Merge f7954617cf into 7083c47998

litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py

@@ -14,6 +14,7 @@ from datetime import datetime
 from re import Pattern
 from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypedDict, cast
 
+import openai
 import yaml
 from fastapi import HTTPException
 
@@ -2056,6 +2057,13 @@ class ContentFilterGuardrail(CustomGuardrail):
         except HTTPException:
             status = "guardrail_intervened"
             raise
+        except openai.OpenAIError as e:
+            # Upstream provider/stream error propagating through the iterator.
+            # The guardrail never evaluated content, so it must not be recorded
+            # as a guardrail failure. "not_run" is the canonical status here.
+            status = "not_run"
+            exception_str = str(e)
+            raise
         except Exception as e:
             status = "guardrail_failed_to_respond"
             exception_str = str(e)

tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_upstream_error_not_run.py

@@ -0,0 +1,42 @@
+import openai
+from unittest.mock import MagicMock
+
+import pytest
+
+from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import (
+    ContentFilterGuardrail,
+)
+from litellm.types.guardrails import GuardrailEventHooks
+
+
+@pytest.mark.asyncio
+async def test_streaming_hook_upstream_error_is_not_run():
+    """
+    When the upstream stream raises a provider error (openai.OpenAIError
+    subclass), the content filter never evaluates content. The error must
+    propagate unchanged and be logged as "not_run", not as a guardrail
+    failure. Regression test for issue #31004.
+    """
+    guardrail = ContentFilterGuardrail(
+        guardrail_name="test-streaming-upstream-error",
+        patterns=[],
+        event_hook=GuardrailEventHooks.during_call,
+    )
+
+    async def mock_stream():
+        raise openai.APIError("Provider returned error", request=MagicMock(), body=None)
+        yield  # pragma: no cover - makes this an async generator
+
+    user_api_key_dict = MagicMock()
+    request_data: dict = {}
+
+    with pytest.raises(openai.APIError):
+        async for _ in guardrail.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=user_api_key_dict,
+            response=mock_stream(),
+            request_data=request_data,
+        ):
+            pass
+
+    info = request_data["metadata"]["standard_logging_guardrail_information"][0]
+    assert info["guardrail_status"] == "not_run"