diff --git a/litellm/litellm_core_utils/exception_mapping_utils.py b/litellm/litellm_core_utils/exception_mapping_utils.py index 4ee726b67de..b76c97ad2de 100644 --- a/litellm/litellm_core_utils/exception_mapping_utils.py +++ b/litellm/litellm_core_utils/exception_mapping_utils.py @@ -2486,6 +2486,18 @@ def exception_type( exception_provider=exception_provider, extra_information=extra_information, ) + from litellm.llms.base_llm.chat.transformation import BaseLLMException + + if custom_llm_provider and isinstance(original_exception, BaseLLMException): + _map_openai_like_exception( + model=model, + original_exception=mappable_exception, + custom_llm_provider=custom_llm_provider, + error_str=error_str, + exception_type=exception_type, + exception_provider=exception_provider, + extra_information=extra_information, + ) if "BadRequestError.__init__() missing 1 required positional argument: 'param'" in str( original_exception ): # deal with edge-case invalid request error bug in openai-python sdk diff --git a/litellm/llms/anthropic/experimental_pass_through/messages/handler.py b/litellm/llms/anthropic/experimental_pass_through/messages/handler.py index f4d24bb933c..7459d1b2da5 100644 --- a/litellm/llms/anthropic/experimental_pass_through/messages/handler.py +++ b/litellm/llms/anthropic/experimental_pass_through/messages/handler.py @@ -12,6 +12,7 @@ from functools import partial from typing import Any, Final, cast import litellm +from litellm.litellm_core_utils.exception_mapping_utils import exception_type from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.llms.anthropic.common_utils import ( flatten_unencrypted_web_search_results_in_anthropic_messages, @@ -382,13 +383,18 @@ async def anthropic_messages( ) ctx: Final = contextvars.copy_context() func_with_context: Final = partial(ctx.run, func) - init_response: Final = await loop.run_in_executor(None, func_with_context) - - if asyncio.iscoroutine(init_response): - response = await init_response - else: - response = init_response - return response + try: + init_response: Final = await loop.run_in_executor(None, func_with_context) + if asyncio.iscoroutine(init_response): + return await init_response + return init_response + except Exception as e: # noqa: BLE001 # the mapping boundary must see every provider-layer failure, like acompletion + raise exception_type( + model=model, + custom_llm_provider=custom_llm_provider, + original_exception=e, + extra_kwargs=kwargs, + ) def validate_anthropic_api_metadata(metadata: dict | None = None) -> dict | None: diff --git a/tests/e2e/logging/test_otel_trace_e2e.py b/tests/e2e/logging/test_otel_trace_e2e.py index d7b28c170c2..52cb691e2b7 100644 --- a/tests/e2e/logging/test_otel_trace_e2e.py +++ b/tests/e2e/logging/test_otel_trace_e2e.py @@ -743,3 +743,61 @@ class TestOtelTraceCompleteness: ) genai = next(span for span in hits[0].spans if span.operation_name == genai_span) _assert_error_span_contract(genai) + + @pytest.mark.covers("logging.otel.failure.exports_metric", exercised_on=["messages"]) + def test_failed_messages_error_span_attributes( + self, client: LoggingClient, otel_reader: OtelReader, resources: ResourceManager + ) -> None: + """A failed `/v1/messages` request must carry the same error-span + contract as a failed `/chat/completions` request (LIT-6164). The + async messages entrypoint used to surface the provider handler's raw + BaseLLMException to the failure logger, so the model-call span came + out with error.type=BaseLLMException and no + litellm.provider.error.llm_provider attribute. + + Same setup as the chat sibling: a deployment with an invalid upstream + API key passes proxy auth and fails at the provider with a real 401, + and failed requests are not billed, so no cost-write span.""" + route = "/v1/messages" + _assert_otel_destination_configured(client) + + model_name = f"otel-err-{unique_marker()}" + model_id = client.create_model( + model_name, + LiteLLMParamsBody(model="anthropic/claude-haiku-4-5", api_key=INVALID_UPSTREAM_API_KEY), + ) + resources.defer(lambda: client.delete_model(model_id)) + key = client.key_with_alias(f"otel-err-{unique_marker()}", models=[model_name]) + resources.defer(lambda: client.delete_key(key)) + + deadline = time.monotonic() + client.proxy.poll_timeout + while True: + outcome = client.messages_raw(key, model_name, "trigger an upstream auth failure", max_tokens=16) + assert not outcome.ok, "the call must fail; the deployment's upstream key is invalid" + if "AnthropicException" in outcome.body or time.monotonic() >= deadline: + break + time.sleep(client.proxy.poll_interval) + assert "AnthropicException" in outcome.body, ( + "never saw the mapped upstream provider failure before the deadline; either the key is " + "still propagating or the messages route surfaced the raw unmapped provider error - " + f"last outcome {outcome.status_code}: {outcome.body[:200]}" + ) + assert outcome.status_code == 401, ( + f"an upstream