Root cause fix - migrate all logging update to use 1 function - for centralized kwarg updates

[ghost]

Relevant issues

Root cause fix for #23185

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
• 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

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

🆕 New Feature
🐛 Bug Fix
🧹 Refactoring
📖 Documentation
🚄 Infrastructure
✅ Test

Changes

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 14, 2026 8:24pm

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bac05524a2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Include explicit metadata when logging responses calls

responses() takes metadata as an explicit parameter, but this pre-call logging now passes only kwargs into update_from_kwargs; since explicit params are not inside kwargs, request metadata is dropped from litellm_params for this path. Any callback/cost logic that depends on model_call_details["litellm_params"]["metadata"] (the behavior preserved before this change) will silently lose metadata when users call litellm.responses(..., metadata=...).

Useful? React with 👍 / 👎.

[greptile-apps]

Greptile Summary

This PR introduces a centralized update_from_kwargs helper on the Logging class that automatically extracts metadata and litellm_metadata from the raw kwargs dict, then delegates to the existing update_environment_variables. All 22 call sites across the SDK's non-chat endpoints (batches, containers, evals, images, responses, videos, vector stores, etc.) are migrated to the new helper, eliminating the per-call boilerplate of manually plumbing "metadata": kwargs.get("metadata", {}) into litellm_params. This is the root-cause fix for #23185, where routes like /messages that stored pricing info under kwargs["litellm_metadata"] were not having that data forwarded to use_custom_pricing_for_model.

Key observations:

• Core logic is sound — update_from_kwargs first collects metadata/litellm_metadata from kwargs, then merges explicitly-passed litellm_params on top (so callers can always override), and finally calls the unchanged update_environment_variables. The merge order is tested thoroughly.
• Test suite is comprehensive at the unit level — TestUpdateFromKwargs in test_litellm_logging.py covers extraction, backfill, priority, custom-pricing detection, and empty-kwargs scenarios.
• The regression test for [Bug]: Custom Pricing is broken for /messages and /responses endpoints (streaming) #23185 has been weakened — test_async_anthropic_messages_handler_passes_litellm_metadata in test_llm_http_handler.py now mocks out update_from_kwargs entirely and only asserts that litellm_metadata is present in the forwarded kwargs dict. It no longer verifies that the metadata ends up in litellm_params (where use_custom_pricing_for_model reads it). The dedicated unit tests compensate partially but the integration guard is weaker.
• builtins.list fix in vector_stores/main.py — a subtle but correct fix for a local function named list shadowing the Python builtin in an isinstance check.
• Pre-submission checklist is entirely unchecked — no CI links or test evidence is provided in the PR description.

Confidence Score: 4/5

• Safe to merge with minor concern about the weakened regression test for [Bug]: Custom Pricing is broken for /messages and /responses endpoints (streaming) #23185.
• The core fix is well-designed and broadly applied. Unit tests for the new update_from_kwargs method are thorough. The one deduction is that the original integration regression test for [Bug]: Custom Pricing is broken for /messages and /responses endpoints (streaming) #23185 was weakened by mocking out update_from_kwargs, meaning a future extraction-logic bug would no longer be caught by the handler-level test. The pre-submission checklist is also entirely unchecked with no CI links provided.
• tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py — the regression guard for [Bug]: Custom Pricing is broken for /messages and /responses endpoints (streaming) #23185 is weaker than before; consider strengthening it.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/litellm_logging.py	Adds new update_from_kwargs convenience method that automatically extracts metadata/litellm_metadata from the raw kwargs dict, centralising the metadata forwarding logic that was previously scattered across all callers.
tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py	The regression test for issue #23185 was updated to mock update_from_kwargs instead of update_environment_variables, but the new assertion only checks that litellm_metadata is in the kwargs dict — not that it is correctly extracted into litellm_params, weakening the original regression guard.
tests/test_litellm/litellm_core_utils/test_litellm_logging.py	Adds comprehensive TestUpdateFromKwargs test class covering metadata extraction, backfill logic, priority order, custom-pricing detection, and empty-kwargs edge cases — good unit coverage of the new method.
litellm/llms/custom_httpx/llm_http_handler.py	Migrates the Anthropic messages handler to update_from_kwargs, removing the manual metadata/litellm_metadata extraction from litellm_params — the direct call site for issue #23185.
litellm/responses/main.py	Migrates all six response API entry points (responses, delete_responses, get_responses, list_input_items, cancel_responses, compact_responses) to update_from_kwargs, correctly using local_vars (which already includes kwargs) as the source dictionary.
litellm/videos/main.py	Migrates video API logging to update_from_kwargs; also adds seconds, extra_headers, extra_query, extra_body to overloads and removes api_key/api_base/api_version from typed overloads (these remain accessible via **kwargs).
litellm/batches/main.py	Migrates create_batch and retrieve_batch to update_from_kwargs; removes the explicit "metadata": metadata key from the litellm_params dict, delegating extraction to the new helper.
litellm/vector_stores/main.py	Migrates all vector store operations; also adds import builtins to safely use builtins.list in the mock-response isinstance check where the local list function shadowed the builtin.
tests/test_litellm/images/test_image_edit_utils.py	Updates capturing mock to intercept update_from_kwargs; manually mirrors the metadata-extraction logic inside the capturing stub so the downstream assertion on captured_litellm_params remains valid.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Call site\n(batches, responses, images, videos, etc.)"] -->|"update_from_kwargs(kwargs=kwargs, litellm_params=..., ...)"| B["update_from_kwargs"]

