From 8bb8ac7f188e8cd0c2b8a73d60402d0438083e9f Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 16 Oct 2026 21:19:58 +0000 Subject: [PATCH 001/259] fix(proxy): keep a retried /v1/messages stream's spend log by logging the superseded attempt as a failure (#45136) * fix(proxy): keep a retried /v1/messages stream's spend log by logging the superseded attempt as a failure A /v1/messages stream the router retries before any content shares one logging object across both attempts. The superseded attempt's end-of-stream logging ran as a success (a usage-only or placeholder result) and claimed the request's one-shot success log, so the recovered stream's spend row was either missing or written under the superseded attempt's message id. The passthrough logging path now logs an Anthropic stream whose frames carry an `event: error` as a failure with whatever usage the provider reported, the way a dropped stream already does, and the router closes the superseded stream before it opens the retry so that logging is queued first. * fix(proxy): log a provider error frame through an attempt-scoped logging copy The error-frame failure stashed the superseded attempt's usage on the logging object the retry shares, so the retry's success log carried the failed attempt's token counts. The failure now logs through a copy of the logging object with its own model_call_details, and the helper moves into the Anthropic passthrough logging handler so the proxy streaming handler only routes by endpoint type. * fix(proxy): scope the error-frame failure log to streams closed at the frame A stream that ran to its end at the provider's error frame kept logging through the success route before this PR, and the attempt-scoped failure log is now taken only for a stream the router closed at that frame, with the copy made at the close so the retry's deployment never reaches it * chore: drop duplicate litellm import and mark the stream_ended rebind * fix(proxy): log only router-superseded attempts as error-frame failures A client that hangs up at a forwarded error frame closed the stream the same way the router closes a superseded attempt, so a non-retried stream lost its spend. The router now closes a superseded attempt through aclose_superseded_attempt, and only that close routes the attempt to the failure log * test: type the _settle helper's recorded events --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/litellm_core_utils/litellm_logging.py | 7 + .../messages/streaming_iterator.py | 17 ++ .../anthropic_passthrough_logging_handler.py | 22 ++ .../streaming_handler.py | 37 ++- litellm/router.py | 15 +- ...thropic_messages_pre_content_retry_wire.py | 32 +- .../proxy/hooks/test_prompt_cache_observer.py | 4 + ...t_anthropic_passthrough_logging_handler.py | 55 +++- .../test_streaming_handler_interrupt.py | 289 ++++++++++++++++++ tests/unit/test_router/test_router.py | 50 ++- 10 files changed, 516 insertions(+), 12 deletions(-) diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py index 67d1573dbc4..241108b4c2f 100644 --- a/litellm/litellm_core_utils/litellm_logging.py +++ b/litellm/litellm_core_utils/litellm_logging.py @@ -2300,6 +2300,13 @@ class Logging(LiteLLMLoggingBaseClass): self.model_call_details["combined_usage_object"] = usage self.model_call_details["response_cost"] = response_cost + def attempt_scoped_copy(self) -> "Logging": + """A copy with its own model_call_details, so logging one attempt of a request through it leaves + the state the request's other attempts share untouched.""" + attempt: Final = copy.copy(self) + attempt.model_call_details = dict(self.model_call_details) + return attempt + def carry_failed_attempt_usage(self) -> None: """Move what a failed earlier attempt stashed off the retry sharing this logging object, so the retry logs its own usage while the request's failure log still bills every attempt.""" diff --git a/litellm/llms/anthropic/pass_through/messages/streaming_iterator.py b/litellm/llms/anthropic/pass_through/messages/streaming_iterator.py index 5fb2546a5ae..e65fe596da7 100644 --- a/litellm/llms/anthropic/pass_through/messages/streaming_iterator.py +++ b/litellm/llms/anthropic/pass_through/messages/streaming_iterator.py @@ -2,6 +2,7 @@ import asyncio import json import re from collections.abc import AsyncGenerator, AsyncIterator, Mapping, Sequence +from contextvars import ContextVar from datetime import datetime from typing import TYPE_CHECKING, Any, Final, Protocol, runtime_checkable @@ -352,6 +353,22 @@ async def aclose_if_supported(stream: object) -> None: await stream.aclose() +_CLOSING_SUPERSEDED_ATTEMPT: Final[ContextVar[bool]] = ContextVar("closing_superseded_attempt", default=False) + + +async def aclose_superseded_attempt(stream: object) -> None: + """Close an attempt a retry or fallback is taking over, so its end-of-stream logging records a failure.""" + token: Final = _CLOSING_SUPERSEDED_ATTEMPT.set(True) + try: + await aclose_if_supported(stream) + finally: + _CLOSING_SUPERSEDED_ATTEMPT.reset(token) + + +def closing_superseded_attempt() -> bool: + return _CLOSING_SUPERSEDED_ATTEMPT.get() + + _RESPONSE_HEADERS_ADAPTER: Final[TypeAdapter[dict[str, str]]] = TypeAdapter(dict[str, str]) diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py index 23ef89b3c5c..a7f3a4f84a7 100644 --- a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py @@ -10,12 +10,14 @@ from pydantic import ConfigDict, TypeAdapter import litellm from litellm._logging import verbose_proxy_logger from litellm.constants import ANTHROPIC_BATCHES_ROUTE +from litellm.litellm_core_utils.asyncify import asyncify from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.litellm_logging import use_custom_pricing_for_model from litellm.litellm_core_utils.prompt_templates.common_utils import ( get_content_from_model_response, ) from litellm.llms.anthropic import get_anthropic_config +from litellm.llms.anthropic.common_utils import anthropic_error_frame_exception from litellm.llms.anthropic.pass_through.stream_assembly import ( build_complete_streaming_response, build_complete_streaming_response_legacy, @@ -254,6 +256,26 @@ class AnthropicPassthroughLoggingHandler: ), ) + @staticmethod + async def log_error_frame_as_failure( + attempt_logging_obj: LiteLLMLoggingObj, + request_body: Mapping[str, object], + all_chunks: Sequence[str | bytes], + error_event: tuple[str, str, int], + ) -> None: + """A stream closed at the provider's `event: error` frame is one failed attempt of the request: it logs + through the attempt-scoped copy its caller took at the close, so the usage it consumed never reaches + the success log of the attempt the router opens in its place.""" + error_type, message, status_code = error_event + await asyncify(AnthropicPassthroughLoggingHandler.record_partial_usage_for_failure)( + litellm_logging_obj=attempt_logging_obj, request_body=request_body, all_chunks=all_chunks + ) + await attempt_logging_obj.dispatch_failure_handlers( + anthropic_error_frame_exception(error_type, message, status_code, str(attempt_logging_obj.model)), + "", + prefer_async_handlers=True, + ) + @staticmethod def _cost_partial_stream_or_zero( partial_response: ModelResponse | TextCompletionResponse, model: str, logging_obj: LiteLLMLoggingObj diff --git a/litellm/proxy/pass_through_endpoints/streaming_handler.py b/litellm/proxy/pass_through_endpoints/streaming_handler.py index 8f6a688fecd..6799b82bf06 100644 --- a/litellm/proxy/pass_through_endpoints/streaming_handler.py +++ b/litellm/proxy/pass_through_endpoints/streaming_handler.py @@ -165,6 +165,7 @@ class PassThroughStreamingHandler: ) logging_scheduled = False + stream_ended = False # rebind-ok: flipped once the upstream ran to its end, so the finally can tell a close-at-frame from a