", + intent=None, + guardrails=None, + user_api_key_dict=UserAPIKeyAuth(models=["*"]), + ) + + websocket.accept.assert_awaited_once() + assert [name for name, _ in events] == ["send_text", "close"] + + error_event = json.loads(events[0][1]) + assert error_event["type"] == "error" + assert error_event["error"]["type"] == "server_error" + assert "vertex token refresh exploded" in error_event["error"]["message"] + + close_kwargs = events[1][1] + assert close_kwargs["code"] == 1011 + assert "vertex token refresh exploded" in close_kwargs["reason"] + assert len(close_kwargs["reason"].encode("utf-8")) <= 123 + + +@pytest.mark.asyncio +async def test_realtime_websocket_phase2_failure_on_closed_socket_does_not_escape(): + """The lower handler layer may have already closed the client socket before + the phase-2 handler runs (it closes on backend failures itself, then can + re-raise). Send and close must each be guarded: the close is still + attempted after a failed send, and neither failure escapes to the ASGI + layer.""" + from litellm.proxy import proxy_server + + websocket = MagicMock() + websocket.headers = {} + websocket.scope = {"headers": []} + websocket.accept = AsyncMock() + websocket.send_text = AsyncMock(side_effect=RuntimeError('Cannot call "send" once a close message has been sent.')) + websocket.close = AsyncMock(side_effect=RuntimeError('Cannot call "send" once a close message has been sent.')) + + mock_processor = MagicMock() + mock_processor.common_processing_pre_call_logic = AsyncMock( + return_value=({"model": "gpt-4o-realtime-preview"}, MagicMock()) + ) + + with ( + patch( + "litellm.proxy.proxy_server.can_key_call_resolved_model", + new=AsyncMock(return_value=None), + ), + patch( + "litellm.proxy.proxy_server.ProxyBaseLLMRequestProcessing", + return_value=mock_processor, + ), + patch( + "litellm.proxy.proxy_server.route_request", + new=AsyncMock(side_effect=RuntimeError("vertex token refresh exploded")), + ), + ): + await proxy_server.realtime_websocket_endpoint( + websocket=websocket, + model="gpt-4o-realtime-preview", + intent=None, + guardrails=None, + user_api_key_dict=UserAPIKeyAuth(models=["*"]), + ) + + websocket.close.assert_awaited_once() + _, close_kwargs = websocket.close.call_args + assert close_kwargs["code"] == 1011 + assert "vertex token refresh exploded" in close_kwargs["reason"] + + @pytest.mark.asyncio async def test_transcription_sessions_encrypts_client_secret( proxy_app, diff --git a/tests/test_litellm/realtime_api/test_main.py b/tests/test_litellm/realtime_api/test_main.py index 406f5ef56d9..8ed7fb06e84 100644 --- a/tests/test_litellm/realtime_api/test_main.py +++ b/tests/test_litellm/realtime_api/test_main.py @@ -1,6 +1,8 @@ import asyncio import os import sys +import time +from unittest.mock import MagicMock sys.path.insert(0, os.path.abspath("../../..")) @@ -91,6 +93,67 @@ def test_client_secret_session_model_takes_priority_over_top_level(monkeypatch): assert captured["request_data"]["session"]["model"] == "gpt-realtime-session" +@pytest.mark.asyncio +async def test_arealtime_vertex_hung_credential_resolution_raises_promptly(monkeypatch): + """Regression for the realtime accept-then-silence hang: a stalled Google + OAuth token refresh used to block _arealtime's vertex branch unbounded + (minutes of zero frames for the client). It must instead raise a clear, + prompt error naming the credential-resolution timeout.""" + + async def hanging_token_refresh(**kwargs): + await asyncio.sleep(30) + + def mock_get_llm_provider(model, api_base, api_key): + return model, "vertex_ai", None, api_base + + monkeypatch.setattr(realtime_main, "get_llm_provider", mock_get_llm_provider) + monkeypatch.setattr(realtime_main.vertex_llm_base, "_ensure_access_token_async", hanging_token_refresh) + monkeypatch.setattr(realtime_main, "REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS", 0.05) + + start = time.monotonic() + with pytest.raises(ValueError, match="timed out fetching Google OAuth access token"): + await realtime_main._arealtime.__wrapped__( + model="gemini-live-2.5-flash", + websocket=MagicMock(), + litellm_logging_obj=FakeLogging(), + vertex_credentials="fake-credentials", + vertex_project="fake-project", + vertex_location="us-central1", + ) + assert time.monotonic() - start < 5 + + +@pytest.mark.asyncio +async def test_arealtime_vertex_credential_timeout_survives_thread_offloaded_refresh(monkeypatch): + """The real stall is a blocking google-auth refresh that runs in a worker + thread via asyncify, not a plain awaitable sleep. A timeout that only bounds + cancellable awaits would leave that shape hanging, so bound the shape the + proxy actually runs.""" + from litellm.litellm_core_utils.asyncify import asyncify + + async def thread_offloaded_hanging_refresh(**kwargs): + return await asyncify(time.sleep)(30) + + def mock_get_llm_provider(model, api_base, api_key): + return model, "vertex_ai", None, api_base + + monkeypatch.setattr(realtime_main, "get_llm_provider", mock_get_llm_provider) + monkeypatch.setattr(realtime_main.vertex_llm_base, "_ensure_access_token_async", thread_offloaded_hanging_refresh) + monkeypatch.setattr(realtime_main, "REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS", 0.05) + + start = time.monotonic() + with pytest.raises(ValueError, match="timed out fetching Google OAuth access token"): + await realtime_main._arealtime.__wrapped__( + model="gemini-live-2.5-flash", + websocket=MagicMock(), + litellm_logging_obj=FakeLogging(), + vertex_credentials="fake-credentials", + vertex_project="fake-project", + vertex_location="us-central1", + ) + assert time.monotonic() - start < 5 + + def test_client_secret_forwards_nested_transcription_model_untouched(monkeypatch): captured = _run_client_secret( session={ From ef6af5c615815592da85e26315b6212046ba82ba Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:03:27 -0700 Subject: [PATCH 033/220] test(e2e): accept both model-not-found phrasings on a shared proxy /audio/transcriptions answers a model-less request with one of two 400s depending on whether any wildcard deployment is registered at the time, and every suite shares one proxy, so run order decided which message came back. The assertion pinned only the no-wildcard wording, so it went red whenever the model-access-group suite had registered its wildcards first. It now accepts either message and still holds the error to naming the model Verified against a live proxy in both states: with a wildcard registered (the message CI was seeing) and with none (the message the assertion expected), the suite passes 3/3 either way --- .../llm_translation/test_audio_transcriptions_e2e.py | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-) diff --git a/tests/e2e/llm_translation/test_audio_transcriptions_e2e.py b/tests/e2e/llm_translation/test_audio_transcriptions_e2e.py index 92b33fef85f..735f1a4a703 100644 --- a/tests/e2e/llm_translation/test_audio_transcriptions_e2e.py +++ b/tests/e2e/llm_translation/test_audio_transcriptions_e2e.py @@ -3,7 +3,10 @@ Registers an OpenAI speech-to-text deployment at runtime and uploads a spoken weather question (the realtime suite's 24kHz WAV fixture) as multipart, asserting the returned transcript is non-empty and mentions the word it was asked about. -Also pins missing file/model negatives. +Also pins missing file/model negatives. A model-less request comes back as one of +two 400s depending on whether any wildcard deployment happens to be registered on +the shared proxy, so the assertion accepts either phrasing and holds both to naming +the model as the problem. """ from __future__ import annotations @@ -25,6 +28,8 @@ WEATHER_WAV = ( Path(__file__).resolve().parent / "realtime" / "fixtures" / "weather_question_24k.wav" ) +MISSING_MODEL_PHRASES: Final = ("model=none", "invalid model", "model is required") + class _OptionalTranscriptionForm(BaseModel): model: str | None = None @@ -105,8 +110,8 @@ class TestAudioTranscriptions: match result: case UnknownApiError(status_code=400, body=body): lowered: Final = body.lower() - assert "model" in lowered and ("required" in lowered or "invalid model" in lowered), ( - f"missing model error must identify the required model: {body[:300]}" + assert any(phrase in lowered for phrase in MISSING_MODEL_PHRASES), ( + f"missing model error must name the model as the problem: {body[:300]}" ) case other: pytest.fail(f"missing model expected a model-specific 400, got {other!r}") From b6fef179ff151e2cb88990ed24fe2613a1824704 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:08:04 -0700 Subject: [PATCH 034/220] fix(cli): stop a repeat logout from retracting its own keychain warning A logout that could not reach the keychain deleted the token file whenever it still held its own secret, and the next logout read that missing file as proof the keychain was clean. It answered the warning the first run had just issued with "Logged out successfully" while the entry an earlier login left behind was still live. The file is the only record that something may still be in there, which is what `_nothing_left_behind` already says it relies on, so keep it and take only the secret out. A keychain that did answer is a different case. `SecretStranded` means the entry is confirmed there and would not delete, and that needs no note in the file, while keeping one lets every later command read the credential straight back out of the keychain, which makes "Logged out locally" untrue. That one drops the file, as it did before. The secret still goes first either way: a copy that cannot be replaced with a secret-free one is removed rather than kept. --- litellm/litellm_core_utils/cli_token_utils.py | 24 ++++++++++--- .../test_cli_token_utils.py | 34 +++++++++++++++++-- .../proxy/client/cli/test_auth_commands.py | 2 +- 3 files changed, 52 insertions(+), 8 deletions(-) diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 15b1390f337..4e01dd723ce 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -153,19 +153,33 @@ def _keep_the_secret_in_the_file(record: CliTokenRecord, outcome: SecretWrite) - def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretErase: """Remove the credential from both stores. Reports whether the keychain is now free of it. - A file that holds no secret of its own is kept when the keychain will not confirm the entry is - gone, because it is the only remaining record that something is still in there to remove. That - is what lets a later run tell a machine with a credential it cannot reach apart from one that - never had a login at all. Anything still holding a secret is removed either way. + A logout the keychain never answered keeps the token file, with its secret taken out, because + that file is the only remaining record that something may still be in there to remove. It is + what lets a later run tell a machine with a credential it cannot reach apart from one that never + had a login at all, and taking it away would leave the next logout answering the warning this + one just issued with a false all-clear. The secret goes either way. """ outcome: Final = vault.erase() record: Final = _read_token_file() settled: Final = _nothing_left_behind(outcome, record) - if settled or record is None or record.key is not None: + if settled or not _keep_the_unchecked_keychain_on_record(outcome, record): Path(get_cli_token_file_path()).unlink(missing_ok=True) return SecretErased() if settled else outcome +def _keep_the_unchecked_keychain_on_record(outcome: SecretErase, record: CliTokenRecord | None) -> bool: + """Whether the token file, stripped of its secret, is worth keeping as the note that says so. + + Only a keychain that could not be reached leaves the question open. One that answered for itself + is remembered without any help from the file, and a file it can still pair a live entry with + would leave the machine signed in to the login that was just ended. A copy that cannot be + replaced with a secret-free one is not kept either, because the secret goes first. + """ + if record is None or isinstance(outcome, SecretStranded): + return False + return _scrub_file_secret(record) + + def _nothing_left_behind(outcome: SecretErase, record: CliTokenRecord | None) -> bool: """Whether the keychain can be trusted to hold no credential of ours once the file is gone. diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 0cf1e55d363..162e5dd4b67 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -477,6 +477,18 @@ class TestClearCliToken: assert clear_cli_token(vault=vault) == SecretStranded() assert not _token_file(isolated_home).exists() + def test_a_keychain_that_will_not_release_the_secret_still_ends_the_local_login( + self, isolated_home, secret_vault_factory + ): + """The warning this returns says the machine is logged out locally and the keychain entry is + what is left over. Keeping the file that names that entry makes the first half untrue: every + later command reads the credential straight back out of the keychain and keeps working.""" + _write_metadata_only_file(isolated_home) + vault = secret_vault_factory(blob=_blob(), erasable=False) + + assert clear_cli_token(vault=vault) == SecretStranded() + assert load_cli_token(vault=vault) is None + @pytest.mark.parametrize( "failure", [KeyringDisabled(), KeyringUnreachable(), KeyringDiscardsWrites()] ) @@ -491,7 +503,7 @@ class TestClearCliToken: vault = secret_vault_factory(available=False, failure=failure) assert clear_cli_token(vault=vault) == failure - assert not _token_file(isolated_home).exists() + assert json.loads(_token_file(isolated_home).read_text()).get("key") is None def test_a_second_logout_still_reports_the_keychain_it_could_not_clear( self, isolated_home, secret_vault_factory @@ -539,7 +551,25 @@ class TestClearCliToken: clear_cli_token(vault=vault) - assert not _token_file(isolated_home).exists() + assert "sk-legacy" not in _token_file(isolated_home).read_text() + + def test_a_repeat_logout_never_answers_its_own_warning_with_an_all_clear( + self, isolated_home, secret_vault_factory + ): + """Sign in while the keychain works, sign in again once it has gone out of reach so the + second secret lands in the file, then log out twice. The first logout cannot say the first + login's entry is gone, and says so. If the second one reads the file the first one took + away as proof of a clean keychain, it retracts that warning while the credential behind it + is still live.""" + vault = secret_vault_factory() + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-first"), vault=vault) + vault.available = False + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-second"), vault=vault) + + assert clear_cli_token(vault=vault) == KeyringUnreachable() + assert clear_cli_token(vault=vault) == KeyringUnreachable() + assert vault.blob is not None + assert "sk-second" not in _token_file(isolated_home).read_text() @pytest.mark.parametrize("failure", [KeyringNotInstalled(), KeyringDisabled(), KeyringUnreachable()]) def test_logging_out_of_a_machine_that_never_logged_in_invents_nothing_to_warn_about( diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index e491dbd5aca..8e1551c0720 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -1074,7 +1074,7 @@ class TestKeychainBackedCommands: assert result.exit_code == 0 assert "could not be removed" in result.output - assert json.loads((isolated_home / ".litellm" / "token.json").read_text()).get("key") is None + assert not (isolated_home / ".litellm" / "token.json").exists() def test_print_token_explains_a_locked_keychain_instead_of_printing_nothing( self, isolated_home, secret_vault_factory From b9d977aeeefc94e249b0ea63106ee81c399ecf1a Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:15:17 -0700 Subject: [PATCH 035/220] fix: guard vertex live passthrough provider lookup and close-code relay --- .../llm_passthrough_endpoints.py | 9 ++-- .../pass_through_endpoints.py | 9 +++- .../test_pass_through_endpoints.py | 43 +++++++++++++++++++ 3 files changed, 56 insertions(+), 5 deletions(-) diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index 050ad0fd627..c2b221b1f7a 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -2397,8 +2397,8 @@ def _resolve_vertex_live_credentials( model: str | None, ) -> VertexPassThroughCredentials | None: """ - Resolution order: credentials registered for the requested project/location, then any DB model entry - flagged ``use_in_pass_through``, then ``default_vertex_config`` and the ``DEFAULT_VERTEXAI_*`` env vars + Resolution order: an explicit project/location registration or ``default_vertex_config``, then any DB model + entry flagged ``use_in_pass_through``, then the ``DEFAULT_VERTEXAI_*`` env vars """ keyed: Final = passthrough_endpoint_router.get_vertex_credentials( project_id=vertex_project, @@ -2457,7 +2457,10 @@ def _resolve_alias_to_upstream_model(setup_model: str, llm_router: "Router | Non ) if upstream is None: return setup_model - _, provider, _, _ = litellm.get_llm_provider(model=upstream) + try: + _, provider, _, _ = litellm.get_llm_provider(model=upstream) + except litellm.exceptions.BadRequestError: + return upstream return upstream.removeprefix(f"{provider}/") diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py index d68ef8019d2..5608e8b384d 100644 --- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py @@ -32,7 +32,7 @@ from websockets.exceptions import ( ConnectionClosedOK, InvalidStatus, ) -from websockets.frames import Close +from websockets.frames import EXTERNAL_CLOSE_CODES, Close import litellm from litellm._logging import verbose_proxy_logger @@ -1930,13 +1930,18 @@ def _truncated_close_reason(reason: str) -> str: def _upstream_close_to_relay(task_results: Iterable[object]) -> Close | None: """ - The upstream close worth telling the client about: anything other than a plain, reasonless normal close + The upstream close worth telling the client about: anything other than a plain, reasonless normal close. + + Codes outside ``EXTERNAL_CLOSE_CODES`` and the private range never travel on the wire (1006 for a socket that + died without a close frame, 1005 for one that sent no code), so relaying them would build an invalid frame """ upstream_close: Final = next((result for result in task_results if isinstance(result, Close)), None) if upstream_close is None: return None if upstream_close.code == 1000 and upstream_close.reason == "": return None + if upstream_close.code not in EXTERNAL_CLOSE_CODES and not 3000 <= upstream_close.code < 5000: + return None return upstream_close diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index 844aa099541..2663396bf53 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -5188,6 +5188,49 @@ async def test_websocket_passthrough_rewrites_gateway_alias_setup_model(): assert sent_setup["model"] == "projects/proj-db/locations/global/publishers/google/models/gemini-live-2.5-flash" +@pytest.mark.asyncio +@pytest.mark.parametrize("rcvd_close", [None, "abnormal", "no_status"]) +async def test_websocket_passthrough_does_not_relay_unsendable_upstream_close(rcvd_close): + from websockets.exceptions import ConnectionClosedError + from websockets.frames import Close + + rcvd = { + None: None, + "abnormal": Close(1006, "connection died"), + "no_status": Close(1005, ""), + }[rcvd_close] + upstream_ws = ClosingUpstreamWebSocket( + ConnectionClosedError(rcvd=rcvd, sent=None, rcvd_then_sent=None) + ) + websocket = _client_websocket(_pending_receive) + + with _patched_websocket_passthrough_environment(upstream_ws): + await websocket_passthrough_request( + websocket=websocket, + target="wss://aiplatform.googleapis.com/ws/google.cloud.aiplatform.v1.LlmBidiService/BidiGenerateContent", + custom_headers={"Authorization": "Bearer token"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/vertex_ai/live", + accept_websocket=False, + ) + + websocket.close.assert_awaited_once_with() + + +@pytest.mark.asyncio +async def test_websocket_passthrough_rewrites_alias_of_unrecognised_upstream_model(): + llm_router = MagicMock() + llm_router.get_model_list.return_value = [ + {"model_name": "gemini-live", "litellm_params": {"model": "self-hosted-live-endpoint"}} + ] + + sent_frame = await _run_setup_rewrite_passthrough("gemini-live", llm_router=llm_router) + + sent_setup = json.loads(sent_frame)["setup"] + assert sent_setup["model"] == "projects/proj-db/locations/global/publishers/google/models/self-hosted-live-endpoint" + + @pytest.mark.asyncio async def test_websocket_passthrough_leaves_full_resource_setup_model_untouched(): full_resource = "projects/other/locations/us-central1/publishers/google/models/gemini-2.0-flash-live-preview-04-09" From d434787a20e5e170bf94cfb89393c691f1ad0054 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:19:38 -0700 Subject: [PATCH 036/220] fix: refuse to guess a vertex project when live passthrough has no model hint --- .../passthrough_endpoint_router.py | 17 ++++++++++--- .../test_passthrough_endpoint_router.py | 25 +++++++++++++++---- 2 files changed, 33 insertions(+), 9 deletions(-) diff --git a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py index e7887e33ca6..28067c842cd 100644 --- a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py +++ b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py @@ -126,6 +126,9 @@ class PassthroughEndpointRouter: ``deployment_key_to_vertex_credentials`` is only reachable when the caller names a project and location, which WebSocket clients never do, so DB-stored deployments need this lookup to be usable at all. + + With no model to go on, only deployments that agree on a project and location answer: guessing between + two Vertex projects would mint a token for one and later send the other one's model name """ llm_router: Final = self.llm_router_getter() if llm_router is None: @@ -135,16 +138,22 @@ class PassthroughEndpointRouter: for deployment in (llm_router.get_model_list() or ()) if (credentials := self._resolve_vertex_deployment_credentials(deployment["litellm_params"])) is not None ) - if len(resolved) == 0: - return None - return next( + matched: Final = next( ( credentials for deployment, credentials in resolved if model is not None and self._deployment_matches_model(deployment, model) ), - resolved[0][1], + None, ) + if matched is not None: + return matched + targets: Final = frozenset( + (credentials.vertex_project, credentials.vertex_location) for _, credentials in resolved + ) + if len(targets) != 1: + return None + return resolved[0][1] def _resolve_vertex_deployment_credentials( self, litellm_params: LiteLLMParamsTypedDict diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py index 7816178471f..f86bfb8bb1b 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py @@ -268,14 +268,29 @@ def test_vertex_model_hint_prefers_matching_deployment(): assert by_upstream_id is not None and by_upstream_id.vertex_project == "proj-live" -def test_vertex_unmatched_hint_falls_back_to_first_flagged_deployment(): +def test_vertex_without_usable_hint_refuses_to_guess_between_projects(): passthrough_router = _passthrough_router(_two_vertex_deployments_router()) - unmatched = passthrough_router.get_vertex_credentials_from_router_deployments(model="unknown-model") - no_hint = passthrough_router.get_vertex_credentials_from_router_deployments(model=None) + assert passthrough_router.get_vertex_credentials_from_router_deployments(model="unknown-model") is None + assert passthrough_router.get_vertex_credentials_from_router_deployments(model=None) is None - assert unmatched is not None and unmatched.vertex_project == "proj-first" - assert no_hint is not None and no_hint.vertex_project == "proj-first" + +def test_vertex_without_hint_falls_back_when_deployments_share_a_target(): + llm_router = litellm.Router( + model_list=[ + _vertex_deployment( + "gemini-flash", "vertex_ai/gemini-2.5-flash", vertex_project="proj-one", vertex_location="global" + ), + _vertex_deployment( + "gemini-live", "vertex_ai/gemini-live-2.5-flash", vertex_project="proj-one", vertex_location="global" + ), + ] + ) + passthrough_router = _passthrough_router(llm_router) + + resolved = passthrough_router.get_vertex_credentials_from_router_deployments(model=None) + + assert resolved is not None and resolved.vertex_project == "proj-one" def test_no_flagged_vertex_deployment_returns_none(): From f86aeba1e7b1ed1395a5b6d5b5c47c2e6c94d7ca Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:21:21 -0700 Subject: [PATCH 037/220] fix(cli): stop a file-held secret from vouching for an unreadable keychain A logout run from an install without the keyring package treated a token file holding its own secret as proof that no keychain entry could exist. That only holds for the login which wrote the file. A login before it may have had the package and put its credential in the keychain, where it outlives both the uninstall and the file that replaced it, so logout reported a clean sweep over a live credential. Every keychain that cannot be reached is now treated the same way, and the message says the keychain went unchecked rather than asserting what is in it. --- litellm/litellm_core_utils/cli_token_utils.py | 15 +++++------- litellm/proxy/client/cli/commands/auth.py | 2 +- .../test_cli_token_utils.py | 23 ++++++++++++------- .../proxy/client/cli/test_auth_commands.py | 16 ++++++++----- 4 files changed, 32 insertions(+), 24 deletions(-) diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 4e01dd723ce..78b52f33e2d 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -185,22 +185,19 @@ def _nothing_left_behind(outcome: SecretErase, record: CliTokenRecord | None) -> A machine with no token file has no stored login to end, and `clear_cli_token` keeps one behind whenever the keychain is left unconfirmed, so a missing file is real evidence rather than the - absence of it. Past that, a keychain that exists but is out of reach right now is - never trusted, whatever the file looks like: the login that stored a secret there and the - logout that cannot remove it are separate runs, free to differ in whether the keychain was - usable at the time. The exception is a missing `keyring` package, which had to be missing when - the credential was stored too, so a file still holding its own secret proves no keychain was - ever involved. `SecretStranded` is the keychain answering for itself and outranks the file. + absence of it. Past that, a keychain that could not be reached is never trusted, whatever the + file looks like. Even a file holding its own secret says only that the login which wrote it had + no keychain to write to, and the login before it may well have had one: the entry that login + left outlives both the uninstalled package and the file that replaced it. `SecretStranded` is + the keychain answering for itself and outranks the file. """ match outcome: case SecretErased(): return True case SecretStranded(): return False - case KeyringDisabled() | KeyringUnreachable(): + case KeyringDisabled() | KeyringNotInstalled() | KeyringUnreachable(): return record is None - case KeyringNotInstalled(): - return record is None or record.key is not None def get_litellm_gateway_api_key( diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index eb956cd536c..5fcc53e69eb 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -803,7 +803,7 @@ def logout(ctx: click.Context): click.echo(STRANDED_CREDENTIAL_MESSAGE) click.echo("Unlock your keychain and run 'lite logout' again to clear it.") case KeyringNotInstalled(): - click.echo(STRANDED_CREDENTIAL_MESSAGE) + click.echo(UNCHECKED_KEYCHAIN_MESSAGE) click.echo(f"Install the keyring package with: {KEYRING_INSTALL_HINT}, then run 'lite logout' again.") case KeyringDisabled(): click.echo(UNCHECKED_KEYCHAIN_MESSAGE) diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 162e5dd4b67..143c9e738a9 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -582,15 +582,22 @@ class TestClearCliToken: assert clear_cli_token(vault=vault) == SecretErased() - def test_a_file_backed_login_logs_out_quietly_without_keyring(self, isolated_home, secret_vault_factory): - """The complement, and the one inference the file does support: nothing here can reach a - keychain without the package, so an install that lacks it and a file that still holds its - own secret between them account for the whole credential.""" - _write_legacy_file(isolated_home) - vault = secret_vault_factory(available=False, failure=KeyringNotInstalled()) + def test_a_file_backed_login_cannot_vouch_for_a_keychain_no_package_can_reach( + self, isolated_home, secret_vault_factory + ): + """Sign in with the keyring package installed, lose the package, then sign in again so the + second secret lands in the file. The first login's entry outlives both, and the file that + replaced it holds a secret of its own, which is the shape a logout must not read as proof + that no keychain was ever involved.""" + vault = secret_vault_factory() + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-keychain"), vault=vault) + vault.available = False + vault.failure = KeyringNotInstalled() + save_cli_token(CliTokenRecord(base_url=SERVER, key="sk-in-file"), vault=vault) - assert clear_cli_token(vault=vault) == SecretErased() - assert not _token_file(isolated_home).exists() + assert clear_cli_token(vault=vault) == KeyringNotInstalled() + assert vault.blob is not None + assert "sk-in-file" not in _token_file(isolated_home).read_text() def test_is_safe_when_nothing_was_ever_stored(self, isolated_home, secret_vault_factory): assert clear_cli_token(vault=secret_vault_factory()) == SecretErased() diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 8e1551c0720..9e4ede88337 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -459,7 +459,7 @@ class TestLogoutCommand: assert result.exit_code == 0 assert "Logged out successfully" not in result.output - assert "still in the OS keychain" in result.output + assert "could not be checked" in result.output assert "pip install 'litellm[cli]'" in result.output def test_logout_does_not_call_an_unusable_keychain_clean(self, isolated_home, secret_vault_factory): @@ -492,9 +492,12 @@ class TestLogoutCommand: assert "still in the OS keychain" in result.output assert "Unlock your keychain" in result.output - def test_logout_from_a_file_only_login_stays_quiet(self, isolated_home, secret_vault_factory): - """The credential never went to a keychain, so removing the file is the whole logout and - warning about a keychain entry would send the user chasing one that cannot exist.""" + def test_logout_without_the_keyring_package_still_warns_about_a_file_held_secret( + self, isolated_home, secret_vault_factory + ): + """A file holding its own secret only says the login that wrote it had no keychain to write + to. An earlier login on this machine may have had one, and no install without the package + can look, so the honest answer is that the keychain went unchecked.""" _write_token_file(isolated_home, key="sk-in-file") result = self.runner.invoke( @@ -502,8 +505,9 @@ class TestLogoutCommand: ) assert result.exit_code == 0 - assert "Logged out successfully" in result.output - assert "still in the OS keychain" not in result.output + assert "Logged out successfully" not in result.output + assert "could not be checked" in result.output + assert "pip install 'litellm[cli]'" in result.output class TestWhoamiCommand: From ef104acdaf9ec791f1b1674fd7d6f7eabadeb6fc Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:28:48 -0700 Subject: [PATCH 038/220] fix(cli): report a token file logout cannot remove instead of crashing A ~/.litellm that has gone read-only, or one left root-owned by a sudo login, refuses both the scrubbed rewrite and the removal. The removal was unguarded, so 'lite logout' ended in a PermissionError traceback with the credential still readable in the file. It now comes back as an outcome the command reports, naming the file and what to do about it, and a file that holds no secret is still not worth alarming anyone over. --- litellm/litellm_core_utils/cli_token_utils.py | 34 ++++++- litellm/proxy/client/cli/commands/auth.py | 5 + .../test_cli_token_utils.py | 91 +++++++++++++++++++ .../proxy/client/cli/test_auth_commands.py | 17 ++++ 4 files changed, 143 insertions(+), 4 deletions(-) diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 