    B --> C{"'metadata'\nin kwargs?"}
    C -->|Yes| D["base['metadata'] = kwargs['metadata']"]
    C -->|No| E{"'litellm_metadata'\nin kwargs?"}
    D --> E

    E -->|"Yes (and is dict)"| F["base['litellm_metadata'] = kwargs['litellm_metadata']"]
    F --> G{"'metadata' already\nset in base?"}
    G -->|No| H["base['metadata'] = litellm_metadata.copy()"]
    G -->|Yes| I["skip backfill"]
    E -->|No| J["no-op"]

    H --> K["Merge caller litellm_params\nover base (caller wins)"]
    I --> K
    J --> K

    K --> L["update_environment_variables(\n  litellm_params=merged,\n  optional_params=...,\n  model=..., user=...\n)"]

    L --> M["self.litellm_params updated\n(metadata + litellm_metadata\nnow present)"]
    M --> N["use_custom_pricing_for_model\nchecks litellm_params\n→ detects custom pricing ✓"]

Loading

Comments Outside Diff (1)

1. tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py, line 186-233 (link)

   Regression test weakened for issue [Bug]: Custom Pricing is broken for /messages and /responses endpoints (streaming) #23185

   The original regression test for [Bug]: Custom Pricing is broken for /messages and /responses endpoints (streaming) #23185 verified the complete end-to-end data flow: that litellm_metadata ended up inside litellm_params as seen by use_custom_pricing_for_model. The new test only verifies the first hop — that litellm_metadata is present in the kwargs dict forwarded to update_from_kwargs. Since update_from_kwargs is fully mocked (Mock()), the test no longer validates that the metadata is correctly extracted and placed into litellm_params at all.

   If a future regression broke the extraction logic inside update_from_kwargs, this test would still pass. The unit tests in TestUpdateFromKwargs partially compensate, but the integration regression from [Bug]: Custom Pricing is broken for /messages and /responses endpoints (streaming) #23185 is no longer guarded here.

   Consider removing the mock for update_from_kwargs and instead asserting on logging_obj.litellm_params after the call, similar to how TestUpdateFromKwargs.test_custom_pricing_detected_via_litellm_metadata works:

   # Instead of:
   mock_logging_obj.update_from_kwargs = Mock()
   # ...
   mock_logging_obj.update_from_kwargs.assert_called_once()
   kwargs_arg = ...
   assert "litellm_metadata" in kwargs_arg
   
   # Prefer a real logging obj + assert the downstream effect:
   assert "litellm_metadata" in mock_logging_obj.litellm_params
   assert mock_logging_obj.litellm_params["litellm_metadata"]["model_info"] == custom_model_info

   Rule Used: # Code Review Rule: Mock Test Integrity

   What:... (source)

_{Last reviewed commit: 90252ee}