completed stream model_name: Final = PassThroughStreamingHandler._extract_model_for_cost_injection( request_body=request_body, url_route=url_route, @@ -224,6 +225,7 @@ class PassThroughStreamingHandler: ) if pending: yield pending + stream_ended = True # rebind-ok: the finally reads it to spot a close-at-frame # Stream completed cleanly. When the proxy armed deferred # dispatch (post-call guardrails active), park the logging # coroutine on logging_obj instead of enqueueing now, so @@ -267,13 +269,46 @@ class PassThroughStreamingHandler: # a success would double-log the same request. if not logging_scheduled and raw_bytes and response.status_code < 400: logging_scheduled = True + superseded_error_event: Final = ( + None + if stream_ended + else PassThroughStreamingHandler._error_frame_closed_at(endpoint_type, raw_bytes) + ) try: - GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(async_coroutine=_build_logging_coroutine()) + GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue( + async_coroutine=( + _build_logging_coroutine() + if superseded_error_event is None + else AnthropicPassthroughLoggingHandler.log_error_frame_as_failure( + attempt_logging_obj=litellm_logging_obj.attempt_scoped_copy(), + request_body=resolved_request_body, + all_chunks=tuple(raw_bytes), + error_event=superseded_error_event, + ) + ) + ) except Exception as e: verbose_proxy_logger.error("Error scheduling chunk_processor logging: %s", e) else: bind_budget_reservation_to_callbacks(litellm_logging_obj.litellm_params) + @staticmethod + def _error_frame_closed_at(endpoint_type: EndpointType, raw_bytes: Sequence[bytes]) -> tuple[str, str, int] | None: + """The provider `event: error` frame of a stream the router closed because a retry or fallback took the + request over. That attempt is logged as its own failure, on a copy taken now so the retry's rebinding of + the shared logging object never reaches it. A stream that ran to its end or that the client closed keeps + logging through the success route, so its usage is still billed.""" + from litellm.llms.anthropic.pass_through.messages.streaming_iterator import ( + closing_superseded_attempt, + parse_anthropic_error_event, + ) + + if endpoint_type != EndpointType.ANTHROPIC or not closing_superseded_attempt(): + return None + + complete_frames, _ = split_complete_sse_frames(b"".join(raw_bytes)) + return parse_anthropic_error_event(complete_frames) + @staticmethod async def route_streaming_logging_to_handler( litellm_logging_obj: LiteLLMLoggingObj, diff --git a/litellm/router.py b/litellm/router.py index 4f2358f9008..316c7f27ec0 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -6035,6 +6035,7 @@ class Router: """ from litellm.llms.anthropic.pass_through.messages.streaming_iterator import ( aclose_if_supported, + aclose_superseded_attempt, parse_anthropic_error_event, parse_anthropic_refusal_stop_details, ) @@ -6044,6 +6045,14 @@ class Router: source_iterator: Final = response + async def close_source() -> None: + with anyio.CancelScope(shield=True), contextlib.suppress(BaseException): + await aclose_if_supported(source_iterator) + + async def close_superseded() -> None: + with anyio.CancelScope(shield=True), contextlib.suppress(BaseException): + await aclose_superseded_attempt(source_iterator) + async def stream_with_fallbacks() -> AsyncGenerator[bytes, None]: from litellm.exceptions import MidStreamFallbackError @@ -6130,11 +6139,11 @@ class Router: model, initial_kwargs, wrapper, + close_superseded, ): yield item finally: - with anyio.CancelScope(shield=True), contextlib.suppress(BaseException): - await aclose_if_supported(source_iterator) + await close_source() # Referenced by stream_with_fallbacks via closure - assigned here, before # the generator body ever runs, so the reference resolves fine despite @@ -6150,6 +6159,7 @@ class Router: model: str, initial_kwargs: dict[str, Any], # mutable-ok: handed to _aanthropic_messages_fallback_attempt, which mutates it wrapper: "FallbackAwareAnthropicMessagesStream", + close_superseded: Callable[[], Awaitable[None]], ) -> AsyncGenerator[bytes, None]: """Turns a source-iterator failure into a fallback attempt or the error reaching the caller.""" from litellm.exceptions import MidStreamFallbackError @@ -6169,6 +6179,7 @@ class Router: ) if fallback_error is None: raise stream_error + await close_superseded() outcome: Final = await self._astream_retry_same_group( fallback_error, initial_kwargs, diff --git a/tests/integration/messages_endpoint/providers/anthropic/test_anthropic_messages_pre_content_retry_wire.py b/tests/integration/messages_endpoint/providers/anthropic/test_anthropic_messages_pre_content_retry_wire.py index cb7f4c2b9ce..e0b2e4a336f 100644 --- a/tests/integration/messages_endpoint/providers/anthropic/test_anthropic_messages_pre_content_retry_wire.py +++ b/tests/integration/messages_endpoint/providers/anthropic/test_anthropic_messages_pre_content_retry_wire.py @@ -40,6 +40,7 @@ FirstAttempt = Literal[ "drop_after_message_start", "error_frame", "error_frame_after_message_start", + "error_frame_after_content", "http_status", "drop_after_content", ] @@ -66,6 +67,8 @@ class _Failure: return status_reply(self.status) case "drop_after_content": return dropping_reply(chunks, abort_after=3) + case "error_frame_after_content": + return stream_reply((*chunks[:3], error_frame(self.status, f"scripted {self.status}"))) _MID_STREAM_FAILURES: Final = ( @@ -124,12 +127,12 @@ def _stream(gateway: Gateway, body: Mapping[str, JsonValue]) -> tuple[int, tuple return response.status_code, parse_sse(response.text) +def _rows(request_id: str) -> list[dict[str, JsonValue]]: + return read_rows('SELECT status, model_group FROM "LiteLLM_SpendLogs" WHERE request_id=%s', (request_id,)) + + def _success_rows(request_id: str) -> list[dict[str, JsonValue]]: - return eventually( - lambda: read_rows('SELECT status, model_group FROM "LiteLLM_SpendLogs" WHERE request_id=%s', (request_id,)), - lambda rows: len(rows) >= 1, - seconds=70, - ) + return eventually(lambda: _rows(request_id), lambda rows: len(rows) >= 1, seconds=70) def _assert_completed_by_retry( @@ -141,6 +144,8 @@ def _assert_completed_by_retry( assert delta_text(events) == _TEXT, events assert [request.target for request in wire.drain()] == ["/v1/messages"] * attempts assert _success_rows(_served_id(prompt, attempts)) == [{"status": "success", "model_group": model}] + superseded_rows: Final = {attempt: _rows(_served_id(prompt, attempt)) for attempt in range(1, attempts)} + assert not any(superseded_rows.values()), superseded_rows def _assert_rejected_before_content( @@ -302,6 +307,23 @@ def test_stream_dropping_after_content_is_not_retried(gateway: Gateway) -> None: assert [request.target for request in wire.drain()] == ["/v1/messages"] +def test_stream_error_frame_after_content_is_not_retried_and_keeps_its_spend_row(gateway: Gateway) -> None: + prompt: Final = _prompt() + attempts: Final = Attempts() + with ( + wire_server(_upstream(prompt, _Failure("error_frame_after_content", 529), attempts)) as wire, + gateway.scenario() as scenario, + ): + model: Final = scenario.model(model=f"anthropic/{_MODEL}", api_base=wire.url, api_key=_API_KEY, num_retries=1) + status, events = _stream(gateway, _body(model, prompt, "deployment")) + assert status == 200, events + assert event_types(events) == (*LIFECYCLE[:3], "error"), events + assert message_id(events) == _served_id(prompt, 1), events + assert delta_text(events) == _TEXT, events + assert [request.target for request in wire.drain()] == ["/v1/messages"] + assert [row["model_group"] for row in _success_rows(_served_id(prompt, 1))] == [model] + + def test_stream_rejected_with_401_before_it_opens_is_not_retried(gateway: Gateway) -> None: prompt: Final = _prompt() attempts: Final = Attempts() diff --git a/tests/unit/proxy/hooks/test_prompt_cache_observer.py b/tests/unit/proxy/hooks/test_prompt_cache_observer.py index 3354aa457b7..ea89bb8d4a2 100644 --- a/tests/unit/proxy/hooks/test_prompt_cache_observer.py +++ b/tests/unit/proxy/hooks/test_prompt_cache_observer.py @@ -201,6 +201,10 @@ class RecordingObserver(PromptCacheObserver): await super().async_log_success_event(kwargs, response_obj, start_time, end_time) self.finished.set() + async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time): + await super().async_log_failure_event(kwargs, response_obj, start_time, end_time) + self.finished.set() + def native_response(): return { diff --git a/tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py b/tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py index 88c5e2d2c1b..4770ecd5485 100644 --- a/tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py +++ b/tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py @@ -5,11 +5,10 @@ from typing import Any, Dict, Final, List from unittest.mock import AsyncMock, MagicMock, Mock, patch import httpx -import litellm import pytest from pydantic import ValidationError - +import litellm from litellm.integrations.custom_logger import CustomLogger from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.llms.anthropic.pass_through.stream_assembly import extract_sse_data @@ -2644,6 +2643,58 @@ class TestRecordPartialUsageForFailure: assert "response_cost" not in logging_obj.model_call_details +class _FailureRecorder(CustomLogger): + def __init__(self): + super().__init__() + self.failure_kwargs = [] + + async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time): + self.failure_kwargs.append(kwargs) + + +class TestLogErrorFrameAsFailure: + """A stream closed at the provider's `event: error` frame is one failed attempt: its failure log bills the + usage streamed before the frame on the attempt-scoped copy its caller hands it, and the request's shared + logging state stays as it was, so the attempt the router opens next logs with nothing left behind.""" + + @pytest.mark.asyncio + async def test_logs_a_failure_with_the_usage_streamed_before_the_frame_and_leaves_the_request_state_alone(self): + recorder = _FailureRecorder() + logging_obj = LiteLLMLoggingObj( + model="claude-sonnet-5", + messages=[{"role": "user", "content": "hello"}], + stream=True, + call_type="anthropic_messages", + start_time=datetime.now(), + litellm_call_id="test-error-frame-failure", + function_id="test-error-frame-failure", + dynamic_async_failure_callbacks=[recorder], + ) + chunks = [ + *TestRecordPartialUsageForFailure()._interrupted_chunks()[:1], + TestRecordPartialUsageForFailure._sse( + "error", {"type": "error", "error": {"type": "overloaded_error", "message": "Overloaded"}} + ), + ] + + await AnthropicPassthroughLoggingHandler.log_error_frame_as_failure( + attempt_logging_obj=logging_obj.attempt_scoped_copy(), + request_body={"model": "claude-sonnet-5", "stream": True}, + all_chunks=chunks, + error_event=("overloaded_error", "Overloaded", 529), + ) + + assert len(recorder.failure_kwargs) == 1 + failure_payload = recorder.failure_kwargs[0]["standard_logging_object"] + assert failure_payload["status"] == "failure" + assert failure_payload["prompt_tokens"] == 52 + assert failure_payload["response_cost"] > 0 + assert isinstance(recorder.failure_kwargs[0]["exception"], litellm.InternalServerError) + assert "combined_usage_object" not in logging_obj.model_call_details + assert "response_cost" not in logging_obj.model_call_details + assert logging_obj.should_run_logging("async_failure") + + @pytest.mark.parametrize( ("all_chunks", "interrupted"), [ diff --git a/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py b/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py index b278d7dea01..8ce41129aec 100644 --- a/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py +++ b/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py @@ -18,6 +18,7 @@ from litellm.proxy.pass_through_endpoints.llm_provider_handlers.tinyfish_passthr from litellm.integrations.custom_logger import CustomLogger from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER +from litellm.llms.anthropic.pass_through.messages.streaming_iterator import aclose_superseded_attempt from litellm.proxy.pass_through_endpoints.streaming_handler import ( PassThroughStreamingHandler, ) @@ -749,6 +750,294 @@ async def test_chunk_processor_logs_failure_not_success_on_mid_stream_exception( assert isinstance(recorder.failure_kwargs[0]["exception"], httpx.ReadTimeout) +def _anthropic_message_start_frame(message_id: str = "msg_1", input_tokens: int = 52) -> bytes: + return _anthropic_sse( + "message_start", + { + "type": "message_start", + "message": { + "id": message_id, + "type": "message", + "role": "assistant", + "model": "claude-sonnet-5", + "content": [], + "stop_reason": None, + "stop_sequence": None, + "usage": {"input_tokens": input_tokens, "output_tokens": 1}, + }, + }, + ) + + +def _anthropic_completed_stream_tail() -> tuple[bytes, ...]: + return ( + _anthropic_sse( + "content_block_start", + {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}, + ), + _anthropic_sse( + "content_block_delta", + {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "pong"}}, + ), + _anthropic_sse("content_block_stop", {"type": "content_block_stop", "index": 0}), + _anthropic_sse( + "message_delta", + {"type": "message_delta", "delta": {"stop_reason": "end_turn", "stop_sequence": None}, "usage": {"output_tokens": 3}}, + ), + _anthropic_sse("message_stop", {"type": "message_stop"}), + ) + + +async def _settle(recorded: list[dict[str, object]]) -> None: + for _ in range(300): + if recorded: + return + await asyncio.sleep(0.01) + + +def _anthropic_overloaded_error_frame() -> bytes: + return _anthropic_sse("error", {"type": "error", "error": {"type": "overloaded_error", "message": "Overloaded"}}) + + +def _anthropic_stream_of(frames: tuple[bytes, ...]): + mock = MagicMock(spec=httpx.Response) + mock.status_code = 200 + + async def _aiter_bytes(): + for frame in frames: + yield frame + + mock.aiter_bytes = _aiter_bytes + return mock + + +@pytest.mark.parametrize( + "frames, billed", + [ + pytest.param( + (_anthropic_message_start_frame(), _anthropic_overloaded_error_frame()), True, id="error-after-message-start" + ), + pytest.param((_anthropic_overloaded_error_frame(),), False, id="error-frame-only"), + ], +) +@pytest.mark.asyncio +async def test_chunk_processor_logs_a_provider_error_frame_as_a_failure_when_closed_at_the_frame(frames, billed): + """The router retries a /v1/messages stream whose provider sent `event: error` before any content by closing + it at the frame. That attempt never completed: it logs once as a failure carrying whatever usage the provider + had reported, and the request's one-shot success log stays free for the retry that shares the logging object.""" + recorder = _EventRecorder() + logging_obj = LiteLLMLoggingObj( + model="claude-sonnet-5", + messages=[{"role": "user", "content": "hi"}], + stream=True, + call_type="anthropic_messages", + start_time=datetime.now(), + litellm_call_id="test-error-frame", + function_id="test-error-frame", + dynamic_async_success_callbacks=[recorder], + dynamic_async_failure_callbacks=[recorder], + ) + stream = PassThroughStreamingHandler.chunk_processor( + response=_anthropic_stream_of(frames), + request_body={"model": "claude-sonnet-5", "stream": True}, + litellm_logging_obj=logging_obj, + endpoint_type=EndpointType.ANTHROPIC, + start_time=datetime.now(), + passthrough_success_handler_obj=MagicMock(), + url_route="/v1/messages", + ) + received = [] + async for chunk in stream: + received.append(chunk) + if b"event: error" in chunk: + break + await aclose_superseded_attempt(stream) + await _settle(recorder.failure_kwargs) + + assert received == list(frames) + assert recorder.success_kwargs == [] + assert logging_obj.model_call_details.get("has_dispatched_final_stream_success") is not True + assert "combined_usage_object" not in logging_obj.model_call_details + assert "response_cost" not in logging_obj.model_call_details + assert logging_obj.should_run_logging("async_failure") + assert len(recorder.failure_kwargs) == 1 + failure_payload = recorder.failure_kwargs[0]["standard_logging_object"] + assert failure_payload["status"] == "failure" + assert isinstance(recorder.failure_kwargs[0]["exception"], litellm.InternalServerError) + if billed: + assert failure_payload["prompt_tokens"] == 52 + assert failure_payload["response_cost"] > 0 + else: + assert failure_payload["response_cost"] == 0 + + +@pytest.mark.asyncio +async def test_chunk_processor_logs_a_stream_that_ended_at_the_provider_error_frame_through_the_success_route(): + """A provider error frame that ends a stream nobody closed early is the request's final answer, and that + stream keeps logging the way it always has: once, through the success route, with the usage it carried.""" + recorder = _EventRecorder() + logging_obj = LiteLLMLoggingObj( + model="claude-sonnet-5", + messages=[{"role": "user", "content": "hi"}], + stream=True, + call_type="anthropic_messages", + start_time=datetime.now(), + litellm_call_id="test-final-error-frame", + function_id="test-final-error-frame", + dynamic_async_success_callbacks=[recorder], + dynamic_async_failure_callbacks=[recorder], + ) + frames = (_anthropic_message_start_frame(), _anthropic_overloaded_error_frame()) + received = [ + chunk + async for chunk in PassThroughStreamingHandler.chunk_processor( + response=_anthropic_stream_of(frames), + request_body={"model": "claude-sonnet-5", "stream": True}, + litellm_logging_obj=logging_obj, + endpoint_type=EndpointType.ANTHROPIC, + start_time=datetime.now(), + passthrough_success_handler_obj=MagicMock(), + url_route="/v1/messages", + ) + ] + await _settle(recorder.success_kwargs) + + assert received == list(frames) + assert recorder.failure_kwargs == [] + success_payload = recorder.success_kwargs[0]["standard_logging_object"] + assert success_payload["status"] == "success" + assert success_payload["prompt_tokens"] == 52 + assert logging_obj.model_call_details["prompt_cache_response_complete"] is False + + +@pytest.mark.asyncio +async def test_chunk_processor_bills_a_stream_the_client_closed_at_the_provider_error_frame(): + """The Anthropic SDK raises on a provider `event: error` frame and drops the connection, so a non-retried + stream is closed at the frame by the client, not by a retry taking over. Its usage must still reach the + success route, or the tokens the provider billed never reach the key's spend.""" + recorder = _EventRecorder() + logging_obj = LiteLLMLoggingObj( + model="claude-sonnet-5", + messages=[{"role": "user", "content": "hi"}], + stream=True, + call_type="anthropic_messages", + start_time=datetime.now(), + litellm_call_id="test-client-closed-at-error-frame", + function_id="test-client-closed-at-error-frame", + dynamic_async_success_callbacks=[recorder], + dynamic_async_failure_callbacks=[recorder], + ) + stream = PassThroughStreamingHandler.chunk_processor( + response=_anthropic_stream_of((_anthropic_message_start_frame(), _anthropic_overloaded_error_frame())), + request_body={"model": "claude-sonnet-5", "stream": True}, + litellm_logging_obj=logging_obj, + endpoint_type=EndpointType.ANTHROPIC, + start_time=datetime.now(), + passthrough_success_handler_obj=MagicMock(), + url_route="/v1/messages", + ) + async for chunk in stream: + if b"event: error" in chunk: + break + await stream.aclose() + await _settle(recorder.success_kwargs) + + assert recorder.failure_kwargs == [] + success_payload = recorder.success_kwargs[0]["standard_logging_object"] + assert success_payload["status"] == "success" + assert success_payload["prompt_tokens"] == 52 + + +@pytest.mark.asyncio +async def test_superseded_attempt_failure_log_keeps_its_own_deployment_when_the_retry_rebinds_the_logging_object(): + """The router rebinds the shared logging object to the retry's deployment right after closing the superseded + stream, before the logging worker runs. The failure log of the superseded attempt must still name the + deployment that served it, or the router's cooldown lands on the deployment serving the retry.""" + recorder = _EventRecorder() + logging_obj = LiteLLMLoggingObj( + model="claude-sonnet-5", + messages=[{"role": "user", "content": "hi"}], + stream=True, + call_type="anthropic_messages", + start_time=datetime.now(), + litellm_call_id="test-superseded-deployment", + function_id="test-superseded-deployment", + dynamic_async_failure_callbacks=[recorder], + ) + logging_obj.update_environment_variables( + model="claude-sonnet-5", litellm_params={"model_info": {"id": "deployment-1"}}, optional_params={} + ) + superseded = PassThroughStreamingHandler.chunk_processor( + response=_anthropic_stream_of((_anthropic_message_start_frame(), _anthropic_overloaded_error_frame())), + request_body={"model": "claude-sonnet-5", "stream": True}, + litellm_logging_obj=logging_obj, + endpoint_type=EndpointType.ANTHROPIC, + start_time=datetime.now(), + passthrough_success_handler_obj=MagicMock(), + url_route="/v1/messages", + ) + async for chunk in superseded: + if b"event: error" in chunk: + break + await aclose_superseded_attempt(superseded) + logging_obj.update_environment_variables( + model="claude-sonnet-5-retry", litellm_params={"model_info": {"id": "deployment-2"}}, optional_params={} + ) + await _settle(recorder.failure_kwargs) + + failure_kwargs = recorder.failure_kwargs[0] + assert failure_kwargs["litellm_params"]["model_info"]["id"] == "deployment-1" + assert failure_kwargs["model"] == "claude-sonnet-5" + assert logging_obj.model_call_details["litellm_params"]["model_info"]["id"] == "deployment-2" + + +@pytest.mark.asyncio +async def test_retry_success_log_carries_its_own_usage_after_a_provider_error_frame(): + """Both attempts of a retried /v1/messages stream log through the request's one logging object. The + superseded attempt's failure log must not leave the usage it consumed behind, or the retry's success + log bills the failed attempt's tokens instead of its own.""" + recorder = _EventRecorder() + logging_obj = LiteLLMLoggingObj( + model="claude-sonnet-5", + messages=[{"role": "user", "content": "hi"}], + stream=True, + call_type="anthropic_messages", + start_time=datetime.now(), + litellm_call_id="test-retry-own-usage", + function_id="test-retry-own-usage", + dynamic_async_success_callbacks=[recorder], + dynamic_async_failure_callbacks=[recorder], + ) + + def attempt(frames: tuple[bytes, ...]): + return PassThroughStreamingHandler.chunk_processor( + response=_anthropic_stream_of(frames), + request_body={"model": "claude-sonnet-5", "stream": True}, + litellm_logging_obj=logging_obj, + endpoint_type=EndpointType.ANTHROPIC, + start_time=datetime.now(), + passthrough_success_handler_obj=MagicMock(), + url_route="/v1/messages", + ) + + superseded = attempt((_anthropic_message_start_frame("msg_a1", 52), _anthropic_overloaded_error_frame())) + async for chunk in superseded: + if b"event: error" in chunk: + break + await aclose_superseded_attempt(superseded) + await _settle(recorder.failure_kwargs) + retry = attempt((_anthropic_message_start_frame("msg_a2", 7), *_anthropic_completed_stream_tail())) + async for _ in retry: + pass + await _settle(recorder.success_kwargs) + + failure_payload = recorder.failure_kwargs[0]["standard_logging_object"] + assert failure_payload["prompt_tokens"] == 52 + success_payload = recorder.success_kwargs[0]["standard_logging_object"] + assert success_payload["status"] == "success" + assert (success_payload["prompt_tokens"], success_payload["completion_tokens"]) == (7, 3) + + def _google_sse(prompt_tokens: int, completion_tokens: int, text: str) -> bytes: payload = { "candidates": [{"content": {"parts": [{"text": text}], "role": "model"}, "index": 0}], diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index 6bc52ab3dfc..c8ecf901602 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -38,6 +38,7 @@ from litellm.integrations.custom_guardrail import CustomGuardrail from litellm.integrations.custom_logger import CustomLogger from litellm.integrations.otel.plumbing.context import set_request_root_span from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging +from litellm.llms.anthropic.pass_through.messages.streaming_iterator import closing_superseded_attempt from litellm.llms.anthropic.pass_through.messages.agentic_streaming_iterator import ( SERVER_FULFILLED_TOOL_LEAK_ERROR_SSE_BYTES, ) @@ -16004,6 +16005,7 @@ class _AnthropicMessagesFakeByteStream: self._chunks = list(chunks) self._hidden_params = {"additional_headers": {"x-amzn-requestid": "req-1"}} self.closed = False + self.closed_as_superseded = False def __aiter__(self): return self @@ -16014,6 +16016,7 @@ class _AnthropicMessagesFakeByteStream: return self._chunks.pop(0) async def aclose(self) -> None: + self.closed_as_superseded = self.closed_as_superseded or closing_superseded_attempt() self.closed = True @@ -17256,6 +17259,30 @@ async def test_anthropic_messages_stream_dropped_before_content_is_retried_withi assert [(attempted, budget) for _, attempted, budget in provider.calls] == [(0, 2), (1, 2)] +@pytest.mark.asyncio +async def test_anthropic_messages_superseded_stream_is_closed_before_the_retry_opens(): + """Both attempts log through the request's one logging object, so the superseded attempt's end-of-stream + logging must be queued before the retry starts producing its own: the router closes the dropped stream + before it asks the group for another.""" + router = _anthropic_messages_retry_router(num_retries=1) + superseded = _AnthropicMessagesFakeByteStream( + [_anthropic_messages_message_start_chunk(), _anthropic_messages_overloaded_error_chunk()] + ) + superseded_closed_when_retry_opened: list[bool] = [] + + def retry_stream(): + superseded_closed_when_retry_opened.append(superseded.closed_as_superseded) + return _anthropic_messages_retried_stream() + + provider = _AnthropicMessagesScriptedProvider(lambda: superseded, retry_stream) + + stream = await _anthropic_messages_stream_through_router(router, provider) + body = [chunk async for chunk in stream] + + assert body == [_anthropic_messages_message_start_chunk(), _anthropic_messages_content_chunk("pong")] + assert superseded_closed_when_retry_opened == [True] + + @pytest.mark.asyncio async def test_a_ptu_share_holders_messages_stream_dropped_before_content_is_retried_on_its_shared_group(monkeypatch): """The same-group retry has to look the group's health up as the calling team; looked up with no @@ -18387,7 +18414,7 @@ async def test_aanthropic_messages_recover_stream_error_flushes_buffered_frames_ with patch.object(router, "_aanthropic_messages_fallback_attempt") as mock_attempt: recovery = router._aanthropic_messages_recover_stream_error( - declined, True, buffered, "primary", {"model": "primary"}, _anthropic_messages_make_wrapper() + declined, True, buffered, "primary", {"model": "primary"}, _anthropic_messages_make_wrapper(), AsyncMock() ) with pytest.raises(BedrockError) as exc_info: await drain(recovery) @@ -18408,7 +18435,7 @@ async def test_aanthropic_messages_recover_stream_error_hands_converted_raise_to with patch.object(router, "_aanthropic_messages_fallback_attempt", new=fake_attempt): recovery = router._aanthropic_messages_recover_stream_error( - raised, False, (), "primary", {"model": "primary"}, _anthropic_messages_make_wrapper() + raised, False, (), "primary", {"model": "primary"}, _anthropic_messages_make_wrapper(), AsyncMock() ) collected = [chunk async for chunk in recovery] assert collected == [b"fallback"] @@ -18466,6 +18493,25 @@ async def test_anthropic_messages_post_first_chunk_error_skips_fallback(): mock_fallback.assert_not_awaited() +@pytest.mark.asyncio +async def test_anthropic_messages_stream_the_client_closes_at_a_forwarded_error_is_not_closed_as_superseded(): + """An error frame forwarded to the client ends that request, and the SDK hangs up on it. No retry took the + request over, so the source must not be closed as a superseded attempt, which would log it as a failure and + drop the usage the provider billed for it.""" + router = _anthropic_messages_make_router() + error_chunk = _anthropic_messages_overloaded_error_chunk() + source = _AnthropicMessagesFakeByteStream([_anthropic_messages_content_chunk("partial answer"), error_chunk]) + + wrapped = await router._aanthropic_messages_streaming_iterator(response=source, initial_kwargs={"model": "primary"}) + async for chunk in wrapped: + if chunk == error_chunk: + break + await wrapped.aclose() + + assert source.closed is True + assert source.closed_as_superseded is False + + @pytest.mark.asyncio async def test_anthropic_messages_non_retriable_error_flushes_buffered_lifecycle_frames(): """A non-retriable error arriving while lifecycle frames are still From ac6300f3800ad9fcf12ee0f572105cf277b05a60 Mon Sep 17 00:00:00 2001 From: yuneng-jiang Date: Fri, 16 Oct 2026 14:30:04 -0700 Subject: [PATCH 002/259] fix(mcp): stop proxy-started legacy listeners from consuming server rpm (#46968) * fix(mcp): stop proxy-started legacy listeners from consuming server rpm The listener the proxy opens on notifications/initialized for legacy MCP sessions ran the server rate-limit admission once per target, so every session spent an rpm slot before the client sent anything. Admission is now injected into the listener and applied only to client-initiated subscriptions (subscriptions/listen and resources/subscribe). Authorization and its periodic recheck still run on every path. * test(mcp): key the fake admission window by caller and server * fix(mcp): stop subscription catalog refreshes from consuming server rpm A subscription listener lists the caller's catalog to diff list_changed events, at setup and on every upstream event. Those listings went through the normal per-server admission, so against an upstream that declares listChanged each legacy session spent several rpm slots before the client asked for anything. The listener now marks its targets as already admitted for its own lifetime, so its refreshes reuse the subscription's admission. Client requests on the session are still metered. * test(mcp): exercise subscription admission through the proxy rate limiter --- .../_experimental/mcp_server/subscriptions.py | 36 ++++- .../mcp_server/test_subscriptions.py | 130 ++++++++++++++++++ 2 files changed, 162 insertions(+), 4 deletions(-) diff --git a/litellm/proxy/_experimental/mcp_server/subscriptions.py b/litellm/proxy/_experimental/mcp_server/subscriptions.py index dcf24f974fe..db622ea3cbf 100644 --- a/litellm/proxy/_experimental/mcp_server/subscriptions.py +++ b/litellm/proxy/_experimental/mcp_server/subscriptions.py @@ -55,6 +55,7 @@ from litellm.proxy._experimental.mcp_server.catalog import ( ) from litellm.proxy._experimental.mcp_server.contracts import OperationContext from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError from litellm.types.mcp import MCP_LEGACY_VERSIONS from litellm.types.mcp_server.mcp_server_manager import MCPServer @@ -300,6 +301,29 @@ async def _listing(event: ServerEvent, context: OperationContext) -> str | None: raise MCPError(code=-32603, message="Subscription