fix(logging): handle ResponseCompletedEvent in anthropic_messages spe…

[V-3604]

Relevant issues

Fixes 28595

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Type

🐛 Bug Fix

Changes

When /v1/messages routes to a non-Anthropic backend via the Responses API, the streaming success handler delivers a ResponseCompletedEvent to _handle_anthropic_messages_response_logging. That function had no branch for this type and fell through to AnthropicResponse.model_validate(result), which raised a ValidationError because ResponseCompletedEvent is not an anthropic-shaped object. The exception was caught by _success_handler_helper_fn as "Non-Blocking", so client received the correct response but the request was silently dropped from LiteLLM_SpendLogs.

The fix adds two guards before the existing httpx_response branch -
If result is a ResponseCompletedEvent, ResponseIncompleteEvent, or ResponseFailedEvent, unwrap result.response (the ResponsesAPIResponse); if it is already a ResponsesAPIResponse, return it directly. Both types flow through the downstream logging pipeline correctly. _transform_usage_objects already handles ResponsesAPIResponse and converts its usage format for spend logging.

[greptile-apps]

Greptile Summary

This PR adds guards in _handle_anthropic_messages_response_logging so that ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent, and ResponsesAPIResponse objects are recognized and returned instead of falling through to AnthropicResponse.model_validate, which was throwing a ValidationError and silently dropping LiteLLM_SpendLogs entries.

• The ValidationError crash is fixed for all four event/response types; four unit tests cover each new branch in isolation.
• For the non-streaming anthropic_messages path the fix is complete — _process_hidden_params_and_response_cost builds the standard logging payload correctly.
• For the streaming path, _handle_anthropic_messages_response_logging unwraps ResponseCompletedEvent → ResponsesAPIResponse before _get_assembled_streaming_response runs; that function has no ResponsesAPIResponse branch and returns None, so complete_streaming_response is never set and spend-log rows are still dropped.

Confidence Score: 4/5

Safe to merge for the crash fix; the streaming spend-logging gap was already broken before and remains so.

The fix correctly prevents the ValidationError crash for all four event types, and the non-streaming logging path works end-to-end. The streaming path still does not produce a complete_streaming_response because _get_assembled_streaming_response has no branch for an already-unwrapped ResponsesAPIResponse, so spend-log rows for streaming anthropic_messages Responses API calls continue to be silently dropped — the same symptom the PR states it fixes.

litellm/litellm_core_utils/litellm_logging.py — specifically the _get_assembled_streaming_response method which needs a ResponsesAPIResponse passthrough branch to complete the streaming fix.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/litellm_logging.py	Adds ResponseCompletedEvent/ResponseIncompleteEvent/ResponseFailedEvent/ResponsesAPIResponse guards to _handle_anthropic_messages_response_logging; fixes the ValidationError crash but does not fix spend logging for streaming because _get_assembled_streaming_response still returns None for ResponsesAPIResponse, leaving complete_streaming_response unset.
tests/test_litellm/litellm_core_utils/test_litellm_logging.py	Adds four unit tests covering all new branches in _handle_anthropic_messages_response_logging; all tests are properly mocked with no network calls, but they test the helper in isolation and do not verify that _get_assembled_streaming_response produces a non-None complete_streaming_response for the streaming path.

_{Reviews (2): Last reviewed commit: "test(logging): add ResponseFailedEvent c..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Vigilans]

Fixes 28595 in the PR body is missing # — should be Fixes #28595.

This also fixes the AnthropicResponse validation crash (Error 2) reported in #27091 — same root cause, ResponseIncompleteEvent variant.

@greptileai request re-review

[greptile-apps]

Streaming spend-log still dropped after this fix

_handle_anthropic_messages_response_logging is called inside _success_handler_helper_fn before _get_assembled_streaming_response. After the fix, a ResponseCompletedEvent is converted to ResponsesAPIResponse here and returned to the caller. _get_assembled_streaming_response then receives a ResponsesAPIResponse, falls through to its else: return None branch (lines 3424-3425), and returns None. Because complete_streaming_response is None, neither response_cost nor standard_logging_object is ever set for the streaming path, so the LiteLLM_SpendLogs row is still silently dropped.

The non-streaming path is fine (self.stream is not True triggers _process_hidden_params_and_response_cost at line 1891 which builds the payload). For the streaming path, _get_assembled_streaming_response needs a ResponsesAPIResponse passthrough branch identical to the one that already exists for ModelResponse (line 3398).

[V-3604]

Closing - superseded by #29394 and #28985, which landed this exact fix in _handle_anthropic_messages_response_logging. #28985 covers the same issue (#28595).