From 77406957da0ebcd3c48d08b8b60d867c1082e036 Mon Sep 17 00:00:00 2001
From: PRABHU KIRAN VANDRANKI <72809214+VANDRANKI@users.noreply.github.com>
Date: Wed, 6 May 2026 15:35:29 -0400
Subject: [PATCH] test(logging): add unit tests for anthropic_messages logging
 Responses API event guard

---
 ...ropic_messages_logging_responses_events.py | 110 ++++++++++++++++++
 1 file changed, 110 insertions(+)
 create mode 100644 tests/unit/litellm_core_utils/test_anthropic_messages_logging_responses_events.py

diff --git a/tests/unit/litellm_core_utils/test_anthropic_messages_logging_responses_events.py b/tests/unit/litellm_core_utils/test_anthropic_messages_logging_responses_events.py
new file mode 100644
index 00000000000..5d4b6ee1455
--- /dev/null
+++ b/tests/unit/litellm_core_utils/test_anthropic_messages_logging_responses_events.py
@@ -0,0 +1,110 @@
+"""
+Unit tests for _handle_anthropic_messages_response_logging guard against
+ResponseIncompleteEvent / ResponseCompletedEvent (issue #27091).
+
+These cover the regression: when the experimental Anthropic pass-through routes
+through the Responses API backend the success handler receives a Responses API
+event object, not a plain dict or AnthropicResponse. Before the fix, Pydantic
+raised a ValidationError in background logging.
+"""
+from unittest.mock import MagicMock, patch
+import pytest
+
+
+def _make_logging_obj(stream: bool = False):
+    """Build a minimal Logging instance without touching live infrastructure."""
+    from litellm.litellm_core_utils.litellm_logging import Logging
+    from litellm.utils import CallTypes
+
+    obj = Logging.__new__(Logging)
+    obj.stream = stream
+    obj.call_type = CallTypes.anthropic_messages.value
+    obj.model = "claude-sonnet-4-6"
+    obj.model_call_details = {}  # no httpx_response stored
+    return obj
+
+
+def _make_incomplete_event():
+    """Return a minimal ResponseIncompleteEvent with a stub .response."""
+    from litellm.types.llms.openai import ResponseIncompleteEvent
+
+    stub_response = MagicMock()
+    stub_response.status = "incomplete"
+    return ResponseIncompleteEvent(
+        type="response.incomplete",
+        response=stub_response,
+    )
+
+
+def _make_completed_event():
+    """Return a minimal ResponseCompletedEvent with a stub .response."""
+    from litellm.types.llms.openai import ResponseCompletedEvent
+
+    stub_response = MagicMock()
+    stub_response.status = "completed"
+    return ResponseCompletedEvent(
+        type="response.completed",
+        response=stub_response,
+    )
+
+
+class TestHandleAnthropicMessagesLoggingResponsesEvents:
+    """
+    _handle_anthropic_messages_response_logging must not raise a ValidationError
+    when it receives a Responses API event (ResponseIncompleteEvent,
+    ResponseCompletedEvent, or any unexpected type).
+    """
+
+    def test_incomplete_event_does_not_raise(self):
+        """ResponseIncompleteEvent must not trigger AnthropicResponse.model_validate."""
+        obj = _make_logging_obj()
+        event = _make_incomplete_event()
+        # Should return without raising ValidationError
+        result = obj._handle_anthropic_messages_response_logging(result=event)
+        # Result passes through unchanged
+        assert result is event
+
+    def test_completed_event_does_not_raise(self):
+        """ResponseCompletedEvent must not trigger AnthropicResponse.model_validate."""
+        obj = _make_logging_obj()
+        event = _make_completed_event()
+        result = obj._handle_anthropic_messages_response_logging(result=event)
+        assert result is event
+
+    def test_unknown_type_does_not_raise(self):
+        """Arbitrary non-dict, non-AnthropicResponse types must not crash."""
+        obj = _make_logging_obj()
+        random_obj = object()
+        # Should not raise
+        result = obj._handle_anthropic_messages_response_logging(result=random_obj)
+        assert result is random_obj
+
+    def test_model_response_still_returned_as_is(self):
+        """ModelResponse inputs should continue to pass through unchanged."""
+        from litellm import ModelResponse
+
+        obj = _make_logging_obj()
+        mr = ModelResponse(model="claude-sonnet-4-6")
+        result = obj._handle_anthropic_messages_response_logging(result=mr)
+        assert isinstance(result, ModelResponse)
+
+    def test_streaming_model_response_returned_as_is(self):
+        """Streaming ModelResponse should still short-circuit at the first guard."""
+        from litellm import ModelResponse
+
+        obj = _make_logging_obj(stream=True)
+        mr = ModelResponse(model="claude-sonnet-4-6")
+        result = obj._handle_anthropic_messages_response_logging(result=mr)
+        assert isinstance(result, ModelResponse)
+
+    def test_debug_log_emitted_for_incomplete_event(self, caplog):
+        """A debug message must be emitted when a Responses API event is encountered."""
+        import logging
+
+        obj = _make_logging_obj()
+        event = _make_incomplete_event()
+        with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
+            obj._handle_anthropic_messages_response_logging(result=event)
+        assert any(
+            "ResponseIncompleteEvent" in rec.message for rec in caplog.records
+        ), "Expected debug log mentioning ResponseIncompleteEvent"