78b52f33e2d..f2c0f8ed001 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -63,8 +63,22 @@ class CredentialNotRecorded: """ +@dataclass(frozen=True, slots=True) +class CredentialNotCleared: + """The token file still holds the secret, because it could not be removed or rewritten. + + Logging out of the keychain is only half of it. A `~/.litellm` that refuses both the scrubbed + rewrite and the removal leaves the credential readable on disk, which is the one thing a logout + is for, so it is reported instead of being counted as a clean sweep. + """ + + detail: str + + SecretSave: TypeAlias = SecretWrite | CredentialNotSaved | CredentialNotRecorded +SecretClear: TypeAlias = SecretErase | CredentialNotCleared + class CliTokenRecord(BaseModel): """A stored CLI credential. @@ -150,23 +164,35 @@ def _keep_the_secret_in_the_file(record: CliTokenRecord, outcome: SecretWrite) - return outcome -def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretErase: +def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretClear: """Remove the credential from both stores. Reports whether the keychain is now free of it. A logout the keychain never answered keeps the token file, with its secret taken out, because that file is the only remaining record that something may still be in there to remove. It is what lets a later run tell a machine with a credential it cannot reach apart from one that never had a login at all, and taking it away would leave the next logout answering the warning this - one just issued with a false all-clear. The secret goes either way. + one just issued with a false all-clear. The secret goes either way, and a file that will give up + neither its copy nor itself outranks whatever the keychain had to say. """ outcome: Final = vault.erase() record: Final = _read_token_file() settled: Final = _nothing_left_behind(outcome, record) - if settled or not _keep_the_unchecked_keychain_on_record(outcome, record): - Path(get_cli_token_file_path()).unlink(missing_ok=True) + if not settled and _keep_the_unchecked_keychain_on_record(outcome, record): + return outcome + removal: Final = _remove_token_file() + if removal is not None and record is not None and record.key is not None: + return removal return SecretErased() if settled else outcome +def _remove_token_file() -> CredentialNotCleared | None: + try: + Path(get_cli_token_file_path()).unlink(missing_ok=True) + except OSError as error: + return CredentialNotCleared(str(error)) + return None + + def _keep_the_unchecked_keychain_on_record(outcome: SecretErase, record: CliTokenRecord | None) -> bool: """Whether the token file, stripped of its secret, is worth keeping as the note that says so. diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index 5fcc53e69eb..b3a7db4bed4 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -27,6 +27,7 @@ from litellm.litellm_core_utils.cli_keyring import ( ) from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, + CredentialNotCleared, CredentialNotRecorded, CredentialNotSaved, SecretSave, @@ -796,9 +797,13 @@ def login(ctx: click.Context, config_claude: bool): @click.pass_context def logout(ctx: click.Context): """Logout and clear stored authentication""" + path: Final = get_cli_token_file_path() match clear_cli_token(vault=context_secret_vault(ctx)): case SecretErased(): click.echo("Logged out successfully. Authentication token cleared.") + case CredentialNotCleared(detail=detail): + click.echo(f"Your credential is still in {path}, which could not be removed: {detail}.") + click.echo("Delete that file, or make the directory writable and run 'lite logout' again.") case SecretStranded(): click.echo(STRANDED_CREDENTIAL_MESSAGE) click.echo("Unlock your keychain and run 'lite logout' again to clear it.") diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 143c9e738a9..2830fad58b4 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -27,6 +27,7 @@ from litellm.litellm_core_utils.cli_keyring import ( from litellm.litellm_core_utils.cli_token_utils import ( CliTokenRecord, CredentialNotRecorded, + CredentialNotCleared, CredentialNotSaved, clear_cli_token, get_cli_token_file_path, @@ -81,6 +82,26 @@ def _blob(base_url=SERVER, key="sk-vault", jwt_token=""): return json.dumps({"base_url": base_url, "key": key, "jwt_token": jwt_token}) +_REAL_REPLACE = os.replace + + +def _refuse_replace(*args, **kwargs): + raise OSError("device or resource busy") + + +class _ReplaceThatStartsRefusing: + """`os.replace` standing in for a path that cannot be replaced yet: a file another process holds + open on Windows, a directory that went read-only between staging and the rewrite.""" + + def __init__(self): + self.allowed = False + + def __call__(self, src, dst): + if not self.allowed: + raise OSError("device or resource busy") + _REAL_REPLACE(src, dst) + + class TestGetCliTokenFilePath: def test_points_at_the_home_config_file(self, isolated_home): assert get_cli_token_file_path() == str(isolated_home / ".litellm" / "token.json") @@ -459,6 +480,43 @@ class TestScrubFailure: assert json.loads(path.read_text())["key"] == "sk-legacy" assert list(path.parent.glob(".tmp-*")) == [] + def test_a_rewrite_that_fails_after_the_keychain_took_the_secret_hands_it_back( + self, isolated_home, secret_vault_factory, monkeypatch + ): + """Staging can succeed and the rewrite still fail afterwards, which is the one window where + both stores hold the credential. The keychain copy goes back, so the file is left exactly as + it was found and the move can be tried again.""" + path = _write_legacy_file(isolated_home) + vault = secret_vault_factory() + monkeypatch.setattr("litellm.litellm_core_utils.private_json.os.replace", _refuse_replace) + + record = load_cli_token(vault=vault) + + assert record.key == "sk-legacy" + assert vault.blob is None + assert json.loads(path.read_text())["key"] == "sk-legacy" + assert list(path.parent.glob(".tmp-*")) == [] + + def test_a_rollback_the_keychain_refuses_is_finished_by_the_next_read( + self, isolated_home, secret_vault_factory, monkeypatch + ): + """A keychain that will not give back what it just took leaves the credential in both stores. + Nothing is lost by that, and nothing is abandoned either: the next read carries the move the + rest of the way, so the duplicate outlives only the condition that caused it.""" + path = _write_legacy_file(isolated_home) + vault = secret_vault_factory(erasable=False) + replace = _ReplaceThatStartsRefusing() + monkeypatch.setattr("litellm.litellm_core_utils.private_json.os.replace", replace) + + assert load_cli_token(vault=vault).key == "sk-legacy" + assert vault.blob is not None + assert json.loads(path.read_text())["key"] == "sk-legacy" + + replace.allowed = True + + assert load_cli_token(vault=vault).key == "sk-legacy" + assert json.loads(path.read_text()).get("key") is None + class TestClearCliToken: def test_removes_the_credential_from_both_stores(self, isolated_home, secret_vault_factory): @@ -599,6 +657,39 @@ class TestClearCliToken: assert vault.blob is not None assert "sk-in-file" not in _token_file(isolated_home).read_text() + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") + def test_a_file_that_can_be_neither_scrubbed_nor_removed_is_reported_not_raised( + self, isolated_home, secret_vault_factory + ): + """A `~/.litellm` gone read-only, or one left root-owned by a `sudo lite login`, refuses the + scrubbed rewrite and the removal alike. The credential is still readable on disk, which is + the one thing logging out is for, so it has to come back as an answer rather than as a + traceback the user has to read the code to understand.""" + path = _write_legacy_file(isolated_home) + path.parent.chmod(0o500) + try: + outcome = clear_cli_token(vault=secret_vault_factory()) + finally: + path.parent.chmod(0o700) + + assert isinstance(outcome, CredentialNotCleared) + assert json.loads(path.read_text())["key"] == "sk-legacy" + + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") + def test_a_metadata_file_that_will_not_go_is_not_worth_alarming_the_user_over( + self, isolated_home, secret_vault_factory + ): + """The secret was in the keychain and the keychain gave it up. What is stuck on disk names a + credential that no longer exists, so the logout it describes really did happen.""" + path = _write_metadata_only_file(isolated_home) + path.parent.chmod(0o500) + try: + outcome = clear_cli_token(vault=secret_vault_factory(blob=_blob())) + finally: + path.parent.chmod(0o700) + + assert outcome == SecretErased() + def test_is_safe_when_nothing_was_ever_stored(self, isolated_home, secret_vault_factory): assert clear_cli_token(vault=secret_vault_factory()) == SecretErased() diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 9e4ede88337..71fc1cd3e6a 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -492,6 +492,23 @@ class TestLogoutCommand: assert "still in the OS keychain" in result.output assert "Unlock your keychain" in result.output + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") + def test_logout_reports_a_token_file_it_cannot_remove(self, isolated_home, secret_vault_factory): + """`lite logout` on a read-only ~/.litellm used to end in a PermissionError traceback with + the credential still sitting in the file. The user has to be told what is left and where.""" + _write_token_file(isolated_home, key="sk-in-file") + config_dir = isolated_home / ".litellm" + config_dir.chmod(0o500) + try: + result = self.runner.invoke(logout, obj={"secret_vault": secret_vault_factory()}) + finally: + config_dir.chmod(0o700) + + assert result.exit_code == 0 + assert "Logged out successfully" not in result.output + assert "still in" in result.output + assert str(config_dir / "token.json") in result.output + def test_logout_without_the_keyring_package_still_warns_about_a_file_held_secret( self, isolated_home, secret_vault_factory ): From 58844d3bda3c74ba35c3a571de0a8d2fcf2b6a79 Mon Sep 17 00:00:00 2001 From: mateo-berri Date: Thu, 20 Aug 2026 02:36:47 -0700 Subject: [PATCH 039/220] refactor(realtime): inject the vertex access token resolver Take the resolver and its timeout as parameters of the bounded helper and bind the vertex one once at module level, so the timeout tests drive an injected fake instead of patching a shared singleton. --- litellm/realtime_api/main.py | 15 ++- litellm/types/llms/vertex_ai.py | 13 ++- tests/test_litellm/realtime_api/test_main.py | 99 +++++++++++++------- 3 files changed, 86 insertions(+), 41 deletions(-) diff --git a/litellm/realtime_api/main.py b/litellm/realtime_api/main.py index 8fde7cb75c5..56b3931711e 100644 --- a/litellm/realtime_api/main.py +++ b/litellm/realtime_api/main.py @@ -15,7 +15,7 @@ from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler from litellm.llms.xai.common_utils import XAIModelInfo from litellm.secret_managers.main import get_secret_str -from litellm.types.llms.vertex_ai import VERTEX_CREDENTIALS_TYPES +from litellm.types.llms.vertex_ai import VERTEX_CREDENTIALS_TYPES, VertexAccessTokenResolver from litellm.types.realtime import ( RealtimeClientSecretRequest, RealtimeExpiresAfter, @@ -43,6 +43,7 @@ openai_realtime: Final = OpenAIRealtime() bedrock_realtime: Final = BedrockRealtime() xai_realtime: Final = XAIRealtime() vertex_llm_base: Final = VertexBase() +vertex_access_token_resolver: Final[VertexAccessTokenResolver] = vertex_llm_base._ensure_access_token_async base_llm_http_handler = BaseLLMHTTPHandler() @@ -290,20 +291,22 @@ async def arealtime_calls( async def _resolve_vertex_access_token_bounded( credentials: VERTEX_CREDENTIALS_TYPES | None, project_id: str | None, + resolver: VertexAccessTokenResolver, + timeout_seconds: float, ) -> tuple[str, str]: try: return await asyncio.wait_for( - vertex_llm_base._ensure_access_token_async( + resolver( credentials=credentials, project_id=project_id, custom_llm_provider="vertex_ai", ), - timeout=REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS, + timeout=timeout_seconds, ) except asyncio.TimeoutError as e: raise ValueError( "Vertex AI realtime: timed out fetching Google OAuth access token after " - f"{REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS}s; check network egress from the proxy " + f"{timeout_seconds}s; check network egress from the proxy " "to the OAuth token endpoint (oauth2.googleapis.com)" ) from e @@ -508,6 +511,8 @@ async def _arealtime( ) = await _resolve_vertex_access_token_bounded( credentials=vertex_credentials, project_id=vertex_project, + resolver=vertex_access_token_resolver, + timeout_seconds=REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS, ) vertex_realtime_config: Final = VertexAIRealtimeConfig( @@ -588,6 +593,8 @@ async def _realtime_health_check( ) = await _resolve_vertex_access_token_bounded( credentials=VertexBase.safe_get_vertex_ai_credentials(vertex_model_params), project_id=VertexBase.safe_get_vertex_ai_project(vertex_model_params), + resolver=vertex_access_token_resolver, + timeout_seconds=REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS, ) vertex_realtime_config: Final = VertexAIRealtimeConfig( access_token=access_token, diff --git a/litellm/types/llms/vertex_ai.py b/litellm/types/llms/vertex_ai.py index b750563432e..3b95b786631 100644 --- a/litellm/types/llms/vertex_ai.py +++ b/litellm/types/llms/vertex_ai.py @@ -1,5 +1,5 @@ from enum import Enum -from typing import Any, Final, Literal +from typing import Any, Final, Literal, Protocol from typing_extensions import ( Required, @@ -747,6 +747,17 @@ class VertexVideoGenerationResponse(TypedDict, total=False): VERTEX_CREDENTIALS_TYPES = str | dict[str, str] +class VertexAccessTokenResolver(Protocol): + """Resolves a Google OAuth access token and the project id it belongs to.""" + + async def __call__( + self, + credentials: VERTEX_CREDENTIALS_TYPES | None, + project_id: str | None, + custom_llm_provider: Literal["vertex_ai", "vertex_ai_beta", "gemini"], + ) -> tuple[str, str]: ... + + class VertexPartnerProvider(str, Enum): mistralai = "mistralai" llama = "llama" diff --git a/tests/test_litellm/realtime_api/test_main.py b/tests/test_litellm/realtime_api/test_main.py index 8ed7fb06e84..9f48d4d427b 100644 --- a/tests/test_litellm/realtime_api/test_main.py +++ b/tests/test_litellm/realtime_api/test_main.py @@ -93,12 +93,70 @@ def test_client_secret_session_model_takes_priority_over_top_level(monkeypatch): assert captured["request_data"]["session"]["model"] == "gpt-realtime-session" +async def _hanging_resolver(credentials, project_id, custom_llm_provider) -> tuple[str, str]: + await asyncio.sleep(30) + return "", "" + + +async def _thread_offloaded_hanging_resolver(credentials, project_id, custom_llm_provider) -> tuple[str, str]: + from litellm.litellm_core_utils.asyncify import asyncify + + await asyncify(time.sleep)(30) + return "", "" + + +async def _instant_resolver(credentials, project_id, custom_llm_provider) -> tuple[str, str]: + return "token-abc", "resolved-project" + + @pytest.mark.asyncio -async def test_arealtime_vertex_hung_credential_resolution_raises_promptly(monkeypatch): +async def test_vertex_credential_resolution_returns_the_resolved_token_and_project(): + assert await realtime_main._resolve_vertex_access_token_bounded( + credentials="fake-credentials", + project_id="fake-project", + resolver=_instant_resolver, + timeout_seconds=5, + ) == ("token-abc", "resolved-project") + + +@pytest.mark.asyncio +async def test_vertex_credential_resolution_times_out_instead_of_hanging(): """Regression for the realtime accept-then-silence hang: a stalled Google - OAuth token refresh used to block _arealtime's vertex branch unbounded - (minutes of zero frames for the client). It must instead raise a clear, - prompt error naming the credential-resolution timeout.""" + OAuth token refresh used to block the vertex branch unbounded (minutes of + zero frames for the client). It must raise promptly and name the timeout.""" + start = time.monotonic() + with pytest.raises(ValueError, match="timed out fetching Google OAuth access token"): + await realtime_main._resolve_vertex_access_token_bounded( + credentials="fake-credentials", + project_id="fake-project", + resolver=_hanging_resolver, + timeout_seconds=0.05, + ) + assert time.monotonic() - start < 5 + + +@pytest.mark.asyncio +async def test_vertex_credential_resolution_bounds_a_thread_offloaded_refresh(): + """The real stall is a blocking google-auth refresh that runs in a worker + thread via asyncify, not a plain awaitable sleep. A timeout that only bounds + cancellable awaits would leave that shape hanging, so bound the shape the + proxy actually runs.""" + start = time.monotonic() + with pytest.raises(ValueError, match="timed out fetching Google OAuth access token"): + await realtime_main._resolve_vertex_access_token_bounded( + credentials="fake-credentials", + project_id="fake-project", + resolver=_thread_offloaded_hanging_resolver, + timeout_seconds=0.05, + ) + assert time.monotonic() - start < 5 + + +@pytest.mark.asyncio +async def test_arealtime_vertex_branch_resolves_credentials_under_a_bound(monkeypatch): + """The wiring half of the regression: the vertex branch of _arealtime must + go through the bounded resolver, so a hung token refresh surfaces as a + prompt error there rather than as an accepted-then-silent websocket.""" async def hanging_token_refresh(**kwargs): await asyncio.sleep(30) @@ -107,38 +165,7 @@ async def test_arealtime_vertex_hung_credential_resolution_raises_promptly(monke return model, "vertex_ai", None, api_base monkeypatch.setattr(realtime_main, "get_llm_provider", mock_get_llm_provider) - monkeypatch.setattr(realtime_main.vertex_llm_base, "_ensure_access_token_async", hanging_token_refresh) - monkeypatch.setattr(realtime_main, "REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS", 0.05) - - start = time.monotonic() - with pytest.raises(ValueError, match="timed out fetching Google OAuth access token"): - await realtime_main._arealtime.__wrapped__( - model="gemini-live-2.5-flash", - websocket=MagicMock(), - litellm_logging_obj=FakeLogging(), - vertex_credentials="fake-credentials", - vertex_project="fake-project", - vertex_location="us-central1", - ) - assert time.monotonic() - start < 5 - - -@pytest.mark.asyncio -async def test_arealtime_vertex_credential_timeout_survives_thread_offloaded_refresh(monkeypatch): - """The real stall is a blocking google-auth refresh that runs in a worker - thread via asyncify, not a plain awaitable sleep. A timeout that only bounds - cancellable awaits would leave that shape hanging, so bound the shape the - proxy actually runs.""" - from litellm.litellm_core_utils.asyncify import asyncify - - async def thread_offloaded_hanging_refresh(**kwargs): - return await asyncify(time.sleep)(30) - - def mock_get_llm_provider(model, api_base, api_key): - return model, "vertex_ai", None, api_base - - monkeypatch.setattr(realtime_main, "get_llm_provider", mock_get_llm_provider) - monkeypatch.setattr(realtime_main.vertex_llm_base, "_ensure_access_token_async", thread_offloaded_hanging_refresh) + monkeypatch.setattr(realtime_main, "vertex_access_token_resolver", hanging_token_refresh) monkeypatch.setattr(realtime_main, "REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS", 0.05) start = time.monotonic() From d643136895d6d3c00f2339b75e63162098bc0802 Mon Sep 17 00:00:00 2001 From: Mateo Date: Thu, 20 Aug 2026 02:49:01 -0700 Subject: [PATCH 040/220] fix(realtime): resolve the vertex token resolver at call time Binding the bound method at import froze the module-level VertexBase instance, so callers that swap it no longer reached their replacement. --- litellm/realtime_api/main.py | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-) diff --git a/litellm/realtime_api/main.py b/litellm/realtime_api/main.py index 56b3931711e..4e02be36daa 100644 --- a/litellm/realtime_api/main.py +++ b/litellm/realtime_api/main.py @@ -2,7 +2,7 @@ import asyncio import os -from typing import Any, Final, cast +from typing import Any, Final, Literal, cast import litellm from litellm.constants import ( @@ -43,7 +43,6 @@ openai_realtime: Final = OpenAIRealtime() bedrock_realtime: Final = BedrockRealtime() xai_realtime: Final = XAIRealtime() vertex_llm_base: Final = VertexBase() -vertex_access_token_resolver: Final[VertexAccessTokenResolver] = vertex_llm_base._ensure_access_token_async base_llm_http_handler = BaseLLMHTTPHandler() @@ -288,6 +287,18 @@ async def arealtime_calls( ) +async def vertex_access_token_resolver( + credentials: VERTEX_CREDENTIALS_TYPES | None, + project_id: str | None, + custom_llm_provider: Literal["vertex_ai", "vertex_ai_beta", "gemini"], +) -> tuple[str, str]: + return await vertex_llm_base._ensure_access_token_async( + credentials=credentials, + project_id=project_id, + custom_llm_provider=custom_llm_provider, + ) + + async def _resolve_vertex_access_token_bounded( credentials: VERTEX_CREDENTIALS_TYPES | None, project_id: str | None, From 4ca1f3148a9df608eda0a4b562badc2c90d25f28 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:58:53 -0700 Subject: [PATCH 041/220] fix(cli): finish a refused token file rewrite in place Taking the secret out of ~/.litellm/token.json stages a replacement and moves it into place, which needs room for a second file and a directory that will accept a new entry. A full disk refuses the first and a read-only ~/.litellm the second, and logout gave up there: it removed the file when it could, dropping the record that the keychain had never been confirmed clear, so the logout after it reported a clean keychain it never checked Shortening the file already in place needs neither, so the logout scrub and the legacy migration now fall back to overwriting it where it lies. On a read-only ~/.litellm the logout the user asked for now happens, instead of coming back with instructions to delete the file by hand --- litellm/litellm_core_utils/cli_token_utils.py | 45 ++++++++++---- litellm/litellm_core_utils/private_json.py | 15 +++++ .../test_cli_token_utils.py | 58 ++++++++++++++----- .../litellm_core_utils/test_private_json.py | 37 ++++++++++++ .../proxy/client/cli/test_auth_commands.py | 38 +++++++++--- 5 files changed, 162 insertions(+), 31 deletions(-) create mode 100644 tests/test_litellm/litellm_core_utils/test_private_json.py diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index f2c0f8ed001..2f45742ce03 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -38,6 +38,7 @@ from litellm.litellm_core_utils.private_json import ( commit_staged_json, discard_staged_json, ensure_private_dir, + overwrite_private_json, stage_private_json, write_private_json, ) @@ -180,7 +181,7 @@ def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretClear: if not settled and _keep_the_unchecked_keychain_on_record(outcome, record): return outcome removal: Final = _remove_token_file() - if removal is not None and record is not None and record.key is not None: + if removal is not None and record is not None and not _scrub_file_secret(record): return removal return SecretErased() if settled else outcome @@ -198,8 +199,9 @@ def _keep_the_unchecked_keychain_on_record(outcome: SecretErase, record: CliToke Only a keychain that could not be reached leaves the question open. One that answered for itself is remembered without any help from the file, and a file it can still pair a live entry with - would leave the machine signed in to the login that was just ended. A copy that cannot be - replaced with a secret-free one is not kept either, because the secret goes first. + would leave the machine signed in to the login that was just ended. A copy that will give up + its secret neither to a staged replacement nor to an overwrite is not kept either, because the + secret goes first. """ if record is None or isinstance(outcome, SecretStranded): return False @@ -210,11 +212,12 @@ def _nothing_left_behind(outcome: SecretErase, record: CliTokenRecord | None) -> """Whether the keychain can be trusted to hold no credential of ours once the file is gone. A machine with no token file has no stored login to end, and `clear_cli_token` keeps one behind - whenever the keychain is left unconfirmed, so a missing file is real evidence rather than the - absence of it. Past that, a keychain that could not be reached is never trusted, whatever the - file looks like. Even a file holding its own secret says only that the login which wrote it had - no keychain to write to, and the login before it may well have had one: the entry that login - left outlives both the uninstalled package and the file that replaced it. `SecretStranded` is + whenever the keychain is left unconfirmed, taking the secret out in place when it cannot stage a + replacement, so a missing file is real evidence rather than the absence of it. Past that, a + keychain that could not be reached is never trusted, whatever the file looks like. Even a file + holding its own secret says only that the login which wrote it had no keychain to write to, and + the login before it may well have had one: the entry that login left outlives both the + uninstalled package and the file that replaced it. `SecretStranded` is the keychain answering for itself and outranks the file. """ match outcome: @@ -323,6 +326,11 @@ def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliToken before the keychain is handed anything. Copying the credential into a second store and only then discovering the first one cannot be cleaned would widen exposure instead of narrowing it, which is the opposite of what moving it into the keychain is for. + + A staged file that will not go into place is overwritten where it lies before the keychain is + asked to take the new entry back, so the migration finishes on a directory that would only ever + have refused it. Rolling back is the last resort, and a rollback the keychain also refuses + leaves the secret in both stores until the next read, which retries this same migration. """ if record.key is None: return None @@ -332,7 +340,7 @@ def _migrate_file_secret(record: CliTokenRecord, vault: SecretVault) -> CliToken if not isinstance(vault.write(_encode_secret(record.base_url, record.key, record.jwt_token)), SecretStored): discard_staged_json(staged) return record - if not _commit_token_file(staged): + if not _commit_token_file(staged) and not _overwrite_file_secret(record): vault.erase() return record @@ -342,7 +350,24 @@ def _scrub_file_secret(record: CliTokenRecord) -> bool: if record.key is None and not record.jwt_token: return True staged: Final = _stage_scrubbed_file(record) - return staged is not None and _commit_token_file(staged) + if staged is not None and _commit_token_file(staged): + return True + return _overwrite_file_secret(record) + + +def _overwrite_file_secret(record: CliTokenRecord) -> bool: + """Take the secret out of the token file where it lies, when no replacement can be put in place. + + The atomic rewrite wants room for a second file and a directory that will accept it. A full disk + refuses the first and a read-only `~/.litellm` the second, and neither stands in the way of + shortening the file that is already there. It is worth the loss of atomicity because a partial + write reads as no login at all, which is where the refused rewrite left the next run anyway. + """ + try: + overwrite_private_json(get_cli_token_file_path(), _without_secret(record).model_dump(exclude_none=True)) + except OSError: + return False + return True def _stage_scrubbed_file(record: CliTokenRecord) -> str | None: diff --git a/litellm/litellm_core_utils/private_json.py b/litellm/litellm_core_utils/private_json.py index fbeb74aab5a..30f64c8fc27 100644 --- a/litellm/litellm_core_utils/private_json.py +++ b/litellm/litellm_core_utils/private_json.py @@ -45,6 +45,21 @@ def commit_staged_json(staged: str, path: str) -> None: raise +def overwrite_private_json(path: str, data: Mapping[str, object]) -> None: + """Rewrite a file that is already there, in place, keeping the mode it was created with. + + `write_private_json` needs room for a second file and a directory that will accept it, which is + what a full disk and a read-only `~/.litellm` respectively refuse. Shortening the file already + in place needs neither. It is not atomic, so an interrupted write leaves a partial file, and it + never creates one, so it cannot put a world-readable file where a private one was. + """ + fd: Final = os.open(path, os.O_WRONLY | os.O_TRUNC) + with os.fdopen(fd, "w") as f: + json.dump(data, f, indent=2) + f.flush() + os.fsync(f.fileno()) + + def discard_staged_json(staged: str) -> None: """Throw a staged file away when the change it was part of is abandoned""" Path(staged).unlink(missing_ok=True) diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 2830fad58b4..e3d14c6da46 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -480,12 +480,13 @@ class TestScrubFailure: assert json.loads(path.read_text())["key"] == "sk-legacy" assert list(path.parent.glob(".tmp-*")) == [] - def test_a_rewrite_that_fails_after_the_keychain_took_the_secret_hands_it_back( + def test_a_rewrite_the_directory_refuses_is_finished_in_place( self, isolated_home, secret_vault_factory, monkeypatch ): """Staging can succeed and the rewrite still fail afterwards, which is the one window where - both stores hold the credential. The keychain copy goes back, so the file is left exactly as - it was found and the move can be tried again.""" + both stores hold the credential. Shortening the file already there needs neither a second + file nor a cooperative directory, so the move finishes rather than handing the keychain copy + back and leaving the cleartext where it was.""" path = _write_legacy_file(isolated_home) vault = secret_vault_factory() monkeypatch.setattr("litellm.litellm_core_utils.private_json.os.replace", _refuse_replace) @@ -493,26 +494,30 @@ class TestScrubFailure: record = load_cli_token(vault=vault) assert record.key == "sk-legacy" - assert vault.blob is None - assert json.loads(path.read_text())["key"] == "sk-legacy" + assert vault.blob is not None + assert json.loads(path.read_text()).get("key") is None assert list(path.parent.glob(".tmp-*")) == [] + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores file permissions") def test_a_rollback_the_keychain_refuses_is_finished_by_the_next_read( self, isolated_home, secret_vault_factory, monkeypatch ): - """A keychain that will not give back what it just took leaves the credential in both stores. - Nothing is lost by that, and nothing is abandoned either: the next read carries the move the - rest of the way, so the duplicate outlives only the condition that caused it.""" + """A file that will take neither a replacement nor an overwrite, and a keychain that will not + give back what it just took, leave the credential in both stores. Nothing is lost by that, + and nothing is abandoned either: the next read carries the move the rest of the way, so the + duplicate outlives only the conditions that caused it.""" path = _write_legacy_file(isolated_home) vault = secret_vault_factory(erasable=False) replace = _ReplaceThatStartsRefusing() monkeypatch.setattr("litellm.litellm_core_utils.private_json.os.replace", replace) + path.chmod(0o400) assert load_cli_token(vault=vault).key == "sk-legacy" assert vault.blob is not None assert json.loads(path.read_text())["key"] == "sk-legacy" replace.allowed = True + path.chmod(0o600) assert load_cli_token(vault=vault).key == "sk-legacy" assert json.loads(path.read_text()).get("key") is None @@ -657,24 +662,49 @@ class TestClearCliToken: assert vault.blob is not None assert "sk-in-file" not in _token_file(isolated_home).read_text() - @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") - def test_a_file_that_can_be_neither_scrubbed_nor_removed_is_reported_not_raised( + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores file permissions") + def test_a_file_that_gives_up_neither_its_secret_nor_itself_is_reported_not_raised( self, isolated_home, secret_vault_factory ): - """A `~/.litellm` gone read-only, or one left root-owned by a `sudo lite login`, refuses the - scrubbed rewrite and the removal alike. The credential is still readable on disk, which is - the one thing logging out is for, so it has