From a604c890590f88729b98e02df1d57d6518e00003 Mon Sep 17 00:00:00 2001 From: yuneng-jiang Date: Wed, 28 Oct 2026 13:25:53 -0700 Subject: [PATCH 1/7] fix(ui): bundle the Aria avatar so it loads in every deployment (#48890) AriaAvatar pointed at /ui/assets/aria-agent.png, a path that only exists when the proxy serves the static export under /ui. The dashboard dev server and the split-chart nginx image 404 it, so the header Aria button and every other AriaAvatar showed a broken image. Import the PNG statically like the provider logos (#34125) so the bundle always serves it under /_next/static/media --- .../components/liteadmin/AriaAvatar.test.tsx | 21 +++++++++++++++++++ .../src/components/liteadmin/AriaAvatar.tsx | 4 ++-- 2 files changed, 23 insertions(+), 2 deletions(-) create mode 100644 ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.test.tsx diff --git a/ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.test.tsx b/ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.test.tsx new file mode 100644 index 00000000000..5c9cf27470e --- /dev/null +++ b/ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.test.tsx @@ -0,0 +1,21 @@ +import { render, screen } from "@testing-library/react"; +import { afterEach, describe, expect, it } from "vitest"; +import { setServerRootPath } from "@/lib/serverRootPath"; +import { AriaAvatar } from "./AriaAvatar"; + +describe("AriaAvatar", () => { + afterEach(() => { + setServerRootPath("/"); + }); + + it.each(["/", "/litellm"])("loads the bundled image when the server root is %s", (root) => { + setServerRootPath(root); + + render(); + + expect(screen.getByRole("presentation", { hidden: true })).toHaveAttribute( + "src", + "/_next/static/media/aria-agent.png", + ); + }); +}); diff --git a/ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.tsx b/ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.tsx index 8c78908d3a9..e0f014624fe 100644 --- a/ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.tsx +++ b/ui/litellm-dashboard/src/components/liteadmin/AriaAvatar.tsx @@ -1,10 +1,10 @@ -import { resolveLogoSrc } from "@/lib/assetPaths"; +import ariaAgentImage from "../../../public/assets/aria-agent.png"; import { cn } from "@/lib/cva.config"; export function AriaAvatar({ size = 28, className }: { size?: number; className?: string }) { return ( Date: Wed, 28 Oct 2026 13:32:09 -0700 Subject: [PATCH 2/7] fix(responses)!: bill partial usage on mid-stream failures and report bridged /v1/messages failures to the proxy (#47537) A Responses API stream that fails partway now bills the usage it already consumed, ends with an error instead of a clean stop, and honors stream_timeout between chunks. Bridged /v1/messages calls report those failures to the proxy's failure hooks, so spend, budgets and logs see them --- litellm/litellm_core_utils/litellm_logging.py | 7 + .../responses_adapters/streaming_iterator.py | 65 ++- litellm/proxy/common_request_processing.py | 89 +++- litellm/responses/streaming_iterator.py | 78 +++- litellm/router.py | 3 + .../test_exception_mapping_utils.py | 4 +- ...t_responses_adapters_streaming_iterator.py | 196 +++++++++ .../proxy/test_common_request_processing.py | 174 +++++++- .../mcp/test_aresponses_api_with_mcp.py | 59 +++ .../test_responses_streaming_iterator.py | 391 +++++++++++++++++- .../unit/responses/test_streaming_iterator.py | 10 +- tests/unit/test_router/test_router.py | 126 ++++++ 12 files changed, 1155 insertions(+), 47 deletions(-) diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py index 48e8a44b9a4..3b8157b2249 100644 --- a/litellm/litellm_core_utils/litellm_logging.py +++ b/litellm/litellm_core_utils/litellm_logging.py @@ -842,6 +842,7 @@ class Logging(LiteLLMLoggingBaseClass): self._async_success_enqueued: bool = False self._sync_success_scheduled: bool = False self.on_detached_stream_failure: Callable[[Exception], Awaitable[None]] | None = None + self._stream_failure_claimed: bool = False self.shadow_eval_request_snapshot: GuardrailRequestSnapshot | None = None def copy_for_new_call(self, litellm_call_id: str, start_time: datetime.datetime) -> "Logging": @@ -2561,6 +2562,12 @@ class Logging(LiteLLMLoggingBaseClass): }, ) + def claim_stream_failure_report(self) -> bool: + if self._stream_failure_claimed: + return False + self._stream_failure_claimed = True + return True + def record_assembled_response_for_failure(self, assembled: ModelResponse) -> None: """Bill a fully streamed response on the failure log when a post-call hook rejects it.""" usage: Final = getattr(assembled, "usage", None) diff --git a/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py b/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py index 614a7c96848..44f233f2254 100644 --- a/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py +++ b/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py @@ -2,15 +2,16 @@ ## Translates OpenAI call to Anthropic `/v1/messages` format import asyncio from collections import deque -from collections.abc import AsyncGenerator, Iterator, Mapping +from collections.abc import AsyncGenerator, Awaitable, Callable, Iterator, Mapping from typing import TYPE_CHECKING, Any, Final +import anyio from pydantic import ConfigDict, field_validator from litellm import verbose_logger from litellm._logging import redact_internal_details_from_client_message from litellm._uuid import uuid -from litellm.exceptions import MidStreamFallbackError +from litellm.exceptions import APIError, MidStreamFallbackError from litellm.litellm_core_utils.prompt_templates.common_utils import ( encrypted_reasoning_signature, ) @@ -22,6 +23,7 @@ from litellm.llms.anthropic.messages.native.utils import ( from litellm.llms.anthropic.messages.safeguards import StreamedSafeguardResults from litellm.llms.anthropic.messages.utils import anthropic_sse_frame from litellm.responses.streaming_iterator import ( + map_stream_error_to_exception, partial_responses_incomplete_event_on_disconnect, stream_error_status_and_message, ) @@ -130,10 +132,12 @@ class AnthropicResponsesStreamWrapper: ) -> None: self.responses_stream = responses_stream self.model = model + self.litellm_logging_obj = litellm_logging_obj self.safeguards_evaluator: Final = safeguards_evaluator self._message_id: str = f"msg_{uuid.uuid4()}" if litellm_logging_obj is not None: litellm_logging_obj.record_streamed_anthropic_message_id(self._message_id) + self._unreported_failure: Exception | None = None self._current_block_index: int = -1 # Map item_id -> content_block_index so we can stop the right block later self._item_id_to_block_index: dict[str, int] = {} @@ -377,16 +381,24 @@ class AnthropicResponsesStreamWrapper: if event_type == "response.failed": failed: Final = _FailedResponseEvent.model_validate(event) - status_code, message = stream_error_status_and_message( - failed.response.error if failed.response is not None else None - ) + upstream_error: Final = failed.response.error if failed.response is not None else None + status_code, message = stream_error_status_and_message(upstream_error) verbose_logger.error( "AnthropicResponsesStreamWrapper: upstream Responses stream for %s failed (%s): %s", self.model, status_code, message, ) - self._fail_stream(status_code, message) + self._fail_stream( + status_code, + message, + failure=map_stream_error_to_exception( + upstream_error, + self.model, + getattr(self.responses_stream, "custom_llm_provider", None) or "", + logging_obj=self.litellm_logging_obj, + ), + ) return # ---- response completed -> message_delta + message_stop ---- @@ -439,10 +451,36 @@ class AnthropicResponsesStreamWrapper: self._sent_message_stop = True return - def _fail_stream(self, status_code: int, message: str) -> None: + def _fail_stream(self, status_code: int, message: str, failure: Exception | None = None) -> None: self._stream_failed = True + self._unreported_failure = failure or APIError( + status_code=status_code, + message=message, + llm_provider=getattr(self.responses_stream, "custom_llm_provider", None), + model=self.model, + ) self._chunk_queue.append(_anthropic_error_chunk(status_code, message)) + async def _report_unreported_failure(self) -> None: + failure: Final = self._unreported_failure + if failure is None: + return + self._unreported_failure = None + on_failure: Final[Callable[[Exception], Awaitable[None]] | None] = getattr( + self.litellm_logging_obj, "_on_detached_stream_failure", None + ) + if on_failure is None: + return + try: + with anyio.CancelScope(shield=True): + await on_failure(failure) + except Exception as hook_failure: # noqa: BLE001 # a failing proxy hook must not hide the error event from the client + verbose_logger.warning( + "AnthropicResponsesStreamWrapper: stream failure hook raised: %s(%s)", + type(hook_failure).__name__, + hook_failure, + ) + def __aiter__(self) -> "AnthropicResponsesStreamWrapper": return self @@ -450,6 +488,7 @@ class AnthropicResponsesStreamWrapper: if self._chunk_queue: return self._chunk_queue.popleft() if self._stream_failed: + await self._report_unreported_failure() raise StopAsyncIteration if not self._sent_message_start: @@ -478,7 +517,7 @@ class AnthropicResponsesStreamWrapper: verbose_logger.exception( "AnthropicResponsesStreamWrapper: upstream Responses stream for %s failed", self.model ) - self._fail_stream(*_failure_status_and_message(e)) + self._fail_stream(*_failure_status_and_message(e), failure=_original_failure(e)) if not self._chunk_queue and not self._sent_message_stop and not self._stream_failed: verbose_logger.error( @@ -498,8 +537,14 @@ class AnthropicResponsesStreamWrapper: async def _sse_frames(self) -> AsyncGenerator[bytes, None]: safeguard_results: Final = StreamedSafeguardResults(self.safeguards_evaluator) - async for chunk in self: - yield anthropic_sse_frame(await safeguard_results.observe(chunk)) + try: + async for chunk in self: + yield anthropic_sse_frame(await safeguard_results.observe(chunk)) + finally: + from litellm.llms.anthropic.messages.native.streaming_iterator import closing_superseded_attempt + + if not closing_superseded_attempt(): + await self._report_unreported_failure() def partial_response_on_disconnect(self) -> "ResponseIncompleteEvent | None": return partial_responses_incomplete_event_on_disconnect(self.responses_stream) diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index 24a36a7f2dc..f5d6417fb79 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -4,6 +4,7 @@ import json import logging import math from collections.abc import AsyncGenerator, AsyncIterator, Awaitable, Callable, Coroutine, Mapping, Sequence +from dataclasses import dataclass from datetime import datetime from functools import lru_cache from itertools import chain @@ -136,7 +137,7 @@ from litellm.proxy.utils import ( # noqa: F401 # legacy module exports _check_and_merge_model_level_guardrails, # pyright: ignore[reportPrivateUsage,reportUnusedImport] # backwards-compatible package export check_and_merge_model_level_guardrails, ) -from litellm.responses.streaming_iterator import partial_responses_incomplete_event_on_disconnect +from litellm.responses.streaming_iterator import partial_responses_incomplete_event_on_disconnect, spawn_logging_task from litellm.router import Router from litellm.router_utils.add_retry_fallback_headers import get_hidden_params_dict from litellm.router_utils.auto_router_model_naming import AUTO_ROUTER_MODEL_PREFIX @@ -2006,6 +2007,49 @@ def _content_text(fields: Mapping[str, object]) -> str: return str(fields.get("content", "")) +async def _no_stream_failure_to_release() -> None: + return None + + +@dataclass(slots=True) +class _ClientStreamFailureReport: + logging_obj: LiteLLMLoggingObj + report: Callable[[Exception], Awaitable[None]] + held: Exception | None = None + + @classmethod + def hold_while_streaming(cls, logging_obj: LiteLLMLoggingObj | None) -> Callable[[], Awaitable[None]]: + if logging_obj is None or logging_obj.on_detached_stream_failure is None: + return _no_stream_failure_to_release + return cls(logging_obj, logging_obj.on_detached_stream_failure)._start_holding() + + def _start_holding(self) -> Callable[[], Awaitable[None]]: + self.logging_obj.on_detached_stream_failure = self.hold + return self.release + + async def hold(self, failure: Exception) -> None: + if self.held is None: + self.held = failure + + async def release(self) -> None: + self.logging_obj.on_detached_stream_failure = self.report + if self.held is None: + return + await asyncio.shield( + spawn_logging_task( + asyncio.get_running_loop(), self._report_held(self.held), task_name="stream failure report" + ) + ) + + async def _report_held(self, failure: Exception) -> None: + try: + await self.report(failure) + except Exception as hook_failure: # noqa: BLE001 # a failing failure hook must not skip the stream cleanup + verbose_proxy_logger.warning( + "Stream failure hook raised: %s(%s)", type(hook_failure).__name__, sanitize_for_log(hook_failure) + ) + + class ProxyBaseLLMRequestProcessing: def __init__(self, data: dict, served_model: str | None = None): self.data = data @@ -3498,6 +3542,8 @@ class ProxyBaseLLMRequestProcessing: request_data: Final = self.data async def _on_detached_stream_failure(exc: Exception) -> None: + if not logging_obj.claim_stream_failure_report(): + return try: await proxy_logging_obj.post_call_failure_hook( user_api_key_dict=user_api_key_dict, @@ -4448,11 +4494,15 @@ class ProxyBaseLLMRequestProcessing: client_disconnected = False delivered_chunk = False recent_tail = SSE_STREAM_START_TAIL # rebind-ok: rolling window over the yielded bytes + stream_logging_obj: Final[object] = request_data.get("litellm_logging_obj") # pyright: ignore[reportUnknownVariableType, reportUnknownMemberType] # legacy request data is narrowed to Logging below guarded_stream: Final[AsyncGenerator[object, None]] = proxy_logging_obj.async_post_call_streaming_iterator_hook( user_api_key_dict=user_api_key_dict, response=response, request_data=request_data, ) + release_stream_failure: Final = _ClientStreamFailureReport.hold_while_streaming( + stream_logging_obj if isinstance(stream_logging_obj, LiteLLMLoggingObj) else None + ) try: str_so_far = "" async for chunk in guarded_stream: @@ -4523,10 +4573,18 @@ class ProxyBaseLLMRequestProcessing: verbose_proxy_logger.exception( "litellm.proxy.proxy_server.async_data_generator(): Exception occured - %s", e ) - transformed_exception: Final = await proxy_logging_obj.post_call_failure_hook( - user_api_key_dict=user_api_key_dict, - original_exception=e, - request_data=request_data, + already_reported: Final = ( + isinstance(stream_logging_obj, LiteLLMLoggingObj) + and not stream_logging_obj.claim_stream_failure_report() + ) + transformed_exception: Final = ( + None + if already_reported + else await proxy_logging_obj.post_call_failure_hook( + user_api_key_dict=user_api_key_dict, + original_exception=e, + request_data=request_data, + ) ) if transformed_exception is not None: e = transformed_exception @@ -4549,15 +4607,18 @@ class ProxyBaseLLMRequestProcessing: seal: Final = "" if seal_open_frame is None else seal_open_frame(recent_tail) yield seal + error_frame if seal else error_frame finally: - await ProxyBaseLLMRequestProcessing.finalize_streaming_generator_cleanup( - request=request, - request_data=request_data, - response=response, - stream_completed=stream_completed, - client_disconnected=client_disconnected, - user_api_key_dict=user_api_key_dict, - proxy_logging_obj=proxy_logging_obj, - ) + try: + await release_stream_failure() + finally: + await ProxyBaseLLMRequestProcessing.finalize_streaming_generator_cleanup( + request=request, + request_data=request_data, + response=response, + stream_completed=stream_completed, + client_disconnected=client_disconnected, + user_api_key_dict=user_api_key_dict, + proxy_logging_obj=proxy_logging_obj, + ) @staticmethod def async_sse_data_generator( diff --git a/litellm/responses/streaming_iterator.py b/litellm/responses/streaming_iterator.py index c1f422c7d7b..4e418a91156 100644 --- a/litellm/responses/streaming_iterator.py +++ b/litellm/responses/streaming_iterator.py @@ -41,7 +41,11 @@ from litellm.exceptions import MidStreamFallbackError, RateLimitError from litellm.litellm_core_utils.asyncify import run_async_function from litellm.litellm_core_utils.core_helpers import process_response_headers from litellm.litellm_core_utils.exception_mapping_utils import with_deployment_api_base -from litellm.litellm_core_utils.internal_call_metadata import without_budget_reservation +from litellm.litellm_core_utils.internal_call_metadata import ( + is_background_response, + requests_background, + without_budget_reservation, +) from litellm.litellm_core_utils.litellm_logging import BilledRoundsSpend from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.llm_response_utils.get_api_base import get_api_base @@ -65,7 +69,7 @@ from litellm.types.llms.openai import ( ResponsesAPIStreamingResponse, ) from litellm.types.responses.main import GenericResponseOutputItem, OutputText -from litellm.types.utils import CallTypes +from litellm.types.utils import CallTypes, ModelResponse from litellm.utils import async_post_call_success_deployment_hook if TYPE_CHECKING: @@ -288,7 +292,7 @@ def stream_error_status_and_message(error_obj: object) -> tuple[int, str]: return _status_code_for_error_fields(error_type, error_code), message -def _map_stream_error_to_exception( +def map_stream_error_to_exception( error_obj: object, model: str, custom_llm_provider: str, @@ -578,6 +582,7 @@ class BaseResponsesAPIStreamingIterator: self.request_data.get("input"), self.request_data, self._generated_content + self._generated_tool_arguments, + use_default_image_token_count=False, ) if _estimate_wanted else None @@ -731,7 +736,7 @@ class BaseResponsesAPIStreamingIterator: self._earlier_rounds_spend = earlier def _map_error_event_exception(self, error_obj: object) -> Exception: - return _map_stream_error_to_exception( + return map_stream_error_to_exception( error_obj, self.model or "", self.custom_llm_provider or "", @@ -981,6 +986,39 @@ class BaseResponsesAPIStreamingIterator: except Exception: pass + def _record_partial_usage_for_failure(self) -> None: + generated_text: Final = self._generated_content + self._generated_tool_arguments + if ( + not generated_text + or requests_background(self.request_data) + or is_background_response(self._created_response) + or self._completed_response_logged + or self.logging_obj.model_call_details.get("combined_usage_object") is not None + ): + return + estimate: Final = _estimate_usage_safely( + self.model or "", + self.request_data.get("input"), + self.request_data, + generated_text, + use_default_image_token_count=True, + ) + if estimate is None: + return + usage: Final = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage( # pyright: ignore[reportPrivateUsage] # same shared transform the success path above uses + estimate + ) + served_tier: Final[object] = getattr(self._created_response, "service_tier", None) + assembled: Final = ModelResponse( + model=self.model, usage=usage, service_tier=served_tier if isinstance(served_tier, str) else None + ) + if self._earlier_rounds_spend is None: + self.logging_obj.record_assembled_response_for_failure(assembled) + return + self.logging_obj.record_failure_after_earlier_rounds( + self._earlier_rounds_spend, usage, self.logging_obj.response_cost_calculator(result=assembled) or 0.0 + ) + def _handle_failure(self, exception: Exception): """ Trigger failure handlers before bubbling the exception. @@ -990,6 +1028,7 @@ class BaseResponsesAPIStreamingIterator: if self._failure_handled: return self._failure_handled = True + self._record_partial_usage_for_failure() traceback_exception: Final = traceback.format_exc() end_time: Final = datetime.now() @@ -1060,15 +1099,20 @@ class BaseResponsesAPIStreamingIterator: the same shape the success handler assembles for a natural end of stream. Its response is the response.created payload (it carries the served service_tier) plus the text and tool arguments relayed so far, with counted - usage since the provider never sent any. None when nothing was relayed yet - or the stream already reached its terminal event and billed itself. + usage since the provider never sent any. None when nothing was relayed yet, + the stream already reached its terminal event and billed itself, or its failure + log already billed the relayed output. """ created: Final = self._created_response generated: Final = self._generated_content + self._generated_tool_arguments - if self.completed_response is not None or created is None or not generated: + if created is None or not generated or self._billed_without_disconnect(): return None usage: Final = _estimate_usage_safely( - self.model or "", self.request_data.get("input"), self.request_data, generated + self.model or "", + self.request_data.get("input"), + self.request_data, + generated, + use_default_image_token_count=False, ) if usage is None: return None @@ -1082,6 +1126,9 @@ class BaseResponsesAPIStreamingIterator: stamp_responses_usage_cost(partial, self.logging_obj) return ResponseIncompleteEvent(type=ResponsesAPIStreamEvents.RESPONSE_INCOMPLETE, response=partial) + def _billed_without_disconnect(self) -> bool: + return self.completed_response is not None or self._failure_handled + def _note_yielded_event(self, event: ResponsesAPIStreamingResponse) -> None: self._yielded_first_chunk = True if event.type not in PRE_OUTPUT_LIFECYCLE_EVENT_TYPES: @@ -1218,6 +1265,9 @@ class ResponsesAPIStreamingIterator(BaseResponsesAPIStreamingIterator): raise StopAsyncIteration from e self._raise_for_transport_error(e) except httpx.TimeoutException as e: + if self.completed_response is not None: + self.finished = True + raise StopAsyncIteration from e self._raise_for_transport_timeout(e) except httpx.HTTPError as e: # Handle HTTP errors @@ -1303,6 +1353,9 @@ class SyncResponsesAPIStreamingIterator(BaseResponsesAPIStreamingIterator): raise StopIteration from e self._raise_for_transport_error(e) except httpx.TimeoutException as e: + if self.completed_response is not None: + self.finished = True + raise StopIteration from e self._raise_for_transport_timeout(e) except httpx.HTTPError as e: # Handle HTTP errors @@ -1727,6 +1780,8 @@ def _estimate_usage_from_text( request_input: object, responses_api_request: Mapping[str, object], generated_text: str, + *, + use_default_image_token_count: bool = False, ) -> ResponseAPIUsage: messages: Final = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages( # pyright: ignore[reportUnknownMemberType] # the transformer's signature is partially untyped input=request_input, # pyright: ignore[reportArgumentType] # the raw Responses API input is a str or ResponseInputParam list, matching the helper's declared union @@ -1734,7 +1789,7 @@ def _estimate_usage_from_text( only_since_compaction=True, ) input_tokens: Final = litellm.token_counter( # pyright: ignore[reportUnknownMemberType] # token_counter's public signature is untyped - model=model, messages=messages + model=model, messages=messages, use_default_image_token_count=use_default_image_token_count ) output_tokens: Final = litellm.token_counter( # pyright: ignore[reportUnknownMemberType] # token_counter's public signature is untyped model=model, text=generated_text, count_response_tokens=True @@ -1751,6 +1806,8 @@ def _estimate_usage_safely( request_input: object, responses_api_request: Mapping[str, object], generated_text: str, + *, + use_default_image_token_count: bool, ) -> ResponseAPIUsage | None: try: return _estimate_usage_from_text( @@ -1758,6 +1815,7 @@ def _estimate_usage_safely( request_input=request_input, responses_api_request=responses_api_request, generated_text=generated_text, + use_default_image_token_count=use_default_image_token_count, ) except Exception as e: verbose_logger.debug("Could not estimate usage from stream text, billing $0: %s", e) @@ -2276,7 +2334,7 @@ class _ResponsesWsTurn: ) if failed_event is None: return None - return _map_stream_error_to_exception( + return map_stream_error_to_exception( _ws_event_error(failed_event), model, custom_llm_provider, logging_obj=connection_logging_obj ) diff --git a/litellm/router.py b/litellm/router.py index d712ed4ac52..5b18d36316f 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -6456,6 +6456,9 @@ class Router: hop_kwargs: Final = mid_stream_fallback_hop_kwargs( model=model, original_generic_function=original_generic_function, controls=controls, kwargs=kwargs ) + shared_logging: Final = kwargs.get("litellm_logging_obj") + if isinstance(shared_logging, LiteLLMLogging): + shared_logging.carry_failed_attempt_usage() response: Final = await self._ageneric_api_call_with_fallbacks_helper( model=model, original_generic_function=original_generic_function, **kwargs ) diff --git a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py index 9e8e9cf2a80..55e7c47e2c9 100644 --- a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py +++ b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py @@ -1516,9 +1516,9 @@ def test_bedrock_mantle_openai_envelope_context_overflow_maps_to_context_window_ def test_bedrock_mantle_streamed_context_overflow_event_maps_to_context_window_exceeded(): - from litellm.responses.streaming_iterator import _map_stream_error_to_exception + from litellm.responses.streaming_iterator import map_stream_error_to_exception - mapped_exception = _map_stream_error_to_exception( + mapped_exception = map_stream_error_to_exception( { "code": "context_length_exceeded", "message": "Your input exceeds the context window of this model. Please adjust your input and try again.", diff --git a/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_streaming_iterator.py b/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_streaming_iterator.py index 5af8e69a0ea..4626f76c8fc 100644 --- a/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_streaming_iterator.py +++ b/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_streaming_iterator.py @@ -4,20 +4,25 @@ Tests for AnthropicResponsesStreamWrapper """ import asyncio +import datetime import json import os import sys +from collections.abc import AsyncIterator, Awaitable, Callable from types import SimpleNamespace +import anyio import pytest sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../../.."))) import litellm from litellm.exceptions import MidStreamFallbackError +from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging from litellm.litellm_core_utils.prompt_templates.common_utils import ( encrypted_reasoning_signature, ) +from litellm.llms.anthropic.messages.native.streaming_iterator import aclose_superseded_attempt from litellm.llms.anthropic.messages.native.utils import INCOMPLETE_STREAM_ERROR_MESSAGE from litellm.llms.anthropic.messages.responses_adapters.streaming_iterator import ( AnthropicResponsesStreamWrapper, @@ -622,3 +627,194 @@ class TestUpstreamFailureEndsStreamWithErrorEvent: chunks = _collect(iter(())) assert [chunk["type"] for chunk in chunks] == ["message_start", "error"] assert chunks[1]["error"] == {"type": "api_error", "message": INCOMPLETE_STREAM_ERROR_MESSAGE} + + +class _FailureHookRecorder: + def __init__(self) -> None: + self.received: tuple[Exception, ...] = () + + async def __call__(self, failure: Exception) -> None: + self.received = (*self.received, failure) + + +def _logging_obj_with_failure_hook(on_failure: Callable[[Exception], Awaitable[None]]) -> LitellmLogging: + logging_obj = LitellmLogging( + model="m", + messages=[{"role": "user", "content": "hi"}], + stream=True, + call_type="anthropic_messages", + start_time=datetime.datetime(2026, 1, 1), + litellm_call_id="lit-6871-test", + function_id="lit-6871-test", + ) + logging_obj._on_detached_stream_failure = on_failure + return logging_obj + + +def _collect_with_logging( + stream: AsyncIterator[dict[str, object]], logging_obj: LitellmLogging +) -> list[dict[str, object]]: + async def _run() -> list[dict[str, object]]: + wrapper = AnthropicResponsesStreamWrapper(responses_stream=stream, model="m", litellm_logging_obj=logging_obj) + return [chunk async for chunk in wrapper] + + return asyncio.run(_run()) + + +class _DeliveryCountingConsumer: + def __init__(self) -> None: + self.delivered: tuple[dict[str, object], ...] = () + self.delivered_when_hook_fired: tuple[int, ...] = () + + async def hook(self, failure: Exception) -> None: + self.delivered_when_hook_fired = (*self.delivered_when_hook_fired, len(self.delivered)) + + async def drain(self, wrapper: AnthropicResponsesStreamWrapper) -> None: + async for chunk in wrapper: + self.delivered = (*self.delivered, chunk) + + +def _failing_after_one_delta() -> AsyncIterator[dict[str, object]]: + async def _gen() -> AsyncIterator[dict[str, object]]: + yield {"type": "response.created"} + yield {"type": "response.output_item.added", "item": {"type": "message", "id": "msg_1"}} + yield {"type": "response.output_text.delta", "item_id": "msg_1", "delta": "Hi"} + raise litellm.RateLimitError(message="You have no credits remaining.", llm_provider="openai", model="m") + + return _gen() + + +class TestUpstreamFailureReachesTheProxyFailureHook: + def test_raised_upstream_failure_reaches_the_hook_unwrapped_once_the_error_event_was_consumed(self): + rate_limit = litellm.RateLimitError(message="You have no credits remaining.", llm_provider="openai", model="m") + wrapped = MidStreamFallbackError( + message=str(rate_limit), + model="m", + llm_provider="openai", + original_exception=rate_limit, + is_pre_first_chunk=False, + ) + recorder = _FailureHookRecorder() + + async def _gen() -> AsyncIterator[dict[str, object]]: + yield {"type": "response.created"} + yield {"type": "response.output_item.added", "item": {"type": "message", "id": "msg_1"}} + yield {"type": "response.output_text.delta", "item_id": "msg_1", "delta": "Hi"} + raise wrapped + + chunks = _collect_with_logging(_gen(), _logging_obj_with_failure_hook(recorder)) + assert [chunk["type"] for chunk in chunks] == [ + "message_start", + "content_block_start", + "content_block_delta", + "error", + ] + assert recorder.received == (rate_limit,) + + def test_response_failed_event_reaches_the_hook_as_an_exception_carrying_the_upstream_status(self): + recorder = _FailureHookRecorder() + failed = SimpleNamespace( + status="failed", + output=[], + usage=None, + error={"code": "rate_limit_exceeded", "message": "Rate limit reached for gpt-5.5, try again in 20s."