fix(logging): preserve provider response headers in StandardLoggingPayload

[ishaan-berri]

Relevant issues

Fixes #22341

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

Type

🐛 Bug Fix

Changes

hidden_params.additional_headers in StandardLoggingPayload was always null for every request. Two root causes:

1. get_additional_headers() only iterated over the 4 annotated fields in StandardLoggingAdditionalHeaders, dropping everything else — including llm_provider-x-request-id and other provider-specific headers.
2. StandardLoggingAdditionalHeaders was missing x_ratelimit_reset_requests and x_ratelimit_reset_tokens which are already in OPENAI_RESPONSE_HEADERS.

Fix: after populating the typed fields, copy all remaining headers verbatim. Also add the two missing reset fields to the TypedDict.

Screenshots / Proof of Fix

E2E test: local proxy (openai/fake-openai-endpoint) with a custom logging callback capturing StandardLoggingPayload.

Before fix — additional_headers was always null.

After fix — callback receives:

{
  "additional_headers": {
    "llm_provider-x-railway-request-id": "hyvzt6mcTS2hjQDs9I3ezw",
    "llm_provider-server": "railway-edge",
    "llm_provider-date": "Wed, 15 Apr 2026 18:26:38 GMT",
    "llm_provider-content-type": "application/json",
    "llm_provider-x-railway-cdn-edge": "fastly/cache-pao-kpao1770021-PAO",
    "llm_provider-x-cache": "MISS",
    "x-litellm-model-group": "fake-model"
  }
}

Provider request IDs (e.g. OpenAI's x-request-id) now persist to S3, SpendLogs, and all other logging backends.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 15, 2026 7:25pm

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This PR fixes StandardLoggingPayload.hidden_params.additional_headers always being null by overhauling get_additional_headers() in two ways: (1) after populating the annotated typed fields, all remaining headers are now copied verbatim (preserving provider headers like llm_provider-x-request-id), and (2) the two missing reset fields (x_ratelimit_reset_requests, x_ratelimit_reset_tokens) are added to StandardLoggingAdditionalHeaders to align with OPENAI_RESPONSE_HEADERS. The bulk of the diff is Black reformatting with no semantic impact.

Confidence Score: 5/5

Safe to merge — clean bug fix with appropriate test coverage and no regressions.

All findings are P2 or lower. The core logic is correct: typed fields are populated first with int coercion, and remaining headers are copied verbatim using a case-sensitive key lookup that mirrors the pre-existing behavior. The TypedDict is now aligned with OPENAI_RESPONSE_HEADERS. Tests adequately cover the happy path, None input, and the reset-field case.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/litellm_logging.py	Core fix: get_additional_headers() now builds a typed_keys map then copies all unrecognised headers verbatim; remaining diff is Black reformatting with no semantic changes.
litellm/types/utils.py	Adds x_ratelimit_reset_requests: str and x_ratelimit_reset_tokens: str to StandardLoggingAdditionalHeaders, completing alignment with OPENAI_RESPONSE_HEADERS.
tests/logging_callback_tests/test_standard_logging_payload.py	Existing test updated from exact-equality assertion (which was encoding the broken behavior) to individual field checks plus new provider-header assertions; test input was already rich but old assertion masked the bug.
tests/test_litellm/litellm_core_utils/test_litellm_logging.py	Adds three focused unit tests for get_additional_headers (provider-header preservation, None input, reset fields) alongside Black import-style cleanup; no real network calls.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["get_additional_headers(additiona_headers)"] --> B{"Is None?"}
    B -- Yes --> C["return None"]
    B -- No --> D["Build typed_keys map from annotations"]
    D --> E["Loop over annotated keys"]
    E --> F{"_key present in additiona_headers?"}
    F -- Yes --> G["Coerce to int, fallback to str"]
    G --> H["Store as typed field"]
    F -- No --> I["Skip"]
    H --> J["Next annotated key"]
    I --> J
    J -- more --> E
    J -- done --> K["Loop over all additiona_headers items"]
    K --> L{"k.lower() already in typed_keys?"}
    L -- Yes --> M["Skip - already handled"]
    L -- No --> N["Copy verbatim e.g. llm_provider-x-request-id"]
    M --> O["Next header"]
    N --> O
    O -- more --> K
    O -- done --> P["return merged result"]

Loading

_{Reviews (2): Last reviewed commit: "fix(logging): update test_get_additional..." | Re-trigger Greptile}

[greptile-apps]

Removed debug log reduces int-parse observability

The original code emitted verbose_logger.debug(...) when a typed header value couldn't be cast to int, which helped diagnose unexpected provider response header formats. The new fallback silently stores the raw string instead — correct behavior, but the diagnostic signal is gone. Consider keeping a debug-level log alongside the fallback:

Suggested change

	except (ValueError, TypeError):
	verbose_logger.debug(
	f"Could not convert {additiona_headers[_key]} to int for key {key}."
	)
	additional_logging_headers[key] = additiona_headers[_key] # type: ignore
	except (ValueError, TypeError):
	verbose_logger.debug(
	f"Could not convert {additiona_headers[_key]} to int for key {key}; storing as string."
	)
	additional_logging_headers[key] = additiona_headers[_key] # type: ignore

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!

[ryan-crabbe-berri]

lgtm