to come back as an answer rather than as a - traceback the user has to read the code to understand.""" + """A `~/.litellm` gone read-only refuses the staged rewrite and the removal, and a token file + left read-only with it, as a `sudo lite login` leaves both, refuses the overwrite too. The + credential is still readable on disk, which is the one thing logging out is for, so it has to + come back as an answer rather than as a traceback the user has to read the code to + understand.""" path = _write_legacy_file(isolated_home) + path.chmod(0o400) path.parent.chmod(0o500) try: outcome = clear_cli_token(vault=secret_vault_factory()) finally: path.parent.chmod(0o700) + path.chmod(0o600) assert isinstance(outcome, CredentialNotCleared) assert json.loads(path.read_text())["key"] == "sk-legacy" + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") + def test_a_directory_that_takes_no_new_file_still_gives_up_the_secret_in_the_old_one( + self, isolated_home, secret_vault_factory + ): + """A read-only `~/.litellm` accepts no replacement token file and no removal of the one it + has, and still lets that one be shortened. The secret goes, the file stays as the note that + the keychain went unchecked, and the logout after it warns again instead of reading the gap + the removal would have left as a clean keychain. + + The key is a realistic length so the file genuinely shrinks: a rewrite in place that leaves + the tail of the old contents behind hands the next run a file it cannot parse.""" + path = _write_legacy_file(isolated_home, key="sk-" + "a" * 700) + vault = secret_vault_factory(available=False, failure=KeyringUnreachable()) + path.parent.chmod(0o500) + try: + assert clear_cli_token(vault=vault) == KeyringUnreachable() + assert clear_cli_token(vault=vault) == KeyringUnreachable() + finally: + path.parent.chmod(0o700) + + assert json.loads(path.read_text()).get("key") is None + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") def test_a_metadata_file_that_will_not_go_is_not_worth_alarming_the_user_over( self, isolated_home, secret_vault_factory diff --git a/tests/test_litellm/litellm_core_utils/test_private_json.py b/tests/test_litellm/litellm_core_utils/test_private_json.py new file mode 100644 index 00000000000..cedff61959f --- /dev/null +++ b/tests/test_litellm/litellm_core_utils/test_private_json.py @@ -0,0 +1,37 @@ +import json +import os +import stat + +import pytest + +from litellm.litellm_core_utils.private_json import overwrite_private_json, write_private_json + + +class TestOverwritePrivateJson: + def test_replaces_the_contents_of_the_file_already_there(self, tmp_path): + path = tmp_path / "token.json" + write_private_json(str(path), {"key": "sk-" + "a" * 700}) + + overwrite_private_json(str(path), {"user_id": "u-1"}) + + assert json.loads(path.read_text()) == {"user_id": "u-1"} + + def test_refuses_to_create_the_file_it_was_asked_to_rewrite(self, tmp_path): + """This is the one writer that does not go through a private temp file, so a path it creates + would land with whatever the umask allows. Refusing keeps it unable to put a world-readable + file where the caller believed a private one already was.""" + path = tmp_path / "token.json" + + with pytest.raises(FileNotFoundError): + overwrite_private_json(str(path), {"user_id": "u-1"}) + + assert not path.exists() + + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores file permissions") + def test_keeps_the_owner_only_mode_the_file_was_created_with(self, tmp_path): + path = tmp_path / "token.json" + write_private_json(str(path), {"key": "sk-live"}) + + overwrite_private_json(str(path), {"user_id": "u-1"}) + + assert stat.S_IMODE(path.stat().st_mode) == 0o600 diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 71fc1cd3e6a..8191a8edca7 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -492,12 +492,37 @@ class TestLogoutCommand: assert "still in the OS keychain" in result.output assert "Unlock your keychain" in result.output - @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") - def test_logout_reports_a_token_file_it_cannot_remove(self, isolated_home, secret_vault_factory): - """`lite logout` on a read-only ~/.litellm used to end in a PermissionError traceback with - the credential still sitting in the file. The user has to be told what is left and where.""" + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores file permissions") + def test_logout_reports_a_token_file_it_cannot_clear(self, isolated_home, secret_vault_factory): + """`lite logout` on a read-only ~/.litellm holding a read-only token file used to end in a + PermissionError traceback with the credential still sitting in the file. The user has to be + told what is left and where.""" _write_token_file(isolated_home, key="sk-in-file") config_dir = isolated_home / ".litellm" + path = config_dir / "token.json" + path.chmod(0o400) + config_dir.chmod(0o500) + try: + result = self.runner.invoke(logout, obj={"secret_vault": secret_vault_factory()}) + finally: + config_dir.chmod(0o700) + path.chmod(0o600) + + assert result.exit_code == 0 + assert "Logged out successfully" not in result.output + assert "still in" in result.output + assert str(config_dir / "token.json") in result.output + + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") + def test_logout_on_a_read_only_directory_still_takes_the_secret_out_of_the_file( + self, isolated_home, secret_vault_factory + ): + """A ~/.litellm that will accept no replacement file and no removal still lets the file it + has be shortened, so the logout the user asked for happens rather than being handed back to + them with instructions.""" + _write_token_file(isolated_home, key="sk-in-file") + config_dir = isolated_home / ".litellm" + path = config_dir / "token.json" config_dir.chmod(0o500) try: result = self.runner.invoke(logout, obj={"secret_vault": secret_vault_factory()}) @@ -505,9 +530,8 @@ class TestLogoutCommand: config_dir.chmod(0o700) assert result.exit_code == 0 - assert "Logged out successfully" not in result.output - assert "still in" in result.output - assert str(config_dir / "token.json") in result.output + assert "Logged out successfully" in result.output + assert "sk-in-file" not in path.read_text() def test_logout_without_the_keyring_package_still_warns_about_a_file_held_secret( self, isolated_home, secret_vault_factory From b142d1d76576221e394893c8e7abaf3254e167d3 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 02:59:20 -0700 Subject: [PATCH 042/220] fix(cli): stop whoami calling an unreadable credential authenticated `lite whoami` led with "Authenticated" whenever a token file was on disk, even when the keychain holding the credential would not give it up. The notice about that sat below the account lines, so the session read as a working one and sent the user looking for the problem anywhere but the keychain --- litellm/proxy/client/cli/commands/auth.py | 2 +- .../proxy/client/cli/test_auth_commands.py | 10 ++++++++-- 2 files changed, 9 insertions(+), 3 deletions(-) diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index b3a7db4bed4..ac330ad6892 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -867,7 +867,7 @@ def whoami(ctx: click.Context): click.echo("Not authenticated. Run 'lite login' to authenticate.") return - click.echo("Authenticated") + click.echo("Authenticated" if token_data.key is not None else "Signed in, but the credential cannot be read") click.echo(f"User Email: {token_data.user_email or 'Unknown'}") click.echo(f"User ID: {token_data.user_id or 'Unknown'}") click.echo(f"User Role: {token_data.user_role or 'Unknown'}") diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py index 8191a8edca7..1dd8a7ead92 100644 --- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py @@ -1137,7 +1137,12 @@ class TestKeychainBackedCommands: assert "could not be read" in result.output assert "lite login" in result.output - def test_whoami_flags_a_locked_keychain(self, isolated_home, secret_vault_factory): + def test_whoami_does_not_call_a_credential_it_cannot_read_authenticated( + self, isolated_home, secret_vault_factory + ): + """A login whose secret is stuck in an unreachable keychain authenticates nothing. Leading + with "Authenticated" and a token age reads as a working session, and sends the user looking + for the problem somewhere other than the keychain the notice underneath names.""" _write_home_json( isolated_home, "token.json", @@ -1147,7 +1152,8 @@ class TestKeychainBackedCommands: result = self.runner.invoke(whoami, obj=obj) - assert "Authenticated" in result.output + assert "Authenticated" not in result.output + assert "the credential cannot be read" in result.output assert "could not be read" in result.output def test_whoami_names_the_kill_switch_rather_than_a_missing_package( From 8b7c801d61be5e4d02127ff6e86d743b157678f9 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 03:03:35 -0700 Subject: [PATCH 043/220] test(e2e): pin openai_passthrough routing, cost logging, and file list isolation Five e2e tests over routes a customer drives through the gateway, each one pinning a fix that currently has no live coverage. The dedicated /openai_passthrough prefix used to be swallowed by the provider-scoped /{provider}/v1/files and /{provider}/v1/batches routes, which bound "openai_passthrough" as a provider name and failed inside the gateway before ever reaching OpenAI. Two tests now upload a file and list batches through that prefix and assert OpenAI's own objects come back. Streamed /openai_passthrough/v1/responses and /openai_passthrough/v1/embeddings are relayed to OpenAI but still have to be costed, since the customer budgets against this traffic. Both used to land a row the gateway could not use: the streamed responses call logged a zero-cost row under a random id, and embeddings wrote no row at all. Each test now reconciles the logged spend and token counts against the response the caller was actually served. GET /v1/files narrowed its data to the caller's own rows but left first_id and last_id addressing the shared provider account's page, handing any caller raw provider file ids belonging to other tenants. The new test asserts both cursors address rows in the page the caller can see. ResourceManager.defer now accepts any callable rather than one returning None, so a delete that answers with a response model can be deferred as-is. --- tests/e2e/batches/batch_client.py | 7 + tests/e2e/batches/test_batches_e2e.py | 34 +++++ .../coverage_registry/llm_conversational.yaml | 1 + .../llm_nonconversational.yaml | 4 + tests/e2e/lifecycle.py | 9 +- .../e2e/llm_translation/passthrough_client.py | 132 +++++++++++++++++- .../llm_translation/test_passthrough_e2e.py | 124 +++++++++++++++- 7 files changed, 305 insertions(+), 6 deletions(-) diff --git a/tests/e2e/batches/batch_client.py b/tests/e2e/batches/batch_client.py index 5cc5d1dae3b..968a357e8af 100644 --- a/tests/e2e/batches/batch_client.py +++ b/tests/e2e/batches/batch_client.py @@ -40,8 +40,15 @@ class FileObject(BaseModel): class FileList(BaseModel): + """GET /v1/files page. The cursors are modelled because they are part of the + page's isolation contract: they must address rows in `data`, never rows the + caller was not allowed to see.""" + object: str | None = None data: list[FileObject] = [] + first_id: str | None = None + last_id: str | None = None + has_more: bool | None = None class BatchObject(BaseModel): diff --git a/tests/e2e/batches/test_batches_e2e.py b/tests/e2e/batches/test_batches_e2e.py index 53bf9739983..536bc113a25 100644 --- a/tests/e2e/batches/test_batches_e2e.py +++ b/tests/e2e/batches/test_batches_e2e.py @@ -572,6 +572,40 @@ class TestOpenAIFiles: f"listed file must round-trip the upload purpose, got {match.purpose!r}" ) + @pytest.mark.covers( + "llm.files.openai.list_isolation.nonstream.works", + exercised_on=["files"], + ) + def test_list_page_cursors_address_only_the_callers_own_files( + self, client: BatchClient, resources: ResourceManager + ) -> None: + """A list page's pagination cursors must address rows in that page. + + The proxy fronts one shared provider account, so the upstream page is the + whole organization's. The gateway narrows `data` to the files the caller + owns, and `first_id` / `last_id` have to be narrowed with it: left as the + upstream org's, they hand any caller raw provider file ids belonging to + other tenants, which is the handle the file routes accept. + """ + key = resources.key(user_id=f"e2e-file-list-{unique_marker()}") + + listed = unwrap(client.list_files(key=key)) + + expected_first = listed.data[0].id if listed.data else None + expected_last = listed.data[-1].id if listed.data else None + assert listed.first_id == expected_first, ( + f"first_id {listed.first_id!r} is not the first row this caller can see " + f"({expected_first!r}); the page leaked another caller's file id" + ) + assert listed.last_id == expected_last, ( + f"last_id {listed.last_id!r} is not the last row this caller can see " + f"({expected_last!r}); the page leaked another caller's file id" + ) + assert listed.has_more is not True, ( + "the page advertises another page, but the proxy never forwards a cursor " + "upstream, so following it re-serves this same page forever" + ) + @pytest.mark.covers( "llm.files.openai.retrieve.nonstream.works", exercised_on=["files"], diff --git a/tests/e2e/coverage_registry/llm_conversational.yaml b/tests/e2e/coverage_registry/llm_conversational.yaml index 82bee39b9b2..1ddc146c12d 100644 --- a/tests/e2e/coverage_registry/llm_conversational.yaml +++ b/tests/e2e/coverage_registry/llm_conversational.yaml @@ -63,6 +63,7 @@ - {id: llm.responses.openai.input_validation.nonstream.works, module: llm, tier: P1, subject_endpoint: responses, route: openai, capability: input_validation, streaming: nonstream, assertions: [works], source: "vendor strategy §9.9 / LIT-4778", rationale: "Responses missing/empty input and missing model are rejected"} - {id: llm.responses.openai.basic.stream.works, module: llm, tier: P0, subject_endpoint: responses, route: openai, capability: basic, streaming: stream, assertions: [works], source: "response_api_endpoints/endpoints.py:26", rationale: "Streaming via /v1/responses"} - {id: llm.responses.openai.basic.nonstream.cost_logged, module: llm, tier: P0, subject_endpoint: responses, route: openai, capability: basic, streaming: nonstream, assertions: [works, cost_logged], source: "response_api_endpoints/endpoints.py:26", rationale: "Cost logged on responses"} +- {id: llm.responses.openai.passthrough.stream.cost_logged, module: llm, tier: P0, subject_endpoint: responses, route: openai, capability: basic, streaming: stream, assertions: [cost_logged], source: "test_passthrough_e2e.py", rationale: "A streamed POST /openai_passthrough/v1/responses is costed and keyed by the provider response id; it used to log a zero-cost row under a random id (LIT-5870)"} - {id: llm.responses.openai.tool_use.nonstream.works, module: llm, tier: P0, subject_endpoint: responses, route: openai, capability: tool_use, streaming: nonstream, assertions: [works], source: "model_prices json", rationale: "Tool calls via Responses API"} - {id: llm.responses.openai.vision.nonstream.works, module: llm, tier: P0, subject_endpoint: responses, route: openai, capability: vision, streaming: nonstream, assertions: [works], source: "model_prices json", rationale: "Vision via Responses API"} - {id: llm.responses.anthropic.basic.nonstream.works, module: llm, tier: P1, subject_endpoint: responses, route: anthropic, capability: basic, streaming: nonstream, assertions: [works], source: "response_api_endpoints/endpoints.py:26", rationale: "Responses w/ Anthropic translation (smoke)"} diff --git a/tests/e2e/coverage_registry/llm_nonconversational.yaml b/tests/e2e/coverage_registry/llm_nonconversational.yaml index bb7169509eb..674d369b49f 100644 --- a/tests/e2e/coverage_registry/llm_nonconversational.yaml +++ b/tests/e2e/coverage_registry/llm_nonconversational.yaml @@ -3,6 +3,7 @@ - {id: llm.embeddings.openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_embeddings_endpoint_e2e.py:23", rationale: "Core endpoint, live vector response"} - {id: llm.embeddings.openai.input_validation.nonstream.works, module: llm, tier: P1, subject_endpoint: embeddings, route: openai, capability: input_validation, streaming: nonstream, assertions: [works], source: "vendor strategy §9.3 / LIT-4778", rationale: "Missing model/input on /embeddings return client errors"} - {id: llm.embeddings.openai.basic.nonstream.cost_logged, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [cost_logged], source: "SPEND_TRACKING_COVERAGE_MATRIX.md:34", rationale: "Cost tracking on embeddings"} +- {id: llm.embeddings.openai.passthrough.nonstream.cost_logged, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [cost_logged], source: "test_passthrough_e2e.py", rationale: "POST /openai_passthrough/v1/embeddings is costed; the route wrote no spend row at all, so budgets never saw traffic OpenAI was billing for (LIT-5870)"} - {id: llm.embeddings.azure_openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: embeddings, route: azure_openai, capability: basic, streaming: nonstream, assertions: [works], source: "llms/azure/azure.py", rationale: "Azure embeddings via translation"} - {id: llm.embeddings.bedrock.basic.nonstream.works, module: llm, tier: P1, subject_endpoint: embeddings, route: bedrock_converse, capability: basic, streaming: nonstream, assertions: [works], source: "llms/bedrock/embed/embedding.py", rationale: "Bedrock Titan embeddings"} - {id: llm.embeddings.vertex.basic.nonstream.works, module: llm, tier: P1, subject_endpoint: embeddings, route: vertex, capability: basic, streaming: nonstream, assertions: [works], source: "vertex_embeddings/embedding_handler.py", rationale: "Vertex embeddings"} @@ -13,6 +14,7 @@ - {id: llm.batches.openai.cancel.nonstream.works, module: llm, tier: P0, subject_endpoint: batches, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_batches_e2e.py", rationale: "Batch cancel"} - {id: llm.batches.openai.list.nonstream.works, module: llm, tier: P0, subject_endpoint: batches, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_batches_e2e.py", rationale: "Batch list envelope"} - {id: llm.batches.openai.file_lifecycle.nonstream.works, module: llm, tier: P0, subject_endpoint: batches, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_batches_e2e.py", rationale: "File upload/retrieve/delete for batch flow"} +- {id: llm.batches.openai.passthrough.nonstream.works, module: llm, tier: P1, subject_endpoint: batches, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_passthrough_e2e.py", rationale: "GET /openai_passthrough/v1/batches relays OpenAI's own batch page; the dedicated prefix must not bind as a provider name on the /{provider}/v1/batches route (LIT-5870)"} - {id: llm.batches.openai_encoded.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: batches, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "batches/capabilities.py", rationale: "Encoded scenario lifecycle"} - {id: llm.batches.openai_unified.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: batches, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "batches/capabilities.py", rationale: "Unified/managed-id scenario"} - {id: llm.batches.openai_model_param.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: batches, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "batches/capabilities.py", rationale: "Model-param scenario"} @@ -29,6 +31,8 @@ - {id: llm.files.openai.retrieve.nonstream.works, module: llm, tier: P0, subject_endpoint: files, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "files_endpoints.py", rationale: "File retrieve by id"} - {id: llm.files.openai.delete.nonstream.works, module: llm, tier: P0, subject_endpoint: files, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "files_endpoints.py", rationale: "File delete returns deleted=true"} - {id: llm.files.openai.list.nonstream.works, module: llm, tier: P0, subject_endpoint: files, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "files_endpoints.py", rationale: "File list paginated"} +- {id: llm.files.openai.list_isolation.nonstream.works, module: llm, tier: P0, subject_endpoint: files, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_batches_e2e.py", rationale: "GET /v1/files pagination cursors address only rows the caller owns; on a shared provider account the upstream cursors otherwise hand out other tenants' raw provider file ids (LIT-5870)"} +- {id: llm.files.openai.passthrough.nonstream.works, module: llm, tier: P1, subject_endpoint: files, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_passthrough_e2e.py", rationale: "POST/DELETE /openai_passthrough/v1/files relay OpenAI's own file object; the dedicated prefix must not bind as a provider name on the /{provider}/v1/files route (LIT-5870)"} - {id: llm.files.azure_openai.upload.nonstream.works, module: llm, tier: P0, subject_endpoint: files, route: azure_openai, capability: basic, streaming: nonstream, assertions: [works], source: "batches/capabilities.py:45", rationale: "Azure file upload managed backend"} - {id: llm.files.vertex.upload.nonstream.works, module: llm, tier: P0, subject_endpoint: files, route: vertex, capability: basic, streaming: nonstream, assertions: [works], source: "batches/capabilities.py:52", rationale: "Vertex file upload to GCS"} - {id: llm.files.bedrock.upload.nonstream.works, module: llm, tier: P0, subject_endpoint: files, route: bedrock_converse, capability: basic, streaming: nonstream, assertions: [works], source: "batches/capabilities.py:59", rationale: "Bedrock file upload to S3"} diff --git a/tests/e2e/lifecycle.py b/tests/e2e/lifecycle.py index 4ef25509905..c9a67ebdb8c 100644 --- a/tests/e2e/lifecycle.py +++ b/tests/e2e/lifecycle.py @@ -52,7 +52,7 @@ class ResourceManager: """ client: ResourceClient - _cleanups: List[Callable[[], None]] = field( + _cleanups: List[Callable[[], object]] = field( default_factory=list ) # mutable-ok: append-only teardown registry @@ -60,8 +60,11 @@ class ResourceManager: """No global setup needed today; present for lifecycle symmetry.""" return None - def defer(self, cleanup: Callable[[], None]) -> None: - """Register a teardown action for any resource the test just created.""" + def defer(self, cleanup: Callable[[], object]) -> None: + """Register a teardown action for any resource the test just created. + + Whatever the action returns is discarded, so a delete that answers with a + response model can be deferred directly.""" self._cleanups.append(cleanup) def key(self, models: list[str] | None = None, user_id: str | None = "e2e-test-user") -> str: diff --git a/tests/e2e/llm_translation/passthrough_client.py b/tests/e2e/llm_translation/passthrough_client.py index 439594f3624..e0dfae679a9 100644 --- a/tests/e2e/llm_translation/passthrough_client.py +++ b/tests/e2e/llm_translation/passthrough_client.py @@ -15,7 +15,7 @@ from dataclasses import dataclass from pydantic import BaseModel, Field from proxy_client import ProxyClient -from e2e_http import Headers, StreamingResponse +from e2e_http import FileUploadForm, Headers, NoBody, Result, StreamingResponse from models import ChatMessage @@ -113,6 +113,76 @@ class OpenAIChatBody(BaseModel): max_completion_tokens: int = 64 +class PassthroughFileObject(BaseModel): + id: str + object: str | None = None + purpose: str | None = None + filename: str | None = None + bytes: int | None = None + + +class PassthroughFileDeleted(BaseModel): + id: str + deleted: bool + + +class PassthroughListEntry(BaseModel): + id: str + + +class ResponsesUsage(BaseModel): + input_tokens: int + output_tokens: int + + +class ResponsesObject(BaseModel): + id: str + usage: ResponsesUsage | None = None + + +class ResponsesStreamEvent(BaseModel): + """One SSE frame of a native Responses stream. Only the terminal frames carry a + `response`, so it stays optional and the deltas validate as themselves.""" + + type: str + response: ResponsesObject | None = None + + +def completed_responses_object(result: StreamingResponse) -> ResponsesObject | None: + """The `response.completed` frame's response object, or None if the stream never + completed. Its `id` is what the spend row is keyed by on this route, and its + usage is what the row is priced from.""" + events = ( + ResponsesStreamEvent.model_validate_json(payload) + for payload in result.stream_events + ) + completed = tuple( + event.response + for event in events + if event.type == "response.completed" and event.response is not None + ) + return completed[-1] if completed else None + + +class OpenAIResponsesBody(BaseModel): + model: str + input: str + stream: bool = False + + +class OpenAIEmbeddingBody(BaseModel): + model: str + input: str + + +class PassthroughBatchList(BaseModel): + """OpenAI's own batch page, relayed verbatim. `object` is required so a body + that is not an OpenAI list fails validation instead of passing vacuously.""" + + object: str + data: list[PassthroughListEntry] + + def _tags_header(tags: list[str] | None) -> str | None: return ",".join(tags) if tags else None @@ -196,6 +266,66 @@ class PassthroughClient: stream=stream, ) + # ---- OpenAI file/batch routes under /openai_passthrough ------------- + # + # Relayed to OpenAI untouched, which is the whole point of the prefix: the + # customer opts out of the gateway's managed-file handling here. + + def openai_passthrough_upload_file( + self, key: str, *, content: bytes, filename: str + ) -> Result[PassthroughFileObject]: + return self.proxy.transport.upload( + "/openai_passthrough/v1/files", + headers=self.proxy.transport.bearer(key), + form=FileUploadForm(purpose="batch"), + filename=filename, + content=content, + response_type=PassthroughFileObject, + ) + + def openai_passthrough_delete_file( + self, key: str, file_id: str + ) -> Result[PassthroughFileDeleted]: + return self.proxy.transport.delete( + f"/openai_passthrough/v1/files/{file_id}", + headers=self.proxy.transport.bearer(key), + json=NoBody(), + response_type=PassthroughFileDeleted, + ) + + def openai_passthrough_list_batches(self, key: str) -> Result[PassthroughBatchList]: + return self.proxy.transport.get( + "/openai_passthrough/v1/batches", + headers=self.proxy.transport.bearer(key), + params=NoBody(), + response_type=PassthroughBatchList, + ) + + # ---- OpenAI inference routes under /openai_passthrough ------------- + # + # Relayed to OpenAI verbatim, but still costed by the gateway: the customer + # budgets against this traffic, so a 200 that logs no spend is money the + # gateway never sees. + + def openai_passthrough_responses( + self, key: str, model: str, text: str, *, stream: bool = False + ) -> StreamingResponse: + return self.proxy.transport.send( + "/openai_passthrough/v1/responses", + headers=self.proxy.transport.bearer(key), + json=OpenAIResponsesBody(model=model, input=text, stream=stream), + stream=stream, + ) + + def openai_passthrough_embed( + self, key: str, model: str, text: str + ) -> StreamingResponse: + return self.proxy.transport.send( + "/openai_passthrough/v1/embeddings", + headers=self.proxy.transport.bearer(key), + json=OpenAIEmbeddingBody(model=model, input=text), + ) + def openai_chat( self, key: str, model: str, text: str, *, max_completion_tokens: int = 64 ) -> StreamingResponse: diff --git a/tests/e2e/llm_translation/test_passthrough_e2e.py b/tests/e2e/llm_translation/test_passthrough_e2e.py index b57164df9bb..b084c711a88 100644 --- a/tests/e2e/llm_translation/test_passthrough_e2e.py +++ b/tests/e2e/llm_translation/test_passthrough_e2e.py @@ -13,8 +13,8 @@ A passthrough call returning non-2xx fails hard (never a skip); once it returns import pytest -from e2e_config import unique_marker -from e2e_http import StreamingResponse, require_successful_call +from e2e_config import CHEAP_OPENAI_MODEL, unique_marker +from e2e_http import StreamingResponse, require_successful_call, unwrap from lifecycle import ResourceManager from models import KeyGenerateBody, SpendLogRow from passthrough_client import ( @@ -24,8 +24,11 @@ from passthrough_client import ( JsonSchema, JsonSchemaProperty, PassthroughClient, + completed_responses_object, ) +EMBEDDING_MODEL = "text-embedding-3-small" + pytestmark = pytest.mark.e2e @@ -210,3 +213,120 @@ class TestPassthroughModelAllowlist: "a key restricted to gemini-2.5-flash must be denied a claude passthrough call, " f"got {result.status_code}: {result.body[:300]}" ) + + +class TestOpenAIPassthroughPrefix: + """The dedicated `/openai_passthrough` prefix must reach OpenAI, not be + swallowed by the provider-scoped `/{provider}/v1/...