}, + ) + + async def _gen() -> AsyncIterator[dict[str, object]]: + yield {"type": "response.created"} + yield {"type": "response.failed", "response": failed} + + chunks = _collect_with_logging(_gen(), _logging_obj_with_failure_hook(recorder)) + assert chunks[-1]["type"] == "error" + (failure,) = recorder.received + assert isinstance(failure, litellm.RateLimitError) + assert failure.status_code == 429 + assert "Rate limit reached for gpt-5.5, try again in 20s." in failure.message + + def test_a_raising_hook_never_hides_the_error_event_from_the_client(self): + async def _exploding_hook(failure: Exception) -> None: + raise RuntimeError("spend writer is down") + + logging_obj = _logging_obj_with_failure_hook(_exploding_hook) + + async def _gen() -> AsyncIterator[dict[str, object]]: + yield {"type": "response.created"} + raise ConnectionResetError("Response payload is not completed") + + chunks = _collect_with_logging(_gen(), logging_obj) + assert [chunk["type"] for chunk in chunks] == ["message_start", "error"] + assert chunks[-1]["error"] == {"type": "api_error", "message": "Response payload is not completed"} + + def test_the_hook_fires_only_after_the_consumer_took_the_error_event(self): + consumer = _DeliveryCountingConsumer() + + async def _run() -> None: + wrapper = AnthropicResponsesStreamWrapper( + responses_stream=_failing_after_one_delta(), + model="m", + litellm_logging_obj=_logging_obj_with_failure_hook(consumer.hook), + ) + await consumer.drain(wrapper) + with pytest.raises(StopAsyncIteration): + await wrapper.__anext__() + + asyncio.run(_run()) + assert [chunk["type"] for chunk in consumer.delivered] == [ + "message_start", + "content_block_start", + "content_block_delta", + "error", + ] + assert consumer.delivered_when_hook_fired == (len(consumer.delivered),) + + @pytest.mark.parametrize( + ("close", "expected_reports"), + [ + pytest.param(lambda sse: sse.aclose(), 1, id="client_closes"), + pytest.param(aclose_superseded_attempt, 0, id="fallback_takes_over"), + ], + ) + def test_closing_the_stream_on_the_error_event_reports_unless_a_fallback_takes_over(self, close, expected_reports): + recorder = _FailureHookRecorder() + + async def _close_on_the_error_event() -> None: + wrapper = AnthropicResponsesStreamWrapper( + responses_stream=_failing_after_one_delta(), model="m", litellm_logging_obj=_logging_obj_with_failure_hook(recorder) + ) + sse = wrapper.async_anthropic_sse_wrapper() + async for frame in sse: + if frame.startswith(b"event: error\n"): + await close(sse) + return + pytest.fail("the stream ended without an error event") + + asyncio.run(_close_on_the_error_event()) + assert len(recorder.received) == expected_reports + + def test_a_client_disconnect_during_the_failure_hook_lets_the_hook_finish(self): + hook_started = anyio.Event() + hook_released = anyio.Event() + finished: list[Exception] = [] + + async def _held_hook(failure: Exception) -> None: + hook_started.set() + await hook_released.wait() + finished.append(failure) + + async def _disconnect_mid_hook() -> None: + wrapper = AnthropicResponsesStreamWrapper( + responses_stream=_failing_after_one_delta(), model="m", litellm_logging_obj=_logging_obj_with_failure_hook(_held_hook) + ) + + async def _consume() -> None: + async for _ in wrapper.async_anthropic_sse_wrapper(): + pass + + async with anyio.create_task_group() as task_group: + task_group.start_soon(_consume) + await hook_started.wait() + task_group.cancel_scope.cancel() + hook_released.set() + + asyncio.run(_disconnect_mid_hook()) + assert len(finished) == 1 diff --git a/tests/unit/proxy/test_common_request_processing.py b/tests/unit/proxy/test_common_request_processing.py index 52816dc50a8..790790459c3 100644 --- a/tests/unit/proxy/test_common_request_processing.py +++ b/tests/unit/proxy/test_common_request_processing.py @@ -10826,12 +10826,19 @@ class TestResolveLitellmCallId: class _FailureHookRecorder: """Stands in for ProxyLogging.post_call_failure_hook, recording what the detached-failure closure hands it.""" - def __init__(self, raises: Optional[Exception] = None): + def __init__(self, raises: Optional[Exception] = None, gate: Optional[asyncio.Event] = None): self.calls = [] + self.entered = asyncio.Event() + self.reported = asyncio.Event() self._raises = raises + self._gate = gate async def post_call_failure_hook(self, **kwargs): + self.entered.set() + if self._gate is not None: + await self._gate.wait() self.calls.append(kwargs) + self.reported.set() if self._raises is not None: raise self._raises @@ -10932,6 +10939,171 @@ class TestDetachedStreamFailureHook: assert [call["original_exception"] for call in recorder.calls] == [failure] + @staticmethod + def _armed_stream(recorder: _FailureHookRecorder, iterator_hook_for, response: Optional[MagicMock] = None): + from litellm.proxy._types import UserAPIKeyAuth as RealUserAPIKeyAuth + + logging_obj = LiteLLMLoggingObj( + model="gpt-6.1-sol", + messages=[{"role": "user", "content": "Write a story."}], + stream=True, + call_type="anthropic_messages", + start_time=None, + litellm_call_id="call-lit6871", + function_id="function-lit6871", + ) + user_api_key_dict = RealUserAPIKeyAuth(api_key="sk-test") + processing_obj = ProxyBaseLLMRequestProcessing(data={"litellm_logging_obj": logging_obj}) + ProxyLogging._callback_capabilities_cache.clear() + proxy_logging_obj = ProxyLogging(user_api_key_cache=MagicMock()) + proxy_logging_obj.post_call_failure_hook = recorder.post_call_failure_hook + proxy_logging_obj.async_post_call_streaming_iterator_hook = iterator_hook_for(logging_obj) + processing_obj._arm_detached_stream_failure_hook( + logging_obj=logging_obj, + user_api_key_dict=user_api_key_dict, + proxy_logging_obj=proxy_logging_obj, + ) + stream = ProxyBaseLLMRequestProcessing.async_streaming_data_generator( + response=MagicMock() if response is None else response, + user_api_key_dict=user_api_key_dict, + request_data=processing_obj.data, + proxy_logging_obj=proxy_logging_obj, + serialize_chunk=lambda chunk: "data: chunk\n\n", + serialize_error=lambda error: "data: error\n\n", + ) + return logging_obj, stream + + @staticmethod + def _hook_reporting(upstream_failure: Exception, after_report=None): + def iterator_hook_for(logging_obj): + async def iterator_hook(**_kwargs): + yield b"event: message_start\n\n" + await logging_obj.on_detached_stream_failure(upstream_failure) + yield b"event: error\n\n" + yield b"event: message_stop\n\n" + if after_report is not None: + raise after_report + + return iterator_hook + + return iterator_hook_for + + @pytest.mark.asyncio + async def test_iterator_hook_failure_after_a_bridged_failure_is_the_one_the_failure_hook_gets(self): + recorder = _FailureHookRecorder() + upstream_failure = RuntimeError("upstream Responses stream failed") + hook_failure = RuntimeError("iterator hook rejected the stream") + + _, stream = self._armed_stream(recorder, self._hook_reporting(upstream_failure, after_report=hook_failure)) + + frames = [frame async for frame in stream] + + assert frames == ["data: chunk\n\n"] * 3 + ["data: error\n\n"] + assert [call["original_exception"] for call in recorder.calls] == [hook_failure] + + @pytest.mark.asyncio + async def test_failure_hook_replacing_the_iterator_hook_failure_aborts_the_stream(self): + replacement = HTTPException(status_code=418, detail="rejected by a custom failure hook") + recorder = _FailureHookRecorder(raises=replacement) + upstream_failure = RuntimeError("upstream Responses stream failed") + hook_failure = RuntimeError("iterator hook rejected the stream") + + _, stream = self._armed_stream(recorder, self._hook_reporting(upstream_failure, after_report=hook_failure)) + + frames = [await stream.__anext__() for _ in range(3)] + with pytest.raises(HTTPException) as raised: + await stream.__anext__() + + assert raised.value is replacement + assert frames == ["data: chunk\n\n"] * 3 + assert [call["original_exception"] for call in recorder.calls] == [hook_failure] + + @pytest.mark.asyncio + async def test_failure_reported_by_the_stream_reaches_the_failure_hook_once_after_the_stream_ends(self): + recorder = _FailureHookRecorder() + upstream_failure = RuntimeError("upstream Responses stream failed") + + _, stream = self._armed_stream(recorder, self._hook_reporting(upstream_failure)) + + frames = [] + async for frame in stream: + frames.append(frame) + assert recorder.calls == [] + + assert frames == ["data: chunk\n\n"] * 3 + assert [call["original_exception"] for call in recorder.calls] == [upstream_failure] + + @pytest.mark.asyncio + async def test_failure_report_and_stream_cleanup_survive_the_client_task_being_cancelled_mid_report(self): + gate = asyncio.Event() + recorder = _FailureHookRecorder(gate=gate) + upstream_failure = RuntimeError("upstream Responses stream failed") + response = MagicMock() + response.aclose = AsyncMock() + + _, stream = self._armed_stream(recorder, self._hook_reporting(upstream_failure), response=response) + + async def client_reads_the_whole_stream(): + return [frame async for frame in stream] + + client = asyncio.create_task(client_reads_the_whole_stream()) + await asyncio.wait_for(recorder.entered.wait(), timeout=5) + client.cancel() + with pytest.raises(asyncio.CancelledError): + await client + + response.aclose.assert_awaited_once() + gate.set() + await asyncio.wait_for(recorder.reported.wait(), timeout=5) + assert [call["original_exception"] for call in recorder.calls] == [upstream_failure] + + @pytest.mark.asyncio + async def test_failure_reported_by_the_stream_reaches_the_failure_hook_once_when_the_client_leaves_early(self): + recorder = _FailureHookRecorder() + upstream_failure = RuntimeError("upstream Responses stream failed") + + _, stream = self._armed_stream(recorder, self._hook_reporting(upstream_failure)) + + await stream.__anext__() + await stream.__anext__() + await stream.aclose() + + assert [call["original_exception"] for call in recorder.calls] == [upstream_failure] + + @pytest.mark.asyncio + async def test_failure_reported_after_the_stream_ends_still_reaches_the_failure_hook(self): + recorder = _FailureHookRecorder() + upstream_failure = RuntimeError("upstream Responses stream failed after the client stream closed") + + async def iterator_hook(**_kwargs): + yield b"event: message_start\n\n" + + logging_obj, stream = self._armed_stream(recorder, lambda _logging_obj: iterator_hook) + + frames = [frame async for frame in stream] + await logging_obj.on_detached_stream_failure(upstream_failure) + + assert frames == ["data: chunk\n\n"] + assert [call["original_exception"] for call in recorder.calls] == [upstream_failure] + + @pytest.mark.asyncio + async def test_failure_reported_by_an_iterator_hook_is_not_reported_again_when_the_stream_closes(self): + recorder = _FailureHookRecorder() + upstream_failure = RuntimeError("upstream Responses stream failed") + hook_failure = RuntimeError("iterator hook rejected the error frame") + + async def iterator_hook(**_kwargs): + yield b"event: message_start\n\n" + raise hook_failure + + logging_obj, stream = self._armed_stream(recorder, lambda _logging_obj: iterator_hook) + + frames = [frame async for frame in stream] + await logging_obj.on_detached_stream_failure(upstream_failure) + + assert frames == ["data: chunk\n\n", "data: error\n\n"] + assert [call["original_exception"] for call in recorder.calls] == [hook_failure] + class TestPostCallMaskedOutputReachesDeferredLogging: @pytest.mark.asyncio diff --git a/tests/unit/responses/mcp/test_aresponses_api_with_mcp.py b/tests/unit/responses/mcp/test_aresponses_api_with_mcp.py index 4cbcc6c1519..ec0be896095 100644 --- a/tests/unit/responses/mcp/test_aresponses_api_with_mcp.py +++ b/tests/unit/responses/mcp/test_aresponses_api_with_mcp.py @@ -2668,6 +2668,65 @@ async def test_stream_auto_execute_bills_the_paid_first_round_on_its_success_row assert _request_failure_row(logging_obj, recorder.logged[0]).get("response_cost", 0.0) == 0.0 +class _CutAfterStreamedText(httpx.AsyncByteStream): + async def __aiter__(self): + opened: Final = {**json.loads(json.dumps(_ROUND_TWO_BODY, default=dict)), "status": "in_progress"} + yield _sse( + ( + {"type": "response.created", "response": opened, "sequence_number": 0}, + { + "type": "response.output_text.delta", + "item_id": "msg_round_two", + "output_index": 0, + "content_index": 0, + "delta": "PING " * 40, + "sequence_number": 1, + }, + ) + ) + raise httpx.ReadError("connection reset mid-stream") + + +@pytest.mark.asyncio +async def test_stream_auto_execute_bills_a_follow_up_cut_after_streamed_text_with_no_second_fixed_margin( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +): + from litellm.proxy.utils import ProxyLogging + + monkeypatch.setattr(litellm, "cost_margin_config", {"openai": {"fixed_amount": _FIXED_MARGIN}}) + _serve_responses( + respx_mock, + monkeypatch, + (_streamed(_ROUND_ONE_BODY), httpx.Response(200, stream=_CutAfterStreamedText(), headers=SSE_HEADERS)), + ) + recorder: Final = _RoundRecorder() + logging_obj: Final = _rounds_logging_obj(recorder, "mcp-stream-follow-up-cut-after-text") + + with pytest.raises(httpx.ReadError): + await _stream_auto_execute(logging_obj) + ProxyLogging.fire_deferred_stream_logging({"litellm_logging_obj": logging_obj}) + await _settle(lambda: len(recorder.logged) > 0 and len(recorder.failed) > 0) + + _assert_one_row_bills(recorder.logged, _ROUND_ONE_BODY) + assert len(recorder.failed) == 1 + failed: Final = recorder.failed[0] + assert failed["completion_tokens"] > 0 + priced_with_its_own_margin: Final = litellm.completion_cost( + completion_response=ModelResponse( + model="gpt-5.6", + usage=Usage( + prompt_tokens=failed["prompt_tokens"], + completion_tokens=failed["completion_tokens"], + total_tokens=failed["prompt_tokens"] + failed["completion_tokens"], + ), + ), + model="gpt-5.6", + custom_llm_provider="openai", + ) + assert failed["response_cost"] == pytest.approx(priced_with_its_own_margin - _FIXED_MARGIN) + _assert_breakdown_agrees_with_the_billed_cost(failed) + + @pytest.mark.asyncio async def test_stream_auto_execute_blocked_by_a_post_call_guardrail_keeps_both_rounds_on_the_success_row( respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch diff --git a/tests/unit/responses/test_responses_streaming_iterator.py b/tests/unit/responses/test_responses_streaming_iterator.py index 37a58c32f38..ae688ff9173 100644 --- a/tests/unit/responses/test_responses_streaming_iterator.py +++ b/tests/unit/responses/test_responses_streaming_iterator.py @@ -9,20 +9,23 @@ crashed with exit 139 whenever any CustomLogger was registered). import asyncio import json import time -from collections.abc import Mapping -from typing import Final +from collections.abc import AsyncIterator, Mapping +from datetime import datetime +from typing import Final, cast import httpx import pytest import litellm +from litellm.exceptions import MidStreamFallbackError from litellm.integrations.custom_logger import CustomLogger from litellm.litellm_core_utils import thread_pool_executor as thread_pool_executor_module -from litellm.responses import streaming_iterator as responses_streaming_iterator_module from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig +from litellm.responses import streaming_iterator as responses_streaming_iterator_module from litellm.responses.streaming_iterator import ResponsesAPIStreamingIterator -from litellm.types.llms.openai import ResponsesAPIResponse +from litellm.types.llms.openai import ResponsesAPIResponse, ResponsesAPIStreamingResponse +from litellm.types.utils import StandardLoggingPayload class RecordingCustomLogger(CustomLogger): @@ -169,6 +172,240 @@ async def test_sync_callbacks_run_only_after_async_handler_completes(recording_e assert submit_times[0] >= recorder.async_hook_finished +class FailureRecorder(CustomLogger): + def __init__(self) -> None: + super().__init__() + self.failure_payloads: tuple[StandardLoggingPayload, ...] = () + self.success_payloads: tuple[StandardLoggingPayload, ...] = () + + async def async_log_failure_event( + self, kwargs: dict[str, object], response_obj: object, start_time: datetime, end_time: datetime + ) -> None: + self.failure_payloads = ( + *self.failure_payloads, + cast(StandardLoggingPayload, kwargs["standard_logging_object"]), + ) + + async def async_log_success_event( + self, kwargs: dict[str, object], response_obj: object, start_time: datetime, end_time: datetime + ) -> None: + self.success_payloads = ( + *self.success_payloads, + cast(StandardLoggingPayload, kwargs["standard_logging_object"]), + ) + + +class _ChunkRecorder: + def __init__(self) -> None: + self.chunks: tuple[ResponsesAPIStreamingResponse, ...] = () + + async def drain(self, iterator: ResponsesAPIStreamingIterator) -> None: + async for chunk in iterator: + self.chunks = (*self.chunks, chunk) + + +_SEA_DELTAS = ("The sea ", "is wide ", "and deep.") +_SEA_PROMPT = "Write a 300 word story about the sea." +_UNREACHABLE_IMAGE_INPUT: list[dict[str, object]] = [ + { + "role": "user", + "content": [ + {"type": "input_text", "text": "Describe this picture."}, + {"type": "input_image", "image_url": "http://images.invalid/photo.png", "detail": "high"}, + ], + } +] + + +_TOOL_ARGUMENT_DELTAS = ('{"city": ', '"Lisbon", ', '"unit": "celsius"}') + + +class _UpstreamThatTimesOutAfterThreeDeltas: + headers: dict[str, str] = {} + + def __init__(self, event_type: str = "response.output_text.delta", deltas: tuple[str, ...] = _SEA_DELTAS) -> None: + self.event_type = event_type + self.deltas = deltas + + async def aiter_bytes(self) -> AsyncIterator[bytes]: + for sequence_number, delta in enumerate(self.deltas): + event = { + "type": self.event_type, + "item_id": "item_1", + "output_index": 0, + "content_index": 0, + "delta": delta, + "sequence_number": sequence_number, + } + yield f"data: {json.dumps(event)}\n\n".encode() + raise httpx.ReadTimeout("Timeout on reading data from socket") + + +class _UpstreamThatFailsWithUsageAfterThreeDeltas: + headers: dict[str, str] = {} + + async def aiter_bytes(self) -> AsyncIterator[bytes]: + for sequence_number, delta in enumerate(_SEA_DELTAS): + event = { + "type": "response.output_text.delta", + "item_id": "msg_1", + "output_index": 0, + "content_index": 0, + "delta": delta, + "sequence_number": sequence_number, + } + yield f"data: {json.dumps(event)}\n\n".encode() + failed = { + "type": "response.failed", + "sequence_number": len(_SEA_DELTAS), + "response": { + "id": "resp_1", + "object": "response", + "created_at": 1700000000, + "status": "failed", + "model": "gpt-5.4-nano", + "output": [], + "error": {"code": "server_error", "message": "The model had an internal error"}, + "usage": {"input_tokens": 400, "output_tokens": 9, "total_tokens": 409}, + }, + } + yield f"data: {json.dumps(failed)}\n\n".encode() + + +async def _failure_payload_after_mid_stream_timeout( + monkeypatch: pytest.MonkeyPatch, + request_input: str | list[dict[str, object]], + upstream: _UpstreamThatTimesOutAfterThreeDeltas | None = None, + background_fields: Mapping[str, object] | None = None, +) -> tuple[FailureRecorder, StandardLoggingPayload]: + upstream_response: Final = upstream or _UpstreamThatTimesOutAfterThreeDeltas() + recorder = FailureRecorder() + monkeypatch.setattr(litellm, "failure_callback", [recorder]) + monkeypatch.setattr(litellm, "_async_failure_callback", [recorder]) + monkeypatch.setattr(litellm, "success_callback", [recorder]) + monkeypatch.setattr(litellm, "_async_success_callback", [recorder]) + logging_obj = _make_logging_obj() + logging_obj.update_environment_variables( + model="gpt-5.4-nano", + optional_params={"stream": True}, + litellm_params={"custom_llm_provider": "openai", "aresponses": True}, + ) + iterator = ResponsesAPIStreamingIterator( + response=cast(httpx.Response, upstream_response), + model="gpt-5.4-nano", + responses_api_provider_config=OpenAIResponsesAPIConfig(), + logging_obj=logging_obj, + custom_llm_provider="openai", + request_data={"model": "gpt-5.4-nano", "input": request_input, "stream": True, **(background_fields or {})}, + ) + received = _ChunkRecorder() + + with pytest.raises(litellm.Timeout, match="Timeout on reading data from socket"): + await received.drain(iterator) + assert [getattr(chunk, "delta", None) for chunk in received.chunks] == list(upstream_response.deltas) + + await iterator._await_pending_logging() + (payload,) = recorder.failure_payloads + return recorder, payload + + +@pytest.mark.asyncio +async def test_mid_stream_read_timeout_logs_failure_with_partial_usage_and_no_success( + monkeypatch: pytest.MonkeyPatch, +) -> None: + recorder, payload = await _failure_payload_after_mid_stream_timeout(monkeypatch, _SEA_PROMPT) + + assert payload["status"] == "failure" + assert payload["prompt_tokens"] == litellm.token_counter( + model="gpt-5.4-nano", messages=[{"role": "user", "content": _SEA_PROMPT}] + ) + assert payload["completion_tokens"] == litellm.token_counter( + model="gpt-5.4-nano", text="".join(_SEA_DELTAS), count_response_tokens=True + ) + assert payload["response_cost"] > 0 + assert recorder.success_payloads == () + + +@pytest.mark.asyncio +async def test_mid_stream_read_timeout_bills_the_tool_call_arguments_already_streamed( + monkeypatch: pytest.MonkeyPatch, +) -> None: + upstream: Final = _UpstreamThatTimesOutAfterThreeDeltas( + event_type="response.function_call_arguments.delta", deltas=_TOOL_ARGUMENT_DELTAS + ) + _, payload = await _failure_payload_after_mid_stream_timeout(monkeypatch, _SEA_PROMPT, upstream) + + assert payload["status"] == "failure" + assert payload["completion_tokens"] == litellm.token_counter( + model="gpt-5.4-nano", text="".join(_TOOL_ARGUMENT_DELTAS), count_response_tokens=True + ) + assert payload["response_cost"] > 0 + + +@pytest.mark.asyncio +async def test_mid_stream_failure_of_a_background_response_bills_nothing_since_retrieving_it_bills_the_whole_run( + monkeypatch: pytest.MonkeyPatch, +) -> None: + _, payload = await _failure_payload_after_mid_stream_timeout( + monkeypatch, _SEA_PROMPT, background_fields={"background": True} + ) + + assert payload["status"] == "failure" + assert payload["completion_tokens"] == 0 + assert payload["response_cost"] == 0 + + +@pytest.mark.asyncio +async def test_mid_stream_failure_usage_estimate_never_fetches_image_dimensions( + monkeypatch: pytest.MonkeyPatch, +) -> None: + _, payload = await _failure_payload_after_mid_stream_timeout(monkeypatch, _UNREACHABLE_IMAGE_INPUT) + + assert payload["status"] == "failure" + assert payload["prompt_tokens"] == litellm.token_counter( + model="gpt-5.4-nano", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "Describe this picture."