catalog exceeds the pagination limit") +async def _enforce_rate_limit(context: OperationContext, server: MCPServer) -> None: + await operations._enforce_mcp_server_rate_limit(context.user_api_key_auth, server) # pyright: ignore[reportPrivateUsage] # same admission policy as other gateway operations + + +async def _unmetered(_: OperationContext, __: MCPServer) -> None: + return None + + +async def _admitted() -> ProxyRateLimitError | None: + return None + + +@asynccontextmanager +async def _listener_admission(targets: tuple[SubscriptionTarget, ...]) -> AsyncGenerator[None]: + scope: Final = operations._mcp_server_admission_memo.set( # pyright: ignore[reportPrivateUsage] # the listener's own catalog refreshes reuse its admission + {target.server.server_id: asyncio.create_task(_admitted()) for target in targets} + ) + try: + yield + finally: + operations._mcp_server_admission_memo.reset(scope) # pyright: ignore[reportPrivateUsage] # restore the caller's admission scope + + async def _recheck_authorization(check: Callable[[], Awaitable[OperationContext]]) -> None: while True: await asyncio.sleep(15) @@ -361,7 +385,7 @@ class GatewaySubscriptions: if ctx.protocol_version != "2026-07-28": raise MCPError(code=-32601, message="Subscriptions require the modern MCP protocol") async with self._capacity(owner): - return await self._listen(ctx, params, authorize) + return await self._listen(ctx, params, authorize, _enforce_rate_limit) @asynccontextmanager async def _capacity(self, owner: str) -> AsyncGenerator[None]: @@ -414,6 +438,7 @@ class GatewaySubscriptions: ctx, SubscriptionFilter(tools_list_changed=True, prompts_list_changed=True, resources_list_changed=True), authorization, + _unmetered, ) if (ready := self._initializing.get(key)) is not None: ready.set() @@ -427,6 +452,7 @@ class GatewaySubscriptions: ctx: ServerRequestContext, notifications: SubscriptionFilter, authorization: SubscriptionAuthorization, + admit: Callable[[OperationContext, MCPServer], Awaitable[None]], scope: anyio.CancelScope | None = None, *, task_status: TaskStatus[None], @@ -443,7 +469,7 @@ class GatewaySubscriptions: async with self._capacity(authorization.owner): await self._listen( - ctx, SubscriptionsListenRequestParams(notifications=notifications), authorization, ready + ctx, SubscriptionsListenRequestParams(notifications=notifications), authorization, admit, ready ) async def legacy_resource(self, ctx: ServerRequestContext, uri: str, *, subscribe: bool) -> None: @@ -464,6 +490,7 @@ class GatewaySubscriptions: ctx, SubscriptionFilter(resource_subscriptions=[uri]), state.authorization, + _enforce_rate_limit, scope, ) state.resources = {**state.resources, uri: scope} @@ -476,6 +503,7 @@ class GatewaySubscriptions: ctx: ServerRequestContext, params: SubscriptionsListenRequestParams, authorize: Callable[[], Awaitable[OperationContext]], + admit: Callable[[OperationContext, MCPServer], Awaitable[None]], ready: Callable[[SubscriptionFilter], None] | None = None, ) -> SubscriptionsListenResult: context: Final = await authorize() @@ -501,7 +529,7 @@ class GatewaySubscriptions: return fresh async def relay(target: SubscriptionTarget, *, task_status: TaskStatus[SubscriptionFilter]) -> None: - await operations._enforce_mcp_server_rate_limit(context.user_api_key_auth, target.server) # pyright: ignore[reportPrivateUsage] # same admission policy as other gateway operations + await admit(context, target.server) legacy: Final = LegacyNotifications() @@ -567,7 +595,7 @@ class GatewaySubscriptions: finally: self._upstream_slots.release() - async with anyio.create_task_group() as group: + async with _listener_admission(targets), anyio.create_task_group() as group: honored: Final = tuple( [SubscriptionFilter.model_validate(await group.start(relay, target)) for target in targets] ) diff --git a/tests/unit/proxy/_experimental/mcp_server/test_subscriptions.py b/tests/unit/proxy/_experimental/mcp_server/test_subscriptions.py index a5d42e58b99..a97ab04c506 100644 --- a/tests/unit/proxy/_experimental/mcp_server/test_subscriptions.py +++ b/tests/unit/proxy/_experimental/mcp_server/test_subscriptions.py @@ -1,4 +1,5 @@ import asyncio +from types import SimpleNamespace from typing import Final from unittest.mock import AsyncMock, Mock @@ -19,6 +20,21 @@ from litellm.types.mcp import MCPTransport from litellm.types.mcp_server.mcp_server_manager import MCPServer +def allow_one_request(monkeypatch: pytest.MonkeyPatch, user_id: str, server_id: str) -> None: + from litellm.proxy import proxy_server + from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError + + windows: Final = {(user_id, server_id): asyncio.Semaphore(1)} + + async def enforce(user: UserAPIKeyAuth | None, server: MCPServer) -> None: + window: Final = windows[(user.user_id if user is not None else None, server.server_id)] + if window.locked(): + raise ProxyRateLimitError(detail=f"{server_id} RPM exceeded") + await window.acquire() + + monkeypatch.setattr(proxy_server, "proxy_logging_obj", SimpleNamespace(enforce_mcp_server_rate_limits=enforce)) + + def context(session: ServerSession, revision: str = "2026-07-28") -> ServerRequestContext: return ServerRequestContext( session=session, @@ -356,3 +372,117 @@ async def test_proxy_route_rejects_legacy_resource_subscription_before_setup(met await getattr(server, method)(context(Mock(spec=ServerSession)), params) finally: server._mcp_proxy_mode.reset(token) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("client_requests", "tools_list_admitted"), + [((), True), (("subscriptions/listen",), False), (("resources/subscribe",), False)], +) +async def test_only_client_initiated_listeners_consume_server_rate_limit( + monkeypatch: pytest.MonkeyPatch, client_requests: tuple[str, ...], tools_list_admitted: bool +) -> None: + from litellm.proxy._experimental.mcp_server import operations + + allow_one_request(monkeypatch, "alice", "upstream") + opened: Final = asyncio.Queue[None]() + + async def hold_upstream(*_: object, **__: object) -> None: + opened.put_nowait(None) + await asyncio.Event().wait() + + client: Final = Mock(spec=MCPClient) + client.run_with_session = AsyncMock(side_effect=hold_upstream) + server: Final = MCPServer(server_id="upstream", name="upstream", url="http://localhost/mcp", transport="http") + target: Final = subscriptions.SubscriptionTarget(server, client, "synthetic-fingerprint") + + async def resolve(_: OperationContext) -> tuple[subscriptions.SubscriptionTarget, ...]: + return (target,) + + service: Final = subscriptions.GatewaySubscriptions(resolve_targets=resolve) + caller: Final = OperationContext(_caller=UserAPIKeyAuth(user_id="alice")) + authorize: Final = AsyncMock(return_value=caller) + session: Final = Mock(spec=ServerSession) + session.client_params = Mock() + legacy: Final = context(session, "2025-11-25") + requests: Final = { + "subscriptions/listen": lambda: service.listen( + context(Mock(spec=ServerSession)), + SubscriptionsListenRequestParams(notifications=SubscriptionFilter(tools_list_changed=True)), + authorize, + "alice", + ), + "resources/subscribe": lambda: service.legacy_resource(legacy, "test://value", subscribe=True), + } + authorization: Final = subscriptions.SubscriptionAuthorization(caller, authorize, lambda *_: "alice") + starts: Final = (lambda: service.legacy_session(legacy, authorization), *(requests[r] for r in client_requests)) + tasks: Final = tuple(asyncio.create_task(start()) for start in starts) + try: + for _ in tasks: + await asyncio.wait_for(opened.get(), 1) + rejection: Final = await operations._mcp_server_rate_limit_rejection(server, caller.user_api_key_auth) + assert (rejection is None) is tools_list_admitted + finally: + for task in tasks: + task.cancel() + await