` routes. + + The customer fronts OpenAI's own file and batch APIs through this prefix + precisely to opt out of the gateway's managed-file handling. `/v1/files` and + `/v1/batches` also answer `/{provider}/v1/files` and `/{provider}/v1/batches`, + so `openai_passthrough` used to bind as a provider name and the request died + inside the gateway with a provider-lookup error, never reaching OpenAI. + """ + + @pytest.mark.covers("llm.files.openai.passthrough.nonstream.works") + def test_passthrough_prefix_uploads_a_file_to_openai( + self, client: PassthroughClient, resources: ResourceManager, scoped_key: str + ) -> None: + content = f'{{"marker":"{unique_marker()}"}}\n'.encode() + uploaded = unwrap( + client.openai_passthrough_upload_file( + scoped_key, content=content, filename="e2e-passthrough-batch.jsonl" + ) + ) + resources.defer( + lambda: client.openai_passthrough_delete_file(scoped_key, uploaded.id) + ) + + assert uploaded.object == "file", ( + f"/openai_passthrough/v1/files did not relay OpenAI's file object: {uploaded}" + ) + assert uploaded.purpose == "batch" + assert uploaded.bytes == len(content) + + @pytest.mark.covers("llm.batches.openai.passthrough.nonstream.works") + def test_passthrough_prefix_lists_batches_from_openai( + self, client: PassthroughClient, scoped_key: str + ) -> None: + listed = unwrap(client.openai_passthrough_list_batches(scoped_key)) + + assert listed.object == "list", ( + f"/openai_passthrough/v1/batches did not relay OpenAI's batch page: {listed}" + ) + + +class TestOpenAIPassthroughSpend: + """A call relayed to OpenAI's own endpoints must still be costed. + + The customer routes native OpenAI traffic through `/openai_passthrough` and + budgets against it, so a call that returns 200 while logging no spend is money + the gateway never sees and a budget that never trips. Streamed Responses calls + and embeddings each used to land exactly that way, on separate code paths. + """ + + @pytest.mark.covers("llm.responses.openai.passthrough.stream.cost_logged") + def test_streamed_responses_call_logs_its_cost( + self, client: PassthroughClient, scoped_key: str + ) -> None: + result = client.openai_passthrough_responses( + scoped_key, + CHEAP_OPENAI_MODEL, + f"Say hi in one word. {unique_marker()}", + stream=True, + ) + require_successful_call(result) + assert result.chunks > 0, "streamed responses passthrough produced no events" + + completed = completed_responses_object(result) + assert completed is not None, ( + f"the stream never delivered a response.completed frame, so there is no " + f"provider id to reconcile against: last events {result.stream_events[-3:]}" + ) + assert completed.usage is not None, ( + f"the completed response carried no usage to price from: {completed}" + ) + + rows = client.proxy.poll_logs_for_request_id( + completed.id, predicate=lambda rows: (rows[0].spend or 0) > 0 + ) + assert rows, ( + f"no spend row for the response the customer was served ({completed.id}); " + "a streamed passthrough call OpenAI bills them for is invisible to the " + "gateway's own spend and budgets" + ) + row = rows[0] + assert (row.spend or 0) > 0, f"streamed responses passthrough was not costed: {row}" + assert row.prompt_tokens == completed.usage.input_tokens, ( + f"logged {row.prompt_tokens} prompt tokens, the response the customer read " + f"reported {completed.usage.input_tokens}" + ) + assert row.completion_tokens == completed.usage.output_tokens, ( + f"logged {row.completion_tokens} completion tokens, the response the customer " + f"read reported {completed.usage.output_tokens}" + ) + + @pytest.mark.covers("llm.embeddings.openai.passthrough.nonstream.cost_logged") + def test_embeddings_call_logs_its_cost( + self, client: PassthroughClient, scoped_key: str + ) -> None: + result = client.openai_passthrough_embed( + scoped_key, EMBEDDING_MODEL, f"cost this sentence {unique_marker()}" + ) + require_successful_call(result) + assert result.call_id, "embeddings passthrough returned no x-litellm-call-id" + + rows = client.proxy.poll_logs_for_request_id( + result.call_id, predicate=lambda rows: (rows[0].spend or 0) > 0 + ) + assert rows, ( + f"no spend row for embeddings call {result.call_id}; the customer is billed " + "by OpenAI for tokens the gateway never counted against their budget" + ) + row = rows[0] + assert (row.spend or 0) > 0, f"embeddings passthrough was not costed: {row}" + assert (row.prompt_tokens or 0) > 0, ( + f"the embeddings row logged no prompt tokens, so whatever cost it carries " + f"was not computed from the real usage: {row}" + ) From 4f04e59ca036860ed1a83eb6169b4bdd20b31bb1 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 03:05:48 -0700 Subject: [PATCH 044/220] fix: harden vertex live passthrough against client model forms and dict credentials - accept the Live SDK's models/ and LiteLLM's vertex_ai/ when rewriting the setup model - keep a dict service account intact instead of stringifying it - treat same-target deployments holding different credentials as ambiguous - guard both websocket states before every close so a second close cannot raise - build the sendable close codes from the public CloseCode enum --- .../llm_passthrough_endpoints.py | 40 +++++-- .../pass_through_endpoints.py | 32 ++++-- .../passthrough_endpoint_router.py | 28 ++++- .../test_llm_pass_through_endpoints.py | 107 ++++++++++++++++++ .../test_pass_through_endpoints.py | 32 ++++++ .../test_passthrough_endpoint_router.py | 53 +++++++++ 6 files changed, 266 insertions(+), 26 deletions(-) diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index c2b221b1f7a..7ce41c1d5b6 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -2384,6 +2384,21 @@ VERTEX_LIVE_UNCONFIGURED_CLOSE_REASON: Final = ( VERTEX_PUBLISHER_MODEL_PREFIX: Final = "publishers/google/models/" +VERTEX_PUBLISHERS_SEGMENT: Final = "publishers/" + + +def _vertex_publisher_model_suffix(model: str) -> str: + """ + Turn whatever the client named into the ``publishers//models/`` tail of a Vertex resource name. + + Clients send bare ids, LiteLLM ids (``vertex_ai/gemini-live-2.5-flash``), and the Live SDK's ``models/``, + and a publisher model id never contains a slash, so anything ahead of the last one is addressing, not identity + """ + publishers_at: Final = model.find(VERTEX_PUBLISHERS_SEGMENT) + if publishers_at != -1: + return model[publishers_at:] + return f"{VERTEX_PUBLISHER_MODEL_PREFIX}{model.rsplit('/', 1)[-1]}" + def _get_llm_router() -> "Router | None": from litellm.proxy.proxy_server import llm_router @@ -2397,8 +2412,12 @@ def _resolve_vertex_live_credentials( model: str | None, ) -> VertexPassThroughCredentials | None: """ - Resolution order: an explicit project/location registration or ``default_vertex_config``, then any DB model - entry flagged ``use_in_pass_through``, then the ``DEFAULT_VERTEXAI_*`` env vars + Resolution order: an explicit project/location registration, then ``default_vertex_config`` (which the proxy + fills from the ``DEFAULT_VERTEXAI_*`` env vars whenever the yaml leaves it out), then any DB model entry + flagged ``use_in_pass_through``. + + DB entries come last on purpose: an operator who set a global default already said which project + pass-through traffic should bill to, and this route silently ignoring that would be the worse surprise """ keyed: Final = passthrough_endpoint_router.get_vertex_credentials( project_id=vertex_project, @@ -2436,22 +2455,23 @@ def _build_vertex_live_setup_model_rewriter( if setup_model.startswith("projects/"): return setup_model aliased: Final = _resolve_alias_to_upstream_model(setup_model, llm_router) - return ( - f"projects/{vertex_project}/locations/{vertex_location}/" - f"{VERTEX_PUBLISHER_MODEL_PREFIX}{aliased.removeprefix(VERTEX_PUBLISHER_MODEL_PREFIX)}" - ) + return f"projects/{vertex_project}/locations/{vertex_location}/{_vertex_publisher_model_suffix(aliased)}" return rewrite def _resolve_alias_to_upstream_model(setup_model: str, llm_router: "Router | None") -> str: + """ + The Live SDK wraps whatever the caller typed as ``models/``, so a gateway alias arrives prefixed + """ if llm_router is None: return setup_model + candidates: Final = (setup_model, setup_model.rsplit("/", 1)[-1]) upstream: Final = next( ( - deployment["litellm_params"]["model"] + deployment["litellm_params"].get("model") for deployment in (llm_router.get_model_list() or ()) - if deployment.get("model_name") == setup_model + if deployment.get("model_name") in candidates ), None, ) @@ -2500,9 +2520,7 @@ async def vertex_ai_live_websocket_passthrough( vertex_credentials_config.vertex_location if vertex_credentials_config is not None else None ) credentials_value: Final = ( - str(vertex_credentials_config.vertex_credentials) - if vertex_credentials_config is not None and vertex_credentials_config.vertex_credentials is not None - else None + vertex_credentials_config.vertex_credentials if vertex_credentials_config is not None else None ) try: diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py index 5608e8b384d..d45421489e7 100644 --- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py @@ -32,7 +32,7 @@ from websockets.exceptions import ( ConnectionClosedOK, InvalidStatus, ) -from websockets.frames import EXTERNAL_CLOSE_CODES, Close +from websockets.frames import Close, CloseCode import litellm from litellm._logging import verbose_proxy_logger @@ -1928,11 +1928,26 @@ def _truncated_close_reason(reason: str) -> str: return encoded[:WEBSOCKET_CLOSE_REASON_MAX_BYTES].decode("utf-8", errors="ignore") +SENDABLE_CLOSE_CODES: Final = frozenset(CloseCode) - frozenset( + {CloseCode.NO_STATUS_RCVD, CloseCode.ABNORMAL_CLOSURE, CloseCode.TLS_HANDSHAKE} +) + + +def _client_socket_is_open(websocket: WebSocket) -> bool: + """ + Starlette tracks the two halves separately and raises on a second close, so both have to still be live + """ + return ( + websocket.client_state != WebSocketState.DISCONNECTED + and websocket.application_state != WebSocketState.DISCONNECTED + ) + + def _upstream_close_to_relay(task_results: Iterable[object]) -> Close | None: """ The upstream close worth telling the client about: anything other than a plain, reasonless normal close. - Codes outside ``EXTERNAL_CLOSE_CODES`` and the private range never travel on the wire (1006 for a socket that + Codes outside ``SENDABLE_CLOSE_CODES`` and the private range never travel on the wire (1006 for a socket that died without a close frame, 1005 for one that sent no code), so relaying them would build an invalid frame """ upstream_close: Final = next((result for result in task_results if isinstance(result, Close)), None) @@ -1940,7 +1955,7 @@ def _upstream_close_to_relay(task_results: Iterable[object]) -> Close | None: return None if upstream_close.code == 1000 and upstream_close.reason == "": return None - if upstream_close.code not in EXTERNAL_CLOSE_CODES and not 3000 <= upstream_close.code < 5000: + if upstream_close.code not in SENDABLE_CLOSE_CODES and not 3000 <= upstream_close.code < 5000: return None return upstream_close @@ -2268,7 +2283,7 @@ async def websocket_passthrough_request( raise exception upstream_close: Final = _upstream_close_to_relay(task.result() for task in done) - if upstream_close is not None and websocket.application_state != WebSocketState.DISCONNECTED: + if upstream_close is not None and _client_socket_is_open(websocket): await websocket.close( code=upstream_close.code, reason=_truncated_close_reason(upstream_close.reason), @@ -2359,7 +2374,7 @@ async def websocket_passthrough_request( ), ) - if websocket.client_state != WebSocketState.DISCONNECTED: + if _client_socket_is_open(websocket): await websocket.close( code=getattr(exc, "status_code", 1011), reason="Upstream connection rejected", @@ -2387,13 +2402,10 @@ async def websocket_passthrough_request( ), ) - if websocket.client_state != WebSocketState.DISCONNECTED: + if _client_socket_is_open(websocket): await websocket.close(code=1011, reason="WebSocket passthrough error") finally: - if ( - websocket.client_state != WebSocketState.DISCONNECTED - and websocket.application_state != WebSocketState.DISCONNECTED - ): + if _client_socket_is_open(websocket): await websocket.close() diff --git a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py index 28067c842cd..7fd607fc4d0 100644 --- a/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py +++ b/litellm/proxy/pass_through_endpoints/passthrough_endpoint_router.py @@ -1,3 +1,4 @@ +import json from collections.abc import Callable from typing import TYPE_CHECKING, Final @@ -27,6 +28,15 @@ def _get_str_value(values: dict[str, object] | None, key: str) -> str | None: return value if isinstance(value, str) else None +def _credential_identity(credentials: VERTEX_CREDENTIALS_TYPES | None) -> str | None: + """ + A hashable stand-in for a credential, so two deployments can be compared for holding the same one + """ + if isinstance(credentials, dict): + return json.dumps(credentials, sort_keys=True) + return credentials + + class PassthroughEndpointRouter: """ Use this class to Get credentials for pass-through endpoints @@ -127,8 +137,8 @@ class PassthroughEndpointRouter: ``deployment_key_to_vertex_credentials`` is only reachable when the caller names a project and location, which WebSocket clients never do, so DB-stored deployments need this lookup to be usable at all. - With no model to go on, only deployments that agree on a project and location answer: guessing between - two Vertex projects would mint a token for one and later send the other one's model name + With no model to go on, only deployments that agree on a project, a location, and a credential answer: + guessing between two Vertex projects would mint a token for one and later send the other one's model name """ llm_router: Final = self.llm_router_getter() if llm_router is None: @@ -149,7 +159,12 @@ class PassthroughEndpointRouter: if matched is not None: return matched targets: Final = frozenset( - (credentials.vertex_project, credentials.vertex_location) for _, credentials in resolved + ( + credentials.vertex_project, + credentials.vertex_location, + _credential_identity(credentials.vertex_credentials), + ) + for _, credentials in resolved ) if len(targets) != 1: return None @@ -172,9 +187,12 @@ class PassthroughEndpointRouter: vertex_location: Final = _get_str_value(credential_values, "vertex_location") or litellm_params.get( "vertex_location" ) - vertex_credentials: Final = _get_str_value(credential_values, "vertex_credentials") or litellm_params.get( - "vertex_credentials" + stored_credentials: Final = ( + credential_values.get("vertex_credentials") if credential_values is not None else None ) + vertex_credentials: Final = ( + stored_credentials if isinstance(stored_credentials, (str, dict)) else None + ) or litellm_params.get("vertex_credentials") if vertex_project is None or vertex_location is None: return None return VertexPassThroughCredentials( diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py index 5ed974c7a47..f994fba371b 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py @@ -3889,6 +3889,9 @@ class TestComprehendMedicalProxyRoute: assert exc_info.value.status_code == 400 +LIVE_RESOURCE_PATH = "projects/proj-db/locations/global/publishers/google/models/gemini-live-2.5-flash" + + class TestVertexAILiveWebsocketPassthrough: def _websocket(self): from starlette.websockets import WebSocketState @@ -3959,6 +3962,110 @@ class TestVertexAILiveWebsocketPassthrough: ) websocket.close.assert_not_awaited() + @pytest.mark.parametrize( + "setup_model, expected", + [ + ("gemini-live-2.5-flash", LIVE_RESOURCE_PATH), + ("models/gemini-live-2.5-flash", LIVE_RESOURCE_PATH), + ("vertex_ai/gemini-live-2.5-flash", LIVE_RESOURCE_PATH), + ("gemini-live", LIVE_RESOURCE_PATH), + ("models/gemini-live", LIVE_RESOURCE_PATH), + ( + "publishers/meta/models/llama-3.3-70b-instruct-maas", + "projects/proj-db/locations/global/publishers/meta/models/llama-3.3-70b-instruct-maas", + ), + ( + "projects/other/locations/us-central1/publishers/google/models/gemini-2.0-flash", + "projects/other/locations/us-central1/publishers/google/models/gemini-2.0-flash", + ), + ], + ) + def test_setup_model_rewriter_normalises_the_forms_clients_send(self, setup_model, expected): + from litellm.proxy.pass_through_endpoints import ( + llm_passthrough_endpoints as passthrough_module, + ) + + llm_router = litellm.Router( + model_list=[ + { + "model_name": "gemini-live", + "litellm_params": { + "model": "vertex_ai/gemini-live-2.5-flash", + "use_in_pass_through": True, + "vertex_project": "proj-db", + "vertex_location": "global", + }, + } + ] + ) + + rewriter = passthrough_module._build_vertex_live_setup_model_rewriter( + vertex_project="proj-db", + vertex_location="global", + llm_router=llm_router, + ) + + assert rewriter is not None + assert rewriter(setup_model) == expected + + @pytest.mark.asyncio + async def test_default_vertex_config_outranks_db_deployment(self, monkeypatch): + from litellm.proxy.pass_through_endpoints import ( + llm_passthrough_endpoints as passthrough_module, + ) + from litellm.types.passthrough_endpoints.vertex_ai import ( + VertexPassThroughCredentials, + ) + + llm_router = litellm.Router( + model_list=[ + { + "model_name": "gemini-live", + "litellm_params": { + "model": "vertex_ai/gemini-live-2.5-flash", + "use_in_pass_through": True, + "vertex_project": "proj-db", + "vertex_location": "global", + "vertex_credentials": '{"type": "db_account"}', + }, + } + ] + ) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", llm_router) + monkeypatch.setattr( + passthrough_module.passthrough_endpoint_router, + "default_vertex_config", + VertexPassThroughCredentials( + vertex_project="proj-env", + vertex_location="global", + vertex_credentials='{"type": "env_account"}', + ), + ) + self._clear_vertex_env(monkeypatch) + websocket = self._websocket() + ensure_token = AsyncMock(return_value=("token-abc", "proj-env")) + ws_passthrough = AsyncMock() + + with ( + patch.object(passthrough_module.vertex_llm_base, "_ensure_access_token_async", ensure_token), + patch.object(passthrough_module, "websocket_passthrough_request", ws_passthrough), + ): + await passthrough_module.vertex_ai_live_websocket_passthrough( + websocket=websocket, + model="gemini-live", + user_api_key_dict=UserAPIKeyAuth(), + ) + + ensure_token.assert_awaited_once_with( + credentials='{"type": "env_account"}', + project_id="proj-env", + custom_llm_provider="vertex_ai_beta", + ) + rewriter = ws_passthrough.await_args.kwargs["setup_model_rewriter"] + assert rewriter("gemini-live") == ( + "projects/proj-env/locations/global/publishers/google/models/gemini-live-2.5-flash" + ) + @pytest.mark.asyncio async def test_credential_failure_close_names_configuration_options(self, monkeypatch): from litellm.proxy.pass_through_endpoints import ( diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index 2663396bf53..00097166c13 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -5243,6 +5243,38 @@ async def test_websocket_passthrough_leaves_full_resource_setup_model_untouched( ) +@pytest.mark.asyncio +async def test_websocket_passthrough_does_not_close_twice_when_success_logging_fails(): + from websockets.exceptions import ConnectionClosedError + from websockets.frames import Close + + upstream_reason = "Publisher Model `projects/p/locations/global/publishers/google/models/nope` was not found" + upstream_ws = ClosingUpstreamWebSocket( + ConnectionClosedError(rcvd=Close(1008, upstream_reason), sent=Close(1008, ""), rcvd_then_sent=True) + ) + websocket = _client_websocket(_pending_receive) + + with ( + _patched_websocket_passthrough_environment(upstream_ws), + patch( + "litellm.proxy.pass_through_endpoints.pass_through_endpoints." + "GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue", + side_effect=RuntimeError("logging worker down"), + ), + ): + await websocket_passthrough_request( + websocket=websocket, + target="wss://aiplatform.googleapis.com/ws/google.cloud.aiplatform.v1.LlmBidiService/BidiGenerateContent", + custom_headers={"Authorization": "Bearer token"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/vertex_ai/live", + accept_websocket=False, + ) + + websocket.close.assert_awaited_once_with(code=1008, reason=upstream_reason) + + def _passthrough_kwargs_for_reservation( user_api_key_dict: UserAPIKeyAuth, parsed_body: Optional[dict] = None ) -> dict: diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py index f86bfb8bb1b..e3cbc2d507f 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_passthrough_endpoint_router.py @@ -293,6 +293,59 @@ def test_vertex_without_hint_falls_back_when_deployments_share_a_target(): assert resolved is not None and resolved.vertex_project == "proj-one" +def test_vertex_without_hint_refuses_to_guess_between_service_accounts(): + llm_router = litellm.Router( + model_list=[ + _vertex_deployment( + "gemini-flash", + "vertex_ai/gemini-2.5-flash", + vertex_project="proj-one", + vertex_location="global", + vertex_credentials='{"client_email": "flash@proj-one.iam"}', + ), + _vertex_deployment( + "gemini-live", + "vertex_ai/gemini-live-2.5-flash", + vertex_project="proj-one", + vertex_location="global", + vertex_credentials='{"client_email": "live@proj-one.iam"}', + ), + ] + ) + passthrough_router = _passthrough_router(llm_router) + + assert passthrough_router.get_vertex_credentials_from_router_deployments(model=None) is None + + +def test_vertex_named_credential_keeps_dict_service_account(): + service_account = {"type": "service_account", "client_email": "live@proj-db.iam"} + CredentialAccessor.upsert_credentials( + [ + _vertex_credential( + "cred_gcp_dict", + { + "vertex_project": "proj-db", + "vertex_location": "global", + "vertex_credentials": service_account, + }, + ) + ] + ) + llm_router = litellm.Router( + model_list=[ + _vertex_deployment( + "gemini-live", "vertex_ai/gemini-live-2.5-flash", litellm_credential_name="cred_gcp_dict" + ) + ] + ) + passthrough_router = _passthrough_router(llm_router) + + resolved = passthrough_router.get_vertex_credentials_from_router_deployments(model=None) + + assert resolved is not None + assert resolved.vertex_credentials == service_account + + def test_no_flagged_vertex_deployment_returns_none(): llm_router = litellm.Router( model_list=[ From aa8e7278e3d27a483ef8039995bce649dc6c0d88 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 03:08:05 -0700 Subject: [PATCH 045/220] test(e2e): drop the passthrough streaming-cost test, it needs a config flag The final streaming usage frame only carries usage.cost when the proxy runs with litellm_settings.include_cost_in_streaming_usage: true, and that flag is readable only off the module-level litellm setting. There is no header, key, or management route that turns it on per request, so a test cannot ask the shared e2e proxy for it, and the proxy's config does not live in this repo. The registry row stays as an uncovered gap with the reason recorded, rather than being deleted, so the behavior is still on the list of things we want covered once the gateway config is reachable. The StreamOptions model, ChatBody.stream_options, Usage.cost, and AnthropicMessagesResponse.id existed only for that test, so they go with it. --- .../coverage_registry/quota_management.yaml | 2 +- tests/e2e/models.py | 14 ---- .../test_passthrough_stream_cost_e2e.py | 68 ------------------- 3 files changed, 1 insertion(+), 83 deletions(-) delete mode 100644 tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py diff --git a/tests/e2e/coverage_registry/quota_management.yaml b/tests/e2e/coverage_registry/quota_management.yaml index 5438ed8534a..98a45eefb2d 100644 --- a/tests/e2e/coverage_registry/quota_management.yaml +++ b/tests/e2e/coverage_registry/quota_management.yaml @@ -50,4 +50,4 @@ - {id: quota_management.spend_tracking.messages_bridge.keeps_cache_tokens, module: quota_management, tier: P1, behavior: spend_tracking, variant: messages_bridge, assertions: [keeps_cache_tokens], exercised_on: [messages], source: "llms/anthropic/experimental_pass_through/responses_adapters/handler.py", rationale: "A /v1/messages request served by a Responses-only OpenAI model keeps its cache-read tokens and their discounted billing across the bridge (#34957)"} - {id: quota_management.spend_tracking.service_tier.bills_tier_rates, module: quota_management, tier: P1, behavior: spend_tracking, variant: service_tier, assertions: [bills_tier_rates], exercised_on: [chat_completions], source: "cost_calculator.py", rationale: "A priority service_tier call bills input, output, and reasoning at the deployment's *_priority rates and records the tier on the row (#35923, #35925)"} - {id: quota_management.spend_tracking.cost_headers.additive_components, module: quota_management, tier: P1, behavior: spend_tracking, variant: cost_headers, assertions: [additive_components], exercised_on: [chat_completions], source: "proxy/common_request_processing.py", rationale: "The x-litellm-response-cost-* component headers sum to the total, input covers only fresh tokens, and reasoning stays a subset of output (#36965)"} -- {id: quota_management.spend_tracking.passthrough_stream.injects_usage_cost, module: quota_management, tier: P1, behavior: spend_tracking, variant: passthrough_stream, assertions: [injects_usage_cost], exercised_on: [openai_passthrough], source: "proxy/pass_through_endpoints/streaming_handler.py", rationale: "With include_cost_in_streaming_usage on, the /openai passthrough's final streaming usage frame carries the proxy-computed cost (#36503)"} +- {id: quota_management.spend_tracking.passthrough_stream.injects_usage_cost, module: quota_management, tier: P1, behavior: spend_tracking, variant: passthrough_stream, assertions: [injects_usage_cost], exercised_on: [openai_passthrough], source: "proxy/pass_through_endpoints/streaming_handler.py", rationale: "With include_cost_in_streaming_usage on, the /openai passthrough's final streaming usage frame carries the proxy-computed cost (#36503). Uncovered: the flag is only settable in litellm_settings, and the shared e2e proxy's config is not in this repo"} diff --git a/tests/e2e/models.py b/tests/e2e/models.py index fe93b13e0a4..f1d5733253a 100644 --- a/tests/e2e/models.py +++ b/tests/e2e/models.py @@ -216,19 +216,10 @@ class McpChatTool(BaseModel): allowed_tools: list[str] | None = None -class StreamOptions(BaseModel): - """OpenAI `stream_options`: `include_usage` asks for a final usage-only SSE - frame, which is where the proxy's `include_cost_in_streaming_usage` setting - injects `usage.cost`.""" - - include_usage: bool = True - - class ChatBody(BaseModel): model: str messages: list[ChatMessage] stream: bool = False - stream_options: StreamOptions | None = None max_tokens: int | None = None max_completion_tokens: int | None = None temperature: float | None = None @@ -331,9 +322,6 @@ class CompletionTokensDetails(BaseModel): class Usage(BaseModel): - """`cost` exists only on streaming usage frames from a proxy running with - `include_cost_in_streaming_usage: true`; providers never send it.""" - prompt_tokens: int | None = None completion_tokens: int | None = None total_tokens: int | None = None @@ -341,7 +329,6 @@ class Usage(BaseModel): cache_creation_input_tokens: int | None = None prompt_tokens_details: PromptTokensDetails | None = None completion_tokens_details: CompletionTokensDetails | None = None - cost: float | None = None class ChatResponse(BaseModel): @@ -462,7 +449,6 @@ class AnthropicMessagesResponse(BaseModel): for triage.""" model_config = ConfigDict(extra="allow") - id: str | None = None model: str | None = None content: list[AnthropicContentBlock] | None = None choices: list[ChatChoice] | None = None diff --git a/tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py b/tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py deleted file mode 100644 index 4c3a2a4509f..00000000000 --- a/tests/e2e/quota_management/spend_tracking/test_passthrough_stream_cost_e2e.py +++ /dev/null @@ -1,68 +0,0 @@ -"""Live e2e: the /openai passthrough injects usage.cost into streaming usage frames. - -Pins #36503: with the proxy running `include_cost_in_streaming_usage: true`, a -streamed call through the provider passthrough surface must carry the computed -cost inside the final usage-only SSE frame, the same contract the native -/chat/completions stream has. Providers never send `cost` themselves, so a -nonzero value proves the proxy computed and injected it on the passthrough path. - -The row-side spend accounting for passthrough calls is covered elsewhere; this -test pins only the in-stream cost surface, which clients read without ever -touching /spend/logs. -""" - -import pytest -from pydantic import BaseModel - -from e2e_config import unique_marker -from models import ChatBody, ChatMessage, StreamOptions, Usage -from spend_e2e_client import SpendClient - -pytestmark = pytest.mark.e2e - -OPENAI_MODEL = "gpt-5.6-luna" - - -class _StreamFrame(BaseModel): - usage: Usage | None = None - - -class TestPassthroughStreamCost: - @pytest.mark.covers("quota_management.spend_tracking.passthrough_stream.injects_usage_cost") - def test_passthrough_stream_final_usage_frame_carries_cost( - self, client: SpendClient, scoped_key: str - ) -> None: - result = client.proxy.transport.send( - "/openai/v1/chat/completions", - headers=client.proxy.transport.bearer(scoped_key), - json=ChatBody( - model=OPENAI_MODEL, - messages=[ - ChatMessage( - role="user", - content=f"{unique_marker()} Reply with the single word passthrough.", - ) - ], - stream=True, - stream_options=StreamOptions(), - ), - stream=True, - ) - assert result.ok and result.stream_events, ( - f"passthrough stream failed (status {result.status_code}): {result.body[:300]}" - ) - - usage_frames = [ - frame.usage - for frame in (_StreamFrame.model_validate_json(event) for event in result.stream_events) - if frame.usage is not None - ] - assert usage_frames, ( - f"no usage frame in the passthrough stream despite stream_options.include_usage; " - f"last event: {result.stream_events[-1][:300]}" - ) - - final_usage = usage_frames[-1] - assert final_usage.cost is not None and final_usage.cost > 0, ( - f"final passthrough usage frame carries no injected cost: {final_usage}" - ) From 69278ae37bba01a5ad19a83273719794c5b47c91 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 03:25:31 -0700 Subject: [PATCH 046/220] docs(e2e): correct the passthrough-stream registry row's uncovered reason --- tests/e2e/coverage_registry/quota_management.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/e2e/coverage_registry/quota_management.yaml b/tests/e2e/coverage_registry/quota_management.yaml index 98a45eefb2d..e4b7e755bd0 100644 --- a/tests/e2e/coverage_registry/quota_management.yaml +++ b/tests/e2e/coverage_registry/quota_management.yaml @@ -50,4 +50,4 @@ - {id: quota_management.spend_tracking.messages_bridge.keeps_cache_tokens, module: quota_management, tier: P1, behavior: spend_tracking, variant: messages_bridge, assertions: [keeps_cache_tokens], exercised_on: [messages], source: "llms/anthropic/experimental_pass_through/responses_adapters/handler.py", rationale: "A /v1/messages request served by a Responses-only OpenAI model keeps its cache-read tokens and their discounted billing across the bridge (#34957)"} - {id: quota_management.spend_tracking.service_tier.bills_tier_rates, module: quota_management, tier: P1, behavior: spend_tracking, variant: service_tier, assertions: [bills_tier_rates], exercised_on: [chat_completions], source: "cost_calculator.py", rationale: "A priority service_tier call bills input, output, and reasoning at the deployment's *_priority rates and records the tier on the row (#35923, #35925)"} - {id: quota_management.spend_tracking.cost_headers.additive_components, module: quota_management, tier: P1, behavior: spend_tracking, variant: cost_headers, assertions: [additive_components], exercised_on: [chat_completions], source: "proxy/common_request_processing.py", rationale: "The x-litellm-response-cost-* component headers sum to the total, input covers only fresh tokens, and reasoning stays a subset of output (#36965)"} -- {id: quota_management.spend_tracking.passthrough_stream.injects_usage_cost, module: quota_management, tier: P1, behavior: spend_tracking, variant: passthrough_stream, assertions: [injects_usage_cost], exercised_on: [openai_passthrough], source: "proxy/pass_through_endpoints/streaming_handler.py", rationale: "With include_cost_in_streaming_usage on, the /openai passthrough's final streaming usage frame carries the proxy-computed cost (#36503). Uncovered: the flag is only settable in litellm_settings, and the shared e2e proxy's config is not in this repo"} +- {id: quota_management.spend_tracking.passthrough_stream.injects_usage_cost, module: quota_management, tier: P1, behavior: spend_tracking, variant: passthrough_stream, assertions: [injects_usage_cost], exercised_on: [openai_passthrough], source: "proxy/pass_through_endpoints/streaming_handler.py", rationale: "With include_cost_in_streaming_usage on, the /openai passthrough's final streaming usage frame carries the proxy-computed cost (#36503). Uncovered: the flag is only settable in litellm_settings, and the shared e2e stack does not turn it on yet"} From fe11202c2df1a6373ed3230e22847a76e157cb0d Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 03:28:28 -0700 Subject: [PATCH 047/220] fix(cli): write the logout note again when the file holding it had to go When a full disk refuses the replacement file and a read-only token file refuses the rewrite in place, the only way left to get the secret off disk is to remove the file carrying it. That file was also the note saying the keychain went unchecked, so its absence made the next logout read a keychain that was never confirmed as one already known to be clean. Removing it is what frees the room the replacement was refused for, so the note is written again on the way out and the logout after this one still warns. --- litellm/litellm_core_utils/cli_token_utils.py | 32 ++++++++++-- .../test_cli_token_utils.py | 52 ++++++++++++++++++- 2 files changed, 79 insertions(+), 5 deletions(-) diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index 2f45742ce03..69a77a36882 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -173,7 +173,8 @@ def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretClear: what lets a later run tell a machine with a credential it cannot reach apart from one that never had a login at all, and taking it away would leave the next logout answering the warning this one just issued with a false all-clear. The secret goes either way, and a file that will give up - neither its copy nor itself outranks whatever the keychain had to say. + neither its copy nor itself is removed rather than kept, with the note written again afterwards + so the warning still outlives this run. """ outcome: Final = vault.erase() record: Final = _read_token_file() @@ -183,6 +184,8 @@ def clear_cli_token(*, vault: SecretVault = SYSTEM_KEYRING) -> SecretClear: removal: Final = _remove_token_file() if removal is not None and record is not None and not _scrub_file_secret(record): return removal + if removal is None and record is not None and _the_keychain_went_unchecked(outcome): + _write_the_note_the_removal_took_with_it(record) return SecretErased() if settled else outcome @@ -194,6 +197,19 @@ def _remove_token_file() -> CredentialNotCleared | None: return None +def _write_the_note_the_removal_took_with_it(record: CliTokenRecord) -> None: + """Put the secret-free note back after the file carrying it had to go to get the secret off disk. + + Reaching here means neither rewrite would take, so the file went instead, and its absence is + what the next logout would read as a keychain already known to be clean. Removing it is also + what frees the room the rewrite was refused for, so the note usually lands on this second try. + When it does not, the warning this logout printed is the only one the user gets. + """ + staged: Final = _stage_scrubbed_file(record) + if staged is not None: + _commit_token_file(staged) + + def _keep_the_unchecked_keychain_on_record(outcome: SecretErase, record: CliTokenRecord | None) -> bool: """Whether the token file, stripped of its secret, is worth keeping as the note that says so. @@ -203,17 +219,27 @@ def _keep_the_unchecked_keychain_on_record(outcome: SecretErase, record: CliToke its secret neither to a staged replacement nor to an overwrite is not kept either, because the secret goes first. """ - if record is None or isinstance(outcome, SecretStranded): + if record is None or not _the_keychain_went_unchecked(outcome): return False return _scrub_file_secret(record) +def _the_keychain_went_unchecked(outcome: SecretErase) -> bool: + """Whether the keychain neither confirmed the erase nor answered that it still holds the secret""" + match outcome: + case SecretErased() | SecretStranded(): + return False + case KeyringDisabled() | KeyringNotInstalled() | KeyringUnreachable(): + return True + + def _nothing_left_behind(outcome: SecretErase, record: CliTokenRecord | None) -> bool: """Whether the keychain can be trusted to hold no credential of ours once the file is gone. A machine with no token file has no stored login to end, and `clear_cli_token` keeps one behind whenever the keychain is left unconfirmed, taking the secret out in place when it cannot stage a - replacement, so a missing file is real evidence rather than the absence of it. Past that, a + replacement and writing the note again when the file holding it had to go, so a missing file is + real evidence rather than the absence of it. Past that, a keychain that could not be reached is never trusted, whatever the file looks like. Even a file holding its own secret says only that the login which wrote it had no keychain to write to, and the login before it may well have had one: the entry that login left outlives both the diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index e3d14c6da46..b8fcf3618cf 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -1,7 +1,9 @@ +import errno import json import os import stat import sys +import tempfile import threading import time @@ -82,6 +84,25 @@ def _blob(base_url=SERVER, key="sk-vault", jwt_token=""): return json.dumps({"base_url": base_url, "key": key, "jwt_token": jwt_token}) +_REAL_MKSTEMP = tempfile.mkstemp + + +class _MkstempThatNeedsTheOldFileGone: + """A disk with exactly one token file's worth of room left on it. + + Staging a replacement needs room for a second file, which is what a full disk refuses. Removing + the file already there is what gives that room back. + """ + + def __init__(self, path): + self.path = path + + def __call__(self, *args, **kwargs): + if self.path.exists(): + raise OSError(errno.ENOSPC, "No space left on device") + return _REAL_MKSTEMP(*args, **kwargs) + + _REAL_REPLACE = os.replace @@ -553,7 +574,7 @@ class TestClearCliToken: assert load_cli_token(vault=vault) is None @pytest.mark.parametrize( - "failure", [KeyringDisabled(), KeyringUnreachable(), KeyringDiscardsWrites()] + "failure", [KeyringDisabled(), KeyringUnreachable(), KeyringNotInstalled()] ) def test_a_secret_in_the_file_is_no_evidence_about_a_keychain_that_exists( self, isolated_home, secret_vault_factory, failure @@ -561,7 +582,10 @@ class TestClearCliToken: """Store a secret in the keychain, sign in again while the keychain is unusable so the new secret lands in the file, then log out while it is still unusable. The file now carries its own secret and the first login's entry is still there, so reading the file as proof of a - clean keychain reports a logout that did not happen.""" + clean keychain reports a logout that did not happen. + + The three unusable states are the whole of what an erase can answer besides erased and + stranded; a backend that keeps nothing it is given is something only a write finds out.""" _write_legacy_file(isolated_home) vault = secret_vault_factory(available=False, failure=failure) @@ -705,6 +729,30 @@ class TestClearCliToken: assert json.loads(path.read_text()).get("key") is None + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores file permissions") + def test_a_note_the_logout_had_to_remove_is_written_again_for_the_next_one( + self, isolated_home, secret_vault_factory, monkeypatch + ): + """A full disk refuses the replacement file and a read-only token file refuses the rewrite + in place, so the only way left to get the secret off disk is to remove the file carrying it. + That file was also the note saying the keychain went unchecked, and its absence is what the + next logout would read as a keychain already known to be clean. + + Removing it is what frees the room the replacement was refused for, so the note is written + again on the way out and the logout after this one still warns.""" + path = _write_legacy_file(isolated_home) + path.chmod(0o400) + monkeypatch.setattr( + "litellm.litellm_core_utils.private_json.tempfile.mkstemp", + _MkstempThatNeedsTheOldFileGone(path), + ) + vault = secret_vault_factory(available=False, failure=KeyringUnreachable()) + + assert clear_cli_token(vault=vault) == KeyringUnreachable() + assert clear_cli_token(vault=vault) == KeyringUnreachable() + + assert json.loads(path.read_text()).get("key") is None + @pytest.mark.skipif(os.geteuid() == 0, reason="root ignores directory permissions") def test_a_metadata_file_that_will_not_go_is_not_worth_alarming_the_user_over( self, isolated_home, secret_vault_factory From a2928efc75032f9a20a7685fce1ae3ffd8dc4c9d Mon Sep 17 00:00:00 2001 From: Mateo Date: Thu, 20 Aug 2026 03:45:00 -0700 Subject: [PATCH 048/220] test(cli): cover the keyless token record and keep keyring to the cli extra `lite up` treats a token record whose key the keychain would not hand over as no login at all, and that clause had no test: every existing freshness test passed a record carrying a real key, so deleting the clause left the whole suite green The base install smoke check now also asserts keyring is absent, which is what makes the lazy import in cli_keyring meaningful. keyring ships in the cli extra only, so a plain `pip install litellm` must not be able to reach it --- .../base_sdk_tests/check_base_sdk_install.py | 2 +- .../proxy/client/cli/test_up_commands.py | 22 +++++++++++++++++++ 2 files changed, 23 insertions(+), 1 deletion(-) diff --git a/tests/base_sdk_tests/check_base_sdk_install.py b/tests/base_sdk_tests/check_base_sdk_install.py index 723f30cad76..6b38de75e2e 100644 --- a/tests/base_sdk_tests/check_base_sdk_install.py +++ b/tests/base_sdk_tests/check_base_sdk_install.py @@ -11,7 +11,7 @@ import sys import traceback from collections.abc import Callable -EXTRAS_ONLY_MODULES = ("fastapi", "uvicorn") +EXTRAS_ONLY_MODULES = ("fastapi", "uvicorn", "keyring") def _require(condition: bool, message: str) -> None: diff --git a/tests/test_litellm/proxy/client/cli/test_up_commands.py b/tests/test_litellm/proxy/client/cli/test_up_commands.py index aebf441f777..a8d81f1c4bb 100644 --- a/tests/test_litellm/proxy/client/cli/test_up_commands.py +++ b/tests/test_litellm/proxy/client/cli/test_up_commands.py @@ -262,6 +262,28 @@ class TestEnsureFreshLogin: assert login_calls == ["http://proxy-b:4000"] + def test_forces_a_fresh_login_when_the_cached_token_has_no_readable_key(self, monkeypatch): + monkeypatch.setattr(up_module.sys.stdin, "isatty", lambda: True) + tokens = iter( + [ + _token(None, "http://proxy-a:4000"), + _token("sk-a", "http://proxy-a:4000"), + ] + ) + monkeypatch.setattr(up_module, "load_cli_token", lambda **_: next(tokens)) + monkeypatch.setattr(up_module, "is_cli_token_fresh", lambda token_data: True) + login_calls = [] + + @click.pass_context + def fake_login(ctx): + login_calls.append(ctx.obj["base_url"]) + + monkeypatch.setattr(up_module, "login", fake_login) + + _ensure_fresh_login(_make_ctx("http://proxy-a:4000")) + + assert login_calls == ["http://proxy-a:4000"] + def test_fails_cleanly_non_interactively_when_only_a_different_proxys_token_is_cached(self, monkeypatch): monkeypatch.setattr(up_module.sys.stdin, "isatty", lambda: False) monkeypatch.setattr(up_module, "load_cli_token", lambda **_: _token("sk-a", "http://proxy-a:4000")) From cc2013e9660ab722fab9f8097497a121336a034f Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 03:55:25 -0700 Subject: [PATCH 049/220] fix(anthropic): map metadata.user_id to prompt_cache_key on the /v1/messages bridge Both /v1/messages bridges (Responses API adapter for openai/* and the chat-completions adapter) now derive prompt_cache_key from the first 64 characters of metadata.user_id, next to the existing user mapping. The chat bridge only sets it when the resolved provider advertises prompt_cache_key in its supported params, so providers that reject unknown params are unaffected. A prompt_cache_key sent explicitly by the client always wins over the derived value. Fixes #37508 --- .../adapters/handler.py | 10 ++- .../adapters/transformation.py | 31 +++++++- .../responses_adapters/handler.py | 7 +- .../responses_adapters/transformation.py | 6 +- .../experimental_pass_through/utils.py | 9 +++ litellm/types/llms/openai.py | 1 + ...al_pass_through_adapters_transformation.py | 79 +++++++++++++++++++ .../adapters/test_handler_prompt_cache_key.py | 64 +++++++++++++++ .../test_responses_adapters_handler.py | 45 +++++++++++ .../test_responses_adapters_transformation.py | 24 ++++++ 10 files changed, 270 insertions(+), 6 deletions(-) create mode 100644 tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_prompt_cache_key.py create mode 100644 tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py b/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py index 48d8a03d549..89066e33cbc 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py @@ -484,10 +484,14 @@ class LiteLLMMessagesToCompletionTransformationHandler: if "output_config" in extra_kwargs: request_data["output_config"] = extra_kwargs["output_config"] + custom_llm_provider: Final = extra_kwargs.get("custom_llm_provider") ( openai_request, tool_name_mapping, - ) = ANTHROPIC_ADAPTER.translate_completion_input_params_with_tool_mapping(request_data) + ) = ANTHROPIC_ADAPTER.translate_completion_input_params_with_tool_mapping( + request_data, + custom_llm_provider=custom_llm_provider if isinstance(custom_llm_provider, str) else None, + ) if openai_request is None: raise ValueError("Failed to translate request to OpenAI format") @@ -526,6 +530,10 @@ class LiteLLMMessagesToCompletionTransformationHandler: if key not in excluded_keys and key not in completion_kwargs and value is not None: completion_kwargs[key] = value + explicit_prompt_cache_key: Final = extra_kwargs.get("prompt_cache_key") + if explicit_prompt_cache_key is not None: + completion_kwargs["prompt_cache_key"] = explicit_prompt_cache_key + # Normalize reasoning_effort based on model capabilities # (e.g. "max" → "xhigh"/"high", "minimal" → "low" if unsupported) # Must run BEFORE _route_openai_thinking, which prepends "responses/" diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index e45414b4a73..5c72fee25a0 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -4,8 +4,10 @@ import json from collections.abc import AsyncIterator, Iterator, Mapping from typing import TYPE_CHECKING, Any, Final, Literal, cast +import litellm from litellm.llms.anthropic.experimental_pass_through.utils import ( is_reasoning_auto_summary_enabled, + prompt_cache_key_from_user_id, ) # OpenAI has a 64-character limit for function/tool names @@ -148,7 +150,7 @@ class AnthropicAdapter: return result def translate_completion_input_params_with_tool_mapping( - self, kwargs + self, kwargs, *, custom_llm_provider: str | None = None ) -> tuple[ChatCompletionRequest | None, dict[str, str]]: """ Translate Anthropic request params to OpenAI format, returning tool name mapping. @@ -179,7 +181,10 @@ class AnthropicAdapter: ( translated_body, tool_name_mapping, - ) = LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai(anthropic_message_request=request_body) + ) = LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai( + anthropic_message_request=request_body, + custom_llm_provider=custom_llm_provider, + ) return translated_body, tool_name_mapping @@ -907,16 +912,32 @@ class LiteLLMAnthropicMessagesAdapter: ChatCompletionSystemMessage(role="system", content=openai_system_content), ) + @staticmethod + def _supports_prompt_cache_key(model: str | None, custom_llm_provider: str | None) -> bool: + if not model or not custom_llm_provider: + return False + supported_params: Final = litellm.get_supported_openai_params( + model=model, custom_llm_provider=custom_llm_provider + ) + return "prompt_cache_key" in (supported_params or ()) + def _translate_metadata_to_openai( self, anthropic_message_request: AnthropicMessagesRequest, new_kwargs: ChatCompletionRequest, + *, + custom_llm_provider: str | None = None, ) -> None: """Translate metadata fields from Anthropic request to OpenAI request.""" if "metadata" in anthropic_message_request: metadata: Final = anthropic_message_request["metadata"] if metadata and "user_id" in metadata: new_kwargs["user"] = metadata["user_id"] + prompt_cache_key: Final = prompt_cache_key_from_user_id(metadata["user_id"]) + if prompt_cache_key is not None and self._supports_prompt_cache_key( + anthropic_message_request.get("model"), custom_llm_provider + ): + new_kwargs["prompt_cache_key"] = prompt_cache_key if "litellm_metadata" in anthropic_message_request: # metadata will be passed to litellm.acompletion(), it's a litellm_param @@ -1069,7 +1090,10 @@ class LiteLLMAnthropicMessagesAdapter: new_kwargs[k] = v def translate_anthropic_to_openai( - self, anthropic_message_request: AnthropicMessagesRequest + self, + anthropic_message_request: AnthropicMessagesRequest, + *, + custom_llm_provider: str | None = None, ) -> tuple[ChatCompletionRequest, dict[str, str]]: """ This is used by the beta Anthropic Adapter, for translating anthropic `/v1/messages` requests to the openai format. @@ -1103,6 +1127,7 @@ class LiteLLMAnthropicMessagesAdapter: self._translate_metadata_to_openai( anthropic_message_request=anthropic_message_request, new_kwargs=new_kwargs, + custom_llm_provider=custom_llm_provider, ) ## CONVERT TOOL CHOICE self._translate_tool_choice_to_openai( diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py index e6d8686b466..843cda249c5 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py @@ -105,7 +105,8 @@ def _build_responses_kwargs( # Forward litellm-specific kwargs (api_key, api_base, logging obj, etc.) excluded: Final = {"anthropic_messages"} - for key, value in _forwarded_kwargs(extra_kwargs).items(): + forwarded_kwargs: Final = _forwarded_kwargs(extra_kwargs) + for key, value in forwarded_kwargs.items(): if key == "litellm_logging_obj" and value is not None: from litellm.litellm_core_utils.litellm_logging import ( Logging as LiteLLMLoggingObject, @@ -121,6 +122,10 @@ def _build_responses_kwargs( elif key not in excluded and key not in responses_kwargs and value is not None: responses_kwargs[key] = value + explicit_prompt_cache_key: Final = forwarded_kwargs.get("prompt_cache_key") + if explicit_prompt_cache_key is not None: + responses_kwargs["prompt_cache_key"] = explicit_prompt_cache_key + return responses_kwargs diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py index 21a8cb9501e..9238433151a 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py @@ -18,6 +18,7 @@ from litellm.litellm_core_utils.reasoning_effort_utils import ( ) from litellm.llms.anthropic.experimental_pass_through.utils import ( is_reasoning_auto_summary_enabled, + prompt_cache_key_from_user_id, ) from litellm.types.llms.anthropic import ( AllAnthropicPassThroughMessageValues, @@ -452,10 +453,13 @@ class LiteLLMAnthropicToResponsesAPIAdapter: if openai_cm is not None: responses_kwargs["context_management"] = openai_cm - # metadata user_id -> user + # metadata user_id -> user and prompt_cache_key metadata: Final = anthropic_request.get("metadata") if isinstance(metadata, dict) and "user_id" in metadata: responses_kwargs["user"] = str(metadata["user_id"])[:64] + prompt_cache_key: Final = prompt_cache_key_from_user_id(metadata["user_id"]) + if prompt_cache_key is not None: + responses_kwargs["prompt_cache_key"] = prompt_cache_key return responses_kwargs diff --git a/litellm/llms/anthropic/experimental_pass_through/utils.py b/litellm/llms/anthropic/experimental_pass_through/utils.py index 46091cd89a2..c5abcf8c04c 100644 --- a/litellm/llms/anthropic/experimental_pass_through/utils.py +++ b/litellm/llms/anthropic/experimental_pass_through/utils.py @@ -1,8 +1,17 @@ import os +from typing import Final import litellm from litellm.types.utils import ModelInfo +OPENAI_MAX_PROMPT_CACHE_KEY_LENGTH: Final = 64 + + +def prompt_cache_key_from_user_id(user_id: object) -> str | None: + if user_id is None: + return None + return str(user_id)[:OPENAI_MAX_PROMPT_CACHE_KEY_LENGTH] or None + def is_reasoning_auto_summary_enabled() -> bool: """Check whether the default 'summary: detailed' injection is enabled (opt-in).""" diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py index edfc50c99f6..beb6612497c 100644 --- a/litellm/types/llms/openai.py +++ b/litellm/types/llms/openai.py @@ -917,6 +917,7 @@ class ChatCompletionRequest(TypedDict, total=False): seed: int service_tier: str safety_identifier: str + prompt_cache_key: str # writable-ok: the /v1/messages adapter assigns it after construction stop: str | list[str] stream_options: dict temperature: float diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py index 1185893d428..dd1e8a93280 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py @@ -635,6 +635,85 @@ def test_translate_anthropic_to_openai_orders_top_level_and_midturn_system(): ] +def _translate_with_metadata( + model: str, metadata: dict[str, Any], custom_llm_provider: str | None +) -> dict[str, Any]: + openai_request, _ = LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai( + anthropic_message_request={ + "model": model, + "max_tokens": 100, + "metadata": metadata, + "messages": [{"role": "user", "content": "hi"}], + }, + custom_llm_provider=custom_llm_provider, + ) + return cast(dict[str, Any], openai_request) + + +def test_translate_anthropic_to_openai_maps_user_id_to_prompt_cache_key_for_openai(): + openai_request = _translate_with_metadata("openai/gpt-5.6-luna", {"user_id": "session-abc"}, "openai") + assert openai_request["user"] == "session-abc" + assert openai_request["prompt_cache_key"] == "session-abc" + + +def test_translate_anthropic_to_openai_truncates_prompt_cache_key_but_keeps_full_user(): + long_id = "".join(str(i % 10) for i in range(100)) + openai_request = _translate_with_metadata("openai/gpt-5.6-luna", {"user_id": long_id}, "openai") + assert openai_request["user"] == long_id + assert openai_request["prompt_cache_key"] == long_id[:64] + assert len(openai_request["prompt_cache_key"]) == 64 + + +@pytest.mark.parametrize("model", ["azure/my-gpt-5-deployment", "my-gpt-5-deployment"]) +def test_translate_anthropic_to_openai_sets_prompt_cache_key_for_azure(model: str): + openai_request = _translate_with_metadata(model, {"user_id": "session-abc"}, "azure") + assert openai_request["prompt_cache_key"] == "session-abc" + + +@pytest.mark.parametrize( + "model, custom_llm_provider", + [ + ("gemini/gemini-2.5-pro", "gemini"), + ("vertex_ai/gemini-2.5-pro", "vertex_ai"), + ("anthropic/claude-sonnet-4-5", "anthropic"), + ("bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0", "bedrock"), + ("no-such-model-lit5875", "no-such-provider-lit5875"), + ], +) +def test_translate_anthropic_to_openai_skips_prompt_cache_key_when_provider_lacks_it( + model: str, custom_llm_provider: str +): + openai_request = _translate_with_metadata(model, {"user_id": "session-abc"}, custom_llm_provider) + assert openai_request["user"] == "session-abc" + assert "prompt_cache_key" not in openai_request + + +def test_translate_anthropic_to_openai_skips_prompt_cache_key_without_provider(): + openai_request = _translate_with_metadata("openai/gpt-5.6-luna", {"user_id": "session-abc"}, None) + assert openai_request["user"] == "session-abc" + assert "prompt_cache_key" not in openai_request + + +@pytest.mark.parametrize("user_id", ["", None]) +def test_translate_anthropic_to_openai_skips_prompt_cache_key_for_empty_or_null_user_id(user_id: str | None): + openai_request = _translate_with_metadata("openai/gpt-5.6-luna", {"user_id": user_id}, "openai") + assert openai_request["user"] == user_id + assert "prompt_cache_key" not in openai_request + + +def test_translate_anthropic_to_openai_without_metadata_sets_neither_user_nor_prompt_cache_key(): + openai_request, _ = LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai( + anthropic_message_request={ + "model": "openai/gpt-5.6-luna", + "max_tokens": 100, + "messages": [{"role": "user", "content": "hi"}], + }, + custom_llm_provider="openai", + ) + assert "user" not in openai_request + assert "prompt_cache_key" not in openai_request + + def test_translate_openai_content_to_anthropic_empty_function_arguments(): """Test that empty function arguments are handled safely and don't cause JSON parsing errors.""" diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_prompt_cache_key.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_prompt_cache_key.py new file mode 100644 index 00000000000..ad6fc04217f --- /dev/null +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_prompt_cache_key.py @@ -0,0 +1,64 @@ +import os +import sys + +import pytest + +sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../../.."))) + +from litellm.llms.anthropic.experimental_pass_through.adapters.handler import ( + LiteLLMMessagesToCompletionTransformationHandler, +) + +MESSAGES = [{"role": "user", "content": "hello"}] + + +def _prepare(model: str, extra_kwargs: dict[str, object], thinking: dict[str, object] | None = None): + completion_kwargs, _ = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs( + max_tokens=1024, + messages=MESSAGES, + model=model, + metadata={"user_id": "session-abc"}, + thinking=thinking, + extra_kwargs=extra_kwargs, + ) + return completion_kwargs + + +def test_prepare_completion_kwargs_derives_prompt_cache_key_for_openai_provider(): + completion_kwargs = _prepare("openai/gpt-5.6-luna", {"custom_llm_provider": "openai"}) + assert completion_kwargs["user"] == "session-abc" + assert completion_kwargs["prompt_cache_key"] == "session-abc" + + +def test_prepare_completion_kwargs_prefers_explicit_prompt_cache_key_over_derived(): + completion_kwargs = _prepare( + "openai/gpt-5.6-luna", + {"custom_llm_provider": "openai", "prompt_cache_key": "explicit-key"}, + ) + assert completion_kwargs["user"] == "session-abc" + assert completion_kwargs["prompt_cache_key"] == "explicit-key" + + +@pytest.mark.parametrize( + "model, extra_kwargs", + [ + ("gemini/gemini-2.5-pro", {"custom_llm_provider": "gemini"}), + ("openai/gpt-5.6-luna", {}), + ], +) +def test_prepare_completion_kwargs_skips_prompt_cache_key_without_provider_support( + model: str, extra_kwargs: dict[str, object] +): + completion_kwargs = _prepare(model, extra_kwargs) + assert completion_kwargs["user"] == "session-abc" + assert "prompt_cache_key" not in completion_kwargs + + +def test_prepare_completion_kwargs_keeps_prompt_cache_key_through_responses_reroute(): + completion_kwargs = _prepare( + "openai/gpt-5.6-luna", + {"custom_llm_provider": "openai"}, + thinking={"type": "enabled", "budget_tokens": 1024}, + ) + assert completion_kwargs["model"] == "responses/openai/gpt-5.6-luna" + assert completion_kwargs["prompt_cache_key"] == "session-abc" diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py new file mode 100644 index 00000000000..7ef3077f9d7 --- /dev/null +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py @@ -0,0 +1,45 @@ +import os +import sys + +sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../../.."))) + +from litellm.llms.anthropic.experimental_pass_through.responses_adapters.handler import ( + _build_responses_kwargs, +) + +MESSAGES = [{"role": "user", "content": "hello"}] + + +def test_build_responses_kwargs_derives_prompt_cache_key_from_user_id(): + responses_kwargs = _build_responses_kwargs( + max_tokens=1024, + messages=MESSAGES, + model="openai/gpt-5.6-luna", + metadata={"user_id": "session-abc"}, + extra_kwargs={"custom_llm_provider": "openai"}, + ) + assert responses_kwargs["user"] == "session-abc" + assert responses_kwargs["prompt_cache_key"] == "session-abc" + + +def test_build_responses_kwargs_prefers_explicit_prompt_cache_key_over_derived(): + responses_kwargs = _build_responses_kwargs( + max_tokens=1024, + messages=MESSAGES, + model="openai/gpt-5.6-luna", + metadata={"user_id": "session-abc"}, + extra_kwargs={"custom_llm_provider": "openai", "prompt_cache_key": "explicit-key"}, + ) + assert responses_kwargs["user"] == "session-abc" + assert responses_kwargs["prompt_cache_key"] == "explicit-key" + + +def test_build_responses_kwargs_without_metadata_sets_no_prompt_cache_key(): + responses_kwargs = _build_responses_kwargs( + max_tokens=1024, + messages=MESSAGES, + model="openai/gpt-5.6-luna", + extra_kwargs={"custom_llm_provider": "openai"}, + ) + assert "user" not in responses_kwargs + assert "prompt_cache_key" not in responses_kwargs diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py index 876213eda3f..297f2052b83 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py @@ -992,6 +992,29 @@ class TestTranslateRequestBroaderCoverage: kwargs = _ADAPTER.translate_request(req) assert len(kwargs["user"]) == 64 + def test_metadata_user_id_mapped_to_prompt_cache_key(self): + req = _make_request(metadata={"user_id": "user-42"}) + kwargs = _ADAPTER.translate_request(req) + assert kwargs["prompt_cache_key"] == "user-42" + + def test_metadata_user_id_prompt_cache_key_truncated_to_first_64_chars(self): + long_id = "".join(str(i % 10) for i in range(100)) + req = _make_request(metadata={"user_id": long_id}) + kwargs = _ADAPTER.translate_request(req) + assert kwargs["prompt_cache_key"] == long_id[:64] + assert len(kwargs["prompt_cache_key"]) == 64 + + def test_metadata_empty_user_id_sets_no_prompt_cache_key(self): + req = _make_request(metadata={"user_id": ""}) + kwargs = _ADAPTER.translate_request(req) + assert kwargs["user"] == "" + assert "prompt_cache_key" not in kwargs + + def test_metadata_null_user_id_sets_no_prompt_cache_key(self): + req = _make_request(metadata={"user_id": None}) + kwargs = _ADAPTER.translate_request(req) + assert "prompt_cache_key" not in kwargs + def test_no_optional_fields_does_not_add_spurious_keys(self): req = _make_request() kwargs = _ADAPTER.translate_request(req) @@ -1005,6 +1028,7 @@ class TestTranslateRequestBroaderCoverage: "text", "context_management", "user", + "prompt_cache_key", ): assert key not in kwargs, f"unexpected key: {key}" From 2c691d3820e25fa66224b89bccfcaf8476416ae5 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 20 Aug 2026 04:10:10 -0700 Subject: [PATCH 050/220] feat(proxy): native CLI login with OAuth authorization code + PKCE The proxy's OAuth authorization server (dynamic registration, PKCE S256, loopback redirects, single-use codes, refresh rotation) gains a proxy-API audience: /authorize?resource= renders a consent page with team selection and /token mints the same per-user credential lite login mints, so a native CLI can sign a user in through the system browser and call /v1/* with user and team attribution. Adds GET /.well-known/litellm-cli-auth as the versioned discovery contract for non-Python clients, POST /revoke (RFC 7009) for logout, and lite login --pkce, lite logout, and lite auth print-token on the CLI side. Proxy-API grants only ever redirect to a loopback address and the server never picks a team on the user's behalf. Fixes #37332 --- litellm/litellm_core_utils/cli_token_utils.py | 3 + .../mcp_server/bridge_token_flow.py | 11 +- .../mcp_server/discoverable_endpoints.py | 53 +- .../mcp_server/gateway_dcr_flow.py | 500 +++++++++++++--- .../outbound_credentials/session_token.py | 19 +- .../mcp_server/proxy_api_credentials.py | 84 +++ litellm/proxy/_lazy_features.py | 2 + litellm/proxy/client/cli/commands/auth.py | 91 ++- .../proxy/client/cli/commands/pkce_login.py | 471 +++++++++++++++ .../html_forms/native_client_consent.py | 91 +++ litellm/proxy/management_endpoints/ui_sso.py | 22 +- .../test_cli_token_utils.py | 27 + .../test_session_token.py | 52 ++ .../mcp_server/test_discoverable_endpoints.py | 229 +++++++ .../mcp_server/test_gateway_dcr_flow.py | 564 +++++++++++++++++- .../mcp_server/test_proxy_api_credentials.py | 167 ++++++ .../proxy/client/cli/test_auth_commands.py | 323 +++++++++- .../proxy/client/cli/test_pkce_login.py | 553 +++++++++++++++++ .../html_forms/test_native_client_consent.py | 70 +++ .../proxy/management_endpoints/test_ui_sso.py | 20 +- 20 files changed, 3209 insertions(+), 143 deletions(-) create mode 100644 litellm/proxy/_experimental/mcp_server/proxy_api_credentials.py create mode 100644 litellm/proxy/client/cli/commands/pkce_login.py create mode 100644 litellm/proxy/common_utils/html_forms/native_client_consent.py create mode 100644 tests/test_litellm/proxy/_experimental/mcp_server/test_proxy_api_credentials.py create mode 100644 tests/test_litellm/proxy/client/cli/test_pkce_login.py create mode 100644 tests/test_litellm/proxy/common_utils/html_forms/test_native_client_consent.py diff --git a/litellm/litellm_core_utils/cli_token_utils.py b/litellm/litellm_core_utils/cli_token_utils.py index a44ce431f4e..0043bc31204 100644 --- a/litellm/litellm_core_utils/cli_token_utils.py +++ b/litellm/litellm_core_utils/cli_token_utils.py @@ -79,6 +79,9 @@ def is_cli_token_fresh(token_data: Mapping[str, object], buffer_hours: float = 0 `LITELLM_CLI_JWT_EXPIRATION_HOURS`.""" from litellm.constants import CLI_JWT_EXPIRATION_HOURS + expires_at: Final = token_data.get("expires_at") + if isinstance(expires_at, (int, float)): + return time.time() < expires_at - buffer_hours * 3600 timestamp: Final = token_data.get("timestamp") if not isinstance(timestamp, (int, float)): return False diff --git a/litellm/proxy/_experimental/mcp_server/bridge_token_flow.py b/litellm/proxy/_experimental/mcp_server/bridge_token_flow.py index 1d8d545023d..b8c25236b0d 100644 --- a/litellm/proxy/_experimental/mcp_server/bridge_token_flow.py +++ b/litellm/proxy/_experimental/mcp_server/bridge_token_flow.py @@ -15,6 +15,7 @@ from litellm.proxy._experimental.mcp_server.oauth_utils import TOKEN_NO_CACHE_HE from litellm.types.mcp_server.mcp_server_manager import MCPServer if TYPE_CHECKING: + from litellm.models.user import LiteLLM_UserTable from litellm.proxy._experimental.mcp_server.discoverable_endpoints import _BridgeAuthorizationCode from litellm.proxy._experimental.mcp_server.outbound_credentials.envelope import ( EnvelopeIdentity, @@ -181,7 +182,13 @@ async def _reload_active_key_by_hash(key_hash: str) -> "_ResolvedKey | _KeyResol async def _reload_active_user_by_id(user_id: str) -> "_KeyResolutionFailure | None": - """Re-validate a live litellm user by id, returning ``None`` when the user is active or a precise + """``None`` when the user is live, else the precise failure ``load_active_user_by_id`` found.""" + loaded: Final = await load_active_user_by_id(user_id) + return loaded if isinstance(loaded, str) else None + + +async def load_active_user_by_id(user_id: str) -> "LiteLLM_UserTable | _KeyResolutionFailure": + """Load a live litellm user by id, returning the record when the user is active or a precise failure otherwise. The interactive DCR client authenticates via SSO, so its refresh envelope seals a user subject; renewing it must re-check the user is still live (present and not SCIM-deactivated) so a deactivated user cannot keep refreshing, mirroring how admission re-validates the same user subject on @@ -226,7 +233,7 @@ async def _reload_active_user_by_id(user_id: str) -> "_KeyResolutionFailure | No return "no_active_key" if isinstance(user_object.metadata, dict) and