fix(pass-through): keep URL model attribution off routes that name another provider

A built-in route that names a non-Google provider, like /openai pointed at
Gemini's OpenAI-compatible host, no longer takes its failure model from the
Google URL. A guardrail block response now returns the x-litellm-call-id of
its spend row.
This commit is contained in:
mateo-berri 2026-10-30 04:44:50 -07:00
parent dd6d207ccc
commit 0c1c3eaf79
2 changed files with 76 additions and 19 deletions

View file

@ -883,8 +883,8 @@ _GOOGLE_PASSTHROUGH_PROVIDERS: Final = frozenset(
def _passthrough_google_provider(url_route: str, custom_llm_provider: str | None) -> str | None:
if custom_llm_provider in _GOOGLE_PASSTHROUGH_PROVIDERS:
return custom_llm_provider
if custom_llm_provider is not None:
return custom_llm_provider if custom_llm_provider in _GOOGLE_PASSTHROUGH_PROVIDERS else None
if VertexPassthroughLoggingHandler.is_google_ai_url(url_route):
return VertexPassthroughLoggingHandler.custom_llm_provider_from_url(url_route)
return None
@ -2072,6 +2072,7 @@ async def pass_through_request(
content=json.dumps(error_body),
status_code=200,
media_type="application/json",
headers={"x-litellm-call-id": litellm_call_id},
)
except Exception as e:
custom_headers = ProxyBaseLLMRequestProcessing.get_custom_headers(

View file

@ -8650,6 +8650,36 @@ class _FailureHookRecorder(CustomLogger):
)
@dataclass(frozen=True, slots=True)
class _PassthroughRun:
failure: _RecordedFailure
response: Response | None
async def _drained_pass_through_response(
request: Request,
target: str,
custom_llm_provider: str | None,
user_api_key_dict: UserAPIKeyAuth,
stream: bool,
) -> Response | None:
try:
received: Final = await pass_through_request(
request=request,
target=target,
custom_headers={},
user_api_key_dict=user_api_key_dict,
custom_llm_provider=custom_llm_provider,
stream=stream,
)
if isinstance(received, StreamingResponse):
async for _ in received.body_iterator:
pass
except (ProxyException, httpx.HTTPError):
return None
return received if isinstance(received, Response) else None
async def _run_passthrough_failure(
target: str,
custom_llm_provider: str | None,
@ -8660,6 +8690,22 @@ async def _run_passthrough_failure(
guardrail: CustomGuardrail | None = None,
body: bytes = b'{"contents": "not a list"}',
) -> _RecordedFailure:
run: Final = await _run_passthrough(
target, custom_llm_provider, upstream, stream, user_api_key_dict, dispatched_endpoint, guardrail, body
)
return run.failure
async def _run_passthrough(
target: str,
custom_llm_provider: str | None,
upstream: httpx.Response | Exception,
stream: bool = False,
user_api_key_dict: UserAPIKeyAuth | None = None,
dispatched_endpoint: Callable[..., object] | None = None,
guardrail: CustomGuardrail | None = None,
body: bytes = b'{"contents": "not a list"}',
) -> _PassthroughRun:
recorder: Final = _FailureHookRecorder()
with (
pytest.MonkeyPatch.context() as litellm_patch,
@ -8689,24 +8735,17 @@ async def _run_passthrough_failure(
mock_request.body = AsyncMock(return_value=body)
mock_request.headers = Headers({"content-type": "application/json"})
mock_request.query_params = QueryParams({})
try:
received: Final = await pass_through_request(
request=mock_request,
target=target,
custom_headers={},
user_api_key_dict=user_api_key_dict or UserAPIKeyAuth(api_key="sk-test"),
custom_llm_provider=custom_llm_provider,
stream=stream,
)
if isinstance(received, StreamingResponse):
async for _ in received.body_iterator:
pass
except (ProxyException, httpx.HTTPError):
pass
response: Final = await _drained_pass_through_response(
mock_request,
target,
custom_llm_provider,
user_api_key_dict or UserAPIKeyAuth(api_key="sk-test"),
stream,
)
await asyncio.sleep(0)
litellm.in_memory_llm_clients_cache.flush_cache()
assert len(recorder.failures) == 1
return recorder.failures[0]
return _PassthroughRun(recorder.failures[0], response)
def _upstream_400(target: str) -> httpx.Response:
@ -8865,6 +8904,22 @@ async def test_passthrough_guardrail_block_logs_model_from_url(
assert get_spend_logs_id(CallTypes.pass_through.value, {}, failure.request_data) is not None
@pytest.mark.asyncio
@pytest.mark.parametrize("body", [b'{"contents": "not a list"}', b'{"litellm_call_id": "client-chosen-id"}'])
async def test_passthrough_guardrail_block_response_names_the_call_id_of_its_spend_row(body: bytes):
guardrail: Final = _ModifyResponseOnPreCallGuardrail(
guardrail_name="block-policy", event_hook=GuardrailEventHooks.pre_call, default_on=True
)
run: Final = await _run_passthrough(
_GEMINI_FLASH_TARGET, "gemini", _upstream_400(_GEMINI_FLASH_TARGET), guardrail=guardrail, body=body
)
assert isinstance(run.failure.original_exception, ModifyResponseException)
assert run.response is not None
assert run.response.headers["x-litellm-call-id"] == get_spend_logs_id(
CallTypes.pass_through.value, {}, run.failure.request_data
)
@pytest.mark.asyncio
@pytest.mark.parametrize("body", [b'{"contents": "not a list"}', b'{"custom_llm_provider": "vertex_ai"}'])
async def test_passthrough_guardrail_block_on_a_user_defined_route_gets_no_url_attribution(body: bytes):
@ -8956,9 +9011,10 @@ async def _run_websocket_connect_failure(
({"model": "from-body"}, _GEMINI_FLASH_TARGET, "gemini", "from-body", "gemini"),
({"contents": []}, _GEMINI_FLASH_TARGET, "gemini", "gemini-3.8-flash", "gemini"),
({"contents": []}, _VERTEX_FLASH_TARGET, None, "gemini-3.8-flash", "vertex_ai"),
({"contents": []}, _VERTEX_FLASH_TARGET, "anthropic", "gemini-3.8-flash", "anthropic"),
({"contents": []}, _VERTEX_FLASH_TARGET, "anthropic", "", "anthropic"),
({}, "https://generativelanguage.googleapis.com/v1beta/openai/models/gemini-3.8-flash", "openai", "", "openai"),
({"custom_llm_provider": "from-body"}, _VERTEX_FLASH_TARGET, None, "gemini-3.8-flash", "vertex_ai"),
({"custom_llm_provider": "gemini"}, _VERTEX_FLASH_TARGET, "anthropic", "gemini-3.8-flash", "anthropic"),
({"custom_llm_provider": "gemini"}, _VERTEX_FLASH_TARGET, "anthropic", "", "anthropic"),
({"contents": []}, _GEMINI_FLASH_TARGET, None, "gemini-3.8-flash", "gemini"),
(
{},