}, + {"type": "image_url", "image_url": {"url": "http://images.invalid/photo.png", "detail": "high"}}, + ], + } + ], + use_default_image_token_count=True, + ) + assert payload["completion_tokens"] > 0 + + +@pytest.mark.asyncio +async def test_mid_stream_failure_keeps_the_usage_the_provider_reported(monkeypatch: pytest.MonkeyPatch) -> None: + recorder = FailureRecorder() + monkeypatch.setattr(litellm, "failure_callback", [recorder]) + monkeypatch.setattr(litellm, "_async_failure_callback", [recorder]) + logging_obj = _make_logging_obj() + logging_obj.update_environment_variables( + model="gpt-5.4-nano", + optional_params={"stream": True}, + litellm_params={"custom_llm_provider": "openai", "aresponses": True}, + ) + iterator = ResponsesAPIStreamingIterator( + response=cast(httpx.Response, _UpstreamThatFailsWithUsageAfterThreeDeltas()), + model="gpt-5.4-nano", + responses_api_provider_config=OpenAIResponsesAPIConfig(), + logging_obj=logging_obj, + custom_llm_provider="openai", + request_data={"model": "gpt-5.4-nano", "input": _SEA_PROMPT, "stream": True}, + ) + + with pytest.raises(MidStreamFallbackError, match="The model had an internal error"): + await _ChunkRecorder().drain(iterator) + await iterator._await_pending_logging() + + (payload,) = recorder.failure_payloads + assert (payload["prompt_tokens"], payload["completion_tokens"]) == (400, 9) + + _CREATED_RESPONSE_WITH_PRIORITY_TIER: Final[Mapping[str, object]] = { "id": "resp_partial", "object": "response", @@ -211,17 +448,135 @@ def _completed_after(text: str) -> Mapping[str, object]: } -def _live_iterator(logging_obj: LitellmLogging) -> ResponsesAPIStreamingIterator: +def _live_iterator( + logging_obj: LitellmLogging, upstream: object = None, background: bool = False +) -> ResponsesAPIStreamingIterator: return ResponsesAPIStreamingIterator( - response=httpx.Response(200), + response=cast(httpx.Response, upstream or httpx.Response(200)), model="gpt-5.4-nano", responses_api_provider_config=OpenAIResponsesAPIConfig(), logging_obj=logging_obj, - request_data={"model": "gpt-5.4-nano", "input": "tell me a story"}, + request_data={"model": "gpt-5.4-nano", "input": "tell me a story", "background": background}, custom_llm_provider="openai", ) +class _UpstreamThatTimesOutAfterRelayingText: + headers: dict[str, str] = {} + + async def aiter_bytes(self) -> AsyncIterator[bytes]: + for event in (_CREATED_WITH_PRIORITY_TIER, _text_delta("Once upon a time", 1)): + yield f"data: {json.dumps(event)}\n\n".encode() + raise httpx.ReadTimeout("Timeout on reading data from socket") + + +class _UpstreamThatCompletesThenTimesOut: + headers: dict[str, str] = {} + + async def aiter_bytes(self) -> AsyncIterator[bytes]: + events: Final = ( + _CREATED_WITH_PRIORITY_TIER, + *(_text_delta(delta, sequence_number) for sequence_number, delta in enumerate(_SEA_DELTAS, start=1)), + _completed_after("".join(_SEA_DELTAS)), + ) + for event in events: + yield f"data: {json.dumps(event)}\n\n".encode() + raise httpx.ReadTimeout("Timeout on reading data from socket") + + +@pytest.mark.asyncio +async def test_timeout_after_completed_response_ends_the_stream_and_bills_it_once( + monkeypatch: pytest.MonkeyPatch, +) -> None: + recorder: Final = FailureRecorder() + monkeypatch.setattr(litellm, "failure_callback", [recorder]) + monkeypatch.setattr(litellm, "_async_failure_callback", [recorder]) + monkeypatch.setattr(litellm, "success_callback", [recorder]) + monkeypatch.setattr(litellm, "_async_success_callback", [recorder]) + logging_obj: Final = _make_logging_obj() + logging_obj.update_environment_variables( + model="gpt-5.4-nano", + optional_params={"stream": True}, + litellm_params={"custom_llm_provider": "openai", "aresponses": True}, + ) + iterator: Final = _live_iterator(logging_obj, upstream=_UpstreamThatCompletesThenTimesOut()) + + chunks: Final = _ChunkRecorder() + await chunks.drain(iterator) + await iterator._await_pending_logging() + + assert chunks.chunks[-1].type == "response.completed" + assert len(recorder.success_payloads) == 1 + success_payload: Final = recorder.success_payloads[0] + response_payload: Final = success_payload["response"] + assert isinstance(response_payload, dict) + usage: Final = cast(dict[str, object], response_payload["usage"]) + assert (usage["prompt_tokens"], usage["completion_tokens"], usage["total_tokens"]) == (5, 7, 12) + assert recorder.failure_payloads == () + assert logging_obj.model_call_details.get("combined_usage_object") is None + + +class _UpstreamThatTimesOutAfterCreated: + headers: dict[str, str] = {} + + def __init__(self, **created_fields: object) -> None: + self.created_fields = created_fields + + async def aiter_bytes(self) -> AsyncIterator[bytes]: + created: Final = { + **_CREATED_WITH_PRIORITY_TIER, + "response": {**_CREATED_RESPONSE_WITH_PRIORITY_TIER, **self.created_fields}, + } + for event in (created, _text_delta("Once upon a time", 1)): + yield f"data: {json.dumps(event)}\n\n".encode() + raise httpx.ReadTimeout("Timeout on reading data from socket") + + +async def _mid_stream_failure_payload( + monkeypatch: pytest.MonkeyPatch, requested_tier: str | None, **created_fields: object +) -> StandardLoggingPayload: + recorder: Final = FailureRecorder() + monkeypatch.setattr(litellm, "failure_callback", [recorder]) + monkeypatch.setattr(litellm, "_async_failure_callback", [recorder]) + logging_obj: Final = _make_logging_obj() + logging_obj.update_environment_variables( + model="gpt-5.4-nano", + optional_params={"stream": True} + if requested_tier is None + else {"stream": True, "service_tier": requested_tier}, + litellm_params={"custom_llm_provider": "openai", "aresponses": True}, + ) + iterator: Final = _live_iterator(logging_obj, upstream=_UpstreamThatTimesOutAfterCreated(**created_fields)) + + with pytest.raises(litellm.Timeout): + await _ChunkRecorder().drain(iterator) + await iterator._await_pending_logging() + + (payload,) = recorder.failure_payloads + return payload + + +@pytest.mark.asyncio +async def test_mid_stream_failure_bills_the_relayed_text_at_the_tier_the_provider_served( + monkeypatch: pytest.MonkeyPatch, +) -> None: + flex_requested: Final = await _mid_stream_failure_payload(monkeypatch, "flex", service_tier="default") + nothing_requested: Final = await _mid_stream_failure_payload(monkeypatch, None, service_tier="default") + + assert nothing_requested["response_cost"] > 0 + assert flex_requested["response_cost"] == nothing_requested["response_cost"] + + +@pytest.mark.asyncio +async def test_mid_stream_failure_bills_nothing_when_the_provider_runs_the_response_in_background( + monkeypatch: pytest.MonkeyPatch, +) -> None: + payload: Final = await _mid_stream_failure_payload(monkeypatch, None, background=True) + + assert payload["status"] == "failure" + assert (payload["completion_tokens"], payload["response_cost"]) == (0, 0) + + def test_partial_response_on_disconnect_bills_the_relayed_text_at_the_served_service_tier(): iterator: Final = _live_iterator(_make_logging_obj()) for event in (_CREATED_WITH_PRIORITY_TIER, _text_delta("Once upon a time ", 1), _text_delta("a fox slept.", 2)): @@ -263,3 +618,25 @@ async def test_partial_response_on_disconnect_is_none_when_there_is_nothing_to_b assert iterator.partial_response_on_disconnect() is None await iterator._await_pending_logging() + + +@pytest.mark.asyncio +async def test_partial_response_on_disconnect_is_none_once_the_failure_log_billed_the_relayed_text(): + iterator: Final = _live_iterator(_make_logging_obj(), upstream=_UpstreamThatTimesOutAfterRelayingText()) + + with pytest.raises(litellm.Timeout): + await _ChunkRecorder().drain(iterator) + + assert iterator.partial_response_on_disconnect() is None + await iterator._await_pending_logging() + + +def test_partial_response_on_disconnect_reports_the_relayed_text_of_a_background_response(): + iterator: Final = _live_iterator(_make_logging_obj(), background=True) + for event in (_CREATED_WITH_PRIORITY_TIER, _text_delta("Once upon a time", 1)): + iterator._process_chunk(json.dumps(event)) + + incomplete_event: Final = iterator.partial_response_on_disconnect() + + assert incomplete_event is not None + assert incomplete_event.response.output[0].content[0].text == "Once upon a time" diff --git a/tests/unit/responses/test_streaming_iterator.py b/tests/unit/responses/test_streaming_iterator.py index 6c3e6177503..664f4dba163 100644 --- a/tests/unit/responses/test_streaming_iterator.py +++ b/tests/unit/responses/test_streaming_iterator.py @@ -210,9 +210,11 @@ _TRAILING_ERRORS = [ httpx.RemoteProtocolError("peer closed connection without sending complete message body"), ] +_TRAILING_ERRORS_AFTER_COMPLETED = [*_TRAILING_ERRORS, httpx.ReadTimeout("Timeout on reading data from socket")] + @pytest.mark.asyncio -@pytest.mark.parametrize("trailing_error", _TRAILING_ERRORS, ids=type) +@pytest.mark.parametrize("trailing_error", _TRAILING_ERRORS_AFTER_COMPLETED, ids=type) async def test_transport_error_after_completed_event_ends_stream_cleanly(trailing_error): """A sloppy connection close after `response.completed` must not turn a complete stream into an error (regression guard for the transport no longer @@ -243,7 +245,7 @@ async def test_transport_error_before_completed_event_raises(): pass -@pytest.mark.parametrize("trailing_error", _TRAILING_ERRORS, ids=type) +@pytest.mark.parametrize("trailing_error", _TRAILING_ERRORS_AFTER_COMPLETED, ids=type) def test_sync_transport_error_after_completed_event_ends_stream_cleanly(trailing_error): iterator = _make_sync_iterator( sse_events=_COMPLETE_STREAM_EVENTS, @@ -1205,7 +1207,9 @@ async def test_completed_event_survives_a_failing_usage_estimate(): falls back to usage None.""" malformed_input: Final = [{"type": "message", "role": "user", "content": 42}] with pytest.raises(ValueError, match="Invalid content type"): - _estimate_usage_from_text("gpt-4o-mini", malformed_input, {"input": malformed_input}, "hello world") + _estimate_usage_from_text( + "gpt-4o-mini", malformed_input, {"input": malformed_input}, "hello world" + ) response = _responses_api_response_without_usage() iterator = _make_iterator( diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index 566024b5d94..b4a3d207bd4 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -3774,6 +3774,132 @@ async def test_continued_stream_sharing_the_proxy_logging_object_bills_the_cut_t ) +def _responses_sse(*events: Mapping[str, object]) -> str: + return "".join(f"event: {event['type']}\ndata: {json.dumps(event)}\n\n" for event in events) + + +def _responses_envelope(response_id: str, model: str, status: str, **fields: object) -> dict[str, object]: + return { + "id": response_id, + "object": "response", + "created_at": 1700000000, + "status": status, + "model": model, + "output": [], + **fields, + } + + +@pytest.mark.asyncio +async def test_responses_fallback_sharing_the_proxy_logging_object_logs_the_fallback_usage( + monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + recorder: Final = _SpendRecorder() + monkeypatch.setattr(litellm, "callbacks", [recorder]) + respx_mock.post("https://primary.example/v1/responses").mock( + return_value=httpx.Response( + 200, + headers={"content-type": "text/event-stream"}, + text=_responses_sse( + {"type": "response.created", "response": _responses_envelope("resp_primary", "gpt-4o", "in_progress")}, + *( + { + "type": "response.output_text.delta", + "item_id": "msg_primary", + "output_index": 0, + "content_index": 0, + "delta": word, + } + for word in ("Once ", "upon ", "a ", "time ", "there") + ), + { + "type": "response.failed", + "response": _responses_envelope( + "resp_primary", + "gpt-4o", + "failed", + error={"code": "server_error", "message": "The model had an internal error"}, + ), + }, + ), + ) + ) + respx_mock.post("https://backup.example/v1/responses").mock( + return_value=httpx.Response( + 200, + headers={"content-type": "text/event-stream"}, + text=_responses_sse( + {"type": "response.created", "response": _responses_envelope("resp_backup", "gpt-4o-mini", "in_progress")}, + { + "type": "response.completed", + "response": _responses_envelope( + "resp_backup", + "gpt-4o-mini", + "completed", + output=[ + { + "type": "message", + "id": "msg_backup", + "status": "completed", + "role": "assistant", + "content": [{"type": "output_text", "text": "Once upon a time", "annotations": []}], + } + ], + usage={"input_tokens": 1000, "output_tokens": 700, "total_tokens": 1700}, + ), + }, + ), + ) + ) + router: Final = Router( + model_list=[ + { + "model_name": "primary", + "litellm_params": {"model": "openai/gpt-4o", "api_key": "k1", "api_base": "https://primary.example/v1"}, + }, + { + "model_name": "backup", + "litellm_params": { + "model": "openai/gpt-4o-mini", + "api_key": "k2", + "api_base": "https://backup.example/v1", + }, + }, + ], + fallbacks=[{"primary": ["backup"]}], + num_retries=0, + ) + shared_logging, _ = litellm.utils.function_setup( + original_function="aresponses", + rules_obj=litellm.utils.Rules(), + start_time=datetime.now(), + model="primary", + input="Tell me a story", + stream=True, + litellm_call_id="shared-call", + ) + + response: Final = await router.aresponses( + model="primary", + input="Tell me a story", + stream=True, + litellm_call_id="shared-call", + litellm_logging_obj=shared_logging, + ) + _ = [event async for event in response] + await asyncio.wait_for(recorder.succeeded.wait(), timeout=10) + + success: Final = recorder.successes[-1] + assert (success["prompt_tokens"], success["completion_tokens"]) == (1000, 700) + spend_row: Final = recorder.spend_rows[-1] + assert (spend_row["prompt_tokens"], spend_row["completion_tokens"], spend_row["spend"]) == ( + 1000, + 700, + success["response_cost"], + ) + + def test_carry_partial_spend_to_fallback_adds_the_partial_the_cut_stream_inherited(): from litellm.types.utils import Usage From 9007e0fa70ca38b6b1ad78f6fd082d03d32d886c Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 28 Oct 2026 13:42:22 -0700 Subject: [PATCH 3/7] fix(files)!: scope provider and model file lists to the caller's managed files under require_managed_files (#47682) * test(files): cover owner-scoped file listing under require_managed_files * fix(files): list only the caller's managed files on every list route under require_managed_files * fix(files): narrow owner-scoped managed file lists to the named provider or model and page with after_id and before_id * fix(files): narrow scoped lists the way uploads route, and refuse lists that cannot be scoped A model name now matches every deployment that serves it, wildcards and aliases included, and a provider matches deployments by the provider uploads resolve, so anthropic/* and unprefixed Claude models count. The OpenAI and Azure lists also return the caller's pass-through uploads. An unknown provider and different after and after_id cursors answer 400, and a list with no matching deployment skips the database read. * fix(files): match stored files by model name and keep flag-off lists as on main Under require_managed_files, a narrowed list now matches a file's model mappings by deployment id or model name, so files LiteLLM stores in its own database show on their model and provider lists. The Anthropic list defaults to 20 rows when no limit is sent, after_id, before_id and the cursor 400s apply only under the flag, and an admin keeps the provider list when the managed files hook is not loaded. List errors are returned as values and mapped to HTTP errors in one place. * fix(files): keep admin provider lists and index the owner-scoped walk Under require_managed_files, an admin keeps the provider list as on main, and only a non-admin list reads the caller's own and team managed files. A stored file now also shows on the provider its model name routes to through a wildcard, with exact model names first, and a list routed to a model takes the default page size of the provider serving it. A new index on created_by and created_at keeps each chunk of the walk an index read, and the model check runs before the file object is parsed. * fix(files): filter provider and model lists in the query A non-admin list scoped to a provider or model now asks Postgres only for the caller's rows that map to one of that scope's deployments or models, or that live in file storage, instead of reading every owned row and filtering them in Python. A caller with many files on other providers no longer waits for the whole set to be read. Team members' files are matched in the same query. * test(files): type the fake raw query in the files list tests * refactor(proxy): list managed files through one typed database handle * fix(files): list alias and routing-group storage uploads on their provider --- .../proxy/hooks/managed_files.py | 151 ++++- .../migration.sql | 2 + .../litellm_proxy_extras/schema.prisma | 1 + litellm/constants.py | 1 + litellm/llms/base_llm/files/transformation.py | 23 +- .../openai_files_endpoints/files_endpoints.py | 255 +++++++- .../managed_id_rewriter.py | 6 + litellm/proxy/schema.prisma | 1 + schema.prisma | 1 + .../security/test_managed_file_listing.py | 263 ++++++++ .../proxy/test_managed_files_hook.py | 17 + .../test_files_endpoint.py | 569 +++++++++++++++++- ui/litellm-dashboard/src/lib/http/schema.d.ts | 6 + 13 files changed, 1245 insertions(+), 51 deletions(-) create mode 100644 litellm-proxy-extras/litellm_proxy_extras/migrations/20261021000000_managed_file_created_by_created_at_index/migration.sql create mode 100644 tests/integration/security/test_managed_file_listing.py diff --git a/enterprise/litellm_enterprise/proxy/hooks/managed_files.py b/enterprise/litellm_enterprise/proxy/hooks/managed_files.py index 7a07350e2d9..416444664a3 100644 --- a/enterprise/litellm_enterprise/proxy/hooks/managed_files.py +++ b/enterprise/litellm_enterprise/proxy/hooks/managed_files.py @@ -28,7 +28,7 @@ import httpx from fastapi import HTTPException from openai import APIConnectionError, AsyncOpenAI from openai.types.file_deleted import FileDeleted -from pydantic import BaseModel, JsonValue, TypeAdapter, ValidationError +from pydantic import BaseModel, ConfigDict, JsonValue, TypeAdapter, ValidationError from typing_extensions import ReadOnly, Unpack import litellm @@ -53,7 +53,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import ( extract_file_metadata, ) from litellm.llms.base_llm.files.storage_backend_factory import get_storage_backend -from litellm.llms.base_llm.files.transformation import BaseFileEndpoints +from litellm.llms.base_llm.files.transformation import BaseFileEndpoints, FileListModelScope from litellm.llms.base_llm.managed_resources.isolation import ( build_list_page, build_owner_filter, @@ -309,6 +309,7 @@ def _is_xai_batch_output( class _ManagedFileRow(Protocol): unified_file_id: str file_object: OpenAIFileObject + model_mappings: Mapping[str, str] flat_model_file_ids: Sequence[str] storage_backend: Optional[str] storage_url: Optional[str] @@ -362,6 +363,8 @@ class _ManagedResourceDatabase(Protocol): @property def litellm_managedfiletable(self) -> _ManagedFileTableActions: ... + async def query_raw(self, query: str, *args: object) -> object: ... + @property def litellm_managedobjecttable(self) -> _ManagedObjectTableActions: ... @@ -385,10 +388,83 @@ class _RouterFileCallKwargs(TypedDict, total=False): custom_llm_provider: ReadOnly[str | None] +def _managed_database(prisma_client: _ManagedResourcePrismaClient) -> _ManagedResourceDatabase: + return prisma_client.db + + def _managed_file_table(prisma_client: _ManagedResourcePrismaClient) -> _ManagedFileTableActions: return prisma_client.db.litellm_managedfiletable +class _ListedFileRow(Protocol): + @property + def unified_file_id(self) -> str: ... + + @property + def file_object(self) -> object: ... + + @property + def model_mappings(self) -> Mapping[str, object]: ... + + @property + def storage_backend(self) -> str | None: ... + + +class _ScopedFileRow(BaseModel): + model_config = ConfigDict(frozen=True) + + unified_file_id: str + file_object: JsonValue = None + model_mappings: Mapping[str, JsonValue] + storage_backend: str | None = None + + +_SCOPED_FILE_ROWS: Final = TypeAdapter(tuple[_ScopedFileRow, ...]) + + +def _owner_condition(owner_filter: Mapping[str, object]) -> tuple[str, tuple[str, ...]]: + match owner_filter: + case {"OR": [{"created_by": str() as user_id}, {"team_id": str() as team_id}]}: + return '("created_by" = $1 OR "team_id" = $2)', (user_id, team_id) + case {"created_by": str() as user_id}: + return '"created_by" = $1', (user_id,) + case {"team_id": str() as team_id}: + return '"team_id" = $1', (team_id,) + case _: + return ("FALSE", ()) if owner_filter else ("TRUE", ()) + + +async def _scoped_file_rows( + database: _ManagedResourceDatabase, + owner_filter: Mapping[str, object], + model_mapping_keys: frozenset[str], + cursor_id: str | None, + direction: Literal["asc", "desc"], + take: int, +) -> tuple[_ScopedFileRow, ...]: + owner_condition, owner_params = _owner_condition(owner_filter) + keys_param: Final = len(owner_params) + 1 + comparison: Final = "<" if direction == "desc" else ">" + cursor_condition: Final = ( + f'AND ("created_at", "unified_file_id") {comparison} (SELECT "created_at", "unified_file_id" ' + f'FROM "LiteLLM_ManagedFileTable" WHERE "unified_file_id" = ${keys_param + 1}) ' + if cursor_id + else "" + ) + order: Final = direction.upper() + rows: Final = await database.query_raw( + 'SELECT "unified_file_id", "file_object", "model_mappings", "storage_backend" FROM "LiteLLM_ManagedFileTable" ' + f'WHERE {owner_condition} AND ("model_mappings" ?| ${keys_param}::text[] OR "storage_backend" IS NOT NULL) ' + f'{cursor_condition}ORDER BY "created_at" {order}, "unified_file_id" {order} ' + f"LIMIT ${keys_param + (2 if cursor_id else 1)}", + *owner_params, + sorted(model_mapping_keys), + *((cursor_id,) if cursor_id else ()), + take, + ) + return _SCOPED_FILE_ROWS.validate_python(rows) + + def _iter_provider_file_id_pairs( rows: Sequence[_ManagedFileRow], requested_provider_file_ids: frozenset[str], @@ -2265,14 +2341,23 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints): user_api_key_dict: UserAPIKeyAuth, limit: Optional[int] = None, after: Optional[str] = None, + model_scope: Optional[FileListModelScope] = None, + before: Optional[str] = None, **data: Dict, ) -> FileListPage: """List the managed files the caller owns, newest first. Pagination is keyset based on ``unified_file_id`` so a key that owns every file on the proxy still reads one bounded page at a time. - ``purpose`` is applied after parsing, because the managed file table - keeps it inside the ``file_object`` blob instead of a column, and rows + ``after`` pages toward older files and ``before`` toward newer ones, + each page still ordered newest first. + ``model_scope`` is applied in the query, so only rows whose + ``model_mappings`` name one of its deployment ids or model names, and + storage-backed rows, leave the database; a storage-backed row then + stays only when its model name is one the router sends to the listed + provider. ``purpose`` is applied after parsing, because the managed + file table keeps it inside the ``file_object`` blob instead of a + column. Rows whose blob will not parse drop out there too, so a chunk of rows can yield fewer matches than the page holds. Successive chunks are read until the page is full or the caller's rows run out, which keeps @@ -2281,9 +2366,8 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints): scan has to continue past it, the chunk widens to ``FILE_LIST_CONTINUATION_CHUNK_SIZE``, so the walk costs one query per that many rows instead of one per page. That bound is per query, not - per request: the work is still linear in the rows the caller owns, and - a filter matching nothing reads every one of them, with no index - covering either the owner filter or the sort. + per request: a ``purpose`` filter matching nothing still reads every + row that passes the query. """ validate_file_list_limit(limit) validate_file_list_purpose(purpose) @@ -2292,36 +2376,60 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints): if owner_filter is None: return FileListPage.model_validate(build_list_page([])) - if after: - cursor_row = await _managed_file_table(self.prisma_client).find_first( - where={**owner_filter, "unified_file_id": after} + if after and before: + raise ProxyException( + message="Pass either 'after' (or 'after_id') or 'before_id' when listing files, not both.", + type="invalid_request_error", + param="before_id", + code=400, + openai_code="invalid_value", + ) + cursor_param: Final = "before_id" if before else "after" + cursor: Final = before or after + database: Final = _managed_database(self.prisma_client) + if cursor: + cursor_row = await database.litellm_managedfiletable.find_first( + where={**owner_filter, "unified_file_id": cursor} ) if cursor_row is None: raise ProxyException( - message=f"Invalid 'after' cursor: no file found with id '{after}'.", + message=f"Invalid '{cursor_param}' cursor: no file found with id '{cursor}'.", type="invalid_request_error", - param="after", + param=cursor_param, code=400, openai_code="invalid_value", ) + if model_scope is not None and not model_scope.model_mapping_keys: + return FileListPage.model_validate(build_list_page([])) + page_size: Final = min(limit or MAX_FILE_LIST_LIMIT, MAX_FILE_LIST_LIMIT) matches: Final[List[OpenAIFileObject]] = [] - cursor_id = after + cursor_id = cursor chunk_size = page_size + 1 + direction: Final[Literal["asc", "desc"]] = "asc" if before else "desc" while len(matches) <= page_size: cursor_args: _CursorPageArgs = {"cursor": {"unified_file_id": cursor_id}, "skip": 1} if cursor_id else {} - chunk = await _managed_file_table(self.prisma_client).find_many( - where=owner_filter, - take=chunk_size, - order=[{"created_at": "desc"}, {"unified_file_id": "desc"}], - **cursor_args, + chunk: Sequence[_ListedFileRow] = ( + await _scoped_file_rows( + database, owner_filter, model_scope.model_mapping_keys, cursor_id, direction, chunk_size + ) + if model_scope is not None + else await database.litellm_managedfiletable.find_many( + where=owner_filter, + take=chunk_size, + order=[{"created_at": direction}, {"unified_file_id": direction}], + **cursor_args, + ) ) matches.extend( _public_file_object(parsed_file_object, row.unified_file_id) for row in chunk - if (parsed_file_object := _parse_managed_file_object(row.file_object, row.unified_file_id)) is not None + if ( + model_scope is None or model_scope.holds(row.model_mappings, stored=row.storage_backend is not None) + ) + and (parsed_file_object := _parse_managed_file_object(row.file_object, row.unified_file_id)) is not None and (purpose is None or parsed_file_object.purpose == purpose) ) if len(chunk) < chunk_size: @@ -2329,7 +2437,10 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints): cursor_id = chunk[-1].unified_file_id chunk_size = max(chunk_size, FILE_LIST_CONTINUATION_CHUNK_SIZE) - return FileListPage.model_validate(build_list_page(matches[:page_size], has_more=len(matches) > page_size)) + page: Final = matches[:page_size] + return FileListPage.model_validate( + build_list_page(page[::-1] if before else page, has_more=len(matches) > page_size) + ) def _is_batch_polling_enabled(self) -> bool: """ diff --git a/litellm-proxy-extras/litellm_proxy_extras/migrations/20261021000000_managed_file_created_by_created_at_index/migration.sql b/litellm-proxy-extras/litellm_proxy_extras/migrations/20261021000000_managed_file_created_by_created_at_index/migration.sql new file mode 100644 index 00000000000..8bbcb67bddf --- /dev/null +++ b/litellm-proxy-extras/litellm_proxy_extras/migrations/20261021000000_managed_file_created_by_created_at_index/migration.sql @@ -0,0 +1,2 @@ +CREATE INDEX CONCURRENTLY IF NOT EXISTS "LiteLLM_ManagedFileTable_created_by_created_at_idx" +ON "LiteLLM_ManagedFileTable" ("created_by", "created_at" DESC); diff --git a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma index 83a0e6f9e05..4187f76c97c 100644 --- a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma +++ b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma @@ -1163,6 +1163,7 @@ model LiteLLM_ManagedFileTable { @@index([unified_file_id]) @@index([flat_model_file_ids], type: Gin) @@index([team_id, created_at(sort: Desc)]) + @@index([created_by, created_at(sort: Desc)]) } model LiteLLM_ManagedObjectTable { // for batches or finetuning jobs which use the diff --git a/litellm/constants.py b/litellm/constants.py index 2b083b03766..1d5c6404bfd 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -129,6 +129,7 @@ MAX_S3_OBJECT_DOWNLOAD_FILENAME_BYTES: Final = 1024 S3_LOG_PROMPTS_ONLY_ENV_VAR: Final = "S3_LOG_PROMPTS_ONLY" S3_PARTITION_GRANULARITY_ENV_VAR: Final = "S3_PARTITION_GRANULARITY" MAX_FILE_LIST_LIMIT: Final = 10000 +ANTHROPIC_DEFAULT_FILE_LIST_LIMIT: Final = 20 DEFAULT_SQS_FLUSH_INTERVAL_SECONDS: Final = get_env_int("DEFAULT_SQS_FLUSH_INTERVAL_SECONDS", 10) DEFAULT_NUM_WORKERS_LITELLM_PROXY: Final = get_env_int("DEFAULT_NUM_WORKERS_LITELLM_PROXY", 1) GRACEFUL_SHUTDOWN_TEARDOWN_HEADROOM_SECONDS: Final = 10 diff --git a/litellm/llms/base_llm/files/transformation.py b/litellm/llms/base_llm/files/transformation.py index 4346e214dc8..491fb1919b6 100644 --- a/litellm/llms/base_llm/files/transformation.py +++ b/litellm/llms/base_llm/files/transformation.py @@ -1,6 +1,7 @@ from abc import ABC, abstractmethod -from collections.abc import AsyncGenerator, Iterator, Mapping -from typing import TYPE_CHECKING, Any, Union +from collections.abc import AsyncGenerator, Callable, Iterable, Iterator, Mapping +from dataclasses import dataclass +from typing import TYPE_CHECKING, Any, Final, Union import httpx from openai.types.file_deleted import FileDeleted @@ -38,6 +39,22 @@ else: UserAPIKeyAuth = Any +def _routes_nowhere(_model_name: str) -> bool: + return False + + +@dataclass(frozen=True, slots=True) +class FileListModelScope: + model_mapping_keys: frozenset[str] + routes_stored_model_name: Callable[[str], bool] = _routes_nowhere + + def holds(self, model_mapping_keys: Iterable[str], stored: bool) -> bool: + mapped_keys: Final = frozenset(model_mapping_keys) + return not self.model_mapping_keys.isdisjoint(mapped_keys) or ( + stored and any(map(self.routes_stored_model_name, mapped_keys)) + ) + + class BaseFileUploadStream(ABC): """Re-iterable request body that yields an upload's bytes lazily. @@ -274,6 +291,8 @@ class BaseFileEndpoints(ABC): user_api_key_dict: UserAPIKeyAuth, limit: int | None = None, after: str | None = None, + model_scope: FileListModelScope | None = None, + before: str | None = None, **data: dict, ) -> FileListPage: pass diff --git a/litellm/proxy/openai_files_endpoints/files_endpoints.py b/litellm/proxy/openai_files_endpoints/files_endpoints.py index d2ceefffe47..5d63c3db222 100644 --- a/litellm/proxy/openai_files_endpoints/files_endpoints.py +++ b/litellm/proxy/openai_files_endpoints/files_endpoints.py @@ -6,9 +6,12 @@ ###################################################################### import asyncio +import functools +import itertools import traceback -from collections.abc import Mapping, Sequence -from typing import Any, BinaryIO, Final, TypedDict, cast, get_args +from collections.abc import Iterable, Iterator, Mapping, Sequence +from dataclasses import dataclass +from typing import Any, BinaryIO, Final, NoReturn, TypedDict, cast, get_args import httpx from fastapi import ( @@ -28,12 +31,13 @@ from typing_extensions import ReadOnly import litellm from litellm import CreateFileRequest, get_secret_str from litellm._logging import verbose_proxy_logger +from litellm.constants import ANTHROPIC_DEFAULT_FILE_LIST_LIMIT from litellm.litellm_core_utils.cloud_storage_security import ( is_managed_cloud_storage_uri, ) from litellm.litellm_core_utils.core_helpers import get_or_create_metadata_bucket from litellm.llms.base_llm.files.litellm_db_storage_backend import LITELLM_DB_STORAGE_BACKEND_NAME -from litellm.llms.base_llm.files.transformation import BaseFileEndpoints +from litellm.llms.base_llm.files.transformation import BaseFileEndpoints, FileListModelScope from litellm.llms.base_llm.managed_resources.isolation import build_list_page from litellm.proxy._types import * from litellm.proxy.auth.user_api_key_auth import user_api_key_auth @@ -112,6 +116,9 @@ from litellm.types.llms.openai import ( OpenAIFileObject, OpenAIFilesPurpose, ) +from litellm.types.proxy.auth.user_api_key_auth import user_api_key_has_admin_view +from litellm.types.router import Deployment +from litellm.types.utils import LlmProviders router: Final = APIRouter() @@ -1673,6 +1680,185 @@ def _as_file_list_page(response: object) -> object: return FileListPage.model_validate(page) +async def _list_owned_managed_files( + managed_files_obj: BaseFileEndpoints, + user_api_key_dict: UserAPIKeyAuth, + purpose: str | None, + limit: int | None, + after: str | None, + before: str | None, + model_scope: FileListModelScope | None, +) -> FileListPage: + return await managed_files_obj.afile_list( + purpose=purpose, + litellm_parent_otel_span=user_api_key_dict.parent_otel_span, + user_api_key_dict=user_api_key_dict, + limit=limit, + after=after, + before=before, + model_scope=model_scope, + ) + + +@dataclass(frozen=True, slots=True) +class _UnscopedFileList: + status_code: int + detail: str + + +@dataclass(frozen=True, slots=True) +class _ListedFiles: + model_scope: FileListModelScope | None + custom_llm_provider: str | None + + +def _raise_unscoped_file_list(failure: _UnscopedFileList) -> NoReturn: + raise HTTPException(status_code=failure.status_code, detail=failure.detail) + + +def _single_target_model_name(target_model_names: str) -> str | _UnscopedFileList: + target_model_names_list: Final = target_model_names.split(",") + if len(target_model_names_list) != 1: + return _UnscopedFileList( + status_code=400, + detail="target_model_names on list files must be a list of one model name. Example: ['gpt-4o']", + ) + return target_model_names_list[0] + + +async def _requested_custom_llm_provider(request: Request, provider: str | None) -> str | None: + return ( + provider + or get_custom_llm_provider_from_request_headers(request=request) + or get_custom_llm_provider_from_request_query(request=request) + or await get_custom_llm_provider_from_request_body(request=request) + ) + + +def _default_file_list_limit(custom_llm_provider: str | None) -> int | None: + return ANTHROPIC_DEFAULT_FILE_LIST_LIMIT if custom_llm_provider == LlmProviders.ANTHROPIC.value else None + + +def _deployment_model_mapping_keys(deployment: Deployment) -> frozenset[str]: + return frozenset(key for key in (deployment.model_name, deployment.model_info.id) if key is not None) + + +def _model_mapping_keys(deployments: Iterable[Deployment]) -> frozenset[str]: + return frozenset(itertools.chain.from_iterable(map(_deployment_model_mapping_keys, deployments))) + + +def _deployments_serving_model(llm_router: Router, model: str) -> tuple[Deployment, ...]: + deployment: Final = llm_router.get_deployment(model_id=model) if llm_router.has_model_id(model) else None + if deployment is not None: + return (deployment,) + return tuple(Deployment.model_validate(listed) for listed in llm_router.get_model_list(model_name=model) or ()) + + +def _resolved_deployment_provider(deployment: Deployment) -> str | None: + try: + return litellm.get_llm_provider( + model=deployment.litellm_params.model, + custom_llm_provider=deployment.litellm_params.custom_llm_provider, + )[1] + except litellm.BadRequestError: + return None + + +def _deployments_on_provider(llm_router: Router | None, custom_llm_provider: str) -> Iterator[Deployment]: + if llm_router is None: + return iter(()) + return ( + deployment + for deployment_id in llm_router.get_model_ids() + if (deployment := llm_router.get_deployment(model_id=deployment_id)) is not None + and _resolved_deployment_provider(deployment) == custom_llm_provider + ) + + +def _routed_model_name_on_provider( + llm_router: Router, custom_llm_provider: str, provider_model_names: frozenset[str], model_name: str +) -> bool: + if model_name in llm_router.model_names: + return model_name in provider_model_names + return llm_router.pattern_router.routes_to_provider(model_name, custom_llm_provider) + + +def _stored_model_name_routes_to_provider( + llm_router: Router, custom_llm_provider: str, provider_model_names: frozenset[str], model_name: str +) -> bool: + routing_group: Final = llm_router.get_routing_group(model_name) + routed_model_names: Final = ( + tuple(routing_group.models) if routing_group is not None else (llm_router.routable_model_group(model_name),) + ) + return any( + _routed_model_name_on_provider(llm_router, custom_llm_provider, provider_model_names, routed_model_name) + for routed_model_name in routed_model_names + ) + + +def _files_on_provider(llm_router: Router | None, custom_llm_provider: str) -> _ListedFiles | _UnscopedFileList: + from litellm.proxy.pass_through_endpoints.managed_id_rewriter import passthrough_file_model_id + + if custom_llm_provider not in litellm.provider_list: + return _UnscopedFileList(status_code=400, detail=f"Unknown provider '{custom_llm_provider}' for listing files.") + passthrough_model_id: Final = passthrough_file_model_id(custom_llm_provider) + deployments: Final = tuple(_deployments_on_provider(llm_router, custom_llm_provider)) + model_mapping_keys: Final = _model_mapping_keys(deployments) | frozenset( + (passthrough_model_id,) if passthrough_model_id is not None else () + ) + if llm_router is None: + return _ListedFiles(FileListModelScope(model_mapping_keys), custom_llm_provider) + return _ListedFiles( + FileListModelScope( + model_mapping_keys, + routes_stored_model_name=functools.partial( + _stored_model_name_routes_to_provider, + llm_router, + custom_llm_provider, + frozenset(deployment.model_name for deployment in deployments), + ), + ), + custom_llm_provider, + ) + + +def _files_serving_model(llm_router: Router, model: str) -> _ListedFiles: + deployments: Final = _deployments_serving_model(llm_router, model) + providers: Final = frozenset(map(_resolved_deployment_provider, deployments)) + return _ListedFiles( + FileListModelScope(_model_mapping_keys(deployments)), next(iter(providers)) if len(providers) == 1 else None + ) + + +async def _listed_files( + llm_router: Router | None, + user_api_key_dict: UserAPIKeyAuth, + custom_llm_provider: str | None, + target_model_names: str | None, + routed_model: str | None, +) -> _ListedFiles | _UnscopedFileList: + if routed_model is not None and llm_router is not None: + return _files_serving_model(llm_router, routed_model) + if target_model_names: + target_model_name: Final = _single_target_model_name(target_model_names) + if isinstance(target_model_name, _UnscopedFileList): + return target_model_name + if llm_router is None: + return _UnscopedFileList( + status_code=500, detail="LLM Router not initialized. Ensure models added to proxy." + ) + await get_authorized_credentials_for_model( + llm_router=llm_router, + model_id=target_model_name, + user_api_key_dict=user_api_key_dict, + operation_context="file list", + ) + return _files_serving_model(llm_router, target_model_name) + if custom_llm_provider is None: + return _ListedFiles(None, None) + return _files_on_provider(llm_router, custom_llm_provider) + + @router.get( "/{provider}/v1/files", dependencies=[Depends(user_api_key_auth)], @@ -1697,6 +1883,8 @@ async def list_files( purpose: str | None = None, limit: int | None = None, after: str | None = None, + after_id: str | None = None, + before_id: str | None = None, ): """ Returns information about a specific file. that can be used across - Assistants API, Batch API @@ -1739,6 +1927,10 @@ async def list_files( ) response: Any | None = None + owned_files_hook: Final = proxy_logging_obj.get_proxy_hook("managed_files") + list_owned_files_only: Final = litellm.require_managed_files is True and not user_api_key_has_admin_view( + user_api_key_dict + ) # Check for model-based credential routing (no file_id encoding check for list) should_route, model_used, _, credentials = await handle_model_based_routing( @@ -1750,7 +1942,36 @@ async def list_files( check_file_id_encoding=False, ) - if should_route and credentials is not None: + if list_owned_files_only: + if after and after_id and after != after_id: + raise HTTPException( + status_code=400, + detail="Pass the list cursor as either 'after' or 'after_id', not both with different ids.", + ) + if not isinstance(owned_files_hook, BaseFileEndpoints): + raise HTTPException( + status_code=500, + detail="Managed file listing is unavailable, so require_managed_files cannot scope the list to the caller.", + ) + listed_files: Final = await _listed_files( + llm_router=llm_router, + user_api_key_dict=user_api_key_dict, + custom_llm_provider=await _requested_custom_llm_provider(request=request, provider=provider), + target_model_names=target_model_names, + routed_model=model_used if should_route else None, + ) + if isinstance(listed_files, _UnscopedFileList): + _raise_unscoped_file_list(listed_files) + response = await _list_owned_managed_files( # rebind-ok: each dispatch branch binds response + managed_files_obj=owned_files_hook, + user_api_key_dict=user_api_key_dict, + purpose=purpose, + limit=limit if limit is not None else _default_file_list_limit(listed_files.custom_llm_provider), + after=after or after_id, + before=before_id, + model_scope=listed_files.model_scope, + ) + elif should_route and credentials is not None: # Use model-based routing with credentials from config prepare_data_with_credentials(data=data, credentials=credentials, include_internal_credentials=True) response = await litellm.afile_list( @@ -1762,12 +1983,9 @@ async def list_files( verbose_proxy_logger.debug("Listed files using model: %s", model_used) elif target_model_names and isinstance(target_model_names, str): - target_model_names_list: Final = target_model_names.split(",") - if len(target_model_names_list) != 1: - raise HTTPException( - status_code=400, - detail="target_model_names on list files must be a list of one model name. Example: ['gpt-4o']", - ) + target_model_name: Final = _single_target_model_name(target_model_names) + if isinstance(target_model_name, _UnscopedFileList): + _raise_unscoped_file_list(target_model_name) if llm_router is None: raise HTTPException( status_code=500, @@ -1775,7 +1993,7 @@ async def list_files( ) credentials = await get_authorized_credentials_for_model( llm_router=llm_router, - model_id=target_model_names_list[0], + model_id=target_model_name, user_api_key_dict=user_api_key_dict, operation_context="file list", ) @@ -1786,20 +2004,17 @@ async def list_files( **data, ) else: - custom_llm_provider: Final = ( - provider - or get_custom_llm_provider_from_request_headers(request=request) - or get_custom_llm_provider_from_request_query(request=request) - or await get_custom_llm_provider_from_request_body(request=request) - ) + custom_llm_provider: Final = await _requested_custom_llm_provider(request=request, provider=provider) managed_files_obj: Final = proxy_logging_obj.get_proxy_hook("managed_files") if custom_llm_provider is None and isinstance(managed_files_obj, BaseFileEndpoints): - response = await managed_files_obj.afile_list( - purpose=purpose, - litellm_parent_otel_span=user_api_key_dict.parent_otel_span, + response = await _list_owned_managed_files( # rebind-ok: each dispatch branch binds response + managed_files_obj=managed_files_obj, user_api_key_dict=user_api_key_dict, + purpose=purpose, limit=limit, after=after, + before=None, + model_scope=None, ) else: resolved_custom_llm_provider: Final = custom_llm_provider or "openai" diff --git a/litellm/proxy/pass_through_endpoints/managed_id_rewriter.py b/litellm/proxy/pass_through_endpoints/managed_id_rewriter.py index a5a04fb68d2..9790c150ae1 100644 --- a/litellm/proxy/pass_through_endpoints/managed_id_rewriter.py +++ b/litellm/proxy/pass_through_endpoints/managed_id_rewriter.py @@ -242,6 +242,12 @@ def _passthrough_sentinel_model_id(provider: str) -> str: return f"_passthrough_{provider}" +def passthrough_file_model_id(provider: str) -> str | None: + if (provider, "/v1/files") not in _LIST_ROUTE_TABLE: + return None + return _passthrough_sentinel_model_id(provider) + + # Key under which the provider marker is stored in a file row's model_mappings. # Its value lands in flat_model_file_ids (built from model_mappings.values()), # giving the file table a DB-queryable provider scope it otherwise lacks. diff --git a/litellm/proxy/schema.prisma b/litellm/proxy/schema.prisma index 83a0e6f9e05..4187f76c97c 100644 --- a/litellm/proxy/schema.prisma +++ b/litellm/proxy/schema.prisma @@ -1163,6 +1163,7 @@ model LiteLLM_ManagedFileTable { @@index([unified_file_id]) @@index([flat_model_file_ids], type: Gin) @@index([team_id, created_at(sort: Desc)]) + @@index([created_by, created_at(sort: Desc)]) } model LiteLLM_ManagedObjectTable { // for batches or finetuning jobs which use the diff --git a/schema.prisma b/schema.prisma index 83a0e6f9e05..4187f76c97c 100644 --- a/schema.prisma +++ b/schema.prisma @@ -1163,6 +1163,7 @@ model LiteLLM_ManagedFileTable { @@index([unified_file_id]) @@index([flat_model_file_ids], type: Gin) @@index([team_id, created_at(sort: Desc)]) + @@index([created_by, created_at(sort: Desc)]) } model LiteLLM_ManagedObjectTable { // for batches or finetuning jobs which use the diff --git a/tests/integration/security/test_managed_file_listing.py b/tests/integration/security/test_managed_file_listing.py new file mode 100644 index 00000000000..8c736b2b2bf --- /dev/null +++ b/tests/integration/security/test_managed_file_listing.py @@ -0,0 +1,263 @@ +import json +import uuid +from collections.abc import Callable, Iterator, Mapping +from contextlib import contextmanager +from dataclasses import dataclass, field +from itertools import count, islice +from pathlib import Path +from types import MappingProxyType +from typing import Final +from urllib.parse import parse_qs, urlsplit + +import pytest +import yaml +from integration._support.client import Gateway, gateway_from_environment +from integration._support.process import owned_proxy +from integration._support.wire import Reply, Request, Wire, wire_server +from anthropic import Anthropic +from openai import OpenAI +from pydantic import BaseModel, ConfigDict, JsonValue + +pytestmark: Final = pytest.mark.shares_worker_state + +UNMANAGED_FILE: Final = "file-uploaded-straight-to-the-provider" +UPLOAD: Final = b"isolation check\n" + + +@dataclass(frozen=True, slots=True) +class Deployment: + provider: str + model: str + litellm_model: str + api_key: str + api_path: str + + +OPENAI: Final = Deployment("openai", "gpt-4.1-managed-listing", "openai/gpt-4.1", "sk-managed-listing", "/v1") +ANTHROPIC: Final = Deployment( + "anthropic", "claude-managed-listing", "anthropic/claude-haiku-5-5", "sk-ant-managed-listing", "" +) +DEPLOYMENTS: Final = (pytest.param(OPENAI, id="openai"), pytest.param(ANTHROPIC, id="anthropic")) + + +@dataclass(frozen=True, slots=True) +class ListCall: + path: str + params: Mapping[str, str] = field(default_factory=lambda: MappingProxyType({})) + headers: Mapping[str, str] = field(default_factory=lambda: MappingProxyType({})) + + +LIST_FORMS: Final = ( + pytest.param(lambda d: ListCall(f"/{d.provider}/v1/files"), id="provider-path"), + pytest.param(lambda d: ListCall("/v1/files", params={"provider": d.provider}), id="provider-query"), + pytest.param(lambda d: ListCall("/v1/files", headers={"custom-llm-provider": d.provider}), id="provider-header"), + pytest.param(lambda d: ListCall("/v1/files", params={"target_model_names": d.model}), id="target-model-names"), + pytest.param(lambda d: ListCall("/v1/files", headers={"x-litellm-model": d.model}), id="model-header"), +) + + +class FilePage(BaseModel): + model_config = ConfigDict(extra="ignore", frozen=True) + + data: list[dict[str, JsonValue]] + has_more: bool + + @property + def ids(self) -> list[str]: + return [str(entry["id"]) for entry in self.data] + + +def _openai_file(file_id: str) -> dict[str, JsonValue]: + return { + "id": file_id, + "object": "file", + "bytes": len(UPLOAD), + "created_at": 1700000000, + "filename": "notes.txt", + "purpose": "user_data", + "status": "processed", + } + + +def _anthropic_file(file_id: str) -> dict[str, JsonValue]: + return { + "id": file_id, + "type": "file", + "filename": "notes.txt", + "mime_type": "text/plain", + "size_bytes": len(UPLOAD), + "created_at": "2026-01-01T00:00:00Z", + "downloadable": False, + } + + +def _reply(body: Mapping[str, JsonValue]) -> Reply: + return Reply(body=json.dumps(body).encode()) + + +def _list_reply(provider: str, file_ids: tuple[str, ...]) -> Reply: + if provider == ANTHROPIC.provider: + return _reply( + { + "data": [_anthropic_file(file_id) for file_id in file_ids], + "has_more": False, + "first_id": file_ids[0], + "last_id": file_ids[-1], + } + ) + return _reply({"object": "list", "data": [_openai_file(file_id) for file_id in file_ids]}) + + +def _responder(provider_files: Mapping[str, tuple[str, ...]]) -> Callable[[Request], Reply]: + uploads: Final = {provider: count() for provider in provider_files} + + def respond(request: Request) -> Reply: + target: Final = urlsplit(request.target) + path: Final = target.path + provider: Final = ANTHROPIC.provider if "x-api-key" in request.headers else OPENAI.provider + files: Final = provider_files[provider] + if (request.method, path) == ("GET", "/v1/models"): + return _reply({"object": "list", "data": []}) + if (request.method, path) == ("POST", "/v1/files"): + uploaded: Final = files[next(uploads[provider])] + return _reply(_anthropic_file(uploaded) if provider == ANTHROPIC.provider else _openai_file(uploaded)) + if (request.method, path) == ("GET", "/v1/files"): + limit: Final = int(parse_qs(target.query).get("limit", ["10000"])[0]) + return _list_reply(provider, (UNMANAGED_FILE, *reversed(files))[:limit]) + return Reply(status=404, body=json.dumps({"error": {"message": f"unscripted {request.target}"}}).encode()) + + return respond + + +def _config(wire: Wire) -> dict[str, JsonValue]: + return { + "model_list": [ + { + "model_name": deployment.model, + "litellm_params": { + "model": deployment.litellm_model, + "api_key": deployment.api_key, + "api_base": f"{wire.url}{deployment.api_path}", + }, + } + for deployment in (OPENAI, ANTHROPIC) + ], + "litellm_settings": {"require_managed_files": True}, + "general_settings": {"master_key": "os.environ/LITELLM_MASTER_KEY", "database_url": "os.environ/DATABASE_URL"}, + } + + +@dataclass(frozen=True, slots=True) +class Rig: + proxy: Gateway + owner: str + intruder: str + files_by_provider: Mapping[str, tuple[str, ...]] + every_file: tuple[str, ...] + + @property + def base_url(self) -> str: + return str(self.proxy.client.base_url).rstrip("/") + + def list(self, call: ListCall, key: str | None = None) -> FilePage: + response: Final = self.proxy.request("GET", call.path, key=key, params=call.params, headers=call.headers) + assert response.status_code == 200, response.text + return FilePage.model_validate_json(response.text) + + +def _upload(proxy: Gateway, key: str, deployments: tuple[Deployment, ...]) -> str: + with OpenAI(base_url=f"{str(proxy.client.base_url).rstrip('/')}/openai/v1", api_key=key, max_retries=0) as client: + uploaded: Final = client.files.with_raw_response.create( + file=("notes.txt", UPLOAD, "text/plain"), + purpose="user_data", + extra_body={"target_model_names": ",".join(deployment.model for deployment in deployments)}, + ) + assert uploaded.status_code == 200, uploaded.text + return uploaded.parse().id + + +@contextmanager +def _rig(directory: Path) -> Iterator[Rig]: + build: Final = uuid.uuid4().hex + provider_files: Final = { + OPENAI.provider: (f"file-older-{build}", f"file-newer-{build}"), + ANTHROPIC.provider: (f"file_older_{build}", f"file_newer_{build}"), + } + with gateway_from_environment() as shared, wire_server(_responder(provider_files)) as wire: + config: Final = directory / "config.yaml" + config.write_text(yaml.safe_dump(_config(wire))) + with owned_proxy(shared, directory, {}, config=config) as proxy, proxy.scenario() as scenario: + owner: Final = scenario.key(user_id=scenario.user()) + intruder: Final = scenario.key(user_id=scenario.user()) + on_both: Final = _upload(proxy, owner, (OPENAI, ANTHROPIC)) + on_openai: Final = _upload(proxy, owner, (OPENAI,)) + on_anthropic: Final = _upload(proxy, owner, (ANTHROPIC,)) + yield Rig( + proxy, + owner, + intruder, + MappingProxyType( + {OPENAI.provider: (on_openai, on_both), ANTHROPIC.provider: (on_anthropic, on_both)} + ), + (on_anthropic, on_openai, on_both), + ) + + +@pytest.fixture(scope="module") +def rig(tmp_path_factory: pytest.TempPathFactory) -> Iterator[Rig]: + with _rig(tmp_path_factory.mktemp("managed-file-listing")) as built: + yield built + + +@pytest.mark.parametrize("deployment", DEPLOYMENTS) +@pytest.mark.parametrize("form", LIST_FORMS) +def test_another_key_lists_none_of_the_files_on_the_shared_key( + rig: Rig, deployment: Deployment, form: Callable[[Deployment], ListCall] +) -> None: + page: Final = rig.list(form(deployment), key=rig.intruder) + assert (page.ids, page.has_more) == ([], False) + + +@pytest.mark.parametrize("deployment", DEPLOYMENTS) +@pytest.mark.parametrize("form", LIST_FORMS) +def test_owner_lists_their_own_managed_files_on_that_provider_or_model_newest_first( + rig: Rig, deployment: Deployment, form: Callable[[Deployment], ListCall] +) -> None: + page: Final = rig.list(form(deployment), key=rig.owner) + assert (page.ids, page.has_more) == (list(rig.files_by_provider[deployment.provider]), False) + + +def test_owner_lists_every_managed_file_when_no_provider_or_model_is_named(rig: Rig) -> None: + page: Final = rig.list(ListCall("/v1/files"), key=rig.owner) + assert (page.ids, page.has_more) == (list(rig.every_file), False) + + +@pytest.mark.parametrize("cursor_param", ["after", "after_id"]) +def test_owner_pages_through_a_provider_list_with_limit_and_a_cursor(rig: Rig, cursor_param: str) -> None: + first: Final = rig.list(ListCall("/openai/v1/files", params={"limit": "1"}), key=rig.owner) + second: Final = rig.list( + ListCall("/openai/v1/files", params={"limit": "1", cursor_param: first.ids[0]}), key=rig.owner + ) + expected: Final = rig.files_by_provider[OPENAI.provider] + assert [(page.ids, page.has_more) for page in (first, second)] == [ + ([expected[0]], True), + ([expected[1]], False), + ] + + +def test_anthropic_sdk_auto_pages_the_owners_files_to_the_end(rig: Rig) -> None: + with Anthropic(base_url=f"{rig.base_url}/anthropic", api_key=rig.owner, max_retries=0) as client: + listed: Final = [file.id for file in islice(client.beta.files.list(limit=1), 5)] + assert listed == list(rig.files_by_provider[ANTHROPIC.provider]) + + +def test_anthropic_sdk_pages_backwards_to_the_newest_file_and_stops(rig: Rig) -> None: + newest_first: Final = rig.files_by_provider[ANTHROPIC.provider] + with Anthropic(base_url=f"{rig.base_url}/anthropic", api_key=rig.owner, max_retries=0) as client: + listed: Final = [file.id for file in islice(client.beta.files.list(limit=1, before_id=newest_first[-1]), 5)] + assert listed == list(reversed(newest_first[:-1])) + + +def test_master_key_still_lists_the_files_on_the_shared_provider_key(rig: Rig) -> None: + listed: Final = rig.list(ListCall(f"/{ANTHROPIC.provider}/v1/files")).ids + assert UNMANAGED_FILE in listed diff --git a/tests/unit/enterprise/proxy/test_managed_files_hook.py b/tests/unit/enterprise/proxy/test_managed_files_hook.py index 01f20689e89..f9f4e17c760 100644 --- a/tests/unit/enterprise/proxy/test_managed_files_hook.py +++ b/tests/unit/enterprise/proxy/test_managed_files_hook.py @@ -15,6 +15,7 @@ import pytest from typing import Final, Optional from unittest.mock import AsyncMock, MagicMock, patch +from litellm.llms.base_llm.files.transformation import FileListModelScope from litellm.proxy._types import LitellmUserRoles, ProxyException, UserAPIKeyAuth from litellm.proxy.openai_files_endpoints.common_utils import BATCH_CREATE_HIDDEN_PARAM from litellm.types.llms.openai import FileListPage, OpenAIFileObject @@ -649,6 +650,22 @@ async def test_afile_list_denies_a_caller_with_no_identity_at_all(): assert table.find_many_calls == [] +@pytest.mark.asyncio +async def test_afile_list_reads_no_rows_when_no_deployment_can_hold_a_file() -> None: + managed_files, table = _make_managed_files_over_rows([_make_managed_file_row("unified-mine")]) + + response: Final = await managed_files.afile_list( + purpose=None, + litellm_parent_otel_span=None, + user_api_key_dict=_make_user_api_key_dict(), + model_scope=FileListModelScope(frozenset()), + ) + + assert response.data == [] + assert response.has_more is False + assert table.find_many_calls == [] + + @pytest.mark.asyncio async def test_afile_list_filters_by_purpose(): managed_files, _ = _make_managed_files_over_rows( diff --git a/tests/unit/proxy/openai_files_endpoint/test_files_endpoint.py b/tests/unit/proxy/openai_files_endpoint/test_files_endpoint.py index d47db69257b..ff63e37bcd8 100644 --- a/tests/unit/proxy/openai_files_endpoint/test_files_endpoint.py +++ b/tests/unit/proxy/openai_files_endpoint/test_files_endpoint.py @@ -1,3 +1,4 @@ +import itertools import json from typing import Final, List, Optional from urllib.parse import quote @@ -12,6 +13,7 @@ from pytest_mock import MockerFixture import litellm from litellm import Router +from litellm.constants import ANTHROPIC_DEFAULT_FILE_LIST_LIMIT from litellm.files.types import FileContentStreamingResult from litellm.proxy._types import LiteLLM_UserTableFiltered, LitellmUserRoles, UserAPIKeyAuth from litellm.proxy.hooks import get_proxy_hook @@ -4969,6 +4971,8 @@ def _managed_file_row( stored_file_id: Optional[str] = None, created_by: str = "test-user", team_id: Optional[str] = None, + model_mappings: Optional[dict[str, str]] = None, + storage_backend: Optional[str] = None, ) -> MagicMock: file_object = OpenAIFileObject( id=stored_file_id or unified_file_id, @@ -4984,6 +4988,8 @@ def _managed_file_row( file_object=file_object.model_dump(), created_by=created_by, team_id=team_id, + model_mappings=model_mappings or {}, + storage_backend=storage_backend, ) @@ -5010,16 +5016,49 @@ class _ManagedFileTableOverRows: return