user_object.metadata.get("scim_active") is False: return "no_active_key" - return None + return user_object async def _key_owner_scim_deactivated(key: "UserAPIKeyAuth") -> bool: diff --git a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py index 1ca4c657706..f9c5db76fa0 100644 --- a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py +++ b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py @@ -47,8 +47,12 @@ from litellm.proxy._experimental.mcp_server.gateway_dcr_flow import ( aggregate_token, complete_connect_flow, is_gateway_dcr_client_id, + is_proxy_api_resource, + native_client_auth_contract, + native_client_authorize, register_aggregate_client, relative_request_url, + revoke_refresh_token, ) from litellm.proxy._experimental.mcp_server.oauth_utils import ( TOKEN_NO_CACHE_HEADERS, @@ -58,6 +62,10 @@ from litellm.proxy._experimental.mcp_server.oauth_utils import ( validate_trusted_redirect_uri, well_known_root_suffix, ) +from litellm.proxy._experimental.mcp_server.proxy_api_credentials import ( + lookup_consent_teams, + mint_proxy_credential, +) from litellm.proxy.auth.ip_address_utils import IPAddressUtils from litellm.proxy.common_utils.encrypt_decrypt_utils import ( decrypt_value_helper, @@ -1663,6 +1671,18 @@ async def authorize( ) if mcp_server_name is None and client_id and is_gateway_dcr_client_id(client_id): + if is_proxy_api_resource(request, resource): + return await native_client_authorize( + request=request, + client_id=client_id, + redirect_uri=redirect_uri, + state=state, + code_challenge=code_challenge, + code_challenge_method=code_challenge_method, + response_type=response_type, + session_user_id=_session_cookie_user_id(request), + lookup_consent_teams=lookup_consent_teams, + ) return aggregate_authorize( request=request, client_id=client_id, @@ -1764,6 +1784,7 @@ async def token_endpoint( reload_user=_reload_active_user_by_id, cache=user_api_key_cache, resource=resource, + mint_proxy_credential=mint_proxy_credential, ) lookup_name: Final = mcp_server_name or client_id @@ -1793,12 +1814,19 @@ async def token_endpoint( @router.post("/authorize/complete") -async def authorize_complete(request: Request, flow: str = Form(...), delivery: str | None = Form(None)): +async def authorize_complete( + request: Request, + flow: str = Form(...), + delivery: str | None = Form(None), + team_id: str | None = Form(None), + decision: str | None = Form(None), +) -> Response: """Finish an aggregate connect flow: mint the gateway authorization code for the signed-in user and hand it back to the DCR client, by 303 redirect (default) or, for a loopback client on a different machine, as a copyable callback URL (``delivery=manual``). POST plus the per-flow HttpOnly cookie set at /authorize; an - anonymous or bad-flow request just 400s.""" + anonymous or bad-flow request just 400s. The native-client consent page adds + ``decision`` (approve or deny) and the ``team_id`` the credential is attributed to.""" from litellm.proxy.proxy_server import user_api_key_cache # noqa: PLC0415 # circular import at module load return await complete_connect_flow( @@ -1807,9 +1835,30 @@ async def authorize_complete(request: Request, flow: str = Form(...), delivery: session_user_id=_session_cookie_user_id(request), cache=user_api_key_cache, delivery=delivery, + team_id=team_id, + decision=decision, ) +@router.post("/revoke") +async def revoke_endpoint(request: Request, token: str = Form(...), client_id: str = Form(...)) -> Response: + """RFC 7009 revocation for the gateway's refresh tokens (``lite logout``). Always 200 + for a known client, whatever the token's state; access tokens expire on their own.""" + from litellm.proxy.proxy_server import ( # noqa: PLC0415 # circular import at module load + master_key, + user_api_key_cache, + ) + + return await revoke_refresh_token(token=token, client_id=client_id, master_key=master_key, cache=user_api_key_cache) + + +@router.get("/.well-known/litellm-cli-auth") +async def native_client_auth_discovery(request: Request) -> JSONResponse: + """The versioned contract a native client (``lite login --pkce``, or a CLI in any other + language) reads to sign a user in through the browser and obtain a proxy credential.""" + return JSONResponse(native_client_auth_contract(request), headers=TOKEN_NO_CACHE_HEADERS) + + # Per RFC 6749 §4.1.2.1, an IdP that rejects an OAuth authorization request # redirects back to the configured redirect URI with ``error`` / # ``error_description`` / ``error_uri`` query params and no ``code``. The MCP diff --git a/litellm/proxy/_experimental/mcp_server/gateway_dcr_flow.py b/litellm/proxy/_experimental/mcp_server/gateway_dcr_flow.py index 85885fc75f5..53db6fbf9d1 100644 --- a/litellm/proxy/_experimental/mcp_server/gateway_dcr_flow.py +++ b/litellm/proxy/_experimental/mcp_server/gateway_dcr_flow.py @@ -42,15 +42,16 @@ import hmac import html import secrets from base64 import urlsafe_b64encode -from collections.abc import Awaitable, Callable, Mapping +from collections.abc import Awaitable, Callable, Iterable, Mapping from datetime import datetime, timezone -from typing import Final, Literal, TypeVar +from types import MappingProxyType +from typing import Final, Literal, Protocol, TypeVar from urllib.parse import parse_qsl, urlencode, urlparse, urlunparse from fastapi import HTTPException, Request from fastapi.responses import HTMLResponse, JSONResponse, RedirectResponse, Response from pydantic import BaseModel, ConfigDict, Field, ValidationError -from typing_extensions import assert_never +from typing_extensions import ReadOnly, TypedDict, assert_never from litellm._logging import verbose_logger from litellm.caching.caching import DualCache @@ -70,6 +71,7 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.session_credent from litellm.proxy._experimental.mcp_server.outbound_credentials.session_token import ( SESSION_REFRESH_TTL_SECONDS, MintedSessionToken, + SessionAudience, SessionKeys, SessionPrincipal, mint_session_refresh_token, @@ -79,6 +81,9 @@ from litellm.proxy.common_utils.encrypt_decrypt_utils import ( decrypt_value_helper, encrypt_value_helper, ) +from litellm.proxy.common_utils.html_forms.native_client_consent import ( + render_native_client_consent_page, +) from litellm.types.mcp_server.mcp_server_manager import MCPServer GATEWAY_DCR_CLIENT_ID_PREFIX: Final = "llm_dcrc_" @@ -144,6 +149,46 @@ ReloadUser = Callable[[str], Awaitable[ReloadUserFailure | None]] ``None`` means the user is active; ``unavailable`` is a retryable DB outage; anything else fails the grant closed.""" +PROXY_API_AUDIENCE: Final[SessionAudience] = "proxy_api" +"""The audience a native client (``lite login --pkce``, a Go CLI) asks for by sending the +proxy base URL itself as the RFC 8707 ``resource``: the grant then mints the proxy-API CLI +credential that LLM routes accept, instead of the MCP-only session pair.""" + +ProxyCredentialMintFailure = Literal[ReloadUserFailure, "not_a_member"] + + +class MintedProxyCredential(BaseModel): + model_config = ConfigDict(frozen=True) + key: str = Field(min_length=1) + expires_in: int = Field(gt=0) + user_id: str = Field(min_length=1) + team_id: str | None = None + + +class MintProxyCredential(Protocol): + """Injected proxy-API credential minter ``(user_id, team_id)``: reloads the user live, + checks team membership, and mints the same credential ``lite login`` mints.""" + + def __call__( + self, user_id: str, team_id: str | None, / + ) -> Awaitable[MintedProxyCredential | ProxyCredentialMintFailure]: ... + + +class ConsentTeam(BaseModel): + model_config = ConfigDict(frozen=True) + team_id: str = Field(min_length=1) + team_alias: str | None = None + + +class LookupConsentTeams(Protocol): + """Injected lookup of the teams a signed-in user may bind a proxy-API credential to.""" + + def __call__(self, user_id: str, /) -> Awaitable[tuple[ConsentTeam, ...] | ReloadUserFailure]: ... + + +async def _refuse_proxy_credential(user_id: str, team_id: str | None) -> ProxyCredentialMintFailure: + return "unresolvable" + class GatewayDcrClient(BaseModel): """The registration record sealed into a gateway DCR ``client_id``. @@ -173,6 +218,7 @@ class _ConnectFlow(BaseModel): jti: str = Field(min_length=1) exp: int resource_server_id: str | None = None + audience: SessionAudience | None = None class _GatewayAuthCode(BaseModel): @@ -190,6 +236,8 @@ class _GatewayAuthCode(BaseModel): iat: int exp: int resource_server_id: str | None = None + audience: SessionAudience | None = None + team_id: str | None = None def is_gateway_dcr_client_id(client_id: str | None) -> bool: @@ -318,9 +366,9 @@ def _cookie_path_and_secure(request: Request) -> tuple[str, bool]: return parsed.path or "/", parsed.scheme == "https" -def _append_query_params(url: str, params: dict[str, str]) -> str: +def _append_query_params(url: str, params: Iterable[tuple[str, str]]) -> str: parsed: Final = urlparse(url) - query: Final = parse_qsl(parsed.query, keep_blank_values=True) + list(params.items()) + query: Final = (*parse_qsl(parsed.query, keep_blank_values=True), *params) return urlunparse(parsed._replace(query=urlencode(query))) @@ -392,6 +440,155 @@ def aggregate_authorize( section 4.1.2.1 an unvalidated redirect URI must not receive an error redirect, and once the client is at fault there is no trusted place to send the browser. """ + rejected: Final = _rejected_authorize_request( + client_id, redirect_uri, state, code_challenge, code_challenge_method, response_type + ) + if rejected is not None: + return rejected + base_url: Final = get_request_base_url(request) + if session_user_id is None: + return _login_redirect(base_url, request) + scoped_server: Final = resolve_scoped_resource_server(request, resource) + handle: Final = secrets.token_urlsafe(24) + flow: Final = _new_connect_flow( + session_user_id=session_user_id, + client_id=client_id, + redirect_uri=redirect_uri, + state=state, + code_challenge=code_challenge or "", + resource_server_id=scoped_server.server_id if scoped_server is not None else None, + audience=None, + ) + connect_url: Final = _append_query_params( + f"{base_url}/ui/connect", + (("connect_flow", handle), ("connect_client", _origin_only(redirect_uri))), + ) + response: Final = RedirectResponse(connect_url, status_code=303) + _set_flow_cookie(response, request, handle, flow) + return response + + +async def native_client_authorize( + request: Request, + client_id: str, + redirect_uri: str, + state: str, + code_challenge: str | None, + code_challenge_method: str | None, + response_type: str | None, + session_user_id: str | None, + lookup_consent_teams: LookupConsentTeams, +) -> Response: + """The authorize verb for a native client that named the proxy API itself as its + RFC 8707 ``resource``: the same client, redirect, PKCE, and sign-in checks as the + aggregate verb plus a loopback-only redirect (the credential this grant mints is the + user's personal proxy key, which belongs on their own machine and never behind a hosted + callback), then the consent page rendered right here (no connect-page interlude, since + there is no per-server vaulting to do) with the flow sealed into the per-flow cookie + and its handle carried only in the form, never in a URL.""" + rejected: Final = _rejected_authorize_request( + client_id, redirect_uri, state, code_challenge, code_challenge_method, response_type + ) + if rejected is not None: + return rejected + if not is_loopback_redirect_host(urlparse(redirect_uri)): + return _oauth_error(400, "invalid_request", "a proxy-API grant may only redirect to a loopback address") + base_url: Final = get_request_base_url(request) + if session_user_id is None: + return _login_redirect(base_url, request) + teams: Final = await lookup_consent_teams(session_user_id) + if not isinstance(teams, tuple): + return _consent_lookup_failure_response(teams) + handle: Final = secrets.token_urlsafe(24) + flow: Final = _new_connect_flow( + session_user_id=session_user_id, + client_id=client_id, + redirect_uri=redirect_uri, + state=state, + code_challenge=code_challenge or "", + resource_server_id=None, + audience=PROXY_API_AUDIENCE, + ) + page: Final = render_native_client_consent_page( + client_origin=_origin_only(redirect_uri), + user_id=session_user_id, + teams=tuple((team.team_id, team.team_alias or team.team_id) for team in teams), + flow_handle=handle, + complete_url=f"{base_url}/authorize/complete", + ) + response: Final = HTMLResponse(page, headers=_CONSENT_PAGE_HEADERS) + _set_flow_cookie(response, request, handle, flow) + return response + + +_CONSENT_PAGE_HEADERS: Final = MappingProxyType( + { + **TOKEN_NO_CACHE_HEADERS, + "X-Frame-Options": "DENY", + "Content-Security-Policy": "frame-ancestors 'none'", + } +) + +NATIVE_CLIENT_AUTH_CONTRACT_VERSION: Final = 1 +"""The version a native client checks before trusting the rest of the discovery document. +Bump it only when an existing field changes meaning or goes away; adding fields is free.""" + + +class NativeClientAuthContract(TypedDict): + contract_version: ReadOnly[int] + issuer: ReadOnly[str] + authorization_endpoint: ReadOnly[str] + token_endpoint: ReadOnly[str] + registration_endpoint: ReadOnly[str] + revocation_endpoint: ReadOnly[str] + resource: ReadOnly[str] + response_types_supported: ReadOnly[tuple[str, ...]] + grant_types_supported: ReadOnly[tuple[str, ...]] + code_challenge_methods_supported: ReadOnly[tuple[str, ...]] + token_endpoint_auth_methods_supported: ReadOnly[tuple[str, ...]] + revocation_endpoint_auth_methods_supported: ReadOnly[tuple[str, ...]] + + +def native_client_auth_contract(request: Request) -> NativeClientAuthContract: + """The versioned discovery document at ``/.well-known/litellm-cli-auth``: everything a + native client (in any language) needs to run the sign-in without reading LiteLLM + source. ``resource`` is the exact value to send as the RFC 8707 ``resource`` parameter + on authorize and token requests so the grant is issued for the proxy API.""" + base_url: Final = get_request_base_url(request) + contract: Final[NativeClientAuthContract] = { + "contract_version": NATIVE_CLIENT_AUTH_CONTRACT_VERSION, + "issuer": base_url, + "authorization_endpoint": f"{base_url}/authorize", + "token_endpoint": f"{base_url}/token", + "registration_endpoint": f"{base_url}/register", + "revocation_endpoint": f"{base_url}/revoke", + "resource": base_url, + "response_types_supported": ("code",), + "grant_types_supported": ("authorization_code", "refresh_token"), + "code_challenge_methods_supported": ("S256",), + "token_endpoint_auth_methods_supported": ("none",), + "revocation_endpoint_auth_methods_supported": ("none",), + } + return contract + + +def is_proxy_api_resource(request: Request, resource: str | None) -> bool: + """True when the RFC 8707 ``resource`` names the proxy itself (its base URL), which is + how a native client asks for the proxy-API audience rather than an MCP session.""" + if resource is None: + return False + canonical: Final = canonical_resource_uri(resource) + return canonical is not None and canonical == canonicalize_url_identity(get_request_base_url(request)) + + +def _rejected_authorize_request( + client_id: str, + redirect_uri: str, + state: str, + code_challenge: str | None, + code_challenge_method: str | None, + response_type: str | None, +) -> Response | None: client: Final = open_gateway_dcr_client(client_id) if client is None: return _oauth_error(400, "invalid_client", "unknown or malformed client_id") @@ -407,14 +604,25 @@ def aggregate_authorize( ) if len(state) > MAX_STATE_LENGTH: return _oauth_error(400, "invalid_request", f"state must be at most {MAX_STATE_LENGTH} characters") - base_url: Final = get_request_base_url(request) - if session_user_id is None: - login_url: Final = f"{base_url}/sso/key/generate?{urlencode({'return_to': relative_request_url(request)})}" - return RedirectResponse(login_url, status_code=303) + return None + + +def _login_redirect(base_url: str, request: Request) -> Response: + return_to: Final = urlencode((("return_to", relative_request_url(request)),)) + return RedirectResponse(f"{base_url}/sso/key/generate?{return_to}", status_code=303) + + +def _new_connect_flow( + session_user_id: str, + client_id: str, + redirect_uri: str, + state: str, + code_challenge: str, + resource_server_id: str | None, + audience: SessionAudience | None, +) -> _ConnectFlow: now: Final = datetime.now(timezone.utc) - scoped_server: Final = resolve_scoped_resource_server(request, resource) - handle: Final = secrets.token_urlsafe(24) - flow: Final = _ConnectFlow( + return _ConnectFlow( user_id=session_user_id, client_id=client_id, redirect_uri=redirect_uri, @@ -422,13 +630,12 @@ def aggregate_authorize( code_challenge=code_challenge, jti=secrets.token_urlsafe(24), exp=int(now.timestamp()) + CONNECT_FLOW_TTL_SECONDS, - resource_server_id=scoped_server.server_id if scoped_server is not None else None, + resource_server_id=resource_server_id, + audience=audience, ) - connect_url: Final = _append_query_params( - f"{base_url}/ui/connect", - {"connect_flow": handle, "connect_client": _origin_only(redirect_uri)}, - ) - response: Final = RedirectResponse(connect_url, status_code=303) + + +def _set_flow_cookie(response: Response, request: Request, handle: str, flow: _ConnectFlow) -> None: path, secure = _cookie_path_and_secure(request) response.set_cookie( key=_flow_cookie_name(handle), @@ -439,7 +646,18 @@ def aggregate_authorize( httponly=True, samesite="lax", ) - return response + + +def _consent_lookup_failure_response(failure: ReloadUserFailure) -> Response: + match failure: + case "unavailable": + return _oauth_error(503, "temporarily_unavailable", "the gateway database is unavailable; retry") + case "unresolvable": + return _oauth_error(500, "server_error", "the gateway is not configured to resolve users") + case "no_active_key": + return _oauth_error(403, "access_denied", "the signed-in user is not active") + case _: + assert_never(failure) def _origin_only(url: str) -> str: @@ -455,6 +673,8 @@ async def complete_connect_flow( session_user_id: str | None, cache: DualCache, delivery: str | None = None, + team_id: str | None = None, + decision: str | None = None, ) -> Response: """The deliberate finish step of the connect flow: mint the gateway authorization code and send the browser back to the client. @@ -479,9 +699,16 @@ async def complete_connect_flow( party. Unknown ``delivery`` values are rejected rather than defaulted: a client that asked for manual delivery and got a dead redirect instead would silently lose its code. + + ``decision`` and ``team_id`` come from the native-client consent page. ``"deny"`` + burns the flow and sends the client ``error=access_denied`` so it stops waiting; + ``team_id`` is sealed into the code only for proxy-API flows, where it picks which of + the user's teams the minted credential is attributed to. """ if delivery not in (None, "redirect", "manual"): return _oauth_error(400, "invalid_request", "delivery must be 'redirect' or 'manual'") + if decision not in (None, "approve", "deny"): + return _oauth_error(400, "invalid_request", "decision must be 'approve' or 'deny'") sealed_flow: Final = request.cookies.get(_flow_cookie_name(flow_handle)) if sealed_flow is None: return _oauth_error(400, "invalid_request", "unknown or expired connect flow") @@ -499,6 +726,24 @@ async def complete_connect_flow( f"{_USED_FLOW_CACHE_PREFIX}{flow.jti}", CONNECT_FLOW_TTL_SECONDS + _CLAIM_TTL_BUFFER_SECONDS ): return _oauth_error(400, "invalid_request", "this connect flow was already completed; restart the connection") + response: Final = ( + _denied_flow_response(flow) if decision == "deny" else _approved_flow_response(flow, delivery, team_id, now) + ) + path, secure = _cookie_path_and_secure(request) + response.delete_cookie(key=_flow_cookie_name(flow_handle), path=path, secure=secure, httponly=True, samesite="lax") + return response + + +def _state_param(flow: _ConnectFlow) -> tuple[tuple[str, str], ...]: + return (("state", flow.state),) if flow.state else () + + +def _denied_flow_response(flow: _ConnectFlow) -> Response: + params: Final = (("error", "access_denied"), *_state_param(flow)) + return RedirectResponse(_append_query_params(flow.redirect_uri, params), status_code=303) + + +def _approved_flow_response(flow: _ConnectFlow, delivery: str | None, team_id: str | None, now: datetime) -> Response: manual_delivery: Final = delivery == "manual" and is_loopback_redirect_host(urlparse(flow.redirect_uri)) code_ttl: Final = MANUAL_DELIVERY_AUTH_CODE_TTL_SECONDS if manual_delivery else GATEWAY_AUTH_CODE_TTL_SECONDS code: Final = _seal( @@ -512,16 +757,14 @@ async def complete_connect_flow( iat=int(now.timestamp()), exp=int(now.timestamp()) + code_ttl, resource_server_id=flow.resource_server_id, + audience=flow.audience, + team_id=(team_id or None) if flow.audience == PROXY_API_AUDIENCE else None, ), ) - params: Final = {"code": code, **({"state": flow.state} if flow.state else {})} - callback_url: Final = _append_query_params(flow.redirect_uri, params) - response: Final[Response] = ( - _manual_delivery_response(callback_url) if manual_delivery else RedirectResponse(callback_url, status_code=303) - ) - path, secure = _cookie_path_and_secure(request) - response.delete_cookie(key=_flow_cookie_name(flow_handle), path=path, secure=secure, httponly=True, samesite="lax") - return response + callback_url: Final = _append_query_params(flow.redirect_uri, (("code", code), *_state_param(flow))) + if manual_delivery: + return _manual_delivery_response(callback_url) + return RedirectResponse(callback_url, status_code=303) def _manual_delivery_response(callback_url: str) -> Response: @@ -630,6 +873,37 @@ def _session_token_pair(principal: SessionPrincipal, keys: SessionKeys, now: dat ) +class _ProxyCredentialTokenResponse(TypedDict): + access_token: ReadOnly[str] + token_type: ReadOnly[Literal["Bearer"]] + expires_in: ReadOnly[int] + refresh_token: ReadOnly[str] + user_id: ReadOnly[str] + team_id: ReadOnly[str | None] + + +def _proxy_credential_response( + minted: MintedProxyCredential, principal: SessionPrincipal, keys: SessionKeys, now: datetime +) -> Response: + """The proxy-API token response: the access token is the very credential ``lite + login`` stores (accepted on every proxy route with user and team attribution), and + the refresh token is a gateway-sealed rotating token bound to the team the credential + was minted for, so a renewal keeps the team the user consented to.""" + bound_principal: Final = principal.model_copy(update=MappingProxyType({"team_id": minted.team_id})) + refresh: Final = mint_session_refresh_token(bound_principal, keys, now) + if not isinstance(refresh, MintedSessionToken): + return _oauth_error(500, "server_error", "failed to mint the session credential") + body: Final[_ProxyCredentialTokenResponse] = { + "access_token": minted.key, + "token_type": "Bearer", + "expires_in": minted.expires_in, + "refresh_token": refresh.token.get_secret_value(), + "user_id": minted.user_id, + "team_id": minted.team_id, + } + return JSONResponse(status_code=200, content=body, headers=TOKEN_NO_CACHE_HEADERS) + + def _reload_failure_response(failure: ReloadUserFailure) -> Response: """Map the live-user revalidation failure onto its OAuth error, exhaustively, so a new ``ReloadUserFailure`` member is a type error here rather than silently 400ing.""" @@ -644,6 +918,18 @@ def _reload_failure_response(failure: ReloadUserFailure) -> Response: assert_never(failure) +def _mint_failure_response(failure: ProxyCredentialMintFailure) -> Response: + match failure: + case "not_a_member": + return _oauth_error( + 400, "invalid_grant", "the user is no longer a member of the team this grant was issued for" + ) + case "unavailable" | "unresolvable" | "no_active_key": + return _reload_failure_response(failure) + case _: + assert_never(failure) + + def _resource_conflicts_with_scope( request: Request, resource: str | None, sealed_resource_server_id: str | None ) -> bool: @@ -670,15 +956,26 @@ async def aggregate_token( reload_user: ReloadUser, cache: DualCache, resource: str | None = None, + mint_proxy_credential: MintProxyCredential = _refuse_proxy_credential, ) -> Response: """The aggregate token verb: authorization_code and refresh_token grants for the - identity-only session pair. Every path re-validates the litellm user live before - minting, so a deactivated user cannot obtain or renew a session.""" + identity-only session pair, or for the proxy-API credential when the grant was issued + with that audience. Every path re-validates the litellm user live before minting, so a + deactivated user cannot obtain or renew a session.""" if master_key is None: verbose_logger.error("mcp_gateway_dcr token grant rejected: no master_key configured") return _oauth_error(500, "server_error", "the gateway has no master key configured") keys: Final = session_keys_from_master_key(master_key) now: Final = datetime.now(timezone.utc) + issue: Final = _GrantIssuer( + request=request, + resource=resource, + keys=keys, + now=now, + reload_user=reload_user, + mint_proxy_credential=mint_proxy_credential, + guard=_SingleUseGuard(cache), + ) if grant_type == "authorization_code": return await _authorization_code_grant( request=request, @@ -687,10 +984,8 @@ async def aggregate_token( client_id=client_id, code_verifier=code_verifier, resource=resource, - keys=keys, now=now, - reload_user=reload_user, - guard=_SingleUseGuard(cache), + issue=issue, ) if grant_type == "refresh_token": return await _refresh_token_grant( @@ -700,12 +995,71 @@ async def aggregate_token( resource=resource, keys=keys, now=now, - reload_user=reload_user, - guard=_SingleUseGuard(cache), + issue=issue, ) return _oauth_error(400, "unsupported_grant_type", "grant_type must be authorization_code or refresh_token") +class _GrantIssuer: + """The tail every grant shares once its own proof (code + PKCE, or a refresh token) + has checked out: revalidate the user live, claim the single-use marker, mint. The + claim comes AFTER revalidation and minting so a transient DB 503 never burns a + still-valid code or refresh token, and fails closed when it cannot be recorded.""" + + def __init__( + self, + request: Request, + resource: str | None, + keys: SessionKeys, + now: datetime, + reload_user: ReloadUser, + mint_proxy_credential: MintProxyCredential, + guard: _SingleUseGuard, + ) -> None: + self._request: Final = request + self._resource: Final = resource + self._keys: Final = keys + self._now: Final = now + self._reload_user: Final = reload_user + self._mint_proxy_credential: Final = mint_proxy_credential + self._guard: Final = guard + + async def __call__( + self, principal: SessionPrincipal, claim_key: str, claim_ttl_seconds: int, replayed: str + ) -> Response: + match principal.audience: + case None: + return await self._issue_session_pair(principal, claim_key, claim_ttl_seconds, replayed) + case "proxy_api": + return await self._issue_proxy_credential(principal, claim_key, claim_ttl_seconds, replayed) + case _: + assert_never(principal.audience) + + async def _issue_session_pair( + self, principal: SessionPrincipal, claim_key: str, claim_ttl_seconds: int, replayed: str + ) -> Response: + failure: Final = await self._reload_user(principal.user_id) + if failure is not None: + return _reload_failure_response(failure) + if not await self._guard.claim(claim_key, claim_ttl_seconds): + return _oauth_error(400, "invalid_grant", replayed) + return _session_token_pair(principal, self._keys, self._now) + + async def _issue_proxy_credential( + self, principal: SessionPrincipal, claim_key: str, claim_ttl_seconds: int, replayed: str + ) -> Response: + if self._resource is not None and not is_proxy_api_resource(self._request, self._resource): + return _oauth_error( + 400, "invalid_target", "resource does not match the proxy API this grant was issued for" + ) + minted: Final = await self._mint_proxy_credential(principal.user_id, principal.team_id) + if not isinstance(minted, MintedProxyCredential): + return _mint_failure_response(minted) + if not await self._guard.claim(claim_key, claim_ttl_seconds): + return _oauth_error(400, "invalid_grant", replayed) + return _proxy_credential_response(minted, principal, self._keys, self._now) + + async def _authorization_code_grant( request: Request, code: str | None, @@ -713,10 +1067,8 @@ async def _authorization_code_grant( client_id: str, code_verifier: str | None, resource: str | None, - keys: SessionKeys, now: datetime, - reload_user: ReloadUser, - guard: _SingleUseGuard, + issue: _GrantIssuer, ) -> Response: if not code or not redirect_uri or not code_verifier: return _oauth_error(400, "invalid_request", "code, redirect_uri, and code_verifier are required") @@ -733,23 +1085,19 @@ async def _authorization_code_grant( return _oauth_error(400, "invalid_target", "resource does not match the scope this code was issued for") if not _pkce_verifier_matches(code_verifier, parsed.code_challenge): return _oauth_error(400, "invalid_grant", "PKCE verification failed") - # Revalidate the user BEFORE claiming the code, so a transient DB outage (a retryable - # 503) does not consume a still-valid code and force the client to restart sign-in. - failure: Final = await reload_user(parsed.user_id) - if failure is not None: - return _reload_failure_response(failure) - # Atomic single-use claim is the gate: on a concurrent double-redeem exactly one caller - # wins, and a claim that cannot be recorded fails closed. The marker's TTL derives from - # the code's own remaining lifetime so it outlives whichever lifetime the code was minted with. - if not await guard.claim( - f"{_USED_CODE_CACHE_PREFIX}{parsed.jti}", - parsed.exp - int(now.timestamp()) + _CLAIM_TTL_BUFFER_SECONDS, - ): - return _oauth_error(400, "invalid_grant", "the authorization code was already used") - return _session_token_pair( - SessionPrincipal(user_id=parsed.user_id, client_id=client_id, resource_server_id=parsed.resource_server_id), - keys, - now, + # The marker's TTL derives from the code's own remaining lifetime so it outlives + # whichever lifetime the code was minted with. + return await issue( + SessionPrincipal( + user_id=parsed.user_id, + client_id=client_id, + resource_server_id=parsed.resource_server_id, + audience=parsed.audience, + team_id=parsed.team_id, + ), + claim_key=f"{_USED_CODE_CACHE_PREFIX}{parsed.jti}", + claim_ttl_seconds=parsed.exp - int(now.timestamp()) + _CLAIM_TTL_BUFFER_SECONDS, + replayed="the authorization code was already used", ) @@ -760,8 +1108,7 @@ async def _refresh_token_grant( resource: str | None, keys: SessionKeys, now: datetime, - reload_user: ReloadUser, - guard: _SingleUseGuard, + issue: _GrantIssuer, ) -> Response: if not refresh_token: return _oauth_error(400, "invalid_request", "refresh_token is required") @@ -770,16 +1117,33 @@ async def _refresh_token_grant( return _oauth_error(400, "invalid_grant", "the refresh token is invalid for this client") if _resource_conflicts_with_scope(request, resource, opened.principal.resource_server_id): return _oauth_error(400, "invalid_target", "resource does not match the scope this token was issued for") - failure: Final = await reload_user(opened.principal.user_id) - if failure is not None: - return _reload_failure_response(failure) # Refresh-token rotation (OAuth 2.0 Security BCP section 4.13): the presented refresh token is - # single-use. Claim its jti before issuing the replacement pair, so a captured or replayed - # refresh token cannot mint a second pair after the legitimate holder rotated. Claimed AFTER - # user revalidation so a transient DB 503 does not burn a still-valid token; a claim that - # cannot be recorded fails closed, exactly like the authorization-code path. - if not await guard.claim( - f"{_USED_REFRESH_CACHE_PREFIX}{opened.jti}", SESSION_REFRESH_TTL_SECONDS + _CLAIM_TTL_BUFFER_SECONDS - ): - return _oauth_error(400, "invalid_grant", "the refresh token was already used") - return _session_token_pair(opened.principal, keys, now) + # single-use, so a captured or replayed refresh token cannot mint a second pair after the + # legitimate holder rotated. + return await issue( + opened.principal, + claim_key=f"{_USED_REFRESH_CACHE_PREFIX}{opened.jti}", + claim_ttl_seconds=SESSION_REFRESH_TTL_SECONDS + _CLAIM_TTL_BUFFER_SECONDS, + replayed="the refresh token was already used", + ) + + +async def revoke_refresh_token(token: str, client_id: str, master_key: str | None, cache: DualCache) -> Response: + """RFC 7009 revocation for the gateway's refresh tokens: burn the presented token's + ``jti`` so neither the holder nor a thief can rotate it again. Access tokens are + stateless and expire on their own (the proxy-API credential within + ``CLI_JWT_EXPIRATION_HOURS``), so per RFC 7009 section 2.2 an unrecognized or already + dead token still answers 200; only an unknown client is refused.""" + if not is_gateway_dcr_client_id(client_id) or open_gateway_dcr_client(client_id) is None: + return _oauth_error(401, "invalid_client", "unknown or malformed client_id") + if master_key is None: + verbose_logger.error("mcp_gateway_dcr revoke rejected: no master_key configured") + return _oauth_error(500, "server_error", "the gateway has no master key configured") + keys: Final = session_keys_from_master_key(master_key) + now: Final = datetime.now(timezone.utc) + opened: Final = open_session_refresh_bearer(token, keys, now, expected_client_id=client_id) + if isinstance(opened, SessionRefreshOpened): + _ = await _SingleUseGuard(cache).claim( + f"{_USED_REFRESH_CACHE_PREFIX}{opened.jti}", SESSION_REFRESH_TTL_SECONDS + _CLAIM_TTL_BUFFER_SECONDS + ) + return Response(content="{}", media_type="application/json", headers=TOKEN_NO_CACHE_HEADERS) diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/session_token.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/session_token.py index 15f5f82c4b6..d6b0a462062 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/session_token.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/session_token.py @@ -76,6 +76,13 @@ SessionTokenKind = Literal["session", "session_refresh"] on open, so a signature-valid token of one kind cannot be replayed as the other even if its wire prefix is swapped (the prefix is not part of the signed payload; this claim is).""" +SessionAudience = Literal["proxy_api"] +"""The non-MCP audience a session REFRESH token can be minted for. ``None`` (the default and +the only value ever on an MCP wire) means the aggregate MCP gateway; ``"proxy_api"`` means the +refresh grant re-mints the proxy-API CLI credential instead of an MCP session pair. The audience +is read only from the signed claims, never from the request, so a token of one audience can +never be redeemed as the other.""" + class SessionPrincipal(BaseModel): """The litellm user a session token identifies and the DCR client it was issued to. @@ -97,6 +104,8 @@ class SessionPrincipal(BaseModel): user_id: str = Field(min_length=1) client_id: str = Field(min_length=1) resource_server_id: str | None = None + audience: SessionAudience | None = None + team_id: str | None = None class SessionKeys(BaseModel): @@ -194,6 +203,8 @@ class _SessionClaims(BaseModel): user_id: str = Field(min_length=1) client_id: str = Field(min_length=1) resource_server_id: str | None = None + audience: SessionAudience | None = None + team_id: str | None = None def is_session_token(candidate: str) -> bool: @@ -295,6 +306,8 @@ def _mint( user_id=principal.user_id, client_id=principal.client_id, resource_server_id=principal.resource_server_id, + audience=principal.audience, + team_id=principal.team_id, ) token: Final = prefix + jwt.encode( claims.model_dump(exclude_none=True), keys.signing_key.get_secret_value(), algorithm=_SESSION_JWT_ALGORITHM @@ -333,7 +346,11 @@ def _open( return SessionExpired() return OpenedSessionToken( principal=SessionPrincipal( - user_id=claims.user_id, client_id=claims.client_id, resource_server_id=claims.resource_server_id + user_id=claims.user_id, + client_id=claims.client_id, + resource_server_id=claims.resource_server_id, + audience=claims.audience, + team_id=claims.team_id, ), jti=claims.jti, ) diff --git a/litellm/proxy/_experimental/mcp_server/proxy_api_credentials.py b/litellm/proxy/_experimental/mcp_server/proxy_api_credentials.py new file mode 100644 index 00000000000..a24afc529bd --- /dev/null +++ b/litellm/proxy/_experimental/mcp_server/proxy_api_credentials.py @@ -0,0 +1,84 @@ +"""The proxy-API side of the native-client sign-in: turning a consented OAuth grant into +the same per-user credential ``lite login`` stores, so the bearer a CLI obtains through +the browser flow is accepted on every proxy route with user and team attribution.""" + +from __future__ import annotations + +from collections.abc import Sequence +from typing import Final + +from litellm.constants import CLI_JWT_EXPIRATION_HOURS +from litellm.proxy._experimental.mcp_server.bridge_token_flow import load_active_user_by_id +from litellm.proxy._experimental.mcp_server.gateway_dcr_flow import ( + ConsentTeam, + MintedProxyCredential, + ProxyCredentialMintFailure, + ReloadUserFailure, +) +from litellm.proxy._types import LiteLLM_UserTable +from litellm.proxy.auth.auth_checks import ExperimentalUIJWTToken +from litellm.proxy.management_endpoints.ui_sso import ( + CliSsoTeamDetail, + fetch_cli_sso_team_details, + selected_cli_sso_team_detail, +) + + +async def lookup_consent_teams(user_id: str) -> tuple[ConsentTeam, ...] | ReloadUserFailure: + user: Final = await load_active_user_by_id(user_id) + if isinstance(user, str): + return user + details: Final = await _team_details(user.teams) + if details is None: + return "unavailable" + return tuple( + ConsentTeam(team_id=detail.team_id, team_alias=detail.team_alias) + for detail in details + if detail.team_id is not None + ) + + +async def mint_proxy_credential( + user_id: str, team_id: str | None +) -> MintedProxyCredential | ProxyCredentialMintFailure: + """Mint the ``lite login`` credential for a consented grant. Membership is checked + live, so a team the user left between consent and redemption (or between refreshes) + refuses the grant instead of minting a credential attributed to a team they are no + longer on. The team is exactly the one the consent page sealed into the grant; nothing + is picked on the user's behalf here, so a refresh can never move the credential. The + user row handed to the minter carries no team list, exactly like ``lite login``'s, so + the minter's own first-team fallback stays inert.""" + user: Final = await load_active_user_by_id(user_id) + if isinstance(user, str): + return user + if user.user_role is None: + return "no_active_key" + if team_id is not None and team_id not in user.teams: + return "not_a_member" + details: Final = await _team_details(user.teams) if team_id is not None else () + if details is None: + return "unavailable" + selected: Final = selected_cli_sso_team_detail(details, team_id) + if selected is None: + return "not_a_member" + key: Final = ExperimentalUIJWTToken.get_cli_jwt_auth_token( + user_info=LiteLLM_UserTable(user_id=user.user_id, user_role=user.user_role, models=user.models), + team_id=team_id, + team_alias=selected.team_alias, + team_models=selected.team_models, + team_model_aliases=selected.team_model_aliases, + ) + return MintedProxyCredential( + key=key, + expires_in=CLI_JWT_EXPIRATION_HOURS * 3600, + user_id=user.user_id, + team_id=team_id, + ) + + +async def _team_details(teams: Sequence[str]) -> tuple[CliSsoTeamDetail, ...] | None: + from litellm.proxy.proxy_server import prisma_client # noqa: PLC0415 # rebound after startup, so read it per call + + if prisma_client is None: + return None + return await fetch_cli_sso_team_details(prisma_client, teams) diff --git a/litellm/proxy/_lazy_features.py b/litellm/proxy/_lazy_features.py index 262d97e4579..fdd15a89aa5 100644 --- a/litellm/proxy/_lazy_features.py +++ b/litellm/proxy/_lazy_features.py @@ -155,10 +155,12 @@ LAZY_FEATURES: Final[tuple[LazyFeature, ...]] = ( "/.well-known/oauth-", "/.well-known/openid-configuration", "/.well-known/jwks.json", + "/.well-known/litellm-cli-auth", "/authorize", "/token", "/callback", "/register", + "/revoke", ), # Catches the /{mcp_server_name}/authorize|token|register variants. path_suffixes=("/authorize", "/token", "/register"), diff --git a/litellm/proxy/client/cli/commands/auth.py b/litellm/proxy/client/cli/commands/auth.py index 0a0bcf80ee5..aee715686b7 100644 --- a/litellm/proxy/client/cli/commands/auth.py +++ b/litellm/proxy/client/cli/commands/auth.py @@ -11,7 +11,7 @@ import click import requests from rich.console import Console from rich.table import Table -from typing_extensions import NotRequired, TypedDict +from typing_extensions import NotRequired, ReadOnly, TypedDict from litellm.constants import CLI_JWT_EXPIRATION_HOURS from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh @@ -22,6 +22,13 @@ from .claude_settings import ( ClaudeSettingsError, write_claude_settings, ) +from .pkce_login import ( + PkceFailure, + fresh_api_key, + pkce_token_record, + revoke_stored_credential, + run_pkce_login, +) from .private_json import write_private_json @@ -34,6 +41,13 @@ class CliTokenData(TypedDict): auth_header_name: str jwt_token: str timestamp: float + expires_at: ReadOnly[NotRequired[float]] + refresh_token: ReadOnly[NotRequired[str]] + client_id: ReadOnly[NotRequired[str]] + token_endpoint: ReadOnly[NotRequired[str]] + revocation_endpoint: ReadOnly[NotRequired[str]] + resource: ReadOnly[NotRequired[str]] + team_id: ReadOnly[NotRequired[str | None]] class CliTeam(TypedDict, total=False): @@ -79,10 +93,7 @@ class CliAuthResult(TypedDict): # Token storage utilities def get_token_file_path() -> str: """Get the path to store the authentication token""" - home_dir: Final = Path.home() - config_dir: Final = home_dir / ".litellm" - config_dir.mkdir(exist_ok=True) - return str(config_dir / "token.json") + return str(Path.home() / ".litellm" / "token.json") def save_token(token_data: CliTokenData) -> None: @@ -115,11 +126,15 @@ def get_stored_api_key(expected_base_url: str | None = None) -> str | None: If expected_base_url is provided, the key is only returned when it was originally issued for that URL. This prevents credential leakage when the - CLI is pointed at a different (possibly malicious) server. + CLI is pointed at a different (possibly malicious) server. A key obtained by + ``lite login --pkce`` is refreshed here once it nears expiry. """ - from litellm.litellm_core_utils.cli_token_utils import get_litellm_gateway_api_key - - return get_litellm_gateway_api_key(expected_base_url=expected_base_url) + token_data: Final = load_token() + if token_data is None: + return None + if expected_base_url is not None and token_data.get("base_url") != expected_base_url.rstrip("/"): + return None + return fresh_api_key(token_data, save_token, requests.Session(), reload=load_token) # Team selection utilities @@ -645,6 +660,27 @@ def _configure_claude_code(base_url: str) -> None: click.echo("Your other Claude Code settings were left untouched. Restart Claude Code to pick this up.") +def _finish_login(base_url: str, api_key: str, config_claude: bool) -> None: + from litellm.proxy.client.cli.interface import show_commands + + click.echo("\nLogin successful!") + click.echo(f"JWT Token: {api_key[:20]}...") + click.echo("You can now use the CLI without specifying --api-key") + if config_claude: + _configure_claude_code(base_url) + click.echo("\n" + "=" * 60) + show_commands() + + +def _pkce_login(base_url: str, config_claude: bool) -> None: + credential: Final = run_pkce_login(base_url, requests.Session(), echo=click.echo) + if isinstance(credential, PkceFailure): + click.echo(f"Authentication failed: {credential.reason}") + return + save_token(pkce_token_record(base_url, credential)) + _finish_login(base_url, credential.access_token, config_claude) + + @click.command(name="login") @click.option( "--config-claude", @@ -655,16 +691,28 @@ def _configure_claude_code(base_url: str) -> None: "Unrelated settings are preserved." ), ) +@click.option( + "--pkce", + is_flag=True, + default=False, + help=( + "Sign in with OAuth authorization code + PKCE through your system browser (loopback redirect), " + "with a refresh token that renews the key automatically. Requires a proxy that serves " + "/.well-known/litellm-cli-auth." + ), +) @click.pass_context -def login(ctx: click.Context, config_claude: bool): +def login(ctx: click.Context, config_claude: bool, pkce: bool) -> None: """Login to LiteLLM proxy using SSO authentication""" from litellm.constants import LITELLM_CLI_SOURCE_IDENTIFIER - from litellm.proxy.client.cli.interface import show_commands ctx_obj: Final[CliContextObj] = ctx.obj base_url: Final = ctx_obj["base_url"] try: + if pkce: + _pkce_login(base_url, config_claude) + return cli_sso_flow: Final = _start_cli_sso_flow(base_url=base_url) key_id: Final = cli_sso_flow["login_id"] poll_secret: Final = cli_sso_flow["poll_secret"] @@ -704,16 +752,7 @@ def login(ctx: click.Context, config_claude: bool): } ) - click.echo("\nLogin successful!") - click.echo(f"JWT Token: {api_key[:20]}...") - click.echo("You can now use the CLI without specifying --api-key") - - if config_claude: - _configure_claude_code(base_url) - - # Show available commands after successful login - click.echo("\n" + "=" * 60) - show_commands() + _finish_login(base_url, api_key, config_claude) return else: click.echo("Authentication timed out. Please try again.") @@ -738,7 +777,11 @@ def login(ctx: click.Context, config_claude: bool): @click.command(name="logout") def logout(): """Logout and clear stored authentication""" + token_data: Final = load_token() + revocation: Final = revoke_stored_credential(token_data, requests.Session()) if token_data is not None else None clear_token() + if revocation is not None: + click.echo(f"Could not revoke the refresh token on the proxy ({revocation.reason}); it expires on its own.") click.echo("Logged out successfully. Authentication token cleared.") @@ -769,13 +812,13 @@ def print_token(ctx: click.Context): click.echo("Not authenticated for this server. Run 'lite login'.", err=True) sys.exit(1) - if not is_cli_token_fresh(token_data): + if not is_cli_token_fresh(token_data) and "refresh_token" not in token_data: click.echo("Token expired. Run 'lite login' again.", err=True) sys.exit(1) - api_key: Final = token_data.get("key") + api_key: Final = fresh_api_key(token_data, save_token, requests.Session(), reload=load_token) if not api_key: - click.echo("No token available. Run 'lite login'.", err=True) + click.echo("Token expired. Run 'lite login' again.", err=True) sys.exit(1) click.echo(api_key) diff --git a/litellm/proxy/client/cli/commands/pkce_login.py b/litellm/proxy/client/cli/commands/pkce_login.py new file mode 100644 index 00000000000..6d6646a886c --- /dev/null +++ b/litellm/proxy/client/cli/commands/pkce_login.py @@ -0,0 +1,471 @@ +"""Browser sign-in for ``lite login --pkce``: OAuth 2.1 authorization code + PKCE S256 +against the proxy's own authorization server, as a public client on a loopback redirect. +The proxy publishes everything this needs at ``/.well-known/litellm-cli-auth``, so a CLI +in any other language can run the same steps from that document alone.""" + +from __future__ import annotations + +import hashlib +import secrets +import socket +import threading +import time +import webbrowser +from base64 import urlsafe_b64encode +from collections.abc import Callable, Mapping, Sequence +from dataclasses import dataclass +from http.server import BaseHTTPRequestHandler, HTTPServer +from types import MappingProxyType +from typing import TYPE_CHECKING, Final, Literal, Protocol +from urllib.parse import parse_qs, urlencode, urlparse + +import requests +from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError +from typing_extensions import ReadOnly, TypedDict + +if TYPE_CHECKING: + from .auth import CliTokenData + +CLI_AUTH_DISCOVERY_PATH: Final = "/.well-known/litellm-cli-auth" +CALLBACK_PATH: Final = "/callback" +LOGIN_TIMEOUT_SECONDS: Final = 300 +REFRESH_LEEWAY_SECONDS: Final = 60 +_HTTP_TIMEOUT_SECONDS: Final = 15 +_CLIENT_NAME: Final = "litellm-cli" + + +class CliAuthContract(BaseModel): + model_config = ConfigDict(frozen=True) + + contract_version: Literal[1] + issuer: str + authorization_endpoint: str + token_endpoint: str + registration_endpoint: str + revocation_endpoint: str + resource: str + code_challenge_methods_supported: tuple[str, ...] + + +class _RegisteredClient(BaseModel): + model_config = ConfigDict(frozen=True) + + client_id: str = Field(min_length=1) + + +class _TokenResponse(BaseModel): + model_config = ConfigDict(frozen=True) + + access_token: str = Field(min_length=1) + expires_in: int = Field(gt=0) + refresh_token: str = Field(min_length=1) + user_id: str | None = None + team_id: str | None = None + + +@dataclass(frozen=True, slots=True) +class PkceFailure: + reason: str + + +@dataclass(frozen=True, slots=True) +class PkceCredential: + access_token: str + refresh_token: str + expires_at: float + client_id: str + token_endpoint: str + revocation_endpoint: str + resource: str + user_id: str | None + team_id: str | None + + +@dataclass(frozen=True, slots=True) +class CallbackCode: + code: str + + +@dataclass(frozen=True, slots=True) +class CallbackDenied: + error: str + description: str | None + + +CallbackOutcome = CallbackCode | CallbackDenied + + +class Http(Protocol): + def get(self, url: str, *, timeout: float) -> requests.Response: ... + + def post( + self, + url: str, + *, + data: Mapping[str, str] | None = None, + json: Mapping[str, object] | None = None, + timeout: float, + ) -> requests.Response: ... + + +class LoopbackServer(HTTPServer): + """The OS-assigned loopback listener the browser is sent back to. Only the response + carrying the pending sign-in's ``state`` settles it; anything else (a stray request, a + stale tab, an attacker poking the port) gets a 400 and the wait continues. A connection + that opens and then sends nothing is dropped after ``connection_timeout_seconds`` so it + cannot hold the single-threaded wait past its deadline.""" + + def __init__(self, expected_state: str, connection_timeout_seconds: float = 5) -> None: + super().__init__(("127.0.0.1", 0), _CallbackHandler) + self.expected_state: Final = expected_state + self.connection_timeout_seconds: Final = connection_timeout_seconds + self.outcome: CallbackOutcome | None = None + self.timeout = 1 + + @property + def redirect_uri(self) -> str: + return f"http://127.0.0.1:{self.server_address[1]}{CALLBACK_PATH}" + + def get_request(self) -> tuple[socket.socket, object]: + accepted: Final[tuple[socket.socket, object]] = super().get_request() + accepted[0].settimeout(self.connection_timeout_seconds) + return accepted + + def wait( + self, timeout_seconds: float, clock: Callable[[], float] = time.monotonic + ) -> CallbackOutcome | PkceFailure: + deadline: Final = clock() + timeout_seconds + while self.outcome is None: + if clock() >= deadline: + return PkceFailure("timed out waiting for the browser sign-in to finish") + self.handle_request() + return self.outcome + + +class _CallbackHandler(BaseHTTPRequestHandler): + server: LoopbackServer # pyright: ignore[reportIncompatibleVariableOverride] # only ever constructed by LoopbackServer + + def do_GET(self) -> None: + parsed: Final = urlparse(self.path) + if parsed.path != CALLBACK_PATH: + self._respond(404, "Not found.") + return + params: Final = parse_qs(parsed.query) + if _first(params, "state") != self.server.expected_state: + self._respond(400, "This response does not belong to the pending sign-in; still waiting.") + return + error: Final = _first(params, "error") + if error is not None: + self.server.outcome = CallbackDenied(error=error, description=_first(params, "error_description")) + self._respond(200, "Sign-in was not approved. You can close this window.") + return + code: Final = _first(params, "code") + if code is None: + self._respond(400, "The sign-in response carried no authorization code; still waiting.") + return + self.server.outcome = CallbackCode(code=code) + self._respond(200, "Signed in to LiteLLM. You can close this window and return to the terminal.") + + def log_message(self, format: str, *args: object) -> None: + return + + def _respond(self, status: int, text: str) -> None: + body: Final = text.encode("utf-8") + self.send_response(status) + self.send_header("Content-Type", "text/plain; charset=utf-8") + self.send_header("Content-Length", str(len(body))) + self.send_header("Cache-Control", "no-store") + self.end_headers() + self.wfile.write(body) + + +def _first(params: Mapping[str, Sequence[str]], key: str) -> str | None: + values: Final = params.get(key) + return values[0] if values else None + + +def discover_cli_auth(base_url: str, http: Http) -> CliAuthContract | PkceFailure: + url: Final = f"{base_url.rstrip('/')}{CLI_AUTH_DISCOVERY_PATH}" + try: + response: Final = http.get(url, timeout=_HTTP_TIMEOUT_SECONDS) + except requests.RequestException as exc: + return PkceFailure(f"could not reach {url}: {exc}") + if response.status_code != 200: + return PkceFailure( + f"{url} answered {response.status_code}; this proxy version does not support `lite login --pkce`" + ) + try: + contract: Final = CliAuthContract.model_validate(response.json()) + except (ValueError, ValidationError) as exc: + return PkceFailure(f"{url} returned an unsupported discovery document: {exc}") + if "S256" not in contract.code_challenge_methods_supported: + return PkceFailure("the proxy does not support PKCE S256") + return contract + + +class _ClientRegistration(TypedDict): + client_name: ReadOnly[str] + redirect_uris: ReadOnly[tuple[str, ...]] + grant_types: ReadOnly[tuple[str, ...]] + response_types: ReadOnly[tuple[str, ...]] + token_endpoint_auth_method: ReadOnly[Literal["none"]] + + +def _form(**fields: str) -> Mapping[str, str]: + return MappingProxyType(fields) + + +def register_client(contract: CliAuthContract, redirect_uri: str, http: Http) -> str | PkceFailure: + registration: Final[_ClientRegistration] = { + "client_name": _CLIENT_NAME, + "redirect_uris": (redirect_uri,), + "grant_types": ("authorization_code", "refresh_token"), + "response_types": ("code",), + "token_endpoint_auth_method": "none", + } + try: + response: Final = http.post(contract.registration_endpoint, json=registration, timeout=_HTTP_TIMEOUT_SECONDS) + except requests.RequestException as exc: + return PkceFailure(f"client registration failed: {exc}") + if response.status_code not in (200, 201): + return PkceFailure(f"client registration failed with {response.status_code}: {_error_detail(response)}") + try: + return _RegisteredClient.model_validate(response.json()).client_id + except (ValueError, ValidationError) as exc: + return PkceFailure(f"client registration returned an unexpected body: {exc}") + + +def pkce_pair() -> tuple[str, str]: + verifier: Final = secrets.token_urlsafe(64) + digest: Final = hashlib.sha256(verifier.encode("ascii")).digest() + return verifier, urlsafe_b64encode(digest).rstrip(b"=").decode("ascii") + + +def authorize_url(contract: CliAuthContract, client_id: str, redirect_uri: str, state: str, code_challenge: str) -> str: + query: Final = urlencode( + _form( + response_type="code", + client_id=client_id, + redirect_uri=redirect_uri, + state=state, + code_challenge=code_challenge, + code_challenge_method="S256", + resource=contract.resource, + ) + ) + return f"{contract.authorization_endpoint}?{query}" + + +def redeem_code( + contract: CliAuthContract, + client_id: str, + redirect_uri: str, + code: str, + code_verifier: str, + http: Http, + now: Callable[[], float] = time.time, +) -> PkceCredential | PkceFailure: + return _token_request( + token_endpoint=contract.token_endpoint, + revocation_endpoint=contract.revocation_endpoint, + resource=contract.resource, + client_id=client_id, + form=_form( + grant_type="authorization_code", + code=code, + redirect_uri=redirect_uri, + client_id=client_id, + code_verifier=code_verifier, + resource=contract.resource, + ), + http=http, + now=now, + ) + + +def refresh_credential( + token_endpoint: str, + revocation_endpoint: str, + resource: str, + client_id: str, + refresh_token: str, + http: Http, + now: Callable[[], float] = time.time, +) -> PkceCredential | PkceFailure: + return _token_request( + token_endpoint=token_endpoint, + revocation_endpoint=revocation_endpoint, + resource=resource, + client_id=client_id, + form=_form(grant_type="refresh_token", refresh_token=refresh_token, client_id=client_id, resource=resource), + http=http, + now=now, + ) + + +def _token_request( + token_endpoint: str, + revocation_endpoint: str, + resource: str, + client_id: str, + form: Mapping[str, str], + http: Http, + now: Callable[[], float], +) -> PkceCredential | PkceFailure: + try: + response: Final = http.post(token_endpoint, data=form, timeout=_HTTP_TIMEOUT_SECONDS) + except requests.RequestException as exc: + return PkceFailure(f"token request failed: {exc}") + if response.status_code != 200: + return PkceFailure(f"token request failed with {response.status_code}: {_error_detail(response)}") + try: + token: Final = _TokenResponse.model_validate(response.json()) + except (ValueError, ValidationError) as exc: + return PkceFailure(f"token endpoint returned an unexpected body: {exc}") + return PkceCredential( + access_token=token.access_token, + refresh_token=token.refresh_token, + expires_at=now() + token.expires_in, + client_id=client_id, + token_endpoint=token_endpoint, + revocation_endpoint=revocation_endpoint, + resource=resource, + user_id=token.user_id, + team_id=token.team_id, + ) + + +def revoke_credential(revocation_endpoint: str, client_id: str, refresh_token: str, http: Http) -> PkceFailure | None: + try: + response: Final = http.post( + revocation_endpoint, + data=_form(token=refresh_token, token_type_hint="refresh_token", client_id=client_id), + timeout=_HTTP_TIMEOUT_SECONDS, + ) + except requests.RequestException as exc: + return PkceFailure(f"revocation request failed: {exc}") + if response.status_code != 200: + return PkceFailure(f"revocation failed with {response.status_code}: {_error_detail(response)}") + return None + + +_ERROR_BODY: Final = TypeAdapter(Mapping[str, object]) + + +def _error_detail(response: requests.Response) -> str: + try: + body: Final = _ERROR_BODY.validate_json(response.content) + except ValidationError: + return response.text[:200] + return str(body.get("error_description") or body.get("error") or body.get("detail") or body)[:200] + + +def run_pkce_login( + base_url: str, + http: Http, + open_browser: Callable[[str], object] = webbrowser.open, + echo: Callable[[str], None] = print, + timeout_seconds: float = LOGIN_TIMEOUT_SECONDS, +) -> PkceCredential | PkceFailure: + contract: Final = discover_cli_auth(base_url, http) + if isinstance(contract, PkceFailure): + return contract + state: Final = secrets.token_urlsafe(32) + verifier, challenge = pkce_pair() + with LoopbackServer(state) as server: + client_id: Final = register_client(contract, server.redirect_uri, http) + if isinstance(client_id, PkceFailure): + return client_id + url: Final = authorize_url(contract, client_id, server.redirect_uri, state, challenge) + echo(f"Opening browser to: {url}") + echo("Approve the sign-in in your browser. Waiting...") + threading.Thread(target=open_browser, args=(url,), name="lite-login-browser", daemon=True).start() + outcome: Final = server.wait(timeout_seconds) + match outcome: + case PkceFailure(): + return outcome + case CallbackDenied(): + return PkceFailure(f"sign-in was not approved ({outcome.error}): {outcome.description or 'no details'}") + case CallbackCode(): + return redeem_code(contract, client_id, server.redirect_uri, outcome.code, verifier, http) + + +def pkce_token_record(base_url: str, credential: PkceCredential) -> CliTokenData: + record: Final[CliTokenData] = { + "base_url": base_url.rstrip("/"), + "key": credential.access_token, + "user_id": credential.user_id or "cli-user", + "user_email": "unknown", + "user_role": "cli", + "auth_header_name": "Authorization", + "jwt_token": "", + "timestamp": time.time(), + "expires_at": credential.expires_at, + "refresh_token": credential.refresh_token, + "client_id": credential.client_id, + "token_endpoint": credential.token_endpoint, + "revocation_endpoint": credential.revocation_endpoint, + "resource": credential.resource, + "team_id": credential.team_id, + } + return record + + +def fresh_api_key( + token_data: Mapping[str, object], + save: Callable[[CliTokenData], None], + http: Http, + *, + reload: Callable[[], Mapping[str, object] | None], + now: Callable[[], float] = time.time, +) -> str | None: + """The stored key, refreshed first when it is about to expire and a refresh token is + on file. The rotated pair is saved before the new key is returned, so a crash after + this point never strands the CLI with a burned refresh token. A refresh that fails + reads the record again, because a sibling ``lite`` process may have rotated the pair + first, in which case the key it saved is the live one. A record without + ``expires_at`` (the classic ``lite login`` credential) is returned as stored.""" + key: Final = token_data.get("key") + if not isinstance(key, str) or not key: + return None + expires_at: Final = token_data.get("expires_at") + if not isinstance(expires_at, (int, float)): + return key + if now() < expires_at - REFRESH_LEEWAY_SECONDS: + return key + still_valid: Final = key if now() < expires_at else None + refresh_inputs: Final = _refresh_inputs(token_data) + if refresh_inputs is None: + return still_valid + refreshed: Final = refresh_credential(*refresh_inputs, http=http, now=now) + if isinstance(refreshed, PkceFailure): + return _key_rotated_by_a_sibling(reload(), token_data.get("refresh_token")) or still_valid + base_url: Final = token_data.get("base_url") + save(pkce_token_record(base_url if isinstance(base_url, str) else "", refreshed)) + return refreshed.access_token + + +def _key_rotated_by_a_sibling(record: Mapping[str, object] | None, sent_refresh_token: object) -> str | None: + if record is None or record.get("refresh_token") == sent_refresh_token: + return None + key: Final = record.get("key") + return key if isinstance(key, str) and key else None + + +def _refresh_inputs(token_data: Mapping[str, object]) -> tuple[str, str, str, str, str] | None: + values: Final = tuple( + token_data.get(field) + for field in ("token_endpoint", "revocation_endpoint", "resource", "client_id", "refresh_token") + ) + if not all(isinstance(value, str) and value for value in values): + return None + token_endpoint, revocation_endpoint, resource, client_id, refresh_token = values + return str(token_endpoint), str(revocation_endpoint), str(resource), str(client_id), str(refresh_token) + + +def revoke_stored_credential(token_data: Mapping[str, object], http: Http) -> PkceFailure | None: + refresh_inputs: Final = _refresh_inputs(token_data) + if refresh_inputs is None: + return None + _, revocation_endpoint, _, client_id, refresh_token = refresh_inputs + return revoke_credential(revocation_endpoint, client_id, refresh_token, http) diff --git a/litellm/proxy/common_utils/html_forms/native_client_consent.py b/litellm/proxy/common_utils/html_forms/native_client_consent.py new file mode 100644 index 00000000000..dac92c4e787 --- /dev/null +++ b/litellm/proxy/common_utils/html_forms/native_client_consent.py @@ -0,0 +1,91 @@ +from collections.abc import Sequence +from html import escape +from typing import Final + +from litellm.constants import CLI_JWT_EXPIRATION_HOURS + + +def render_native_client_consent_page( + *, + client_origin: str, + user_id: str, + teams: Sequence[tuple[str, str]], + flow_handle: str, + complete_url: str, +) -> str: + """The consent page a native client's sign-in lands on: who is signed in, which + loopback client asked, which team the credential is attributed to, and an explicit + Approve or Deny that POSTs back to ``complete_url``. Every value is client- or + user-influenced and HTML-escaped; the flow handle travels only in the form body.""" + return f""" + + + + + +Authorize CLI access - LiteLLM + + + +
+

Authorize CLI access

+

A command-line client at {escape(client_origin)} wants to call LiteLLM as {escape(user_id)}.