auth failure must map to 401, got {outcome.status_code}: {outcome.body[:200]}" + ) + assert outcome.call_id is not None, "failed responses must still carry x-litellm-call-id" + + genai_span = f"chat {model_name}" + hits = otel_reader.poll_traces_for_call( + call_id=outcome.call_id, + settled_names=_settled_names(route=route, genai_span=genai_span, require_cost_span=False), + settled_prefixes={DB_SPAN_PREFIX}, + ) + _assert_complete_trace(hits, route=route, genai_span=genai_span, require_cost_span=False) + + root = next(span for span in hits[0].spans if not span.references) + assert str(_tag(root, "http.status_code")) == "401", ( + f"the SERVER span must record the 401 the client received, got {_tag(root, 'http.status_code')!r}" + ) + genai = next(span for span in hits[0].spans if span.operation_name == genai_span) + _assert_error_span_contract(genai) diff --git a/tests/test_litellm/litellm_core_utils/test_exception_mapping_utils.py b/tests/test_litellm/litellm_core_utils/test_exception_mapping_utils.py index 6f7ea9da640..3eb8094f914 100644 --- a/tests/test_litellm/litellm_core_utils/test_exception_mapping_utils.py +++ b/tests/test_litellm/litellm_core_utils/test_exception_mapping_utils.py @@ -1092,3 +1092,30 @@ def test_bedrock_mantle_context_overflow_maps_to_context_window_exceeded(): assert excinfo.value.status_code == 400 assert "prompt is too long: 1055489 tokens > 1050000 maximum" in excinfo.value.message + + +@pytest.mark.parametrize( + "status_code, expected_class", + [(401, litellm.AuthenticationError), (429, litellm.RateLimitError)], +) +def test_a_base_llm_exception_without_a_provider_branch_maps_by_status_code( + status_code, expected_class, quiet_exception_mapping +): + """Regression test for LIT-6164. Native /v1/messages handlers raise raw + BaseLLMException, and providers without an exception_type branch (e.g. + minimax) must keep the upstream status instead of collapsing every failure + into a 500 APIConnectionError once that route maps its exceptions.""" + from litellm.llms.base_llm.chat.transformation import BaseLLMException + + original_exception = BaseLLMException(status_code=status_code, message="upstream rejected the call") + + with pytest.raises(expected_class) as excinfo: + exception_type( + model="MiniMax-M2.5", + original_exception=original_exception, + custom_llm_provider="minimax", + ) + + assert excinfo.value.status_code == status_code + assert excinfo.value.llm_provider == "minimax" + assert "MinimaxException" in excinfo.value.message diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py index 9e58ded81bd..c88058bf215 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py @@ -1286,3 +1286,56 @@ class TestMessagesStreamingSuccessLogging: assert payload["call_type"] == "acompletion" assert payload["total_tokens"] > 0 assert payload["response_cost"] > 0 + + +class _FailureCapture(CustomLogger): + def __init__(self): + super().__init__() + self.error_information: List[Dict[str, Any]] = [] + + async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time): + payload = kwargs.get("standard_logging_object") or {} + self.error_information.append(payload.get("error_information") or {}) + + +@pytest.mark.asyncio +async def test_anthropic_messages_maps_provider_exception_before_failure_logging(monkeypatch): + """Regression test for LIT-6164. The async /v1/messages entrypoint awaited the + provider handler without exception_type mapping, so the @client failure + handler (and every logger behind it, e.g. OTel error spans) saw the raw + BaseLLMException: error.type=BaseLLMException and no llm_provider.""" + from litellm.llms.anthropic.experimental_pass_through.messages import handler + + capture = _FailureCapture() + monkeypatch.setattr(litellm, "callbacks", [capture]) + + def upstream_rejects_the_key(request: httpx.Request) -> httpx.Response: + return httpx.Response( + 401, + json={"type": "error", "error": {"type": "authentication_error", "message": "invalid x-api-key"}}, + request=request, + ) + + upstream = AsyncHTTPHandler() + upstream.client = httpx.AsyncClient(transport=httpx.MockTransport(upstream_rejects_the_key)) + + with pytest.raises(litellm.AuthenticationError) as excinfo: + await handler.anthropic_messages( + max_tokens=16, + messages=[{"role": "user", "content": "hi"}], + model="anthropic/claude-haiku-4-5", + custom_llm_provider="anthropic", + api_key="sk-invalid", + client=upstream, + ) + + assert excinfo.value.status_code == 401 + assert excinfo.value.llm_provider == "anthropic" + assert "AnthropicException" in excinfo.value.message + assert '"authentication_error"' in excinfo.value.message + + assert capture.error_information, "the failure handler must have logged the mapped exception" + error_information = capture.error_information[0] + assert error_information.get("error_class") == "AuthenticationError" + assert error_information.get("llm_provider") == "anthropic" + assert error_information.get("error_code") == "401"