next(iter(self._owned_rows(where)), None) async def find_many(self, where, take=None, order=None, cursor=None, skip=0): - rows = self._owned_rows(where) - if cursor is not None: - start = next( - index - for index, row in enumerate(rows) - if row.unified_file_id == cursor["unified_file_id"] - ) - rows = rows[start + skip :] + owned_rows: Final = self._owned_rows(where) + oldest_first: Final = bool(order) and order[0].get("created_at") == "asc" + ordered_rows: Final = owned_rows[::-1] if oldest_first else owned_rows + start: Final = ( + 0 + if cursor is None + else skip + + next(index for index, row in enumerate(ordered_rows) if row.unified_file_id == cursor["unified_file_id"]) + ) + rows: Final = ordered_rows[start:] return rows if take is None else rows[:take] + async def query_raw(self, query: str, *params: object) -> list[dict[str, object]]: + keys_index: Final = next(index for index, param in enumerate(params) if isinstance(param, list)) + owner_ids: Final = frozenset(params[:keys_index]) + model_mapping_keys: Final = frozenset(params[keys_index]) + cursor_id: Final = params[keys_index + 1] if len(params) == keys_index + 3 else None + ordered_rows: Final = [ + row + for row in (self.rows[::-1] if '"created_at" ASC' in query else self.rows) + if not owner_ids or {row.created_by, row.team_id} & owner_ids + ] + start: Final = ( + 0 + if cursor_id is None + else 1 + next(index for index, row in enumerate(ordered_rows) if row.unified_file_id == cursor_id) + ) + listed_rows: Final = [ + row + for row in ordered_rows[start:] + if not model_mapping_keys.isdisjoint(row.model_mappings) + or ('"storage_backend" IS NOT NULL' in query and row.storage_backend is not None) + ] + return [ + { + "unified_file_id": row.unified_file_id, + "file_object": row.file_object, + "model_mappings": row.model_mappings, + "storage_backend": row.storage_backend, + } + for row in listed_rows[: params[-1]] + ] + def _setup_unscoped_list_files_route_over_real_hook( mocker, monkeypatch, llm_router: Router, rows @@ -5041,7 +5080,9 @@ def _setup_unscoped_list_files_route_over_real_hook( managed_files = PROXY_LiteLLMManagedFiles( internal_usage_cache=MagicMock(), prisma_client=MagicMock() ) - managed_files.prisma_client.db.litellm_managedfiletable = _ManagedFileTableOverRows(rows) + managed_file_table = _ManagedFileTableOverRows(rows) + managed_files.prisma_client.db.litellm_managedfiletable = managed_file_table + managed_files.prisma_client.db.query_raw = managed_file_table.query_raw proxy_logging_obj = setup_proxy_logging_object(monkeypatch, llm_router) proxy_logging_obj.proxy_hook_mapping["managed_files"] = managed_files @@ -6263,3 +6304,513 @@ def test_create_file_passthrough_fails_closed_when_guardrails_would_scan_the_bat assert error["param"] == "passthrough" assert "guardrails" in error["message"] assert forwarded_calls == [] + + +def _callers_files_on_every_deployment() -> list[MagicMock]: + return [ + _managed_file_row("unified-gpt-newer", model_mappings={"gpt-3.5-turbo-id": "file-gpt-newer"}), + _managed_file_row("unified-azure", model_mappings={"azure-gpt-3-5-turbo-id": "file-azure"}), + _managed_file_row("unified-gemini", model_mappings={"gemini-2.0-flash-id": "file-gemini"}), + _managed_file_row( + "unified-theirs", created_by="other-user", model_mappings={"gpt-3.5-turbo-id": "file-theirs"} + ), + _managed_file_row("unified-gpt-older", model_mappings={"gpt-3.5-turbo-id": "file-gpt-older"}), + ] + + +def _list_files_as_test_user(path: str, params: dict[str, str], headers: dict[str, str]) -> httpx.Response: + import litellm.proxy.proxy_server as ps + + try: + return client.get(path, params=params, headers={"Authorization": "Bearer test-key", **headers}) + finally: + app.dependency_overrides.pop(ps.user_api_key_auth, None) + + +@pytest.mark.parametrize( + ("path", "params", "headers", "expected_ids"), + [ + ("/openai/v1/files", {}, {}, ["unified-gpt-newer", "unified-gpt-older"]), + ("/v1/files", {"provider": "azure"}, {}, ["unified-azure"]), + ("/v1/files", {}, {"custom-llm-provider": "gemini"}, ["unified-gemini"]), + ("/v1/files", {"target_model_names": "gemini-2.0-flash"}, {}, ["unified-gemini"]), + ("/v1/files", {}, {"x-litellm-model": "gpt-3.5-turbo"}, ["unified-gpt-newer", "unified-gpt-older"]), + ("/v1/files", {}, {}, ["unified-gpt-newer", "unified-azure", "unified-gemini", "unified-gpt-older"]), + ], + ids=["provider-path", "provider-query", "provider-header", "target-model-names", "model-header", "no-provider"], +) +def test_require_managed_files_lists_the_callers_files_on_the_named_provider_or_model( + mocker: MockerFixture, + monkeypatch: pytest.MonkeyPatch, + llm_router: Router, + path: str, + params: dict[str, str], + headers: dict[str, str], + expected_ids: list[str], +) -> None: + _, provider_list = _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user(path, params, headers) + + assert response.status_code == 200, response.text + assert [file["id"] for file in response.json()["data"]] == expected_ids + provider_list.assert_not_awaited() + + +def test_require_managed_files_lists_a_teammates_files_on_the_provider_and_no_other_teams( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch, llm_router: Router +) -> None: + import litellm.proxy.proxy_server as ps + from litellm.proxy._types import LitellmUserRoles + + _setup_unscoped_list_files_route_over_real_hook( + mocker, + monkeypatch, + llm_router, + [ + _managed_file_row( + "unified-teammate", created_by="teammate", team_id="team-a", model_mappings={"gpt-3.5-turbo-id": "f-1"} + ), + _managed_file_row("unified-own", model_mappings={"gpt-3.5-turbo-id": "f-2"}), + _managed_file_row( + "unified-other-team", created_by="outsider", team_id="team-b", model_mappings={"gpt-3.5-turbo-id": "f-3"} + ), + ], + ) + app.dependency_overrides[ps.user_api_key_auth] = lambda: UserAPIKeyAuth( + api_key="test-key", user_role=LitellmUserRoles.INTERNAL_USER, user_id="test-user", team_id="team-a" + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user("/openai/v1/files", {}, {}) + + assert response.status_code == 200, response.text + assert [file["id"] for file in response.json()["data"]] == ["unified-teammate", "unified-own"] + + +def _router_with_wildcard_and_unprefixed_anthropic_deployments() -> Router: + unprefixed_anthropic_model: Final = next(model for model in sorted(litellm.anthropic_models) if "/" not in model) + return Router( + model_list=[ + { + "model_name": "gpt-3.5-turbo", + "litellm_params": {"model": "openai/gpt-3.5-turbo", "api_key": "openai_api_key"}, + "model_info": {"id": "gpt-3.5-turbo-id"}, + }, + { + "model_name": "anthropic/*", + "litellm_params": {"model": "anthropic/*", "api_key": "anthropic_api_key"}, + "model_info": {"id": "anthropic-wildcard-id"}, + }, + { + "model_name": "claude-unprefixed", + "litellm_params": {"model": unprefixed_anthropic_model, "api_key": "anthropic_api_key"}, + "model_info": {"id": "claude-unprefixed-id"}, + }, + ], + model_group_alias={"claude-alias": "claude-unprefixed"}, + routing_groups=[ + {"group_name": "claude-group", "models": ["claude-unprefixed"], "routing_strategy": "simple-shuffle"} + ], + ) + + +def _callers_files_on_wildcard_unprefixed_and_passthrough_deployments() -> list[MagicMock]: + return [ + _managed_file_row("unified-wildcard", model_mappings={"anthropic-wildcard-id": "file-wildcard"}), + _managed_file_row("unified-unprefixed", model_mappings={"claude-unprefixed-id": "file-unprefixed"}), + _managed_file_row("unified-gpt", model_mappings={"gpt-3.5-turbo-id": "file-gpt"}), + _managed_file_row( + "unified-passthrough", + model_mappings={ + "_passthrough_openai": "file-passthrough", + "_passthrough_provider_marker": "_passthrough_provider:openai", + }, + ), + _managed_file_row( + "unified-storage", + model_mappings={"claude-unprefixed": "litellm_db://unified-storage"}, + storage_backend="litellm_db", + ), + _managed_file_row( + "unified-wildcard-storage", + model_mappings={"anthropic/claude-haiku-5-5": "litellm_db://unified-wildcard-storage"}, + storage_backend="litellm_db", + ), + _managed_file_row( + "unified-alias-storage", + model_mappings={"claude-alias": "s3://bucket/unified-alias-storage"}, + storage_backend="s3", + ), + _managed_file_row( + "unified-group-storage", + model_mappings={"claude-group": "s3://bucket/unified-group-storage"}, + storage_backend="s3", + ), + ] + + +@pytest.mark.parametrize( + ("path", "params", "headers", "expected_ids"), + [ + ( + "/v1/files", + {"target_model_names": "anthropic/claude-haiku-5-5"}, + {}, + ["unified-wildcard", "unified-wildcard-storage"], + ), + ( + "/v1/files", + {}, + {"x-litellm-model": "anthropic/claude-haiku-5-5"}, + ["unified-wildcard", "unified-wildcard-storage"], + ), + ( + "/anthropic/v1/files", + {}, + {}, + [ + "unified-wildcard", + "unified-unprefixed", + "unified-storage", + "unified-wildcard-storage", + "unified-alias-storage", + "unified-group-storage", + ], + ), + ("/openai/v1/files", {}, {}, ["unified-gpt", "unified-passthrough"]), + ("/v1/files", {"target_model_names": "claude-unprefixed"}, {}, ["unified-unprefixed", "unified-storage"]), + ], + ids=[ + "wildcard-target-model-names", + "wildcard-model-header", + "unprefixed-provider", + "passthrough-provider", + "storage-backed-model", + ], +) +def test_require_managed_files_lists_the_callers_files_wherever_the_router_put_them( + mocker: MockerFixture, + monkeypatch: pytest.MonkeyPatch, + path: str, + params: dict[str, str], + headers: dict[str, str], + expected_ids: list[str], +) -> None: + _, provider_list = _setup_unscoped_list_files_route_over_real_hook( + mocker, + monkeypatch, + _router_with_wildcard_and_unprefixed_anthropic_deployments(), + _callers_files_on_wildcard_unprefixed_and_passthrough_deployments(), + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user(path, params, headers) + + assert response.status_code == 200, response.text + assert [file["id"] for file in response.json()["data"]] == expected_ids + provider_list.assert_not_awaited() + + +def _router_with_a_catch_all_openai_deployment() -> Router: + unprefixed_anthropic_model: Final = next(model for model in sorted(litellm.anthropic_models) if "/" not in model) + return Router( + model_list=[ + { + "model_name": "gpt-3.5-turbo", + "litellm_params": {"model": "openai/gpt-3.5-turbo", "api_key": "openai_api_key"}, + "model_info": {"id": "gpt-3.5-turbo-id"}, + }, + { + "model_name": "*", + "litellm_params": {"model": "openai/*", "api_key": "openai_api_key"}, + "model_info": {"id": "catch-all-id"}, + }, + { + "model_name": "claude-unprefixed", + "litellm_params": {"model": unprefixed_anthropic_model, "api_key": "anthropic_api_key"}, + "model_info": {"id": "claude-unprefixed-id"}, + }, + ] + ) + + +@pytest.mark.parametrize( + ("path", "expected_ids"), + [ + ("/openai/v1/files", ["unified-gpt", "unified-catch-all-storage"]), + ("/anthropic/v1/files", ["unified-unprefixed", "unified-storage"]), + ], + ids=["catch-all-provider", "explicit-provider"], +) +def test_require_managed_files_keeps_a_catch_all_route_from_claiming_files_on_other_providers( + mocker: MockerFixture, + monkeypatch: pytest.MonkeyPatch, + path: str, + expected_ids: list[str], +) -> None: + _, provider_list = _setup_unscoped_list_files_route_over_real_hook( + mocker, + monkeypatch, + _router_with_a_catch_all_openai_deployment(), + [ + _managed_file_row("unified-unprefixed", model_mappings={"claude-unprefixed-id": "file-unprefixed"}), + _managed_file_row("unified-gpt", model_mappings={"gpt-3.5-turbo-id": "file-gpt"}), + _managed_file_row( + "unified-storage", + model_mappings={"claude-unprefixed": "litellm_db://unified-storage"}, + storage_backend="litellm_db", + ), + _managed_file_row( + "unified-catch-all-storage", + model_mappings={"model-only-the-catch-all-serves": "litellm_db://unified-catch-all-storage"}, + storage_backend="litellm_db", + ), + ], + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user(path, {}, {}) + + assert response.status_code == 200, response.text + assert [file["id"] for file in response.json()["data"]] == expected_ids + provider_list.assert_not_awaited() + + +@pytest.mark.parametrize( + ("path", "params"), + [ + ("/antropic/v1/files", {}), + ("/openai/v1/files", {"after": "unified-gpt", "after_id": "unified-wildcard"}), + ("/v1/files", {"target_model_names": "gpt-3.5-turbo,claude-unprefixed"}), + ], + ids=["unknown-provider", "conflicting-after-cursors", "two-target-model-names"], +) +def test_require_managed_files_rejects_a_list_it_cannot_scope( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch, path: str, params: dict[str, str] +) -> None: + _, provider_list = _setup_unscoped_list_files_route_over_real_hook( + mocker, + monkeypatch, + _router_with_wildcard_and_unprefixed_anthropic_deployments(), + _callers_files_on_wildcard_unprefixed_and_passthrough_deployments(), + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user(path, params, {}) + + assert response.status_code == 400, response.text + provider_list.assert_not_awaited() + + +@pytest.mark.parametrize("cursor_param", ["after", "after_id"]) +def test_require_managed_files_pages_a_provider_list_with_either_cursor_param( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch, llm_router: Router, cursor_param: str +) -> None: + import litellm.proxy.proxy_server as ps + + _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + try: + first_page: Final = client.get( + "/openai/v1/files", params={"limit": "1"}, headers={"Authorization": "Bearer test-key"} + ) + second_page: Final = client.get( + "/openai/v1/files", + params={"limit": "1", cursor_param: first_page.json()["last_id"]}, + headers={"Authorization": "Bearer test-key"}, + ) + finally: + app.dependency_overrides.pop(ps.user_api_key_auth, None) + + assert (first_page.status_code, second_page.status_code) == (200, 200), (first_page.text, second_page.text) + assert [file["id"] for file in first_page.json()["data"]] == ["unified-gpt-newer"] + assert first_page.json()["has_more"] is True + assert [file["id"] for file in second_page.json()["data"]] == ["unified-gpt-older"] + assert second_page.json()["has_more"] is False + + +def test_require_managed_files_pages_toward_newer_files_with_before_id( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch, llm_router: Router +) -> None: + import litellm.proxy.proxy_server as ps + + _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + try: + page_before_oldest: Final = client.get( + "/v1/files", + params={"limit": "2", "before_id": "unified-gpt-older"}, + headers={"Authorization": "Bearer test-key"}, + ) + newest_page: Final = client.get( + "/v1/files", + params={"limit": "2", "before_id": page_before_oldest.json()["first_id"]}, + headers={"Authorization": "Bearer test-key"}, + ) + finally: + app.dependency_overrides.pop(ps.user_api_key_auth, None) + + assert (page_before_oldest.status_code, newest_page.status_code) == (200, 200), ( + page_before_oldest.text, + newest_page.text, + ) + assert [file["id"] for file in page_before_oldest.json()["data"]] == ["unified-azure", "unified-gemini"] + assert page_before_oldest.json()["has_more"] is True + assert [file["id"] for file in newest_page.json()["data"]] == ["unified-gpt-newer"] + assert newest_page.json()["has_more"] is False + + +def test_managed_file_list_rejects_both_cursor_directions( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch, llm_router: Router +) -> None: + _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user( + "/v1/files", {"after": "unified-gpt-newer", "before_id": "unified-gpt-older"}, {} + ) + + assert response.status_code == 400, response.text + assert "not both" in response.json()["error"]["message"] + + +@pytest.mark.parametrize( + ("params", "headers"), + [({"target_model_names": "gemini-2.0-flash"}, {}), ({}, {"x-litellm-model": "gemini-2.0-flash"})], + ids=["target-model-names", "model-header"], +) +def test_require_managed_files_list_rejects_a_model_the_key_cannot_use( + mocker: MockerFixture, + monkeypatch: pytest.MonkeyPatch, + llm_router: Router, + params: dict[str, str], + headers: dict[str, str], +) -> None: + import litellm.proxy.proxy_server as ps + + _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + app.dependency_overrides[ps.user_api_key_auth] = lambda: UserAPIKeyAuth( + api_key="test-key", + user_role=LitellmUserRoles.INTERNAL_USER, + user_id="test-user", + models=["gpt-3.5-turbo"], + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user("/v1/files", params, headers) + + assert response.status_code == 403, response.text + assert "unified-gemini" not in response.text + + +def test_require_managed_files_list_fails_closed_without_the_managed_files_hook( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch, llm_router: Router +) -> None: + proxy_logging_obj, provider_list = _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + proxy_logging_obj.proxy_hook_mapping.pop("managed_files") + provider_list.return_value = [_managed_file("file-on-the-shared-key")] + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user("/openai/v1/files", {}, {}) + + assert response.status_code == 500, response.text + assert "file-on-the-shared-key" not in response.text + assert "require_managed_files cannot scope the list to the caller" in response.json()["error"]["message"] + provider_list.assert_not_awaited() + + +@pytest.mark.parametrize("hook_loaded", [True, False], ids=["hook-loaded", "no-hook"]) +def test_require_managed_files_keeps_the_provider_list_for_an_admin( + mocker: MockerFixture, + monkeypatch: pytest.MonkeyPatch, + llm_router: Router, + hook_loaded: bool, +) -> None: + import litellm.proxy.proxy_server as ps + + proxy_logging_obj, provider_list = _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + if not hook_loaded: + proxy_logging_obj.proxy_hook_mapping.pop("managed_files") + provider_list.return_value = [_managed_file("file-on-the-shared-key")] + app.dependency_overrides[ps.user_api_key_auth] = lambda: UserAPIKeyAuth( + api_key="test-key", user_role=LitellmUserRoles.PROXY_ADMIN, user_id="admin-user" + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user("/anthropic/v1/files", {}, {}) + + assert response.status_code == 200, response.text + assert [file["id"] for file in response.json()["data"]] == ["file-on-the-shared-key"] + + +def test_managed_file_list_without_require_managed_files_ignores_cursor_params_openai_does_not_define( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch, llm_router: Router +) -> None: + _setup_unscoped_list_files_route_over_real_hook( + mocker, monkeypatch, llm_router, _callers_files_on_every_deployment() + ) + monkeypatch.setattr("litellm.require_managed_files", False) + + response: Final = _list_files_as_test_user( + "/v1/files", {"after": "unified-gpt-newer", "after_id": "unified-gemini", "before_id": "file-raw"}, {} + ) + + assert response.status_code == 200, response.text + assert [file["id"] for file in response.json()["data"]] == ["unified-azure", "unified-gemini", "unified-gpt-older"] + + +@pytest.mark.parametrize( + ("path", "headers", "expected_count", "has_more"), + [ + ("/anthropic/v1/files", {}, ANTHROPIC_DEFAULT_FILE_LIST_LIMIT, True), + ("/v1/files", {"x-litellm-model": "claude-unprefixed"}, ANTHROPIC_DEFAULT_FILE_LIST_LIMIT, True), + ("/anthropic/v1/files", {"x-litellm-model": "gpt-3.5-turbo"}, ANTHROPIC_DEFAULT_FILE_LIST_LIMIT + 1, False), + ("/v1/files", {}, 2 * (ANTHROPIC_DEFAULT_FILE_LIST_LIMIT + 1), False), + ], + ids=["anthropic-provider", "anthropic-model-header", "openai-model-header", "every-provider"], +) +def test_require_managed_files_pages_an_anthropic_list_at_the_anthropic_default_size( + mocker: MockerFixture, + monkeypatch: pytest.MonkeyPatch, + path: str, + headers: dict[str, str], + expected_count: int, + has_more: bool, +) -> None: + _setup_unscoped_list_files_route_over_real_hook( + mocker, + monkeypatch, + _router_with_wildcard_and_unprefixed_anthropic_deployments(), + [ + _managed_file_row( + f"unified-{deployment_id}-{index:03d}", model_mappings={deployment_id: f"file-{index:03d}"} + ) + for deployment_id, index in itertools.product( + ("claude-unprefixed-id", "gpt-3.5-turbo-id"), range(ANTHROPIC_DEFAULT_FILE_LIST_LIMIT + 1) + ) + ], + ) + monkeypatch.setattr("litellm.require_managed_files", True) + + response: Final = _list_files_as_test_user(path, {}, headers) + + assert response.status_code == 200, response.text + assert (len(response.json()["data"]), response.json()["has_more"]) == (expected_count, has_more) diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 45f9630983f..70406105193 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -59857,6 +59857,8 @@ export interface operations { purpose?: string | null; limit?: number | null; after?: string | null; + after_id?: string | null; + before_id?: string | null; }; header?: never; path?: never; @@ -79053,6 +79055,8 @@ export interface operations { purpose?: string | null; limit?: number | null; after?: string | null; + after_id?: string | null; + before_id?: string | null; }; header?: never; path?: never; @@ -86388,6 +86392,8 @@ export interface operations { purpose?: string | null; limit?: number | null; after?: string | null; + after_id?: string | null; + before_id?: string | null; }; header?: never; path: { From cf7c35a1174eb9864cf35c1513c2d6d50b5df26a Mon Sep 17 00:00:00 2001 From: moe-berri Date: Wed, 28 Oct 2026 13:48:37 -0700 Subject: [PATCH 4/7] fix(anthropic): map model_context_window_exceeded stop reason to length (#48874) * test(anthropic): cover model_context_window_exceeded finish reason on chat and responses Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(anthropic): map model_context_window_exceeded stop reason to length Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(rust): mirror model_context_window_exceeded finish reason mapping Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../crates/core-utils/src/core_helpers.rs | 2 + litellm/litellm_core_utils/core_helpers.py | 1 + ...test_anthropic_context_window_stop_wire.py | 122 ++++++++++++++++++ .../litellm_core_utils/test_core_helpers.py | 1 + 4 files changed, 126 insertions(+) create mode 100644 tests/integration/providers/test_anthropic_context_window_stop_wire.py diff --git a/litellm-rust/crates/core-utils/src/core_helpers.rs b/litellm-rust/crates/core-utils/src/core_helpers.rs index ada1c3ceb1a..58ee88d301c 100644 --- a/litellm-rust/crates/core-utils/src/core_helpers.rs +++ b/litellm-rust/crates/core-utils/src/core_helpers.rs @@ -13,6 +13,7 @@ const FINISH_REASONS: &[(&str, &str)] = &[ ("max_tokens", "length"), ("refusal", "content_filter"), ("compaction", "length"), + ("model_context_window_exceeded", "length"), ("guardrail_intervened", "content_filter"), ("content_filtered", "content_filter"), ("content_filter", "content_filter"), @@ -74,6 +75,7 @@ mod tests { #[case::end_turn("end_turn", "stop")] #[case::stop_sequence("stop_sequence", "stop")] #[case::max_tokens("max_tokens", "length")] + #[case::model_context_window_exceeded("model_context_window_exceeded", "length")] #[case::refusal("refusal", "content_filter")] #[case::guardrail_intervened("guardrail_intervened", "content_filter")] #[case::content_filtered("content_filtered", "content_filter")] diff --git a/litellm/litellm_core_utils/core_helpers.py b/litellm/litellm_core_utils/core_helpers.py index 00c6ce97e78..55025508e40 100644 --- a/litellm/litellm_core_utils/core_helpers.py +++ b/litellm/litellm_core_utils/core_helpers.py @@ -266,6 +266,7 @@ _FINISH_REASON_MAP: Final[dict[str, OpenAIChatCompletionFinishReason]] = { "tool_use": "tool_calls", "refusal": "content_filter", "compaction": "length", + "model_context_window_exceeded": "length", # Cohere "COMPLETE": "stop", "ERROR_TOXIC": "content_filter", diff --git a/tests/integration/providers/test_anthropic_context_window_stop_wire.py b/tests/integration/providers/test_anthropic_context_window_stop_wire.py new file mode 100644 index 00000000000..38e3e7dffcd --- /dev/null +++ b/tests/integration/providers/test_anthropic_context_window_stop_wire.py @@ -0,0 +1,122 @@ +import json +import uuid +from collections.abc import Mapping +from typing import Final + +import pytest +from integration._support.client import Gateway +from integration._support.wire import Reply, Request, wire_server +from pydantic import JsonValue + +MODEL: Final = "claude-haiku-4-5" + + +def _sse(event: Mapping[str, JsonValue]) -> bytes: + return f"event: {event['type']}\ndata: {json.dumps(event)}\n\n".encode() + + +def _message(identity: str, stop_reason: str | None) -> dict[str, JsonValue]: + return { + "id": "msg_" + identity, + "type": "message", + "role": "assistant", + "model": MODEL, + "content": [{"type": "text", "text": "alpha bravo"}], + "stop_reason": stop_reason, + "stop_sequence": None, + "usage": {"input_tokens": 199878, "output_tokens": 12799}, + } + + +def _stream(identity: str, stop_reason: str) -> tuple[bytes, ...]: + return ( + _sse({"type": "message_start", "message": {**_message(identity, None), "content": []}}), + _sse({"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}), + _sse({"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "alpha bravo"}}), + _sse({"type": "content_block_stop", "index": 0}), + _sse( + { + "type": "message_delta", + "delta": {"stop_reason": stop_reason, "stop_sequence": None}, + "usage": {"output_tokens": 12799}, + } + ), + _sse({"type": "message_stop"}), + ) + + +def _anthropic(identity: str, stop_reason: str): + def respond(request: Request) -> Reply: + assert request.method == "POST" and request.target == "/v1/messages", request.target + if json.loads(request.body).get("stream"): + return Reply(content_type="text/event-stream", chunks=_stream(identity, stop_reason)) + return Reply(body=json.dumps(_message(identity, stop_reason)).encode()) + + return respond + + +def _events(gateway: Gateway, path: str, body: Mapping[str, JsonValue]) -> tuple[dict[str, JsonValue], ...]: + with gateway.client.stream( + "POST", path, json={**body, "stream": True}, headers={"Authorization": f"Bearer {gateway.key}"} + ) as response: + assert response.status_code == 200, response.read().decode() + return tuple( + json.loads(line.removeprefix("data: ")) + for line in response.iter_lines() + if line.startswith("data: ") and line != "data: [DONE]" + ) + + +def _chat_finish_reason(gateway: Gateway, model: str, *, stream: bool) -> JsonValue: + body: Final = {"model": model, "max_tokens": 64000, "messages": [{"role": "user", "content": "copy the list"}]} + if stream: + reasons: Final = tuple( + choice["finish_reason"] + for chunk in _events(gateway, "/v1/chat/completions", body) + for choice in chunk["choices"] # comprehension-ok: chunk -> choices flatten + if choice.get("finish_reason") + ) + assert len(reasons) == 1, reasons + return reasons[0] + response: Final = gateway.request("POST", "/v1/chat/completions", body) + assert response.status_code == 200, response.text + return response.json()["choices"][0]["finish_reason"] + + +def _responses_status(gateway: Gateway, model: str, *, stream: bool) -> JsonValue: + body: Final = {"model": model, "max_output_tokens": 64000, "input": "copy the list"} + if stream: + terminal: Final = tuple( + event["response"] + for event in _events(gateway, "/v1/responses", body) + if event["type"] in ("response.completed", "response.incomplete", "response.failed") + ) + assert len(terminal) == 1, terminal + return terminal[0]["status"] + response: Final = gateway.request("POST", "/v1/responses", body) + assert response.status_code == 200, response.text + return response.json()["status"] + + +@pytest.mark.parametrize("stream", [False, True], ids=["json", "stream"]) +@pytest.mark.parametrize( + ("stop_reason", "finish_reason"), + [("model_context_window_exceeded", "length"), ("max_tokens", "length"), ("end_turn", "stop")], +) +def test_chat_maps_anthropic_stop_reason(gateway: Gateway, stop_reason: str, finish_reason: str, stream: bool) -> None: + identity: Final = uuid.uuid4().hex + with wire_server(_anthropic(identity, stop_reason)) as wire, gateway.scenario() as scenario: + model: Final = scenario.model(model=f"anthropic/{MODEL}", api_base=wire.url, api_key="scripted-anthropic-key") + assert _chat_finish_reason(gateway, model, stream=stream) == finish_reason + + +@pytest.mark.parametrize("stream", [False, True], ids=["json", "stream"]) +@pytest.mark.parametrize( + ("stop_reason", "status"), + [("model_context_window_exceeded", "incomplete"), ("max_tokens", "incomplete"), ("end_turn", "completed")], +) +def test_responses_maps_anthropic_stop_reason(gateway: Gateway, stop_reason: str, status: str, stream: bool) -> None: + identity: Final = uuid.uuid4().hex + with wire_server(_anthropic(identity, stop_reason)) as wire, gateway.scenario() as scenario: + model: Final = scenario.model(model=f"anthropic/{MODEL}", api_base=wire.url, api_key="scripted-anthropic-key") + assert _responses_status(gateway, model, stream=stream) == status diff --git a/tests/unit/litellm_core_utils/test_core_helpers.py b/tests/unit/litellm_core_utils/test_core_helpers.py index d66c37ba8c2..66a8e6bbda6 100644 --- a/tests/unit/litellm_core_utils/test_core_helpers.py +++ b/tests/unit/litellm_core_utils/test_core_helpers.py @@ -256,6 +256,7 @@ class TestMapFinishReasonAnthropic: ("max_tokens", "length"), ("tool_use", "tool_calls"), ("compaction", "length"), + ("model_context_window_exceeded", "length"), ("content_filtered", "content_filter"), ], ) From 4516a251c940cc6ff7977b460a26a0582a0a18ff Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 28 Oct 2026 14:39:25 -0700 Subject: [PATCH 5/7] feat(accounting): add accounting contracts, PricingSnapshot and ledger schema (LIT-10193) (#48862) * feat(accounting): add typed contracts and spend log adapters * refactor(accounting): reuse spend identity constants and register tests * refactor(accounting): remove unused spend log conversion API * feat(accounting): add PricingSnapshot, delivery enums and projection dimensions Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * feat(db): add LiteLLM_AccountingEvent and LiteLLM_AccountingDelivery ledger tables Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * feat(accounting): export JSON schema and semantic fixtures for the Rust cost crate Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * refactor(accounting): drop the schema and fixture export, the Python package is the contract Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(accounting): make tests/unit/accounting a package Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(accounting): validate pricing snapshots and ledger rows * fix(otel): register the accounting ledger tables as span tables Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(accounting): address review findings on dimensions, token subsets, multipliers and agent target Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Joshua Valluru <326636767+joshua-berri@users.noreply.github.com> Co-authored-by: joshua Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .github/workflows/test-unit.yml | 1 + .../migration.sql | 47 ++ .../litellm_proxy_extras/schema.prisma | 70 +++ litellm/accounting/__init__.py | 62 +++ litellm/accounting/measurements.py | 13 + litellm/accounting/protocols.py | 85 +++ litellm/accounting/types.py | 428 ++++++++++++++++ litellm/integrations/otel/model/spans.py | 2 + litellm/proxy/schema.prisma | 70 +++ schema.prisma | 70 +++ .../database/test_accounting_ledger.py | 128 +++++ tests/unit/accounting/__init__.py | 0 tests/unit/accounting/test_measurements.py | 235 +++++++++ tests/unit/accounting/test_types.py | 484 ++++++++++++++++++ 14 files changed, 1695 insertions(+) create mode 100644 litellm-proxy-extras/litellm_proxy_extras/migrations/20261027000000_add_accounting_ledger_tables/migration.sql create mode 100644 litellm/accounting/__init__.py create mode 100644 litellm/accounting/measurements.py create mode 100644 litellm/accounting/protocols.py create mode 100644 litellm/accounting/types.py create mode 100644 tests/integration/database/test_accounting_ledger.py create mode 100644 tests/unit/accounting/__init__.py create mode 100644 tests/unit/accounting/test_measurements.py create mode 100644 tests/unit/accounting/test_types.py diff --git a/.github/workflows/test-unit.yml b/.github/workflows/test-unit.yml index e3980bd8f3b..2ddf21bca17 100644 --- a/.github/workflows/test-unit.yml +++ b/.github/workflows/test-unit.yml @@ -137,6 +137,7 @@ jobs: include: - shard: core-utils test-path: |- + tests/unit/accounting tests/unit/batch_completion tests/unit/decisions tests/unit/extraction diff --git a/litellm-proxy-extras/litellm_proxy_extras/migrations/20261027000000_add_accounting_ledger_tables/migration.sql b/litellm-proxy-extras/litellm_proxy_extras/migrations/20261027000000_add_accounting_ledger_tables/migration.sql new file mode 100644 index 00000000000..c0ad8692d36 --- /dev/null +++ b/litellm-proxy-extras/litellm_proxy_extras/migrations/20261027000000_add_accounting_ledger_tables/migration.sql @@ -0,0 +1,47 @@ +SET LOCAL lock_timeout = '10s'; + +CREATE TYPE "AccountingEventKind" AS ENUM ('inbound_usage', 'outbound_usage', 'cost'); + +CREATE TYPE "AccountingDeliveryState" AS ENUM ('pending', 'applied', 'parked'); + +CREATE TYPE "AccountingDeliveryTarget" AS ENUM ('LiteLLM_VerificationToken', 'LiteLLM_UserTable', 'LiteLLM_TeamTable', 'LiteLLM_TeamMembership', 'LiteLLM_OrganizationTable', 'LiteLLM_OrganizationMembership', 'LiteLLM_EndUserTable', 'LiteLLM_AgentsTable', 'LiteLLM_TagTable', 'LiteLLM_ProjectTable', 'LiteLLM_ModelAccessGroupBudgetTable', 'LiteLLM_BudgetTable', 'LiteLLM_BudgetWindowSpend', 'LiteLLM_DailyUserSpend', 'LiteLLM_DailyTeamSpend', 'LiteLLM_DailyOrganizationSpend', 'LiteLLM_DailyEndUserSpend', 'LiteLLM_DailyAgentSpend', 'LiteLLM_DailyTagSpend', 'LiteLLM_DailyModelUsage', 'LiteLLM_DailyToolSpend', 'LiteLLM_AutoRouterDailySpend'); + +-- CreateTable +CREATE TABLE IF NOT EXISTS "LiteLLM_AccountingEvent" ( + "event_id" TEXT NOT NULL, + "kind" "AccountingEventKind" NOT NULL, + "source_event_id" TEXT, + "schema_version" INTEGER NOT NULL, + "payload_hash" TEXT NOT NULL, + "occurred_at" TIMESTAMP(3) NOT NULL, + "recorded_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP, + "payload" JSONB NOT NULL, + + CONSTRAINT "LiteLLM_AccountingEvent_schema_version_check" CHECK ("schema_version" > 0), + CONSTRAINT "LiteLLM_AccountingEvent_pkey" PRIMARY KEY ("event_id") +); + +-- CreateTable +CREATE TABLE IF NOT EXISTS "LiteLLM_AccountingDelivery" ( + "event_id" TEXT NOT NULL, + "target" "AccountingDeliveryTarget" NOT NULL, + "version" INTEGER NOT NULL, + "state" "AccountingDeliveryState" NOT NULL DEFAULT 'pending', + "attempts" INTEGER NOT NULL DEFAULT 0, + "next_attempt_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP, + "applied_at" TIMESTAMP(3), + "last_error" TEXT, + + CONSTRAINT "LiteLLM_AccountingDelivery_version_check" CHECK ("version" > 0), + CONSTRAINT "LiteLLM_AccountingDelivery_attempts_check" CHECK ("attempts" >= 0), + CONSTRAINT "LiteLLM_AccountingDelivery_pkey" PRIMARY KEY ("event_id","target","version") +); + +-- CreateIndex +CREATE INDEX IF NOT EXISTS "LiteLLM_AccountingEvent_recorded_at_idx" ON "LiteLLM_AccountingEvent"("recorded_at"); + +-- CreateIndex +CREATE INDEX IF NOT EXISTS "LiteLLM_AccountingEvent_kind_source_event_id_idx" ON "LiteLLM_AccountingEvent"("kind", "source_event_id"); + +-- CreateIndex +CREATE INDEX IF NOT EXISTS "LiteLLM_AccountingDelivery_target_state_next_attempt_at_idx" ON "LiteLLM_AccountingDelivery"("target", "state", "next_attempt_at"); diff --git a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma index 4187f76c97c..2965174114b 100644 --- a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma +++ b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma @@ -2061,3 +2061,73 @@ model LiteLLM_TelemetryReport { @@index([window_end]) @@index([created_at]) } + +enum AccountingEventKind { + inbound_usage + outbound_usage + cost +} + +enum AccountingDeliveryState { + pending + applied + parked +} + +enum AccountingDeliveryTarget { + LiteLLM_VerificationToken + LiteLLM_UserTable + LiteLLM_TeamTable + LiteLLM_TeamMembership + LiteLLM_OrganizationTable + LiteLLM_OrganizationMembership + LiteLLM_EndUserTable + LiteLLM_AgentsTable + LiteLLM_TagTable + LiteLLM_ProjectTable + LiteLLM_ModelAccessGroupBudgetTable + LiteLLM_BudgetTable + LiteLLM_BudgetWindowSpend + LiteLLM_DailyUserSpend + LiteLLM_DailyTeamSpend + LiteLLM_DailyOrganizationSpend + LiteLLM_DailyEndUserSpend + LiteLLM_DailyAgentSpend + LiteLLM_DailyTagSpend + LiteLLM_DailyModelUsage + LiteLLM_DailyToolSpend + LiteLLM_AutoRouterDailySpend +} + +// Durable accounting ledger. One row per usage or cost event; payload holds the +// private event envelope versioned by schema_version. Usage rows retain request-scoped +// pricing context; cost rows include resolved rates and the finalized pricing snapshot. +model LiteLLM_AccountingEvent { + event_id String @id + kind AccountingEventKind + source_event_id String? + schema_version Int + payload_hash String + occurred_at DateTime + recorded_at DateTime @default(now()) + payload Json + + @@index([recorded_at]) + @@index([kind, source_event_id]) +} + +// One row per (event, target table, version): pending work for the projector and, +// once applied in the same transaction as the target write, the receipt that stops a retry. +model LiteLLM_AccountingDelivery { + event_id String + target AccountingDeliveryTarget + version Int + state AccountingDeliveryState @default(pending) + attempts Int @default(0) + next_attempt_at DateTime @default(now()) + applied_at DateTime? + last_error String? + + @@id([event_id, target, version]) + @@index([target, state, next_attempt_at]) +} diff --git a/litellm/accounting/__init__.py b/litellm/accounting/__init__.py new file mode 100644 index 00000000000..d4e56f61279 --- /dev/null +++ b/litellm/accounting/__init__.py @@ -0,0 +1,62 @@ +"""Public accounting contracts. Importing these values does not enable accounting.""" + +from litellm.accounting.measurements import measurements_for +from litellm.accounting.protocols import BudgetEnforcer, CostAggregator, CostCompiler, UsageMeter +from litellm.accounting.types import ( + AdmissionDecision, + AggregateQuery, + AggregateRow, + CostEvent, + DeliveryState, + DeliveryTarget, + Dimension, + DimensionFilter, + DimensionName, + EventKind, + Granularity, + InboundUsageEvent, + Measurement, + Metric, + MetricName, + Multiplier, + NamedCost, + OutboundUsageEvent, + Outcome, + PricedQuantity, + PricingSnapshot, + Rate, + Unpriced, + validate_aggregate_rows, +) + +__all__ = [ + "AdmissionDecision", + "AggregateQuery", + "AggregateRow", + "BudgetEnforcer", + "CostAggregator", + "CostCompiler", + "CostEvent", + "DeliveryState", + "DeliveryTarget", + "Dimension", + "DimensionFilter", + "DimensionName", + "EventKind", + "Granularity", + "InboundUsageEvent", + "Measurement", + "Metric", + "MetricName", + "Multiplier", + "NamedCost", + "OutboundUsageEvent", + "Outcome", + "PricedQuantity", + "PricingSnapshot", + "Rate", + "Unpriced", + "UsageMeter", + "measurements_for", + "validate_aggregate_rows", +] diff --git a/litellm/accounting/measurements.py b/litellm/accounting/measurements.py new file mode 100644 index 00000000000..83011a280da --- /dev/null +++ b/litellm/accounting/measurements.py @@ -0,0 +1,13 @@ +from typing import Final + +from litellm.accounting.types import CostEvent, InboundUsageEvent, Measurement, OutboundUsageEvent + + +def measurements_for(event: InboundUsageEvent | OutboundUsageEvent | CostEvent) -> tuple[Measurement, ...]: + """Return the complete contribution set for one event, ready for one apply call.""" + metrics: Final = event.charges if isinstance(event, CostEvent) else event.metrics + occurred_at: Final = event.occurred_at if isinstance(event, CostEvent) else event.started_at + return tuple( + Measurement(source_event_id=event.event_id, occurred_at=occurred_at, dimensions=event.dimensions, metric=metric) + for metric in metrics or () + ) diff --git a/litellm/accounting/protocols.py b/litellm/accounting/protocols.py new file mode 100644 index 00000000000..8955f56e7a0 --- /dev/null +++ b/litellm/accounting/protocols.py @@ -0,0 +1,85 @@ +"""Contracts only. Implementations own persistence, pricing, projections and live budgets. + +The event row's ``payload`` is a private envelope versioned by the row's ``schema_version``. +Usage envelopes retain the public event and request-scoped pricing context: quantities with units, +model/provider, requested and billed tier/region when known, and per-request/deployment overrides. +These inputs must survive request teardown. Cost-map rates are resolved at compilation, not capture. +Cost envelopes retain the public CostEvent, including its required finalized pricing snapshot. +T5/T6 own envelope serialization; this public API does not introduce a second persisted schema. +""" + +from typing import Protocol + +from litellm.accounting.types import ( + AdmissionDecision, + AggregateQuery, + AggregateRow, + CostEvent, + Dimension, + InboundUsageEvent, + Measurement, + Metric, + OutboundUsageEvent, + Unpriced, +) + + +class UsageMeter(Protocol): + async def record(self, event: InboundUsageEvent | OutboundUsageEvent) -> None: + """Acknowledge commit, no-op identical IDs and reject conflicting content. + + Persist the event and its request-scoped pricing context before acknowledging. + HTTP acknowledgement is a separate policy. Unknown usage remains None. + """ + ... + + +class CostCompiler(Protocol): + async def compile(self, event: OutboundUsageEvent) -> CostEvent | Unpriced: + """Reuse the committed charge or finalize one charge and its delivery intent. + + Resolve captured request-scoped context by event identity and cost-map rates + at compilation. Persist the resolved inputs in the required snapshot. Replay + reuses the committed result and never consumes free allowance again. + Missing pricing is Unpriced, never an inferred zero. + """ + ... + + +class CostAggregator(Protocol): + async def apply(self, measurements: tuple[Measurement, ...]) -> None: + """Apply one source event's complete contribution set per target/version. + + Each target commits all contributions and its receipt atomically. A receipt + for the first metric must not suppress the rest. Usage need not wait for cost. + """ + ... + + async def query(self, query: AggregateQuery) -> tuple[AggregateRow, ...]: + """Return disjoint [start, end) rows per group using existing daily projections. + + Filters combine with AND, values within a filter with IN. DAY emits daily + buckets; None emits whole-interval totals. Reject unsupported metric/scope + combinations. Different groups may overlap, including non-additive tags. + """ + ... + + +class BudgetEnforcer(Protocol): + async def reserve( + self, reservation_id: str, dimensions: tuple[Dimension, ...], estimated: tuple[Metric, ...] + ) -> AdmissionDecision: + """Apply existing admission policies and estimates; duplicate IDs add no hold.""" + ... + + async def settle(self, reservation_id: str | None, cost_event: CostEvent) -> None: + """Settle the committed charge with atomic deduplication and hold mutation. + + Preserve scope/window attribution and retry unfinished settlement after + restart. SQL projection receipts do not provide live-counter atomicity. + """ + ... + + async def release(self, reservation_id: str) -> None: + """Idempotently release unused exposure at the request boundary, preserving charges.""" + ... diff --git a/litellm/accounting/types.py b/litellm/accounting/types.py new file mode 100644 index 00000000000..d0be29bf5b9 --- /dev/null +++ b/litellm/accounting/types.py @@ -0,0 +1,428 @@ +"""Accounting values use UTC instants, decimal strings on the wire and USD charges.""" + +from collections.abc import Mapping +from datetime import datetime, timezone +from decimal import Decimal +from enum import Enum +from typing import Annotated, Final, Literal, TypeAlias + +from pydantic import AfterValidator, AwareDatetime, BaseModel, ConfigDict, Field, model_validator +from typing_extensions import Self + +from litellm.types.utils import CallTypes + +_CONFIG: Final = ConfigDict(frozen=True, extra="forbid", validate_default=True) +Identifier: TypeAlias = Annotated[str, Field(strict=True, min_length=1, pattern=r"\S")] +Quantity: TypeAlias = Annotated[Decimal, Field(strict=True, ge=0, allow_inf_nan=False)] + + +def _utc(value: datetime) -> datetime: + return value.astimezone(timezone.utc) + + +Instant: TypeAlias = Annotated[AwareDatetime, AfterValidator(_utc)] + + +class MetricName(str, Enum): + """Reasoning tokens are included in completion tokens; above-1hr writes are included in cache creation tokens.""" + + REQUESTS = "requests" + SUCCESSFUL_REQUESTS = "successful_requests" + FAILED_REQUESTS = "failed_requests" + PROMPT_TOKENS = "prompt_tokens" + COMPLETION_TOKENS = "completion_tokens" + REASONING_TOKENS = "reasoning_tokens" + TOTAL_TOKENS = "total_tokens" + CACHE_READ_INPUT_TOKENS = "cache_read_input_tokens" + CACHE_CREATION_INPUT_TOKENS = "cache_creation_input_tokens" + CACHE_CREATION_INPUT_TOKENS_ABOVE_1HR = "cache_creation_input_tokens_above_1hr" + SPEND = "spend" + + +class DimensionName(str, Enum): + KEY = "key" + USER = "user" + TEAM = "team" + ORGANIZATION = "organization" + END_USER = "end_user" + MODEL = "model" + MODEL_GROUP = "model_group" + DEPLOYMENT = "deployment" + PROVIDER = "provider" + TOOL = "tool" + AGENT = "agent" + PROJECT = "project" + MODEL_ACCESS_GROUP = "model_access_group" + API_BASE = "api_base" + SERVICE_TIER = "service_tier" + TAG = "tag" + + +class Outcome(str, Enum): + SUCCEEDED = "succeeded" + FAILED = "failed" + CANCELLED = "cancelled" + + +class Granularity(str, Enum): + DAY = "day" + + +class EventKind(str, Enum): + INBOUND_USAGE = "inbound_usage" + OUTBOUND_USAGE = "outbound_usage" + COST = "cost" + + +class DeliveryState(str, Enum): + PENDING = "pending" + APPLIED = "applied" + PARKED = "parked" + + +class DeliveryTarget(str, Enum): + """One member per existing table the projector writes; each is applied and receipted independently.""" + + KEY_SPEND = "LiteLLM_VerificationToken" + USER_SPEND = "LiteLLM_UserTable" + TEAM_SPEND = "LiteLLM_TeamTable" + TEAM_MEMBER_SPEND = "LiteLLM_TeamMembership" + ORGANIZATION_SPEND = "LiteLLM_OrganizationTable" + ORGANIZATION_MEMBER_SPEND = "LiteLLM_OrganizationMembership" + END_USER_SPEND = "LiteLLM_EndUserTable" + AGENT_SPEND = "LiteLLM_AgentsTable" + TAG_SPEND = "LiteLLM_TagTable" + PROJECT_SPEND = "LiteLLM_ProjectTable" + MODEL_ACCESS_GROUP_SPEND = "LiteLLM_ModelAccessGroupBudgetTable" + BUDGET_SPEND = "LiteLLM_BudgetTable" + BUDGET_WINDOW_SPEND = "LiteLLM_BudgetWindowSpend" + DAILY_USER_SPEND = "LiteLLM_DailyUserSpend" + DAILY_TEAM_SPEND = "LiteLLM_DailyTeamSpend" + DAILY_ORGANIZATION_SPEND = "LiteLLM_DailyOrganizationSpend" + DAILY_END_USER_SPEND = "LiteLLM_DailyEndUserSpend" + DAILY_AGENT_SPEND = "LiteLLM_DailyAgentSpend" + DAILY_TAG_SPEND = "LiteLLM_DailyTagSpend" + DAILY_MODEL_USAGE = "LiteLLM_DailyModelUsage" + DAILY_TOOL_SPEND = "LiteLLM_DailyToolSpend" + AUTOROUTER_DAILY_SPEND = "LiteLLM_AutoRouterDailySpend" + + +_SUBSET_PAIRS: Final = ( + (MetricName.REASONING_TOKENS, MetricName.COMPLETION_TOKENS), + (MetricName.CACHE_CREATION_INPUT_TOKENS_ABOVE_1HR, MetricName.CACHE_CREATION_INPUT_TOKENS), +) + + +def _check_subsets(values: Mapping[str, Decimal]) -> None: + for part, whole in _SUBSET_PAIRS: + if part.value in values and whole.value in values and values[part.value] > values[whole.value]: + raise ValueError(f"{part.value} cannot exceed {whole.value}") + + +class Metric(BaseModel): + model_config = _CONFIG + + name: MetricName + value: Quantity + unit: Literal["count", "USD"] = "count" + + @model_validator(mode="after") + def validate_unit(self) -> Self: + if (self.name is MetricName.SPEND) != (self.unit == "USD"): + raise ValueError("SPEND requires USD; usage metrics require count") + if self.unit == "count" and self.value != self.value.to_integral_value(): + raise ValueError("Count metrics require integral values") + return self + + +class Dimension(BaseModel): + model_config = _CONFIG + + name: DimensionName + value: Identifier + + +_REPEATABLE_DIMENSIONS: Final = frozenset({DimensionName.TAG, DimensionName.MODEL_ACCESS_GROUP}) + + +def _dimensions(values: tuple[Dimension, ...]) -> tuple[Dimension, ...]: + if len(frozenset(values)) != len(values): + raise ValueError("Duplicate dimension values") + single_names: Final = tuple(value.name for value in values if value.name not in _REPEATABLE_DIMENSIONS) + if len(frozenset(single_names)) != len(single_names): + raise ValueError("Only tag and model access group dimensions may have multiple values") + return values + + +Dimensions: TypeAlias = Annotated[tuple[Dimension, ...], AfterValidator(_dimensions)] +_CALLER_METRICS: Final = frozenset((MetricName.REQUESTS, MetricName.SUCCESSFUL_REQUESTS, MetricName.FAILED_REQUESTS)) + + +def _metrics(values: tuple[Metric, ...]) -> tuple[Metric, ...]: + if len(frozenset(value.name for value in values)) != len(values): + raise ValueError("Duplicate metric names") + return values + + +Metrics: TypeAlias = Annotated[tuple[Metric, ...], AfterValidator(_metrics)] + + +def _interval(start: datetime, end: datetime, *, allow_empty: bool = False) -> None: + if end < start or (end == start and not allow_empty): + raise ValueError("End must follow start") + + +class InboundUsageEvent(BaseModel): + model_config = _CONFIG + + event_id: Identifier + request_id: Identifier + call_type: CallTypes + started_at: Instant + ended_at: Instant + outcome: Outcome + dimensions: Dimensions + metrics: Metrics + + @model_validator(mode="after") + def validate_caller_activity(self) -> Self: + _interval(self.started_at, self.ended_at, allow_empty=True) + if any(metric.name not in _CALLER_METRICS for metric in self.metrics): + raise ValueError("Inbound events own only caller request metrics") + if not any(metric.name is MetricName.REQUESTS and metric.value == 1 for metric in self.metrics): + raise ValueError("Inbound events count exactly one caller request") + for metric in self.metrics: + if metric.name is MetricName.SUCCESSFUL_REQUESTS and metric.value != int(self.outcome is Outcome.SUCCEEDED): + raise ValueError("Successful request count must match outcome") + if metric.name is MetricName.FAILED_REQUESTS and metric.value != int(self.outcome is not Outcome.SUCCEEDED): + raise ValueError("Failed request count must match outcome") + return self + + +class OutboundUsageEvent(BaseModel): + model_config = _CONFIG + + event_id: Identifier + attempt_id: Identifier + request_id: Identifier | None + retry_of_attempt_id: Identifier | None + call_type: CallTypes + started_at: Instant + ended_at: Instant + outcome: Outcome + dimensions: Dimensions + metrics: Metrics | None + + @model_validator(mode="after") + def validate_attempt_usage(self) -> Self: + _interval(self.started_at, self.ended_at, allow_empty=True) + if self.retry_of_attempt_id == self.attempt_id: + raise ValueError("An attempt cannot retry itself") + if self.metrics is not None and any( + metric.name in _CALLER_METRICS or metric.name is MetricName.SPEND for metric in self.metrics + ): + raise ValueError("Outbound events own only attempt usage metrics") + if self.metrics is not None: + _check_subsets({metric.name.value: metric.value for metric in self.metrics}) + return self + + +PricingUnit: TypeAlias = Literal["token", "second", "character", "image", "request"] + + +class Rate(BaseModel): + model_config = _CONFIG + + measure: Identifier + unit: PricingUnit + usd_per_unit: Quantity + + +class PricedQuantity(BaseModel): + """A pricing input in its billing unit. Only seconds may be fractional.""" + + model_config = _CONFIG + + measure: Identifier + unit: PricingUnit + value: Quantity + + @model_validator(mode="after") + def validate_discrete_quantity(self) -> Self: + if self.unit != "second" and self.value != self.value.to_integral_value(): + raise ValueError("Discrete priced quantities require integral values") + return self + + +class Multiplier(BaseModel): + """A regional or tier factor applied on top of the priced key, e.g. data_residency=eu or vertex_location=us-east5.""" + + model_config = _CONFIG + + name: Identifier + value: Identifier + factor: Annotated[Decimal, Field(strict=True, gt=0, allow_inf_nan=False)] + + +class NamedCost(BaseModel): + """A free-form extra charge by name, e.g. azure_model_router_flat_cost.""" + + model_config = _CONFIG + + name: Identifier + amount: Quantity + + +class PricingSnapshot(BaseModel): + """Retained inputs and result of authoritative compilation, reproducible after cost-map rates change. + + Request-scoped inputs and overrides come from the usage row's versioned payload. Cost-map rates are + resolved at compilation, not frozen at capture. Retries reuse the committed charge and snapshot. + Component names match ``CostBreakdown``: reasoning_cost is part of output_cost, not an additional charge. + Component sums and float tolerances belong to reconciliation, not this structural contract. + """ + + model_config = _CONFIG + + schema_version: Literal[1] = 1 + calculator_version: Identifier + model: Identifier + provider: Identifier | None + pricing_source: Literal["cost_map", "deployment", "per_request", "upstream_header"] + priced_model_key: Identifier + multipliers: tuple[Multiplier, ...] = () + rates: tuple[Rate, ...] = () + usage: tuple[PricedQuantity, ...] = () + input_cost: Quantity | None = None + output_cost: Quantity | None = None + cache_read_cost: Quantity | None = None + cache_creation_cost: Quantity | None = None + reasoning_cost: Quantity | None = None + tool_usage_cost: Quantity | None = None + guardrail_cost: Quantity | None = None + additional_costs: tuple[NamedCost, ...] = () + original_cost: Quantity | None = None + discount_percent: Quantity | None = None + discount_amount: Quantity | None = None + margin_percent: Quantity | None = None + margin_fixed_amount: Quantity | None = None + margin_total_amount: Quantity | None = None + upstream_reported_cost: Quantity | None = None + total_cost: Quantity + + @model_validator(mode="after") + def validate_snapshot(self) -> Self: + rate_keys: Final = frozenset((rate.measure, rate.unit) for rate in self.rates) + quantity_keys: Final = frozenset((quantity.measure, quantity.unit) for quantity in self.usage) + if len(rate_keys) != len(self.rates): + raise ValueError("Duplicate rate measures and units") + if len(quantity_keys) != len(self.usage): + raise ValueError("Duplicate priced quantity measures and units") + if not rate_keys.issubset(quantity_keys): + raise ValueError("Every rate requires one priced quantity with the same measure and unit") + if len(frozenset(cost.name for cost in self.additional_costs)) != len(self.additional_costs): + raise ValueError("Duplicate additional cost names") + if len(frozenset(multiplier.name for multiplier in self.multipliers)) != len(self.multipliers): + raise ValueError("Duplicate multiplier names") + _check_subsets({quantity.measure: quantity.value for quantity in self.usage if quantity.unit == "token"}) + return self + + +class CostEvent(BaseModel): + model_config = _CONFIG + + event_id: Identifier + usage_event_id: Identifier + occurred_at: Instant + dimensions: Dimensions + charges: Metrics + snapshot: PricingSnapshot + + @model_validator(mode="after") + def validate_charge(self) -> Self: + if len(self.charges) != 1 or self.charges[0].name is not MetricName.SPEND: + raise ValueError("A cost event requires one authoritative USD spend metric") + if self.event_id == self.usage_event_id: + raise ValueError("Cost and usage events require distinct