+

Approving issues it a personal credential that expires within {CLI_JWT_EXPIRATION_HOURS} hours. lite logout stops it from being renewed. Only approve if you started this sign-in yourself.

+
+ +{_team_field(teams)} +
+ + +
+
+
+ + +""" + + +def _team_field(teams: Sequence[tuple[str, str]]) -> str: + if not teams: + return "" + if len(teams) == 1: + team_id, team_label = teams[0] + return ( + f'' + f"

Requests are attributed to team {escape(team_label)}.

" + ) + options: Final = "".join( + f'' for team_id, team_label in teams + ) + return ( + f'' + ) diff --git a/litellm/proxy/management_endpoints/ui_sso.py b/litellm/proxy/management_endpoints/ui_sso.py index 46af5dd80e1..1ebcb53fd6b 100644 --- a/litellm/proxy/management_endpoints/ui_sso.py +++ b/litellm/proxy/management_endpoints/ui_sso.py @@ -270,7 +270,7 @@ class _TeamRowGrants(BaseModel): litellm_model_table: _TeamModelAliasTable | None = None -class _CliSsoTeamDetail(BaseModel): +class CliSsoTeamDetail(BaseModel): """The per-team snapshot cached in the CLI SSO flow and echoed to the CLI on poll.""" team_id: str | None = None @@ -279,8 +279,8 @@ class _CliSsoTeamDetail(BaseModel): team_model_aliases: Mapping[str, str] | None = None -_CLI_SSO_TEAM_DETAILS_ADAPTER: Final = TypeAdapter(tuple[_CliSsoTeamDetail, ...]) -_TEAMLESS_CLI_SSO_TEAM_DETAIL: Final = _CliSsoTeamDetail(team_models=()) +_CLI_SSO_TEAM_DETAILS_ADAPTER: Final = TypeAdapter(tuple[CliSsoTeamDetail, ...]) +_TEAMLESS_CLI_SSO_TEAM_DETAIL: Final = CliSsoTeamDetail(team_models=()) class _CustomSsoCall(Protocol): @@ -2192,10 +2192,10 @@ async def _build_cli_sso_user_defined_values( ) -def _cli_sso_team_detail(team_row: Mapping[str, object]) -> _CliSsoTeamDetail: +def _cli_sso_team_detail(team_row: Mapping[str, object]) -> CliSsoTeamDetail: team: Final = _TeamRowGrants.model_validate(team_row) alias_table: Final = team.litellm_model_table - return _CliSsoTeamDetail( + return CliSsoTeamDetail( team_id=team.team_id, team_alias=team.team_alias, team_models=team.models, @@ -2203,10 +2203,10 @@ def _cli_sso_team_detail(team_row: Mapping[str, object]) -> _CliSsoTeamDetail: ) -async def _fetch_cli_sso_team_details( +async def fetch_cli_sso_team_details( prisma_client: PrismaClient, teams: Sequence[str], -) -> tuple[_CliSsoTeamDetail, ...] | None: +) -> tuple[CliSsoTeamDetail, ...] | None: """``None`` means the lookup itself failed, which is not the same as the user having no teams.""" if not teams: return () @@ -2221,7 +2221,7 @@ async def _fetch_cli_sso_team_details( return tuple(_cli_sso_team_detail(team_row.model_dump()) for team_row in prisma_teams) -def _cli_sso_session_teams(team_details: Sequence[_CliSsoTeamDetail]) -> list[str]: +def _cli_sso_session_teams(team_details: Sequence[CliSsoTeamDetail]) -> list[str]: """The teams a login may bind to: only those whose row still exists. A team deleted out from under a membership, which is what deleting an organization @@ -2231,7 +2231,7 @@ def _cli_sso_session_teams(team_details: Sequence[_CliSsoTeamDetail]) -> list[st return [detail.team_id for detail in team_details if detail.team_id is not None] -def _selected_cli_sso_team_detail(team_details: object, team_id: str | None) -> _CliSsoTeamDetail | None: +def selected_cli_sso_team_detail(team_details: object, team_id: str | None) -> CliSsoTeamDetail | None: """``None`` means the team's grants are unknown. An empty grant is a real value meaning unrestricted, so an unknown one must not be minted as empty.""" if team_id is None: @@ -2282,7 +2282,7 @@ async def _complete_cli_sso_callback_session( if hasattr(user_info, "teams") and user_info.teams: teams = user_info.teams if isinstance(user_info.teams, list) else [] - team_details: Final = await _fetch_cli_sso_team_details(prisma_client=prisma_client, teams=teams) + team_details: Final = await fetch_cli_sso_team_details(prisma_client=prisma_client, teams=teams) if team_details is None: raise HTTPException( status_code=500, @@ -2483,7 +2483,7 @@ async def cli_poll_key( # If no team_id provided and user has 0 or 1 team, use first team (or None) team_id = user_teams[0] if len(user_teams) > 0 else None - selected_team: Final = _selected_cli_sso_team_detail( + selected_team: Final = selected_cli_sso_team_detail( team_details=user_team_details, team_id=team_id, ) diff --git a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py index 27fc5eb4bd0..5593211ba6f 100644 --- a/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_cli_token_utils.py @@ -5,6 +5,7 @@ Unit tests for CLI token utilities import json import os import tempfile +import time from pathlib import Path from unittest.mock import mock_open, patch @@ -87,3 +88,29 @@ class TestCLITokenUtils: result = get_litellm_gateway_api_key() assert result is None + + +class TestIsCliTokenFreshWithExpiresAt: + """A ``lite login --pkce`` record carries the proxy's own ``expires_at``, which wins + over the age-based guess made from ``timestamp``.""" + + def test_future_expiry_is_fresh(self): + from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh + + assert is_cli_token_fresh({"expires_at": time.time() + 3600, "timestamp": 0}) is True + + def test_expiry_inside_the_buffer_is_stale(self): + from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh + + assert is_cli_token_fresh({"expires_at": time.time() + 100}) is False + assert is_cli_token_fresh({"expires_at": time.time() + 100}, buffer_hours=0) is True + + def test_past_expiry_is_stale_even_with_a_fresh_timestamp(self): + from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh + + assert is_cli_token_fresh({"expires_at": time.time() - 1, "timestamp": time.time()}) is False + + def test_non_numeric_expiry_falls_back_to_the_timestamp(self): + from litellm.litellm_core_utils.cli_token_utils import is_cli_token_fresh + + assert is_cli_token_fresh({"expires_at": "soon", "timestamp": time.time()}) is True diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_session_token.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_session_token.py index a43592ebe18..36280530eac 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_session_token.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_session_token.py @@ -212,3 +212,55 @@ def test_minted_token_repr_never_leaks_value(): minted = mint_session_token(PRINCIPAL, KEYS, NOW) assert isinstance(minted, MintedSessionToken) assert minted.token.get_secret_value() not in repr(minted) + + +def _decoded_claims(token: str, prefix: str) -> dict: + return jwt.decode( + token.removeprefix(prefix), + KEYS.signing_key.get_secret_value(), + algorithms=["HS256"], + options={"verify_exp": False}, + ) + + +def test_mcp_principal_wire_claims_carry_no_audience_or_team_keys(): + access_claims = _decoded_claims(_mint_access(), SESSION_TOKEN_PREFIX) + refresh_claims = _decoded_claims(_mint_refresh(), SESSION_REFRESH_PREFIX) + for claims in (access_claims, refresh_claims): + assert "audience" not in claims + assert "team_id" not in claims + + +def test_legacy_signed_claims_open_with_no_audience_and_no_team(): + opened = open_session_token(_sign_claims(_valid_claims()), KEYS, NOW) + assert isinstance(opened, OpenedSessionToken) + assert opened.principal.audience is None + assert opened.principal.team_id is None + + +def test_proxy_api_audience_and_team_round_trip_through_the_refresh_token(): + principal = SessionPrincipal(user_id="user-123", client_id="llm_client_abc", audience="proxy_api", team_id="team-b") + minted = mint_session_refresh_token(principal, KEYS, NOW) + assert isinstance(minted, MintedSessionToken) + token = minted.token.get_secret_value() + claims = _decoded_claims(token, SESSION_REFRESH_PREFIX) + assert claims["audience"] == "proxy_api" + assert claims["team_id"] == "team-b" + opened = open_session_refresh_token(token, KEYS, NOW) + assert isinstance(opened, OpenedSessionToken) + assert opened.principal == principal + + +def test_signed_claims_with_an_unknown_audience_are_rejected(): + token = _sign_claims(_valid_claims(audience="bogus")) + assert isinstance(open_session_token(token, KEYS, NOW), SessionMalformed) + + +def test_signed_claims_with_a_non_string_team_are_rejected(): + token = _sign_claims(_valid_claims(team_id=42)) + assert isinstance(open_session_token(token, KEYS, NOW), SessionMalformed) + + +def test_principal_rejects_an_unknown_audience_at_construction(): + with pytest.raises(ValidationError): + SessionPrincipal(user_id="user-123", client_id="llm_client_abc", audience="mcp") diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py index 424b993de85..bdaf1458fb0 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py @@ -1,6 +1,9 @@ """Tests for MCP OAuth discoverable endpoints""" +import hashlib import json +import time +from base64 import urlsafe_b64encode from unittest.mock import AsyncMock, MagicMock, patch import pytest @@ -9432,3 +9435,229 @@ async def test_upstream_resource_sent_on_dcr_bridge_relay_authorize(): query = await _authorize_query(server) assert query["resource"] == ["https://mcp.example.com/mcp"] assert query["client_id"] == ["caller-client"] + + +def _s256(verifier: str) -> str: + return urlsafe_b64encode(hashlib.sha256(verifier.encode("ascii")).digest()).rstrip(b"=").decode("ascii") + + +_NATIVE_CLIENT_MASTER_KEY = "sk-test-salt-for-LIT-5874" + + +def _native_client_app(monkeypatch): + """The unauthenticated discoverable router served over TestClient with a signed UI session + cookie available, plus fakes for the two database-backed hooks the native-client flow calls.""" + import jwt + from fastapi import FastAPI + from fastapi.testclient import TestClient + + from litellm.proxy._experimental.mcp_server.discoverable_endpoints import router + from litellm.proxy._experimental.mcp_server.gateway_dcr_flow import ConsentTeam, MintedProxyCredential + from litellm.proxy._experimental.mcp_server.mcp_server_manager import global_mcp_server_manager + + monkeypatch.setenv("LITELLM_SALT_KEY", _NATIVE_CLIENT_MASTER_KEY) + monkeypatch.setattr("litellm.proxy.proxy_server.master_key", _NATIVE_CLIENT_MASTER_KEY, raising=False) + minted = [] + + async def fake_mint(user_id, team_id): + minted.append((user_id, team_id)) + return MintedProxyCredential(key=f"sk-cli-{len(minted)}", expires_in=3600, user_id=user_id, team_id=team_id) + + async def fake_lookup(user_id): + return (ConsentTeam(team_id="team-a", team_alias="Team A"), ConsentTeam(team_id="team-b")) + + monkeypatch.setattr( + "litellm.proxy._experimental.mcp_server.discoverable_endpoints.mint_proxy_credential", fake_mint + ) + monkeypatch.setattr( + "litellm.proxy._experimental.mcp_server.discoverable_endpoints.lookup_consent_teams", fake_lookup + ) + global_mcp_server_manager.registry.clear() + app = FastAPI() + app.include_router(router) + client = TestClient(app) + session_cookie = jwt.encode( + {"user_id": "u1", "login_method": "username_password", "exp": int(time.time()) + 600}, + _NATIVE_CLIENT_MASTER_KEY, + algorithm="HS256", + ) + return client, session_cookie, minted + + +def _consent_flow_handle(page: str) -> str: + import re + + match = re.search(r'name="flow" value="([^"]+)"', page) + assert match is not None, page + return match.group(1) + + +def test_native_client_login_walks_discovery_consent_token_refresh_and_revoke(monkeypatch): + """The whole ``lite login --pkce`` server side over the real router: a Go CLI reads the versioned + discovery document, registers a loopback public client, the signed-in user consents to a team, + the code redeems for the ``lite login`` credential, the refresh token rotates, and revocation + kills it.""" + from http.cookies import SimpleCookie + from urllib.parse import parse_qs, urlparse + + client, session_cookie, minted = _native_client_app(monkeypatch) + redirect_uri = "http://127.0.0.1:51234/callback" + + discovery = client.get("/.well-known/litellm-cli-auth") + assert discovery.status_code == 200 + assert discovery.headers["cache-control"] == "no-store" + contract = discovery.json() + assert contract["contract_version"] == 1 + assert contract["resource"] == "http://testserver" + assert contract["code_challenge_methods_supported"] == ["S256"] + assert contract["token_endpoint_auth_methods_supported"] == ["none"] + for endpoint in ("authorization_endpoint", "token_endpoint", "registration_endpoint", "revocation_endpoint"): + assert contract[endpoint].startswith("http://testserver/") + + registered = client.post( + contract["registration_endpoint"], + json={ + "client_name": "litellm-cli", + "redirect_uris": [redirect_uri], + "grant_types": ["authorization_code", "refresh_token"], + "response_types": ["code"], + "token_endpoint_auth_method": "none", + }, + ) + assert registered.status_code == 201 + client_id = registered.json()["client_id"] + verifier = "v" * 43 + authorize_params = { + "response_type": "code", + "client_id": client_id, + "redirect_uri": redirect_uri, + "state": "cli-state", + "code_challenge": _s256(verifier), + "code_challenge_method": "S256", + "resource": contract["resource"], + } + + anonymous = client.get(contract["authorization_endpoint"], params=authorize_params, follow_redirects=False) + assert anonymous.status_code == 303 + login_target = urlparse(anonymous.headers["location"]) + assert login_target.path == "/sso/key/generate" + assert parse_qs(login_target.query)["return_to"][0].startswith("/authorize?") + + client.cookies.set("token", session_cookie) + consent = client.get(contract["authorization_endpoint"], params=authorize_params, follow_redirects=False) + assert consent.status_code == 200 + assert consent.headers["x-frame-options"] == "DENY" + assert consent.headers["cache-control"] == "no-store" + assert "http://127.0.0.1:51234" in consent.text + assert '' in consent.text + jar = SimpleCookie() + jar.load(consent.headers["set-cookie"]) + assert all(morsel["httponly"] for morsel in jar.values()) + + denied = client.post( + "/authorize/complete", + data={"flow": _consent_flow_handle(consent.text), "decision": "deny", "team_id": "team-a"}, + follow_redirects=False, + ) + assert denied.status_code == 303 + denied_query = parse_qs(urlparse(denied.headers["location"]).query) + assert denied.headers["location"].startswith(redirect_uri) + assert denied_query["error"] == ["access_denied"] + assert denied_query["state"] == ["cli-state"] + assert minted == [] + + consent_again = client.get(contract["authorization_endpoint"], params=authorize_params, follow_redirects=False) + approved = client.post( + "/authorize/complete", + data={"flow": _consent_flow_handle(consent_again.text), "decision": "approve", "team_id": "team-b"}, + follow_redirects=False, + ) + assert approved.status_code == 303 + assert approved.headers["location"].startswith(redirect_uri) + approved_query = parse_qs(urlparse(approved.headers["location"]).query) + assert approved_query["state"] == ["cli-state"] + code = approved_query["code"][0] + + token = client.post( + contract["token_endpoint"], + data={ + "grant_type": "authorization_code", + "code": code, + "redirect_uri": redirect_uri, + "client_id": client_id, + "code_verifier": verifier, + "resource": contract["resource"], + }, + ) + assert token.status_code == 200, token.text + assert token.headers["cache-control"] == "no-store" + body = token.json() + assert body["access_token"] == "sk-cli-1" + assert body["token_type"] == "Bearer" + assert body["expires_in"] == 3600 + assert body["user_id"] == "u1" + assert body["team_id"] == "team-b" + assert body["refresh_token"].startswith("llm_srefresh_") + assert minted == [("u1", "team-b")] + + refreshed = client.post( + contract["token_endpoint"], + data={ + "grant_type": "refresh_token", + "refresh_token": body["refresh_token"], + "client_id": client_id, + "resource": contract["resource"], + }, + ) + assert refreshed.status_code == 200, refreshed.text + assert refreshed.json()["access_token"] == "sk-cli-2" + assert refreshed.json()["team_id"] == "team-b" + assert refreshed.json()["refresh_token"] != body["refresh_token"] + assert minted == [("u1", "team-b"), ("u1", "team-b")] + + revoked = client.post( + contract["revocation_endpoint"], + data={"token": refreshed.json()["refresh_token"], "token_type_hint": "refresh_token", "client_id": client_id}, + ) + assert revoked.status_code == 200 + assert revoked.json() == {} + + after_revoke = client.post( + contract["token_endpoint"], + data={ + "grant_type": "refresh_token", + "refresh_token": refreshed.json()["refresh_token"], + "client_id": client_id, + "resource": contract["resource"], + }, + ) + assert after_revoke.status_code == 400 + assert after_revoke.json()["error"] == "invalid_grant" + + stranger = client.post( + contract["revocation_endpoint"], data={"token": "whatever", "client_id": "llm_dcrc_not_a_client"} + ) + assert stranger.status_code == 401 + assert stranger.json()["error"] == "invalid_client" + + +def test_native_client_authorize_without_the_proxy_resource_keeps_the_mcp_flow(monkeypatch): + """A registered client asking for the MCP resource (or no resource) never sees the consent + page, so existing MCP clients are untouched by the native-client arm.""" + client, session_cookie, minted = _native_client_app(monkeypatch) + registered = client.post("/register", json={"redirect_uris": ["http://127.0.0.1:51234/callback"]}) + client.cookies.set("token", session_cookie) + for resource in (None, "http://testserver/mcp"): + params = { + "response_type": "code", + "client_id": registered.json()["client_id"], + "redirect_uri": "http://127.0.0.1:51234/callback", + "state": "s", + "code_challenge": _s256("v" * 43), + "code_challenge_method": "S256", + **({"resource": resource} if resource else {}), + } + response = client.get("/authorize", params=params, follow_redirects=False) + assert 'name="decision"' not in response.text + assert "team-b" not in response.text + assert minted == [] diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_gateway_dcr_flow.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_gateway_dcr_flow.py index cc65970a180..4e665c552f3 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_gateway_dcr_flow.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_gateway_dcr_flow.py @@ -1,8 +1,8 @@ """Tests for the aggregate gateway DCR flow (register, authorize, complete, token).""" import hashlib -import html import json +import re from base64 import urlsafe_b64encode from datetime import datetime, timedelta, timezone from http.cookies import SimpleCookie @@ -13,12 +13,13 @@ from starlette.requests import Request from litellm.caching.caching import DualCache from litellm.proxy._experimental.mcp_server.gateway_dcr_flow import ( + _AUTH_CODE_DEBUG_KEY, CONNECT_FLOW_COOKIE_PREFIX, GATEWAY_AUTH_CODE_PREFIX, GATEWAY_AUTH_CODE_TTL_SECONDS, - GATEWAY_DCR_CLIENT_ID_PREFIX, MANUAL_DELIVERY_AUTH_CODE_TTL_SECONDS, - _AUTH_CODE_DEBUG_KEY, + ConsentTeam, + MintedProxyCredential, _GatewayAuthCode, _open_sealed, _seal, @@ -26,14 +27,21 @@ from litellm.proxy._experimental.mcp_server.gateway_dcr_flow import ( aggregate_token, complete_connect_flow, is_gateway_dcr_client_id, + is_proxy_api_resource, + native_client_auth_contract, + native_client_authorize, open_gateway_dcr_client, register_aggregate_client, + revoke_refresh_token, ) from litellm.proxy._experimental.mcp_server.outbound_credentials.session_credentials import ( + SessionBearerAdmitted, + SessionRefreshOpened, + open_session_refresh_bearer, resolve_session_bearer, session_keys_from_master_key, - SessionBearerAdmitted, ) +from litellm.proxy._experimental.mcp_server.outbound_credentials.session_token import SESSION_REFRESH_PREFIX MASTER_KEY = "sk-gateway-dcr-flow-tests" REDIRECT_URI = "https://claude.ai/api/mcp/auth_callback" @@ -888,9 +896,14 @@ async def test_scoped_authorize_runs_connect_page_with_sealed_scope(): assert response.status_code == 303 assert "/ui/connect" in response.headers["location"] _, cookies = _flow_cookie_from(response) - assert _sealed_wire_json(next(iter(cookies.values())), "", "gateway_connect_flow")["resource_server_id"] == "github-id" + assert ( + _sealed_wire_json(next(iter(cookies.values())), "", "gateway_connect_flow")["resource_server_id"] == "github-id" + ) code = await _finish_connect_page(response) - assert _sealed_wire_json(code, GATEWAY_AUTH_CODE_PREFIX, "gateway_authorization_code")["resource_server_id"] == "github-id" + assert ( + _sealed_wire_json(code, GATEWAY_AUTH_CODE_PREFIX, "gateway_authorization_code")["resource_server_id"] + == "github-id" + ) token_response = await _redeem(code, client_id) assert token_response.status_code == 200 principal = _opened_principal(json.loads(token_response.body)) @@ -1039,3 +1052,542 @@ async def test_resource_resolution_is_identity_not_ip_filtered_access(): result = resolve_scoped_resource_server(_request(), SCOPED_RESOURCE) assert result is not None manager.get_mcp_server_by_name.assert_called_once_with("github") + + +LOOPBACK_REDIRECT_URI = "http://127.0.0.1:51234/callback" +PROXY_API_RESOURCE = "https://llm.example.com" +CONSENT_TEAMS = (ConsentTeam(team_id="team-a", team_alias="Team A"), ConsentTeam(team_id="team-b")) + + +class _Minter: + def __init__(self, result=None): + self.calls = [] + self.result = result + + async def __call__(self, user_id, team_id): + self.calls.append((user_id, team_id)) + if self.result is not None: + return self.result + return MintedProxyCredential(key=f"sk-cli-{user_id}", expires_in=3600, user_id=user_id, team_id=team_id) + + +class _ConsentTeams: + def __init__(self, result=CONSENT_TEAMS): + self.calls = [] + self.result = result + + async def __call__(self, user_id): + self.calls.append(user_id) + return self.result + + +async def _native_authorize(client_id, session_user_id="u1", lookup=None, **overrides): + arguments = { + "request": _request(query=f"resource={PROXY_API_RESOURCE}"), + "client_id": client_id, + "redirect_uri": LOOPBACK_REDIRECT_URI, + "state": "client-state-123", + "code_challenge": CODE_CHALLENGE, + "code_challenge_method": "S256", + "response_type": "code", + "session_user_id": session_user_id, + "lookup_consent_teams": lookup if lookup is not None else _ConsentTeams(), + } + return await native_client_authorize(**{**arguments, **overrides}) + + +def _consent_cookie_from(response) -> tuple: + match = re.search(r'name="flow" value="([^"]+)"', response.body.decode()) + assert match is not None + handle = match.group(1) + cookie = SimpleCookie() + cookie.load(response.headers["set-cookie"]) + name = f"{CONNECT_FLOW_COOKIE_PREFIX}{handle}" + return handle, {name: cookie[name].value} + + +async def _complete_consent(consent, cache=None, session_user_id="u1", **overrides): + handle, cookies = _consent_cookie_from(consent) + return await complete_connect_flow( + request=_request("/authorize/complete", cookies=cookies, method="POST"), + flow_handle=handle, + session_user_id=session_user_id, + cache=cache or DualCache(), + **overrides, + ) + + +def _code_from(response) -> str: + return parse_qs(urlparse(response.headers["location"]).query)["code"][0] + + +async def _native_code(client_id, team_id="team-b", cache=None) -> str: + approved = await _complete_consent( + await _native_authorize(client_id), cache=cache, decision="approve", team_id=team_id + ) + assert approved.status_code == 303 + return _code_from(approved) + + +async def _redeem_native(code, client_id, minter, cache=None, resource=PROXY_API_RESOURCE, **overrides): + return await _redeem( + code, + client_id, + cache=cache, + redirect_uri=LOOPBACK_REDIRECT_URI, + resource=resource, + mint_proxy_credential=minter, + **overrides, + ) + + +async def _refresh_native(refresh_token, client_id, minter, cache, **overrides): + return await _redeem_native( + None, client_id, minter, cache=cache, grant_type="refresh_token", refresh_token=refresh_token, **overrides + ) + + +def _opened_refresh(refresh_token, client_id): + opened = open_session_refresh_bearer( + refresh_token, + session_keys_from_master_key(MASTER_KEY), + datetime.now(timezone.utc), + expected_client_id=client_id, + ) + assert isinstance(opened, SessionRefreshOpened) + return opened.principal + + +@pytest.mark.asyncio +async def test_native_authorize_renders_consent_page_and_sets_flow_cookie(): + """A native client (RFC 8707 resource = the proxy itself) gets the server-rendered consent + page instead of the MCP connect-page redirect: the flow handle rides only in the hidden + field, the sealed flow in an HttpOnly cookie, and the page can never be framed or cached.""" + client_id = (await _register([LOOPBACK_REDIRECT_URI]))["client_id"] + lookup = _ConsentTeams() + response = await _native_authorize(client_id, lookup=lookup) + assert response.status_code == 200 + assert response.headers["content-type"].startswith("text/html") + assert response.headers["cache-control"] == "no-store" + assert response.headers["x-frame-options"] == "DENY" + assert response.headers["content-security-policy"] == "frame-ancestors 'none'" + assert lookup.calls == ["u1"] + body = response.body.decode() + assert "http://127.0.0.1:51234" in body + assert "/callback" not in body + assert "u1" in body + assert '' in body + assert '' in body + assert 'action="https://llm.example.com/authorize/complete"' in body + handle, cookies = _consent_cookie_from(response) + assert "httponly" in response.headers["set-cookie"].lower() + flow = _sealed_wire_json(next(iter(cookies.values())), "", "gateway_connect_flow") + assert flow["audience"] == "proxy_api" + assert flow["client_id"] == client_id + assert flow["redirect_uri"] == LOOPBACK_REDIRECT_URI + assert flow["user_id"] == "u1" + assert "resource_server_id" not in flow + assert handle not in body.replace(f'value="{handle}"', "") + + +@pytest.mark.asyncio +async def test_native_authorize_without_session_redirects_to_login_before_any_lookup(): + client_id = (await _register([LOOPBACK_REDIRECT_URI]))["client_id"] + lookup = _ConsentTeams() + response = await _native_authorize(client_id, session_user_id=None, lookup=lookup) + assert response.status_code == 303 + location = response.headers["location"] + assert location.startswith("https://llm.example.com/sso/key/generate?return_to=") + assert "return_to=%2Fauthorize%3Fresource%3D" in location + assert lookup.calls == [] + assert "set-cookie" not in response.headers + + +@pytest.mark.asyncio +async def test_native_authorize_validation_failures_never_reach_consent(): + client_id = (await _register([LOOPBACK_REDIRECT_URI]))["client_id"] + lookup = _ConsentTeams() + for presented_client_id, overrides, expected_error in ( + ("llm_dcrc_bogus", {}, "invalid_client"), + (client_id, {"redirect_uri": "http://127.0.0.1:51235/callback"}, "invalid_request"), + (client_id, {"response_type": "token"}, "unsupported_response_type"), + (client_id, {"code_challenge": None}, "invalid_request"), + (client_id, {"code_challenge_method": "plain"}, "invalid_request"), + ): + response = await _native_authorize(presented_client_id, lookup=lookup, **overrides) + assert response.status_code == 400 + assert json.loads(response.body)["error"] == expected_error + assert "set-cookie" not in response.headers + assert lookup.calls == [] + + +@pytest.mark.asyncio +async def test_native_authorize_refuses_a_hosted_redirect_for_the_proxy_api(): + """Registration accepts any https redirect because MCP clients can be hosted, but a + proxy-API grant hands out the user's personal key, so it only ever goes back to loopback.""" + hosted = "https://evil.example/cb" + client_id = (await _register([hosted]))["client_id"] + lookup = _ConsentTeams() + response = await _native_authorize(client_id, redirect_uri=hosted, lookup=lookup) + assert response.status_code == 400 + assert json.loads(response.body) == { + "error": "invalid_request", + "error_description": "a proxy-API grant may only redirect to a loopback address", + } + assert "set-cookie" not in response.headers + assert lookup.calls == [] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "failure, status, error", + [ + ("unavailable", 503, "temporarily_unavailable"), + ("unresolvable", 500, "server_error"), + ("no_active_key", 403, "access_denied"), + ], +) +async def test_native_authorize_consent_lookup_failures_are_oauth_errors_without_a_flow(failure, status, error): + client_id = (await _register([LOOPBACK_REDIRECT_URI]))["client_id"] + response = await _native_authorize(client_id, lookup=_ConsentTeams(failure)) + assert response.status_code == status + assert json.loads(response.body)["error"] == error + assert "set-cookie" not in response.headers + + +@pytest.mark.asyncio +async def test_native_consent_escapes_untrusted_identifiers(): + client_id = (await _register([LOOPBACK_REDIRECT_URI]))["client_id"] + hostile = (ConsentTeam(team_id='t">