fix(logging): dedupe success logging for non-streaming anthropic_messages (/v1/messages)

[anneheartrecord]

Summary

A single non-streaming POST /v1/messages (anthropic_messages) emitted two litellm_request OTEL spans and fired the success/cost callbacks twice. Plain /chat/completions was unaffected.

Closes #31121

Root cause

/v1/messages is always dispatched through the async wrapper, so logging runs via async_success_handler. The async path also schedules the sync success_handler (via handle_sync_success_callbacks_for_async_calls) for legacy sync callbacks. That sync pass is supposed to skip CustomLogger dispatch when async already handled it — but the guard only excludes call_type == pass_through.

Unlike acompletion, anthropic_messages sets no async flag in litellm_params, so _is_sync_litellm_request() returns True and is_sync_request is True in the sync handler. With anthropic_messages not excluded, the sync pass re-dispatches every CustomLogger (OpenTelemetry, _ProxyDBLogger cost tracking, …) on top of the async one → two litellm_request spans and double cost/success callbacks. This is the non-streaming counterpart of #29550 / #29598, whose fix is gated on stream is True.

Change

Extend the existing pass_through skip in success_handler to also cover anthropic_messages, so the sync pass no longer re-dispatches CustomLoggers — logging happens once via the canonical async path, matching acompletion.

Verifying it doesn't drop logging

End-to-end with litellm.anthropic_messages(..., mock_response=...) and a counting CustomLogger:

	sync log_success_event	async async_log_success_event
before	1	1
after	0	1

Callbacks that reach the async path (otel via success_callback: ["otel"], _ProxyDBLogger, any logger in litellm.callbacks) keep their single async invocation and lose only the duplicate. A CustomLogger reachable only via the sync success_callback list is already not invoked on acompletion either, so this change makes anthropic_messages consistent with the other async paths rather than regressing them.

Tests

Added test_success_handler_skips_sync_callbacks_for_anthropic_messages, mirroring the existing sync/async callback tests. It fails before the change (sync callback fires) and passes after. Full tests/test_litellm/litellm_core_utils/test_litellm_logging.py passes (104).

[codspeed-hq]

Congrats! CodSpeed is installed 🎉

🆕 16 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• test_cost_per_token_openai: 560.1 µs
• test_get_model_info_with_provider: 79.7 µs
• test_token_counter_long_content: 1.7 ms
• test_get_llm_provider_anthropic: 142 µs
• test_get_model_info_openai: 80.9 µs
• test_cost_per_token_anthropic: 559.2 µs
• test_token_counter_with_tools: 411.4 µs
• test_token_counter_multi_turn: 565.1 µs
• test_get_llm_provider_azure: 145.4 µs
• test_token_counter_simple_message: 235.1 µs
• test_token_counter_raw_text: 183.3 µs
• test_get_llm_provider_openai: 138.2 µs
• test_get_model_cost_key_exact_match: 79.7 µs
• test_get_llm_provider_with_prefix: 138.1 µs
• test_get_model_info_anthropic: 78.9 µs
• test_get_model_cost_key_case_insensitive: 82.1 µs

---

[greptile-apps]

Greptile Summary

This PR fixes double-invocation of CustomLogger success callbacks (OTEL spans, cost tracking, etc.) for non-streaming POST /v1/messages (anthropic_messages) requests. The anthropic_messages endpoint is always dispatched through the async wrapper, but — unlike acompletion — it does not set an async flag in litellm_params, so _is_sync_litellm_request() returns True and the sync pass in success_handler re-dispatches every CustomLogger on top of the async dispatch.

• Extends the call_type not in (...) guard in success_handler to also include CallTypes.anthropic_messages.value alongside the existing pass_through exclusion, matching the behaviour established for other async-only routes.
• Adds test_success_handler_skips_sync_callbacks_for_anthropic_messages which directly verifies that log_success_event is not called from the sync path when call_type == "anthropic_messages".

Confidence Score: 5/5

Safe to merge — the change is a minimal, targeted guard extension in the sync callback dispatcher with a direct regression test.

The one-line condition extension is well-understood: CallTypes.anthropic_messages is a valid enum value (confirmed at litellm/types/utils.py:351), the async failure path has no equivalent handle_sync_failure_callbacks_for_async_calls so the failure handler is not affected by the same double-dispatch pattern, no existing tests were weakened, and the new test directly exercises the fixed condition.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/litellm_logging.py	Extends the CustomLogger re-dispatch exclusion in success_handler to cover anthropic_messages in addition to pass_through; change is surgical and well-scoped.
tests/test_litellm/litellm_core_utils/test_litellm_logging.py	Adds a new test verifying the sync CustomLogger path is skipped for anthropic_messages; no existing tests modified.

_{Reviews (1): Last reviewed commit: "fix(logging): dedupe success logging for..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[anneheartrecord]

Rebased onto the current litellm_internal_staging and retargeted the base to it (the branch was previously cut from an older integration branch). The change is unchanged in scope — 2 files, +59/-1: anthropic_messages (/v1/messages) is always dispatched through the async success path, so the sync success_handler pass now skips it to avoid double-firing CustomLogger callbacks (duplicate OTEL litellm_request spans + double cost/success), matching how pass_through is already handled. Regression test included and green locally (test_success_handler_skips_sync_callbacks_for_anthropic_messages). Ready for review 🙏 — resolves #31121.

[anneheartrecord]

@Sameerlite could you take a look at this one when you get a chance? It dedupes the double success logging (and double OTEL span) for non-streaming /v1/messages by extending the existing pass_through skip to anthropic_messages. Rebased onto current litellm_internal_staging; the two red checks are the fork-base guard and a pre-existing openapi_compliance failure on base, both unrelated.