identities") + if self.snapshot.total_cost != self.charges[0].value: + raise ValueError("Snapshot total must equal the authoritative spend") + return self + + +class Unpriced(BaseModel): + model_config = _CONFIG + + usage_event_id: Identifier + reason: Identifier + + +class Measurement(BaseModel): + model_config = _CONFIG + + source_event_id: Identifier + occurred_at: Instant + dimensions: Dimensions + metric: Metric + + +class DimensionFilter(BaseModel): + model_config = _CONFIG + + name: DimensionName + values: Annotated[tuple[Identifier, ...], Field(min_length=1)] + + +class AggregateQuery(BaseModel): + model_config = _CONFIG + + metric: MetricName + filters: tuple[DimensionFilter, ...] + group_by: tuple[DimensionName, ...] + start: Instant + end: Instant + granularity: Granularity | None + + @model_validator(mode="after") + def validate_query(self) -> Self: + _interval(self.start, self.end) + if len(frozenset(item.name for item in self.filters)) != len(self.filters): + raise ValueError("Use one IN filter per dimension; filters combine with AND") + if len(frozenset(self.group_by)) != len(self.group_by): + raise ValueError("Duplicate group-by dimensions") + if any( + (instant.hour, instant.minute, instant.second, instant.microsecond) != (0, 0, 0, 0) + for instant in (self.start, self.end) + ): + raise ValueError("Initial reporting queries require UTC day boundaries") + return self + + +class AggregateRow(BaseModel): + model_config = _CONFIG + + start: Instant + end: Instant + dimensions: Dimensions + value: Quantity + + @model_validator(mode="after") + def validate_bounds(self) -> Self: + _interval(self.start, self.end) + return self + + +def validate_aggregate_rows(rows: tuple[AggregateRow, ...]) -> tuple[AggregateRow, ...]: + """Validate [start, end) buckets per group; overlapping tag groups are not additive.""" + groups: Final = frozenset(frozenset(row.dimensions) for row in rows) + for group in groups: + ordered = sorted((row for row in rows if frozenset(row.dimensions) == group), key=lambda row: row.start) + if any(left.end > right.start for left, right in zip(ordered, ordered[1:])): + raise ValueError("Aggregate buckets overlap within a dimension group") + return rows + + +class AdmissionDecision(BaseModel): + model_config = _CONFIG + + allowed: bool + reason: Identifier | None = None diff --git a/litellm/integrations/otel/model/spans.py b/litellm/integrations/otel/model/spans.py index d8be607150c..e423b6698e6 100644 --- a/litellm/integrations/otel/model/spans.py +++ b/litellm/integrations/otel/model/spans.py @@ -338,6 +338,8 @@ _PRISMA_MODELS: Final[frozenset[str]] = frozenset( "LiteLLM_ToolTable", "LiteLLM_TelemetryReport", "LiteLLM_AccessGroupTable", + "LiteLLM_AccountingDelivery", + "LiteLLM_AccountingEvent", "LiteLLM_ClaudeCodePluginTable", "LiteLLM_MemoryTable", "LiteLLM_AdaptiveRouterState", diff --git a/litellm/proxy/schema.prisma b/litellm/proxy/schema.prisma index 4187f76c97c..2965174114b 100644 --- a/litellm/proxy/schema.prisma +++ b/litellm/proxy/schema.prisma @@ -2061,3 +2061,73 @@ model LiteLLM_TelemetryReport { @@index([window_end]) @@index([created_at]) } + +enum AccountingEventKind { + inbound_usage + outbound_usage + cost +} + +enum AccountingDeliveryState { + pending + applied + parked +} + +enum AccountingDeliveryTarget { + LiteLLM_VerificationToken + LiteLLM_UserTable + LiteLLM_TeamTable + LiteLLM_TeamMembership + LiteLLM_OrganizationTable + LiteLLM_OrganizationMembership + LiteLLM_EndUserTable + LiteLLM_AgentsTable + LiteLLM_TagTable + LiteLLM_ProjectTable + LiteLLM_ModelAccessGroupBudgetTable + LiteLLM_BudgetTable + LiteLLM_BudgetWindowSpend + LiteLLM_DailyUserSpend + LiteLLM_DailyTeamSpend + LiteLLM_DailyOrganizationSpend + LiteLLM_DailyEndUserSpend + LiteLLM_DailyAgentSpend + LiteLLM_DailyTagSpend + LiteLLM_DailyModelUsage + LiteLLM_DailyToolSpend + LiteLLM_AutoRouterDailySpend +} + +// Durable accounting ledger. One row per usage or cost event; payload holds the +// private event envelope versioned by schema_version. Usage rows retain request-scoped +// pricing context; cost rows include resolved rates and the finalized pricing snapshot. +model LiteLLM_AccountingEvent { + event_id String @id + kind AccountingEventKind + source_event_id String? + schema_version Int + payload_hash String + occurred_at DateTime + recorded_at DateTime @default(now()) + payload Json + + @@index([recorded_at]) + @@index([kind, source_event_id]) +} + +// One row per (event, target table, version): pending work for the projector and, +// once applied in the same transaction as the target write, the receipt that stops a retry. +model LiteLLM_AccountingDelivery { + event_id String + target AccountingDeliveryTarget + version Int + state AccountingDeliveryState @default(pending) + attempts Int @default(0) + next_attempt_at DateTime @default(now()) + applied_at DateTime? + last_error String? + + @@id([event_id, target, version]) + @@index([target, state, next_attempt_at]) +} diff --git a/schema.prisma b/schema.prisma index 4187f76c97c..2965174114b 100644 --- a/schema.prisma +++ b/schema.prisma @@ -2061,3 +2061,73 @@ model LiteLLM_TelemetryReport { @@index([window_end]) @@index([created_at]) } + +enum AccountingEventKind { + inbound_usage + outbound_usage + cost +} + +enum AccountingDeliveryState { + pending + applied + parked +} + +enum AccountingDeliveryTarget { + LiteLLM_VerificationToken + LiteLLM_UserTable + LiteLLM_TeamTable + LiteLLM_TeamMembership + LiteLLM_OrganizationTable + LiteLLM_OrganizationMembership + LiteLLM_EndUserTable + LiteLLM_AgentsTable + LiteLLM_TagTable + LiteLLM_ProjectTable + LiteLLM_ModelAccessGroupBudgetTable + LiteLLM_BudgetTable + LiteLLM_BudgetWindowSpend + LiteLLM_DailyUserSpend + LiteLLM_DailyTeamSpend + LiteLLM_DailyOrganizationSpend + LiteLLM_DailyEndUserSpend + LiteLLM_DailyAgentSpend + LiteLLM_DailyTagSpend + LiteLLM_DailyModelUsage + LiteLLM_DailyToolSpend + LiteLLM_AutoRouterDailySpend +} + +// Durable accounting ledger. One row per usage or cost event; payload holds the +// private event envelope versioned by schema_version. Usage rows retain request-scoped +// pricing context; cost rows include resolved rates and the finalized pricing snapshot. +model LiteLLM_AccountingEvent { + event_id String @id + kind AccountingEventKind + source_event_id String? + schema_version Int + payload_hash String + occurred_at DateTime + recorded_at DateTime @default(now()) + payload Json + + @@index([recorded_at]) + @@index([kind, source_event_id]) +} + +// One row per (event, target table, version): pending work for the projector and, +// once applied in the same transaction as the target write, the receipt that stops a retry. +model LiteLLM_AccountingDelivery { + event_id String + target AccountingDeliveryTarget + version Int + state AccountingDeliveryState @default(pending) + attempts Int @default(0) + next_attempt_at DateTime @default(now()) + applied_at DateTime? + last_error String? + + @@id([event_id, target, version]) + @@index([target, state, next_attempt_at]) +} diff --git a/tests/integration/database/test_accounting_ledger.py b/tests/integration/database/test_accounting_ledger.py new file mode 100644 index 00000000000..089ed872dee --- /dev/null +++ b/tests/integration/database/test_accounting_ledger.py @@ -0,0 +1,128 @@ +import os +import subprocess +import sys +from collections.abc import Iterator +from pathlib import Path +from typing import Final + +import psycopg +import pytest +from psycopg import sql + +from litellm.accounting import DeliveryState, DeliveryTarget, EventKind +from tests.integration._support.database import scratch_database + +_SCHEMA: Final = Path(__file__).resolve().parents[3] / "litellm-proxy-extras/litellm_proxy_extras/schema.prisma" +_EVENT_INSERT: Final = ( + 'INSERT INTO "LiteLLM_AccountingEvent" ' + "(event_id, kind, schema_version, payload_hash, occurred_at, payload) " + "VALUES ('event', %s, %s, 'hash', '2026-01-01', '{}') RETURNING kind::text, schema_version" +) +_DELIVERY_INSERT: Final = ( + 'INSERT INTO "LiteLLM_AccountingDelivery" (event_id, target, version, state, attempts) ' + "VALUES ('event', %s, %s, %s, %s) RETURNING target::text, version, state::text, attempts" +) + + +@pytest.fixture(scope="module") +def ledger_database() -> Iterator[str]: + with scratch_database("accounting") as database_url: + result: Final = subprocess.run( + [sys.executable, "-I", "-m", "prisma", "migrate", "deploy", "--schema", str(_SCHEMA)], + env={**os.environ, "DATABASE_URL": database_url}, + capture_output=True, + text=True, + timeout=300, + ) + assert result.returncode == 0, result.stdout + result.stderr + yield database_url + + +@pytest.fixture +def ledger(ledger_database: str) -> Iterator[psycopg.Connection[tuple[object, ...]]]: + with psycopg.connect(ledger_database) as connection: + try: + yield connection + finally: + connection.rollback() + + +@pytest.mark.parametrize("kind", tuple(EventKind)) +def test_ledger_accepts_every_public_event_kind( + ledger: psycopg.Connection[tuple[object, ...]], kind: EventKind +) -> None: + row: Final = ledger.execute(_EVENT_INSERT, (kind.value, 1)).fetchone() + assert row == (kind.value, 1) + + +@pytest.mark.parametrize("target", tuple(DeliveryTarget)) +@pytest.mark.parametrize("state", tuple(DeliveryState)) +def test_delivery_accepts_every_public_target_and_state( + ledger: psycopg.Connection[tuple[object, ...]], target: DeliveryTarget, state: DeliveryState +) -> None: + row: Final = ledger.execute(_DELIVERY_INSERT, (target.value, 1, state.value, 0)).fetchone() + assert row == (target.value, 1, state.value, 0) + + +@pytest.mark.parametrize( + ("enum_name", "values"), + [ + ("AccountingEventKind", tuple(value.value for value in EventKind)), + ("AccountingDeliveryState", tuple(value.value for value in DeliveryState)), + ("AccountingDeliveryTarget", tuple(value.value for value in DeliveryTarget)), + ], +) +def test_migrated_enums_match_the_public_contract( + ledger: psycopg.Connection[tuple[object, ...]], enum_name: str, values: tuple[str, ...] +) -> None: + rows: Final = ledger.execute( + "SELECT e.enumlabel FROM pg_enum e JOIN pg_type t ON t.oid = e.enumtypid WHERE t.typname = %s", + (enum_name,), + ).fetchall() + assert frozenset(row[0] for row in rows) == frozenset(values) + + +@pytest.mark.parametrize("version", [0, -1]) +def test_event_rejects_nonpositive_schema_versions( + ledger: psycopg.Connection[tuple[object, ...]], version: int +) -> None: + with pytest.raises(psycopg.errors.CheckViolation, match="AccountingEvent_schema_version_check"): + ledger.execute(_EVENT_INSERT, (EventKind.OUTBOUND_USAGE.value, version)) + + +@pytest.mark.parametrize( + ("version", "attempts", "constraint"), + [ + (0, 0, "AccountingDelivery_version_check"), + (-1, 0, "AccountingDelivery_version_check"), + (1, -1, "AccountingDelivery_attempts_check"), + ], +) +def test_delivery_rejects_invalid_versions_and_attempts( + ledger: psycopg.Connection[tuple[object, ...]], version: int, attempts: int, constraint: str +) -> None: + with pytest.raises(psycopg.errors.CheckViolation, match=constraint): + ledger.execute(_DELIVERY_INSERT, (DeliveryTarget.KEY_SPEND.value, version, "pending", attempts)) + + +@pytest.mark.parametrize("column", ["state", "target"]) +def test_delivery_rejects_unknown_enum_values(ledger: psycopg.Connection[tuple[object, ...]], column: str) -> None: + ledger.execute(_DELIVERY_INSERT, (DeliveryTarget.KEY_SPEND.value, 1, "pending", 0)) + with pytest.raises(psycopg.errors.InvalidTextRepresentation, match="invalid input value for enum"): + ledger.execute( + sql.SQL('UPDATE "LiteLLM_AccountingDelivery" SET {} = %s').format(sql.Identifier(column)), ("invalid",) + ) + + +def test_event_rejects_unknown_kind(ledger: psycopg.Connection[tuple[object, ...]]) -> None: + with pytest.raises(psycopg.errors.InvalidTextRepresentation, match="invalid input value for enum"): + ledger.execute(_EVENT_INSERT, ("invalid", 1)) + + +def test_new_delivery_defaults_to_pending_with_no_attempts(ledger: psycopg.Connection[tuple[object, ...]]) -> None: + row: Final = ledger.execute( + 'INSERT INTO "LiteLLM_AccountingDelivery" (event_id, target, version) ' + "VALUES ('event', %s, 1) RETURNING state::text, attempts", + (DeliveryTarget.KEY_SPEND.value,), + ).fetchone() + assert row == ("pending", 0) diff --git a/tests/unit/accounting/__init__.py b/tests/unit/accounting/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/unit/accounting/test_measurements.py b/tests/unit/accounting/test_measurements.py new file mode 100644 index 00000000000..b65c3076c5d --- /dev/null +++ b/tests/unit/accounting/test_measurements.py @@ -0,0 +1,235 @@ +from datetime import datetime, timezone +from decimal import Decimal +from typing import Final + +import pytest + +from litellm.accounting import ( + CostEvent, + Dimension, + DimensionName, + InboundUsageEvent, + Measurement, + Metric, + MetricName, + Outcome, + OutboundUsageEvent, + PricingSnapshot, + measurements_for, +) +from litellm.types.utils import CallTypes + +_START: Final = datetime(2026, 1, 1, 23, 59, 59, tzinfo=timezone.utc) +_END: Final = datetime(2026, 1, 2, tzinfo=timezone.utc) +_DIMENSIONS: Final = ( + Dimension(name=DimensionName.TEAM, value="team"), + Dimension(name=DimensionName.TAG, value="alpha"), + Dimension(name=DimensionName.TAG, value="beta"), +) + + +def _snapshot(amount: Decimal) -> PricingSnapshot: + return PricingSnapshot( + calculator_version="test", + model="test-model", + provider=None, + pricing_source="upstream_header", + priced_model_key="test-model", + upstream_reported_cost=amount, + total_cost=amount, + ) + + +@pytest.mark.parametrize( + ("outcome", "outcome_metric"), + [ + (Outcome.SUCCEEDED, MetricName.SUCCESSFUL_REQUESTS), + (Outcome.FAILED, MetricName.FAILED_REQUESTS), + (Outcome.CANCELLED, MetricName.FAILED_REQUESTS), + ], +) +def test_caller_measurements_keep_both_request_and_outcome_counts(outcome, outcome_metric): + event = InboundUsageEvent( + event_id="caller-event", + request_id="caller", + call_type=CallTypes.acompletion, + started_at=_START, + ended_at=_END, + outcome=outcome, + dimensions=_DIMENSIONS, + metrics=(Metric(name=MetricName.REQUESTS, value=Decimal(1)), Metric(name=outcome_metric, value=Decimal(1))), + ) + assert measurements_for(event) == ( + Measurement( + source_event_id="caller-event", + occurred_at=_START, + dimensions=_DIMENSIONS, + metric=Metric(name=MetricName.REQUESTS, value=Decimal(1)), + ), + Measurement( + source_event_id="caller-event", + occurred_at=_START, + dimensions=_DIMENSIONS, + metric=Metric(name=outcome_metric, value=Decimal(1)), + ), + ) + + +def test_usage_measurements_keep_every_metric_in_the_start_time_bucket(): + event = OutboundUsageEvent( + event_id="usage-event", + attempt_id="attempt", + request_id="caller", + retry_of_attempt_id=None, + call_type=CallTypes.acompletion, + started_at=_START, + ended_at=_END, + outcome=Outcome.SUCCEEDED, + dimensions=_DIMENSIONS, + metrics=( + Metric(name=MetricName.PROMPT_TOKENS, value=Decimal(10)), + Metric(name=MetricName.COMPLETION_TOKENS, value=Decimal(2)), + Metric(name=MetricName.TOTAL_TOKENS, value=Decimal(12)), + Metric(name=MetricName.CACHE_READ_INPUT_TOKENS, value=Decimal(3)), + Metric(name=MetricName.CACHE_CREATION_INPUT_TOKENS, value=Decimal(1)), + ), + ) + assert measurements_for(event) == ( + Measurement( + source_event_id="usage-event", + occurred_at=_START, + dimensions=_DIMENSIONS, + metric=Metric(name=MetricName.PROMPT_TOKENS, value=Decimal(10)), + ), + Measurement( + source_event_id="usage-event", + occurred_at=_START, + dimensions=_DIMENSIONS, + metric=Metric(name=MetricName.COMPLETION_TOKENS, value=Decimal(2)), + ), + Measurement( + source_event_id="usage-event", + occurred_at=_START, + dimensions=_DIMENSIONS, + metric=Metric(name=MetricName.TOTAL_TOKENS, value=Decimal(12)), + ), + Measurement( + source_event_id="usage-event", + occurred_at=_START, + dimensions=_DIMENSIONS, + metric=Metric(name=MetricName.CACHE_READ_INPUT_TOKENS, value=Decimal(3)), + ), + Measurement( + source_event_id="usage-event", + occurred_at=_START, + dimensions=_DIMENSIONS, + metric=Metric(name=MetricName.CACHE_CREATION_INPUT_TOKENS, value=Decimal(1)), + ), + ) + + +@pytest.mark.parametrize("metrics", [None, ()]) +def test_unknown_or_empty_usage_does_not_fabricate_zero_measurements(metrics): + event = OutboundUsageEvent( + event_id="usage-event", + attempt_id="attempt", + request_id=None, + retry_of_attempt_id=None, + call_type=CallTypes.acompletion, + started_at=_START, + ended_at=_END, + outcome=Outcome.SUCCEEDED, + dimensions=(), + metrics=metrics, + ) + assert measurements_for(event) == () + + +@pytest.mark.parametrize("amount", [Decimal(0), Decimal("0.1234567890123456789")]) +def test_cost_measurements_preserve_authoritative_amount_identity_and_occurrence(amount): + event = CostEvent( + event_id="cost-event", + usage_event_id="usage-event", + occurred_at=_END, + dimensions=_DIMENSIONS, + charges=(Metric(name=MetricName.SPEND, value=amount, unit="USD"),), + snapshot=_snapshot(amount), + ) + assert measurements_for(event) == ( + Measurement( + source_event_id="cost-event", + occurred_at=_END, + dimensions=_DIMENSIONS, + metric=Metric(name=MetricName.SPEND, value=amount, unit="USD"), + ), + ) + + +def test_billed_retry_and_success_contribute_one_caller_request_and_both_charges(): + caller = InboundUsageEvent( + event_id="caller-event", + request_id="caller", + call_type=CallTypes.acompletion, + started_at=_START, + ended_at=_END, + outcome=Outcome.SUCCEEDED, + dimensions=_DIMENSIONS, + metrics=(Metric(name=MetricName.REQUESTS, value=Decimal(1)),), + ) + retry = OutboundUsageEvent( + event_id="retry-usage", + attempt_id="retry", + request_id="caller", + retry_of_attempt_id=None, + call_type=CallTypes.acompletion, + started_at=_START, + ended_at=_END, + outcome=Outcome.FAILED, + dimensions=_DIMENSIONS, + metrics=(Metric(name=MetricName.TOTAL_TOKENS, value=Decimal(4)),), + ) + success = OutboundUsageEvent( + event_id="success-usage", + attempt_id="success", + request_id="caller", + retry_of_attempt_id="retry", + call_type=CallTypes.acompletion, + started_at=_START, + ended_at=_END, + outcome=Outcome.SUCCEEDED, + dimensions=_DIMENSIONS, + metrics=(Metric(name=MetricName.TOTAL_TOKENS, value=Decimal(12)),), + ) + retry_cost = CostEvent( + event_id="retry-cost", + usage_event_id="retry-usage", + occurred_at=_START, + dimensions=_DIMENSIONS, + charges=(Metric(name=MetricName.SPEND, value=Decimal("0.025"), unit="USD"),), + snapshot=_snapshot(Decimal("0.025")), + ) + success_cost = CostEvent( + event_id="success-cost", + usage_event_id="success-usage", + occurred_at=_START, + dimensions=_DIMENSIONS, + charges=(Metric(name=MetricName.SPEND, value=Decimal("0.125"), unit="USD"),), + snapshot=_snapshot(Decimal("0.125")), + ) + measurements = ( + *measurements_for(caller), + *measurements_for(retry), + *measurements_for(success), + *measurements_for(retry_cost), + *measurements_for(success_cost), + ) + assert sum(m.metric.value for m in measurements if m.metric.name is MetricName.REQUESTS) == Decimal(1) + assert sum(m.metric.value for m in measurements if m.metric.name is MetricName.TOTAL_TOKENS) == Decimal(16) + assert sum(m.metric.value for m in measurements if m.metric.name is MetricName.SPEND) == Decimal("0.150") + assert {m.source_event_id for m in measurements} == { + "caller-event", + "retry-usage", + "success-usage", + "retry-cost", + "success-cost", + } diff --git a/tests/unit/accounting/test_types.py b/tests/unit/accounting/test_types.py new file mode 100644 index 00000000000..f645ee6c2cc --- /dev/null +++ b/tests/unit/accounting/test_types.py @@ -0,0 +1,484 @@ +import re +from collections.abc import Mapping +from datetime import datetime, timedelta, timezone +from decimal import Decimal +from enum import Enum +from pathlib import Path +from typing import Final + +import pytest +from pydantic import ValidationError + +from litellm.accounting import ( + AdmissionDecision, + AggregateQuery, + AggregateRow, + CostEvent, + DeliveryState, + DeliveryTarget, + Dimension, + DimensionFilter, + DimensionName, + EventKind, + Granularity, + InboundUsageEvent, + Metric, + MetricName, + Multiplier, + Outcome, + OutboundUsageEvent, + PricedQuantity, + PricingSnapshot, + Rate, + validate_aggregate_rows, +) +from litellm.accounting.types import PricingUnit +from litellm.types.utils import CallTypes + +_START: Final = datetime(2026, 1, 1, tzinfo=timezone.utc) +_END: Final = datetime(2026, 1, 2, tzinfo=timezone.utc) +_REQUEST: Final = Metric(name=MetricName.REQUESTS, value=Decimal(1)) +_TOKEN: Final = Metric(name=MetricName.PROMPT_TOKENS, value=Decimal(10)) +_SPEND: Final = Metric(name=MetricName.SPEND, value=Decimal("0.1234567890123456789"), unit="USD") +_TEAM: Final = Dimension(name=DimensionName.TEAM, value="team") + + +def _inbound(metrics=(_REQUEST,), outcome=Outcome.SUCCEEDED): + return InboundUsageEvent( + event_id="inbound:request", + request_id="request", + call_type=CallTypes.acompletion, + started_at=_START, + ended_at=_START, + outcome=outcome, + dimensions=(_TEAM,), + metrics=metrics, + ) + + +def _outbound(metrics=(_TOKEN,), **changes): + return OutboundUsageEvent.model_validate( + { + "event_id": "outbound:attempt", + "attempt_id": "attempt", + "request_id": "request", + "retry_of_attempt_id": None, + "call_type": CallTypes.acompletion, + "started_at": _START, + "ended_at": _END, + "outcome": Outcome.SUCCEEDED, + "dimensions": (_TEAM,), + "metrics": metrics, + **changes, + } + ) + + +def _query(**changes): + return AggregateQuery.model_validate( + { + "metric": MetricName.SPEND, + "filters": (DimensionFilter(name=DimensionName.TEAM, values=("one", "two")),), + "group_by": (DimensionName.MODEL,), + "start": _START, + "end": _END, + "granularity": Granularity.DAY, + **changes, + } + ) + + +def test_wire_values_preserve_decimal_precision_and_normalize_aware_times(): + offset = timezone(timedelta(hours=-8)) + event = CostEvent( + event_id="cost:attempt", + usage_event_id="outbound:attempt", + occurred_at=datetime(2025, 12, 31, 16, tzinfo=offset), + dimensions=(_TEAM,), + charges=(_SPEND,), + snapshot=_snapshot(), + ) + wire = event.model_dump(mode="json") + assert wire["occurred_at"] == "2026-01-01T00:00:00Z" + assert wire["charges"] == [{"name": "spend", "value": "0.1234567890123456789", "unit": "USD"}] + assert event.occurred_at == _START + assert CostEvent.model_validate_json(event.model_dump_json()).charges[0].value == _SPEND.value + + +@pytest.mark.parametrize( + ("name", "value", "unit", "message"), + [ + (MetricName.SPEND, Decimal(1), "count", "SPEND requires USD"), + (MetricName.REQUESTS, Decimal(1), "USD", "SPEND requires USD"), + (MetricName.PROMPT_TOKENS, Decimal("1.5"), "count", "integral"), + (MetricName.SPEND, Decimal("NaN"), "USD", "finite"), + (MetricName.SPEND, Decimal("Infinity"), "USD", "finite"), + (MetricName.SPEND, Decimal(-1), "USD", "greater than or equal"), + (MetricName.SPEND, 0.1, "USD", "instance of Decimal"), + ], +) +def test_metrics_reject_wrong_units_nonfinite_values_and_binary_float_inputs(name, value, unit, message): + with pytest.raises(ValidationError, match=message): + Metric(name=name, value=value, unit=unit) + + +@pytest.mark.parametrize("metrics", [(_TOKEN,), (_SPEND,), ()]) +def test_inbound_activity_cannot_own_usage_or_cost_or_omit_request_count(metrics): + with pytest.raises(ValidationError, match="Inbound events"): + _inbound(metrics) + + +@pytest.mark.parametrize( + ("outcome", "metric"), + [ + (Outcome.FAILED, Metric(name=MetricName.SUCCESSFUL_REQUESTS, value=Decimal(1))), + (Outcome.SUCCEEDED, Metric(name=MetricName.FAILED_REQUESTS, value=Decimal(1))), + ], +) +def test_inbound_outcome_counts_cannot_contradict_outcome(outcome, metric): + with pytest.raises(ValidationError, match="count must match outcome"): + _inbound((_REQUEST, metric), outcome) + + +@pytest.mark.parametrize("metrics", [(_REQUEST,), (_SPEND,)]) +def test_outbound_attempts_cannot_own_caller_counts_or_charges(metrics): + with pytest.raises(ValidationError, match="Outbound events own only"): + _outbound(metrics) + + +@pytest.mark.parametrize( + ("changes", "message"), + [ + ({"started_at": datetime(2026, 1, 1)}, "timezone"), + ({"ended_at": _START - timedelta(seconds=1)}, "End must follow start"), + ({"retry_of_attempt_id": "attempt"}, "cannot retry itself"), + ({"dimensions": (_TEAM, _TEAM)}, "Duplicate dimension values"), + ({"dimensions": (_TEAM, Dimension(name=DimensionName.TEAM, value="other"))}, "Only tag and model access group"), + ( + { + "metrics": ( + Metric(name=MetricName.COMPLETION_TOKENS, value=Decimal(5)), + Metric(name=MetricName.REASONING_TOKENS, value=Decimal(6)), + ) + }, + "reasoning_tokens cannot exceed completion_tokens", + ), + ( + { + "metrics": ( + Metric(name=MetricName.CACHE_CREATION_INPUT_TOKENS, value=Decimal(5)), + Metric(name=MetricName.CACHE_CREATION_INPUT_TOKENS_ABOVE_1HR, value=Decimal(6)), + ) + }, + "cache_creation_input_tokens_above_1hr cannot exceed cache_creation_input_tokens", + ), + ({"event_id": " "}, "pattern"), + ({"metrics": (_TOKEN, _TOKEN)}, "Duplicate metric names"), + ], +) +def test_outbound_events_reject_invalid_identity_time_and_duplicate_attribution(changes, message): + with pytest.raises(ValidationError, match=message): + _outbound(**changes) + + +def test_unknown_and_known_zero_usage_are_distinct_and_values_are_immutable(): + unknown = _outbound(None) + zero = _outbound((Metric(name=MetricName.PROMPT_TOKENS, value=Decimal(0)),)) + assert unknown.metrics is None + assert zero.metrics is not None and zero.metrics[0].value == 0 + assert _inbound().started_at == _inbound().ended_at + with pytest.raises(ValidationError, match="frozen"): + zero.attempt_id = "different" + with pytest.raises(ValidationError, match="frozen"): + zero.metrics[0].value = Decimal(10) + + +@pytest.mark.parametrize("charges", [(), (_TOKEN,), (_SPEND, _SPEND)]) +def test_cost_events_require_one_authoritative_spend(charges): + with pytest.raises(ValidationError, match=r"authoritative USD spend|Duplicate metric"): + CostEvent( + event_id="cost", + usage_event_id="usage", + occurred_at=_START, + dimensions=(), + charges=charges, + snapshot=_snapshot(), + ) + + +def test_cost_event_cannot_reuse_its_usage_identity(): + with pytest.raises(ValidationError, match="distinct identities"): + CostEvent( + event_id="usage", + usage_event_id="usage", + occurred_at=_START, + dimensions=(), + charges=(_SPEND,), + snapshot=_snapshot(), + ) + + +@pytest.mark.parametrize("granularity", [Granularity.DAY, None]) +def test_daily_and_whole_interval_queries_preserve_filter_and_group_semantics(granularity): + query = _query(granularity=granularity) + assert query.granularity is granularity + assert query.filters[0].values == ("one", "two") + assert query.group_by == (DimensionName.MODEL,) + assert query.start == _START and query.end == _END + + +@pytest.mark.parametrize( + ("changes", "message"), + [ + ({"end": _START}, "End must follow start"), + ({"start": _START + timedelta(hours=1)}, "UTC day boundaries"), + ({"end": _END + timedelta(microseconds=1)}, "UTC day boundaries"), + ({"granularity": "hour"}, "day"), + ({"group_by": (DimensionName.TEAM, DimensionName.TEAM)}, "Duplicate group-by"), + ({"filters": (DimensionFilter(name=DimensionName.TEAM, values=("one",)),) * 2}, "one IN filter"), + ({"filters": ({"name": "team", "values": ()},)}, "at least 1"), + ], +) +def test_queries_reject_unsupported_or_ambiguous_ranges_and_filters(changes, message): + with pytest.raises(ValidationError, match=message): + _query(**changes) + + +def test_rows_allow_adjacent_buckets_and_overlapping_distinct_groups(): + first = AggregateRow(start=_START, end=_END, dimensions=(_TEAM,), value=Decimal(1)) + second = AggregateRow(start=_END, end=_END + timedelta(days=1), dimensions=(_TEAM,), value=Decimal(2)) + other = AggregateRow(start=_START, end=_END, dimensions=(), value=Decimal(3)) + rows = (second, other, first) + assert validate_aggregate_rows(rows) is rows + assert validate_aggregate_rows(()) == () + + +def test_rows_reject_overlap_within_a_group_regardless_of_dimension_order(): + model = Dimension(name=DimensionName.MODEL, value="model") + first = AggregateRow(start=_START, end=_END, dimensions=(_TEAM, model), value=Decimal(1)) + overlapping = AggregateRow(start=_START, end=_END, dimensions=(model, _TEAM), value=Decimal(2)) + with pytest.raises(ValueError, match="overlap within a dimension group"): + validate_aggregate_rows((first, overlapping)) + + +def test_rows_reject_empty_intervals(): + with pytest.raises(ValidationError, match="End must follow start"): + AggregateRow(start=_START, end=_START, dimensions=(), value=Decimal(0)) + + +def test_admission_decision_preserves_a_denial_reason(): + decision = AdmissionDecision(allowed=False, reason="Team budget exhausted") + assert not decision.allowed + assert decision.reason == "Team budget exhausted" + + +def _snapshot(**changes: object) -> PricingSnapshot: + return PricingSnapshot.model_validate( + { + "calculator_version": "1.80.0", + "model": "gpt-4o", + "provider": "openai", + "pricing_source": "cost_map", + "priced_model_key": "openai/gpt-4o", + "multipliers": (Multiplier(name="data_residency", value="eu", factor=Decimal("1.1")),), + "rates": (Rate(measure="input", unit="token", usd_per_unit=Decimal("0.0000025")),), + "usage": (PricedQuantity(measure="input", unit="token", value=_TOKEN.value),), + "input_cost": Decimal("0.000025"), + "total_cost": _SPEND.value, + **changes, + } + ) + + +def test_cost_event_carries_a_snapshot_whose_total_matches_the_spend(): + event = CostEvent( + event_id="cost:attempt", + usage_event_id="outbound:attempt", + occurred_at=_START, + dimensions=(_TEAM,), + charges=(_SPEND,), + snapshot=_snapshot(), + ) + assert event.snapshot is not None + assert event.snapshot.total_cost == _SPEND.value + assert event.snapshot.multipliers[0].factor == Decimal("1.1") + with pytest.raises(ValidationError, match="Snapshot total must equal the authoritative spend"): + CostEvent( + event_id="cost:attempt", + usage_event_id="outbound:attempt", + occurred_at=_START, + dimensions=(_TEAM,), + charges=(_SPEND,), + snapshot=_snapshot(total_cost=Decimal("0.5")), + ) + + +@pytest.mark.parametrize( + ("changes", "message"), + ( + ( + { + "rates": ( + Rate(measure="input", unit="token", usd_per_unit=Decimal(1)), + Rate(measure="input", unit="token", usd_per_unit=Decimal(2)), + ) + }, + "Duplicate rate measures and units", + ), + ({"usage": (_SPEND,)}, "PricedQuantity"), + ({"usage": (_REQUEST,)}, "PricedQuantity"), + ({"pricing_source": "guess"}, "pricing_source"), + ({"multipliers": ({"name": "x", "value": "y", "factor": Decimal(0)},)}, "factor"), + ({"schema_version": 2}, "schema_version"), + ( + {"additional_costs": ({"name": "fee", "amount": Decimal(1)}, {"name": "fee", "amount": Decimal(2)})}, + "Duplicate additional cost names", + ), + ), +) +def test_snapshots_reject_ambiguous_rates_wrong_usage_and_unknown_sources(changes: Mapping[str, object], message: str): + with pytest.raises(ValidationError, match=message): + _snapshot(**changes) + + +@pytest.mark.parametrize( + ("unit", "value"), + [("token", "10"), ("second", "1.25"), ("character", "25"), ("image", "2"), ("request", "1")], +) +def test_snapshot_preserves_matching_rates_and_quantities_for_each_billing_unit(unit: PricingUnit, value: str) -> None: + quantity: Final = PricedQuantity(measure="input", unit=unit, value=Decimal(value)) + rate: Final = Rate(measure="input", unit=unit, usd_per_unit=Decimal("0.003")) + amount: Final = quantity.value * rate.usd_per_unit + snapshot: Final = _snapshot(rates=(rate,), usage=(quantity,), total_cost=amount) + restored: Final = PricingSnapshot.model_validate_json(snapshot.model_dump_json()) + assert restored == snapshot + assert restored.usage[0].value * restored.rates[0].usd_per_unit == amount + + +@pytest.mark.parametrize( + ("usage", "message"), + [ + ((), "Every rate requires"), + (({"measure": "output", "unit": "token", "value": Decimal(10)},), "Every rate requires"), + (({"measure": "input", "unit": "second", "value": Decimal(10)},), "Every rate requires"), + ( + ({"measure": "input", "unit": "token", "value": Decimal(10)},) * 2, + "Duplicate priced quantity", + ), + ], +) +def test_snapshot_rejects_missing_mismatched_or_duplicate_priced_quantities( + usage: tuple[Mapping[str, object], ...], message: str +) -> None: + with pytest.raises(ValidationError, match=message): + _snapshot(usage=usage) + + +@pytest.mark.parametrize("unit", ["token", "character", "image", "request"]) +def test_priced_counts_remain_integral(unit: PricingUnit) -> None: + with pytest.raises(ValidationError, match="integral"): + PricedQuantity(measure="input", unit=unit, value=Decimal("1.25")) + + +@pytest.mark.parametrize( + ("value", "message"), + [ + (Decimal(-1), "greater than or equal"), + (Decimal("NaN"), "finite"), + (Decimal("Infinity"), "finite"), + (1.25, "instance of Decimal"), + ], +) +def test_pricing_values_reject_invalid_amounts(value: Decimal | float, message: str) -> None: + with pytest.raises(ValidationError, match=message): + PricedQuantity.model_validate({"measure": "input", "unit": "second", "value": value}) + with pytest.raises(ValidationError, match=message): + Rate.model_validate({"measure": "input", "unit": "second", "usd_per_unit": value}) + + +@pytest.mark.parametrize("snapshot_fields", [{}, {"snapshot": None}]) +def test_cost_event_cannot_omit_its_derivation(snapshot_fields: Mapping[str, object]) -> None: + with pytest.raises(ValidationError, match="snapshot"): + CostEvent.model_validate( + { + "event_id": "cost", + "usage_event_id": "usage", + "occurred_at": _START, + "dimensions": (), + "charges": (_SPEND,), + **snapshot_fields, + } + ) + + +def test_upstream_charge_snapshot_preserves_reported_cost_without_inventing_rates() -> None: + snapshot: Final = _snapshot( + pricing_source="upstream_header", rates=(), usage=(), upstream_reported_cost=_SPEND.value + ) + restored: Final = PricingSnapshot.model_validate_json(snapshot.model_dump_json()) + assert restored.upstream_reported_cost == restored.total_cost == _SPEND.value + assert restored.rates == restored.usage == () + + +def test_subset_token_counts_round_trip_without_changing_totals() -> None: + metrics: Final = ( + Metric(name=MetricName.COMPLETION_TOKENS, value=Decimal(10)), + Metric(name=MetricName.REASONING_TOKENS, value=Decimal(3)), + Metric(name=MetricName.CACHE_CREATION_INPUT_TOKENS, value=Decimal(8)), + Metric(name=MetricName.CACHE_CREATION_INPUT_TOKENS_ABOVE_1HR, value=Decimal(2)), + ) + event: Final = _outbound(metrics=metrics) + assert OutboundUsageEvent.model_validate_json(event.model_dump_json()).metrics == metrics + + +@pytest.mark.parametrize( + "schema_path", + [ + "schema.prisma", + "litellm/proxy/schema.prisma", + "litellm-proxy-extras/litellm_proxy_extras/schema.prisma", + ], +) +@pytest.mark.parametrize( + ("enum_name", "python_enum"), + [ + ("AccountingEventKind", EventKind), + ("AccountingDeliveryState", DeliveryState), + ("AccountingDeliveryTarget", DeliveryTarget), + ], +) +def test_prisma_accepts_exactly_the_public_accounting_enum_values( + schema_path: str, enum_name: str, python_enum: type[Enum] +) -> None: + schema: Final = (Path(__file__).resolve().parents[3] / schema_path).read_text() + declaration: Final = re.search(r"enum " + enum_name + r"\s*\{([^}]+)\}", schema) + assert declaration is not None, f"Missing persisted enum {enum_name}" + assert frozenset(declaration[1].split()) == frozenset(member.value for member in python_enum) + + +def test_one_event_carries_every_matched_model_access_group(): + groups = ( + Dimension(name=DimensionName.MODEL_ACCESS_GROUP, value="gpt-group"), + Dimension(name=DimensionName.MODEL_ACCESS_GROUP, value="claude-group"), + ) + assert _outbound(dimensions=(_TEAM, *groups)).dimensions == (_TEAM, *groups) + with pytest.raises(ValidationError, match="Duplicate dimension values"): + _outbound(dimensions=(groups[0], groups[0])) + + +def test_snapshots_reject_duplicate_multipliers_and_contradictory_token_counts(): + multiplier = Multiplier(name="data_residency", value="eu", factor=Decimal("1.1")) + with pytest.raises(ValidationError, match="Duplicate multiplier names"): + _snapshot(multipliers=(multiplier, multiplier)) + with pytest.raises(ValidationError, match="reasoning_tokens cannot exceed completion_tokens"): + _snapshot( + usage=( + PricedQuantity(measure="input", unit="token", value=_TOKEN.value), + PricedQuantity(measure="completion_tokens", unit="token", value=Decimal(5)), + PricedQuantity(measure="reasoning_tokens", unit="token", value=Decimal(6)), + ) + ) + + +def test_agent_live_spend_is_a_delivery_target(): + assert DeliveryTarget.AGENT_SPEND.value == "LiteLLM_AgentsTable" From 66ddbaab8624703e0868a6366e994fce73b5cc90 Mon Sep 17 00:00:00 2001 From: Kerry Lu Date: Wed, 28 Oct 2026 14:48:40 -0700 Subject: [PATCH 6/7] fix(ci): spend logs guard requires the approval to cover the head's spend-log lines (#48904) * fix(ci): spend logs guard requires the approval to cover the head's spend-log lines An approving review survives later pushes (dismiss_stale_reviews_on_push is off on main), so a PR approved once could add more LiteLLM_SpendLogs schema or migration changes and the guard kept passing. The guard now reads the review's commit_id and passes only when the approval is on the head, or the head adds no spend-log line the approved commit did not already carry. * fix(ci): keep approved_by untouched, add latest_approval beside it for the stale-approval check --------- Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ci_cd/spend_logs_guard.py | 47 +++++++++++++++++++++++--- tests/unit/test_spend_logs_guard.py | 52 ++++++++++++++++++++++++++++- 2 files changed, 93 insertions(+), 6 deletions(-) diff --git a/ci_cd/spend_logs_guard.py b/ci_cd/spend_logs_guard.py index a04999c47c0..64e33c4fb70 100644 --- a/ci_cd/spend_logs_guard.py +++ b/ci_cd/spend_logs_guard.py @@ -4,7 +4,9 @@ needs an approving review from the spend-log owner before it can merge. CODEOWNERS can only match file paths, and every schema change edits the same schema.prisma copies, so a path rule would put every migration on the owner. This gate reads the diff instead: when no added line in the watched files mentions the table it passes at once, otherwise it passes only once -the owner's latest review on the pull request is an approval. The workflow re-runs on every review. +the owner's latest review on the pull request is an approval that still covers every spend-log line +in the head (approvals survive pushes, so a later commit may not add spend-log lines the owner never +saw). The workflow re-runs on every review. """ from __future__ import annotations @@ -13,7 +15,7 @@ import argparse import re import subprocess import sys -from collections.abc import Sequence +from collections.abc import Callable, Sequence from pathlib import Path from typing import Final @@ -37,6 +39,7 @@ class Review(BaseModel): model_config = ConfigDict(frozen=True, extra="ignore") state: str user: Reviewer + commit_id: str | None = None _REVIEWS: Final = TypeAdapter(list[Review]) @@ -75,6 +78,24 @@ def approved_by(reviews: Sequence[Review], login: str) -> bool: return bool(states) and states[-1] == "APPROVED" +def latest_approval(reviews: Sequence[Review], login: str) -> Review | None: + """`login`'s most recent APPROVED review on the pull request, the one GitHub keeps in force after pushes.""" + own: Final = tuple(review for review in reviews if review.user.login == login and review.state == "APPROVED") + return own[-1] if own else None + + +def approval_covers( + approval: Review, head: str, hits: Sequence[str], hits_at: Callable[[str], Sequence[str] | None] +) -> bool: + """True when the approval is on the head, or the head adds no spend-log line the approved commit did not.""" + if approval.commit_id is None: + return False + if approval.commit_id == head: + return True + reviewed: Final = hits_at(approval.commit_id) + return reviewed is not None and set(hits) <= set(reviewed) + + def added_diff(base: str, head: str) -> str: result: Final = subprocess.run( ("git", "diff", "--unified=0", "--no-color", f"{base}..{head}", "--", *WATCHED, MIGRATIONS_PREFIX), @@ -85,6 +106,15 @@ def added_diff(base: str, head: str) -> str: return result.stdout +def fetched(commit: str) -> bool: + result: Final = subprocess.run(("git", "fetch", "--no-tags", "--depth=1", "origin", commit), check=False) + return result.returncode == 0 + + +def reviewed_hits(base: str, commit: str) -> tuple[str, ...] | None: + return spend_log_hits(added_diff(base, commit)) if fetched(commit) else None + + class Arguments(BaseModel): model_config = ConfigDict(frozen=True) base: str @@ -104,14 +134,21 @@ def main(argv: Sequence[str]) -> int: if not hits: sys.stdout.write("No added line in schema.prisma or a migration names LiteLLM_SpendLogs; nothing to review.\n") return 0 - if approved_by(_REVIEWS.validate_json(args.reviews.read_bytes()), args.approver): + reviews: Final = _REVIEWS.validate_json(args.reviews.read_bytes()) + approval: Final = latest_approval(reviews, args.approver) if approved_by(reviews, args.approver) else None + if approval is not None and approval_covers( + approval, args.head, hits, lambda commit: reviewed_hits(args.base, commit) + ): sys.stdout.write(f"LiteLLM_SpendLogs change approved by {args.approver}: {', '.join(hits)}\n") return 0 listed: Final = "".join(f" {hit}\n" for hit in hits) - sys.stdout.write( + reason: Final = ( f"This pull request changes LiteLLM_SpendLogs and needs an approving review from @{args.approver}:\n" - f"{listed}The check re-runs when a review is submitted.\n" + if approval is None + else f"LiteLLM_SpendLogs lines changed after @{args.approver} approved {approval.commit_id}, " + f"a new approving review is needed:\n" ) + sys.stdout.write(f"{reason}{listed}The check re-runs when a review is submitted.\n") return 1 diff --git a/tests/unit/test_spend_logs_guard.py b/tests/unit/test_spend_logs_guard.py index 9a033045dd0..6a3355e1c52 100644 --- a/tests/unit/test_spend_logs_guard.py +++ b/tests/unit/test_spend_logs_guard.py @@ -42,6 +42,10 @@ class TestSpendLogHits: assert guard.spend_log_hits(diff) == ("litellm/proxy/schema.prisma: a LiteLLM_SpendLogs",) +def _review(login: str, state: str, commit_id: str | None = "head") -> "guard.Review": + return guard.Review(state=state, user=guard.Reviewer(login=login), commit_id=commit_id) + + class TestApprovedBy: def test_the_latest_review_of_the_approver_decides(self): reviews: Final = tuple( @@ -61,6 +65,42 @@ class TestApprovedBy: assert guard.approved_by((), "kerry-berri") is False +class TestLatestApproval: + def test_the_approvers_last_approved_review_is_returned_with_its_commit(self): + reviews: Final = ( + _review("kerry-berri", "APPROVED", "old"), + _review("someone-else", "APPROVED", "head"), + _review("kerry-berri", "APPROVED", "new"), + _review("kerry-berri", "COMMENTED", "head"), + ) + assert guard.latest_approval(reviews, "kerry-berri") is reviews[2] + assert guard.latest_approval(reviews[:2], "kerry-berri") is reviews[0] + assert guard.latest_approval(reviews[1:2], "kerry-berri") is None + assert guard.latest_approval((), "kerry-berri") is None + + +class TestApprovalCovers: + _HITS: Final = ("schema.prisma: a LiteLLM_SpendLogs", "schema.prisma: b LiteLLM_SpendLogs") + + def test_an_approval_on_the_head_covers_it_without_reading_the_approved_commit(self): + def unreadable(commit: str) -> None: + raise AssertionError(f"read {commit}") + + assert guard.approval_covers(_review("kerry-berri", "APPROVED", "head"), "head", self._HITS, unreadable) is True + + def test_an_older_approval_covers_the_head_only_when_it_saw_every_spend_log_line(self): + approval: Final = _review("kerry-berri", "APPROVED", "old") + assert guard.approval_covers(approval, "head", self._HITS, lambda commit: self._HITS) is True + assert guard.approval_covers(approval, "head", self._HITS[:1], lambda commit: self._HITS) is True + assert guard.approval_covers(approval, "head", self._HITS, lambda commit: self._HITS[:1]) is False + + def test_an_approval_whose_commit_cannot_be_read_does_not_cover_the_head(self): + gone: Final = _review("kerry-berri", "APPROVED", "gone") + assert guard.approval_covers(gone, "head", self._HITS, lambda commit: None) is False + without_commit: Final = _review("kerry-berri", "APPROVED", None) + assert guard.approval_covers(without_commit, "head", self._HITS, lambda commit: self._HITS) is False + + class TestExitCode: def test_a_pull_request_without_spend_log_lines_passes_without_reviews(self, tmp_path, monkeypatch, capsys): monkeypatch.setattr(guard, "added_diff", lambda base, head: _diff("schema.prisma", " foo String", 1)) @@ -75,5 +115,15 @@ class TestExitCode: reviews.write_text('[{"user": {"login": "someone-else"}, "state": "APPROVED"}]') assert guard.main(["--base", "a", "--head", "b", "--approver", "kerry-berri", "--reviews", str(reviews)]) == 1 assert "schema.prisma: x LiteLLM_SpendLogs" in capsys.readouterr().out - reviews.write_text('[{"user": {"login": "kerry-berri"}, "state": "APPROVED"}]') + reviews.write_text('[{"user": {"login": "kerry-berri"}, "state": "APPROVED", "commit_id": "b"}]') assert guard.main(["--base", "a", "--head", "b", "--approver", "kerry-berri", "--reviews", str(reviews)]) == 0 + + def test_an_approval_before_a_new_spend_log_line_was_pushed_fails(self, tmp_path, monkeypatch, capsys): + monkeypatch.setattr(guard, "added_diff", lambda base, head: _diff("schema.prisma", " x LiteLLM_SpendLogs", 7)) + monkeypatch.setattr(guard, "reviewed_hits", lambda base, commit: () if commit == "old" else None, raising=False) + reviews: Final = tmp_path / "r.json" + reviews.write_text('[{"user": {"login": "kerry-berri"}, "state": "APPROVED", "commit_id": "old"}]') + assert guard.main(["--base", "a", "--head", "b", "--approver", "kerry-berri", "--reviews", str(reviews)]) == 1 + assert "changed after @kerry-berri approved old" in capsys.readouterr().out + reviews.write_text('[{"user": {"login": "kerry-berri"}, "state": "APPROVED", "commit_id": "gone"}]') + assert guard.main(["--base", "a", "--head", "b", "--approver", "kerry-berri", "--reviews", str(reviews)]) == 1 From c6ca22ebc58baa4398728bdcbba5b5156105ee8b Mon Sep 17 00:00:00 2001 From: yucheng-berri Date: Wed, 28 Oct 2026 14:49:48 -0700 Subject: [PATCH 7/7] fix(guardrails)!: make hide-secrets honor skip_tool_message_in_guardrail (#48689) * fix(guardrails): make hide-secrets honor skip_tool_message_in_guardrail With the flag on, per guardrail or in litellm_settings, hide-secrets leaves role: tool messages and Anthropic tool_result blocks as sent, as the unified guardrails do. Tool call arguments are still masked. * test(guardrails): type the skip-tool walker test and assert the whole request --- .../enterprise_callbacks/secret_detection.py | 8 ++- litellm/proxy/guardrails/_content_utils.py | 35 ++++++---- .../test_hide_secrets_tool_traffic.py | 68 +++++++++++++++++++ .../proxy/guardrails/test_content_utils.py | 29 ++++++++ 4 files changed, 126 insertions(+), 14 deletions(-) diff --git a/enterprise/litellm_enterprise/enterprise_callbacks/secret_detection.py b/enterprise/litellm_enterprise/enterprise_callbacks/secret_detection.py index 65e4d50587f..c190e680347 100644 --- a/enterprise/litellm_enterprise/enterprise_callbacks/secret_detection.py +++ b/enterprise/litellm_enterprise/enterprise_callbacks/secret_detection.py @@ -32,6 +32,7 @@ from litellm.integrations.custom_guardrail import ( CustomGuardrail, log_guardrail_information, ) +from litellm.llms.base_llm.guardrail_translation.utils import effective_skip_tool_message_for_guardrail from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.guardrails._content_utils import iter_tool_json_fields, map_tool_json_text, walk_user_text from litellm.types.guardrails import GuardrailEventHooks @@ -753,7 +754,12 @@ class _ENTERPRISE_SecretDetection(CustomGuardrail): _masked_entity_count.set({}) # Covers multimodal list content + Responses-API input. - inspected = walk_user_text(data, self.redact_text, self.redact_tool_payload) + self._redact_prompt(data) + inspected = walk_user_text( + data, + self.redact_text, + self.redact_tool_payload, + skip_tool_message=effective_skip_tool_message_for_guardrail(self), + ) + self._redact_prompt(data) if inspected == 0: # Image-only, empty-text, and unsupported payloads inspected diff --git a/litellm/proxy/guardrails/_content_utils.py b/litellm/proxy/guardrails/_content_utils.py index b4830b7c771..9bea22acd71 100644 --- a/litellm/proxy/guardrails/_content_utils.py +++ b/litellm/proxy/guardrails/_content_utils.py @@ -305,6 +305,11 @@ def iter_message_text(data: Mapping[str, object]) -> Iterator[str]: yield from _iter_text_parts_in_content(message.get("content")) +def _is_tool_message(message: object) -> bool: + fields: Final = _text_object(message) + return fields is not None and str(fields.get("role") or "").lower() == "tool" + + def _call_item(item: object) -> tuple[str, object] | None: """Return ``(field, value)`` for a Responses tool-call input item, else None.""" fields: Final = _text_object(item) @@ -316,11 +321,14 @@ def walk_user_text( data: dict[str, object], visit: Callable[[str], str], visit_tool: Callable[[JsonValue], JsonValue] | None = None, + *, + skip_tool_message: bool = False, ) -> int: """Rewrite every text fragment in place via ``visit``. Mutates ``data["messages"]`` and ``data["input"]`` (including tool payloads), ``data["system"]`` and - ``data["instructions"]``. Returns the number + ``data["instructions"]``. ``skip_tool_message`` leaves ``role: tool`` messages and ``tool_result`` + blocks as sent. Returns the number of fragments visited so callers can short-circuit when nothing was inspected. """ @@ -353,6 +361,8 @@ def walk_user_text( def _rewrite_part(part: Mapping[str, object], native: bool) -> Mapping[str, object]: nonlocal visited + if skip_tool_message and part.get("type") == "tool_result": + return part text: Final = _part_text(part) rewritten: Final = {**part, **({"text": visit(text)} if text is not None else {})} if text is not None: @@ -384,18 +394,17 @@ def walk_user_text( return content messages: Final = data.get("messages") - if isinstance(messages, list): - for message in messages: - if not isinstance(message, dict): - continue - if "content" in message: - message["content"] = _rewrite_content(message["content"]) - calls = message.get("tool_calls") - if isinstance(calls, list): - for call in calls: - function = call.get("function") if isinstance(call, dict) else None - if isinstance(function, dict) and "arguments" in function: - function["arguments"] = _rewrite_arguments(function["arguments"]) + for message in _object_list(messages) or (): + if (skip_tool_message and _is_tool_message(message)) or not isinstance(message, dict): + continue + if "content" in message: + message["content"] = _rewrite_content(message["content"]) + calls = message.get("tool_calls") + if isinstance(calls, list): + for call in calls: + function = call.get("function") if isinstance(call, dict) else None + if isinstance(function, dict) and "arguments" in function: + function["arguments"] = _rewrite_arguments(function["arguments"]) if "system" in data: data["system"] = _rewrite_content(data["system"]) # rebind-ok: preserve the native hook's in-place contract diff --git a/tests/integration/security/test_hide_secrets_tool_traffic.py b/tests/integration/security/test_hide_secrets_tool_traffic.py index 42fdc0f5a44..60b49520b1f 100644 --- a/tests/integration/security/test_hide_secrets_tool_traffic.py +++ b/tests/integration/security/test_hide_secrets_tool_traffic.py @@ -413,3 +413,71 @@ def test_resent_tool_history_is_scanned_once(single_worker: Rig, shape: str) -> second: Final = timed(lambda: single_worker.proxy.request("POST", path, following)) assert first.value.status_code == 200 and second.value.status_code == 200, (first.value.text, second.value.text) assert_within_factor_of_control(second.seconds, control=first.seconds, factor=0.5) + + +@pytest.mark.parametrize("placement", ("guardrail", "global")) +def test_skip_tool_message_in_guardrail_forwards_tool_results(tmp_path: Path, placement: str) -> None: + def configure(configuration: dict[str, JsonValue]) -> None: + guardrails: Final = configuration["guardrails"] + settings: Final = configuration["litellm_settings"] + assert isinstance(guardrails, list) and isinstance(guardrails[0], dict) and isinstance(settings, dict) + params: Final = guardrails[0]["litellm_params"] + assert isinstance(params, dict) + (params if placement == "guardrail" else settings)["skip_tool_message_in_guardrail"] = True + + with tool_rig(tmp_path, configure=configure) as rig: + rig.wire.drain() + replies: Final = ( + rig.proxy.request( + "POST", + "/v1/messages", + { + "model": "claude", + "max_tokens": 32, + "cache": {"no-cache": True}, + "messages": [ + {"role": "user", "content": TEXT}, + { + "role": "assistant", + "content": [ + {"type": "tool_use", "id": "tool_1", "name": "read_file", "input": {"env": TEXT}} + ], + }, + { + "role": "user", + "content": [{"type": "tool_result", "tool_use_id": "tool_1", "content": TEXT}], + }, + ], + }, + ), + rig.proxy.request( + "POST", + "/v1/chat/completions", + { + "model": "gpt", + "cache": {"no-cache": True}, + "messages": [ + {"role": "user", "content": TEXT}, + { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_1", + "type": "function", + "function": {"name": "read_file", "arguments": json.dumps({"env": TEXT})}, + } + ], + }, + {"role": "tool", "tool_call_id": "call_1", "content": TEXT}, + ], + }, + ), + ) + assert all(reply.status_code == 200 for reply in replies), [reply.text for reply in replies] + sent: Final = rig.wire.drain() + assert len(sent) == 2 + for request in sent: + messages = json.loads(request.body)["messages"] + assert SECRET not in json.dumps(messages[:2]) + assert SECRET in json.dumps(messages[2]) diff --git a/tests/unit/proxy/guardrails/test_content_utils.py b/tests/unit/proxy/guardrails/test_content_utils.py index a668d5f609d..21d96a3ea97 100644 --- a/tests/unit/proxy/guardrails/test_content_utils.py +++ b/tests/unit/proxy/guardrails/test_content_utils.py @@ -2,6 +2,7 @@ import json from functools import reduce +from typing import Final import pytest from pydantic import JsonValue @@ -924,6 +925,34 @@ def test_native_writer_visits_each_tool_and_system_slot(data): assert "payload" not in json.dumps(data) +def test_skip_tool_message_leaves_tool_results_and_masks_the_rest() -> None: + data: Final[dict[str, object]] = { + "messages": [ + {"role": "user", "content": "secret"}, + { + "role": "assistant", + "content": [{"type": "tool_use", "id": "tool", "name": "read", "input": {"env": "secret"}}], + "tool_calls": [{"id": "call", "function": {"name": "read", "arguments": '{"env": "secret"}'}}], + }, + {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "tool", "content": "secret"}]}, + {"role": "tool", "tool_call_id": "call", "content": "secret"}, + ] + } + assert walk_user_text(data, lambda text: text.replace("secret", "[MASKED]"), skip_tool_message=True) == 3 + assert data == { + "messages": [ + {"role": "user", "content": "[MASKED]"}, + { + "role": "assistant", + "content": [{"type": "tool_use", "id": "tool", "name": "read", "input": {"env": "[MASKED]"}}], + "tool_calls": [{"id": "call", "function": {"name": "read", "arguments": '{"env": "[MASKED]"}'}}], + }, + {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "tool", "content": "secret"}]}, + {"role": "tool", "tool_call_id": "call", "content": "secret"}, + ] + } + + @pytest.mark.parametrize( "part", (