asyncio.gather(*tasks, return_exceptions=True) + + +@pytest.mark.asyncio +async def test_legacy_listener_catalog_refreshes_do_not_consume_server_rate_limit( + monkeypatch: pytest.MonkeyPatch, +) -> None: + from mcp.types import ListToolsResult, ServerCapabilities, Tool, ToolsCapability + + from litellm.proxy import proxy_server + from litellm.proxy._experimental.mcp_server import catalog, operations + from litellm.proxy._experimental.mcp_server.faults.list_outcomes import ServerListOk + + server: Final = MCPServer(server_id="list-changed", name="list-changed", transport=MCPTransport.http) + caller: Final = UserAPIKeyAuth(api_key="list-changed-key", user_id="alice") + allow_one_request(monkeypatch, "alice", "list-changed") + listed: Final = asyncio.Queue[None]() + + async def fetch_tools(_: MCPServer, **__: object) -> tuple[ListToolsResult, ServerListOk]: + listed.put_nowait(None) + return ListToolsResult(tools=[Tool(name="add", inputSchema={"type": "object"})]), ServerListOk(tool_count=1) + + monkeypatch.setenv("LITELLM_SALT_KEY", "list-changed-test-key") + monkeypatch.setattr(proxy_server, "prisma_client", None) + monkeypatch.setattr(operations.global_mcp_server_manager, "registry", {server.server_id: server}) + monkeypatch.setattr(operations, "_get_allowed_mcp_servers", AsyncMock(return_value=[server])) + monkeypatch.setattr(catalog, "get_filtered_server_tools", fetch_tools) + + upstream: Final = Mock() + upstream.server_capabilities = ServerCapabilities(tools=ToolsCapability(list_changed=True)) + upstream.protocol_version = "2025-11-25" + upstream.send_ping = AsyncMock() + + async def run_with_session(connected, **_: object) -> None: + await connected(upstream) + + client: Final = Mock(spec=MCPClient) + client.run_with_session = AsyncMock(side_effect=run_with_session) + target: Final = subscriptions.SubscriptionTarget(server, client, "synthetic-fingerprint") + + async def resolve(_: OperationContext) -> tuple[subscriptions.SubscriptionTarget, ...]: + return (target,) + + initial: Final = operations.prepare_context(user_api_key_auth=caller, mcp_servers=[server.server_id]) + authorize: Final = AsyncMock(return_value=initial) + session: Final = Mock(spec=ServerSession) + session.client_params = Mock() + service: Final = subscriptions.GatewaySubscriptions(resolve_targets=resolve) + listener: Final = asyncio.create_task( + service.legacy_session( + context(session, "2025-11-25"), + subscriptions.SubscriptionAuthorization(initial, authorize, lambda *_: "alice"), + ) + ) + try: + for _ in range(2): + await asyncio.wait_for(listed.get(), 5) + assert await operations._mcp_server_rate_limit_rejection(server, caller) is None + finally: + listener.cancel() + await asyncio.gather(listener, return_exceptions=True) From 6530b329f43ef456c502ba2db7656f871d62474e Mon Sep 17 00:00:00 2001 From: yuneng-jiang Date: Fri, 16 Oct 2026 14:36:48 -0700 Subject: [PATCH 003/259] fix(proxy): give /team/delete transactions a 30s timeout (#46979) Both interactive transactions in /team/delete used Prisma's 5s default. A lock wait or a short database stall past 5s expired the transaction after the keys were already deleted, so the call returned 503 and left the team row, memberships and team models behind --- litellm/constants.py | 1 + .../model_management_endpoints.py | 4 +- .../management_endpoints/team_endpoints.py | 5 +- .../test_team_delete_large_membership.py | 76 +++++++++++++++++++ 4 files changed, 82 insertions(+), 4 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index 39363cc02e4..425fc22d340 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1859,6 +1859,7 @@ SPEND_COUNTER_RESEED_LOCKS_MAX_SIZE: Final = get_env_int("SPEND_COUNTER_RESEED_L PROXY_DB_LOOKUP_MAX_CONCURRENCY: Final = max(1, get_env_int("PROXY_DB_LOOKUP_MAX_CONCURRENCY", 25)) PROXY_DB_LOOKUP_DEADLINE_SECONDS: Final = max(0.1, get_env_float("PROXY_DB_LOOKUP_DEADLINE_SECONDS", 10.0)) PROXY_DB_LOOKUP_STALL_WINDOW_SECONDS: Final = max(0.0, get_env_float("PROXY_DB_LOOKUP_STALL_WINDOW_SECONDS", 30.0)) +TEAM_DELETE_TRANSACTION_TIMEOUT_SECONDS: Final = 30 DEFAULT_CRON_JOB_LOCK_TTL_SECONDS: Final = get_env_int("DEFAULT_CRON_JOB_LOCK_TTL_SECONDS", 60) # 1 minute PROXY_BUDGET_RESCHEDULER_MIN_TIME: Final = get_env_int("PROXY_BUDGET_RESCHEDULER_MIN_TIME", 597) RESET_BUDGET_JOB_BATCH_SIZE: Final = max(1, get_env_int("RESET_BUDGET_JOB_BATCH_SIZE", 500)) diff --git a/litellm/proxy/management_endpoints/model_management_endpoints.py b/litellm/proxy/management_endpoints/model_management_endpoints.py index 36071bfe51a..7460522f53e 100644 --- a/litellm/proxy/management_endpoints/model_management_endpoints.py +++ b/litellm/proxy/management_endpoints/model_management_endpoints.py @@ -28,7 +28,7 @@ from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError, import litellm from litellm._logging import verbose_proxy_logger from litellm._uuid import uuid -from litellm.constants import LITELLM_PROXY_ADMIN_NAME +from litellm.constants import LITELLM_PROXY_ADMIN_NAME, TEAM_DELETE_TRANSACTION_TIMEOUT_SECONDS from litellm.litellm_core_utils.credential_accessor import CredentialAccessor from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap from litellm.litellm_core_utils.ptu_pricing import ( @@ -1881,7 +1881,7 @@ async def delete_team_models( """ deleted_model_ids: Final[list[str]] = [] transaction_client: Final = _ModelTransactionClient.model_validate(prisma_client.db) - async with transaction_client.tx() as tx: + async with transaction_client.tx(timeout=datetime.timedelta(seconds=TEAM_DELETE_TRANSACTION_TIMEOUT_SECONDS)) as tx: for team_id in team_ids: rows = await _get_team_deployments(team_id, prisma_client, table=tx.litellm_proxymodeltable) model_ids = [row.model_id for row in rows] diff --git a/litellm/proxy/management_endpoints/team_endpoints.py b/litellm/proxy/management_endpoints/team_endpoints.py index ba30ac8609b..0f40662ad83 100644 --- a/litellm/proxy/management_endpoints/team_endpoints.py +++ b/litellm/proxy/management_endpoints/team_endpoints.py @@ -17,7 +17,7 @@ import traceback from collections.abc import Iterable, Mapping, Sequence from collections.abc import Set as AbstractSet from dataclasses import dataclass -from datetime import datetime, timezone +from datetime import datetime, timedelta, timezone from itertools import chain from types import MappingProxyType from typing import ( @@ -42,6 +42,7 @@ from litellm._internal_context import with_service_target from litellm._logging import verbose_proxy_logger from litellm._uuid import uuid from litellm.caching.redis_batch import active_request_redis_batches +from litellm.constants import TEAM_DELETE_TRANSACTION_TIMEOUT_SECONDS from litellm.integrations.prometheus import PrometheusLogger from litellm.litellm_core_utils.safe_json_dumps import safe_dumps from litellm.proxy._types import ( @@ -4674,7 +4675,7 @@ async def delete_team( # or is still waiting on the lock, in which case its own re-read happens after this commits # and sees the row gone before it writes anything. delete_filter: Final[_TeamIdInFilter] = {"team_id": {"in": data.team_ids}} - async with prisma_client.tx() as tx: + async with prisma_client.tx(timeout=timedelta(seconds=TEAM_DELETE_TRANSACTION_TIMEOUT_SECONDS)) as tx: for team_id in sorted(data.team_ids): await tx.query_raw(TEAM_ADVISORY_LOCK_SQL, team_id) await tx.litellm_teamtable.delete_many(where=delete_filter) diff --git a/tests/integration/management/test_team_delete_large_membership.py b/tests/integration/management/test_team_delete_large_membership.py index 90d204722ac..3e4bb568394 100644 --- a/tests/integration/management/test_team_delete_large_membership.py +++ b/tests/integration/management/test_team_delete_large_membership.py @@ -53,6 +53,13 @@ SELECT encode(sha256(convert_to(%s || '-' || n::text, 'UTF8')), 'hex'), %s, '{}' FROM generate_series(1, %s::int) AS n """ TAKE_TEAM_LOCK_SQL: Final = "SELECT pg_advisory_xact_lock(hashtext(%s))" +LOCK_MODEL_ROW_SQL: Final = 'SELECT model_id FROM "LiteLLM_ProxyModelTable" WHERE model_id = %s FOR UPDATE' +LOCK_HELD_PAST_DEFAULT_TRANSACTION_SECONDS: Final = 7 +OLDEST_TRANSACTION_BLOCKED_BY_SQL: Final = """ +SELECT coalesce(max(extract(epoch FROM clock_timestamp() - xact_start)), 0)::float AS age +FROM pg_stat_activity +WHERE %s::int = ANY(pg_blocking_pids(pid)) +""" # Sessions blocked on the advisory lock the given backend holds, and nothing else on the shared rig. WAITERS_ON_HELD_LOCK_SQL: Final = """ SELECT count(*)::int AS waiting @@ -131,6 +138,13 @@ def _waiters_on_lock_held_by(backend_pid: int) -> int: return waiting +def _oldest_transaction_blocked_by(backend_pid: int) -> float: + rows: Final = read_rows(OLDEST_TRANSACTION_BLOCKED_BY_SQL, (str(backend_pid),)) + age: Final = rows[0]["age"] + assert isinstance(age, float) + return age + + def _remove_team_by_sql(team_id: str) -> None: """Cleanup for a team the test expects to have deleted itself. Whatever a failed delete left behind (row, memberships, `teams` references) goes by SQL so the shared rig stays clean without sending @@ -161,6 +175,38 @@ def _create_team(scenario: Scenario, **fields: JsonValue) -> str: return team_id +def _create_team_model(scenario: Scenario, team_id: str) -> str: + created: Final = scenario.gateway.post( + "/model/new", + { + "model_name": f"integration-{uuid.uuid4().hex}", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-integration"}, + "model_info": {"team_id": team_id}, + }, + ) + model_id: Final = string_value(object_value(created["model_info"])["id"]) + scenario.cleanups.callback(write_rows, 'DELETE FROM "LiteLLM_ProxyModelTable" WHERE model_id = %s', (model_id,)) + return model_id + + +def _team_model_rows(model_id: str) -> list[dict[str, JsonValue]]: + return read_rows('SELECT model_id FROM "LiteLLM_ProxyModelTable" WHERE model_id = %s', (model_id,)) + + +def _delete_while_locked(gateway: Gateway, team_id: str, lock_sql: str, lock_target: str) -> httpx.Response: + with ThreadPoolExecutor(max_workers=1) as pool: + with psycopg.connect(os.environ["DATABASE_URL"]) as holder: + holder.execute(lock_sql, (lock_target,)) + pending: Final = _post_in_background(pool, gateway, "/team/delete", {"team_ids": [team_id]}) + eventually( + lambda: _oldest_transaction_blocked_by(holder.info.backend_pid), + lambda age: age > LOCK_HELD_PAST_DEFAULT_TRANSACTION_SECONDS, + seconds=30, + ) + assert not pending.done(), "delete returned while the lock was still held" + return pending.result(timeout=60) + + def _generate_key(scenario: Scenario, **fields: JsonValue) -> str: """A key the team delete is expected to remove; cleanup only deletes it if it is still live.""" key: Final = string_value(scenario.gateway.post("/key/generate", fields)["key"]) @@ -392,6 +438,36 @@ def test_delete_waits_for_the_team_advisory_lock_and_completes_after_release( assert len(_tombstones(team)) == 1, _tombstones(team) +def test_delete_completes_after_waiting_on_a_locked_team_model_row_past_the_default_transaction_timeout( + gateway: Gateway, +) -> None: + with gateway.scenario() as scenario: + team: Final = _create_team(scenario) + model_id: Final = _create_team_model(scenario, team) + response: Final = _delete_while_locked(gateway, team, LOCK_MODEL_ROW_SQL, model_id) + assert response.status_code == 200, response.text + assert response.json() == {"deleted_teams": [team]} + assert _team_rows(team) == [] + assert _team_model_rows(model_id) == [] + assert len(_tombstones(team)) == 1, _tombstones(team) + + +def test_delete_completes_after_waiting_on_the_team_lock_past_the_default_transaction_timeout( + gateway: Gateway, +) -> None: + with gateway.scenario() as scenario: + user: Final = scenario.user() + team: Final = _create_team(scenario) + _bulk_member_add(gateway, team, [user]) + response: Final = _delete_while_locked(gateway, team, TAKE_TEAM_LOCK_SQL, team) + assert response.status_code == 200, response.text + assert response.json() == {"deleted_teams": [team]} + assert _team_rows(team) == [] + assert _membership_user_ids(team) == [] + assert _user_teams(user) == [] + assert len(_tombstones(team)) == 1, _tombstones(team) + + def test_deleting_two_teams_sharing_a_member_in_one_call_clears_both_from_the_member(gateway: Gateway) -> None: with gateway.scenario() as scenario: user: Final = scenario.user() From 3637dd711f533070046d348c4032d552dc32b30f Mon Sep 17 00:00:00 2001 From: Yassin Kortam Date: Fri, 16 Oct 2026 14:48:42 -0700 Subject: [PATCH 004/259] feat(ui): add Authorization URL, Token URL and Scopes fields for OAuth Delegate and True Passthrough MCP servers (#46941) Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../PassthroughAuthorizeSection.tsx | 38 +++++++++++- .../_components/createServerPayload.test.ts | 16 +++++ .../editServerPayload.differential.cases.ts | 5 -- .../editServerPayload.differential.test.ts | 58 +++++++++++++++++++ .../_components/editServerPayload.ts | 21 ++++++- .../mcp_server_edit.integration.test.tsx | 12 +++- .../src/components/mcp_tools/types.test.tsx | 8 +++ .../src/components/mcp_tools/types.tsx | 17 +++++- 8 files changed, 163 insertions(+), 12 deletions(-) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/PassthroughAuthorizeSection.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/PassthroughAuthorizeSection.tsx index 3eb80d3f587..e6cd9eb54a5 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/PassthroughAuthorizeSection.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/PassthroughAuthorizeSection.tsx @@ -3,9 +3,11 @@ import { Button } from "@/components/ui/button"; import { Checkbox } from "@/components/ui/checkbox"; import { Label } from "@/components/ui/label"; import { PasswordInput } from "@/components/shared/PasswordInput"; +import { Input } from "@/components/ui/input"; +import { MultiSelect } from "@/components/shared/MultiSelect"; import DcrBridgeToggle from "./DcrBridgeToggle"; import { MountedFormField } from "@/components/common_components/MountedFormField"; -import { textControl } from "./mcpFieldRules"; +import { tagsControl, textControl } from "./mcpFieldRules"; import { credentialAuthClass, isClientForwardedTokenMode } from "@/components/mcp_tools/types"; interface PassthroughOAuthFlow { @@ -69,6 +71,7 @@ export default function PassthroughAuthorizeSection({ const clientSecretPlaceholder = classUnchanged ? "Leave blank to keep the currently saved secret (if any)" : "Leave blank for public clients / PKCE"; + const fieldClassName = "rounded-lg border-border focus:border-info focus:ring-ring"; const clientIdExtra = classUnchanged ? "Set this to make everyone authorize through a specific app; required for upstreams without dynamic client registration (e.g. a pre-registered Slack app)." : "Switching the auth type discards the previously saved app; enter a client ID here or leave blank to use dynamic client registration."; @@ -113,6 +116,39 @@ export default function PassthroughAuthorizeSection({ /> )} + Authorization URL (optional)} + name="authorization_url" + help="Set this when the upstream publishes no OAuth metadata, for example an internal server behind Microsoft Entra. Leave blank to discover it from the upstream." + > + {(control) => ( + + )} + + Token URL (optional)} + name="token_url" + help="Token endpoint of the same authorization server. Leave blank to discover it from the upstream." + > + {(control) => ( + + )} + + Scopes (optional)} + name={["credentials", "scopes"]} + help="Scopes requested from the upstream when a user authorizes. Press enter or comma between scopes." + > + {(control) => } + {isEditing && onRemoveStoredAppChange && (