fix(anthropic): pass logging_obj to client.post for litellm_overhead_time_ms

[milan-berri]

Relevant issues

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• 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 (anthropic chat handler tests: poetry run pytest tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py -v)
• 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

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🐛 Bug Fix

Changes

• Fix: Pass logging_obj from the Anthropic chat handler into every client.post() / async_handler.post() call so the @track_llm_api_timing() decorator on the HTTP handler can set model_call_details["llm_api_duration_ms"]. Without it, litellm_overhead_time_ms stayed null when using Anthropic with LITELLM_DETAILED_TIMING=true.
• Files: litellm/llms/anthropic/chat/handler.py (4 call sites: async/sync streaming in make_call/make_sync_call, async/sync non-streaming in acompletion/completion).
• Test: test_make_call_passes_logging_obj_to_client_post in tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py ensures make_call passes logging_obj to client.post.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 19, 2026 1:26am

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing milan-berri:fix/anthropic-litellm-overhead-time-ms (9e2de44) with main (488b93c)}

[greptile-apps]

Greptile Summary

This PR fixes a bug where litellm_overhead_time_ms remained null when using the Anthropic provider with LITELLM_DETAILED_TIMING=true. The root cause was that logging_obj was not being forwarded to client.post() / async_handler.post() calls in litellm/llms/anthropic/chat/handler.py, so the @track_llm_api_timing() decorator on AsyncHTTPHandler.post / HTTPHandler.post couldn't retrieve it from kwargs to record llm_api_duration_ms on the logging object.

Changes:

• logging_obj is now passed as a keyword argument at all four call sites: async streaming (make_call), sync streaming (make_sync_call), async non-streaming (acompletion_function), and sync non-streaming (completion).
• A new pytest.mark.asyncio unit test verifies that make_call correctly forwards logging_obj to client.post, using mocks with no real network calls.

Observations:

• The fix itself is minimal, correct, and consistent across all four paths.
• The test only covers the async streaming path; the sync streaming and non-streaming paths have no corresponding tests, leaving three of the four fixes without test coverage.
• The "Relevant issues" section in the PR description is empty — no issue number is linked despite the checklist item for it.

Confidence Score: 4/5

• Safe to merge; the change is minimal and correct, with the only gap being partial test coverage of the four fixed call sites.
• The core fix — passing logging_obj to client.post() at all four call sites — is straightforward, well-scoped, and aligns exactly with how the @track_llm_api_timing() decorator extracts the object from kwargs. The HTTP handler already accepted logging_obj as an optional parameter, so no interface changes are needed. The score is 4 rather than 5 because only one of the four fixed paths has a unit test; the sync streaming and non-streaming paths lack test coverage, creating a small risk of silent regression.
• No files require special attention; both changed files are correct. The test file could benefit from broader coverage of the sync and non-streaming call sites.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/chat/handler.py	Passes logging_obj to all four client.post() / async_handler.post() call sites so the @track_llm_api_timing() decorator can extract it from kwargs and set llm_api_duration_ms on the logging object; change is minimal and correct.
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py	Adds a mocked async test verifying make_call forwards logging_obj to client.post; test is correct and uses no real network calls, but only covers one of the four fixed call sites (async streaming); sync streaming and non-streaming paths remain untested.

Sequence Diagram

sequenceDiagram
    participant Caller
    participant AnthropicChatCompletion as handler.py
    participant HTTPHandler as AsyncHTTPHandler/HTTPHandler
    participant Decorator as track_llm_api_timing()

    Caller->>AnthropicChatCompletion: completion() / acompletion()
    AnthropicChatCompletion->>HTTPHandler: client.post(..., logging_obj=logging_obj)
    Note over HTTPHandler,Decorator: Decorator intercepts call
    Decorator->>Decorator: start_time = now()<br/>logging_obj = kwargs.get("logging_obj")
    Decorator->>HTTPHandler: execute post()
    HTTPHandler-->>Decorator: response
    Decorator->>Decorator: end_time = now()<br/>logging_obj.model_call_details["llm_api_duration_ms"] = duration_ms
    Decorator-->>AnthropicChatCompletion: response
    AnthropicChatCompletion-->>Caller: ModelResponse / StreamWrapper

Loading

_{Last reviewed commit: "fix(anthropic): pass..."}

[greptile-apps]

Use .kwargs instead of index-based call_args[1]

call_args[1] is the legacy tuple-indexing API for accessing keyword arguments on a call object. The modern, more readable approach is .kwargs. Both work, but the newer form is clearer and less likely to break if positional args are added.

Suggested change

	mock_client.post.assert_called_once()
	call_kwargs = mock_client.post.call_args[1]
	assert call_kwargs.get("logging_obj") is logging_obj
	call_kwargs = mock_client.post.call_args.kwargs
	assert call_kwargs.get("logging_obj") is logging_obj

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

[greptile-apps]

Test only covers one of four fixed call sites

The PR fixes logging_obj being passed at four separate call sites in handler.py:

1. make_call (async streaming) — ✅ covered by this test
2. make_sync_call (sync streaming) — ❌ not tested
3. acompletion_function async non-streaming — ❌ not tested
4. completion sync non-streaming — ❌ not tested

Per the project's testing guidelines (at least 1 test is a hard requirement), this technically satisfies the bar, but adding tests for the sync path and non-streaming paths would prevent regressions where a future refactor accidentally drops logging_obj from one of those call sites.

Consider adding parallel tests for make_sync_call and the non-streaming paths.

[milan-berri]

before and after: