fix(bedrock): strip client_metadata from converse additionalModelRequestFields

[devin-ai-integration]

TLDR

Problem this solves:

• Codex sends client_metadata, Bedrock Converse rejects the call
• drop_params never removed it, so the route was unusable
• Anthropic and Vertex were already fixed in fix(vertex/anthropic): handle namespace tools and strip client_metadata for codex compatibility #29489

How it solves it:

• Drop client_metadata from additionalModelRequestFields in Converse
• Regression test asserts real passthrough fields still survive

Relevant issues

Linear ticket

Resolves LIT-5232

Pre-Submission checklist

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

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

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

Screenshots / Proof of Fix

Codex is the only client that sends client_metadata, so the failure only shows up on Codex -> proxy -> Bedrock. Converse forwards anything it doesn't recognize as additionalModelRequestFields, Bedrock hands that to the model verbatim, and Anthropic answers with client_metadata: Extra inputs are not permitted

Proxy config used for both runs:

model_list:
  - model_name: claude-opus-4-8
    litellm_params:
      model: bedrock/us.anthropic.claude-opus-4-8
      aws_region_name: us-west-2
litellm_settings:
  drop_params: true

Same curl for both runs, the exact body Codex sends:

curl -s -X POST http://localhost:4000/v1/responses \
  -H 'Content-Type: application/json' -H 'Authorization: Bearer sk-1234' \
  -d '{"model":"claude-opus-4-8","input":"say hi","max_output_tokens":16,"client_metadata":{"originator":"codex_cli_rs"}}'

Before, at 0659738, the request the proxy builds for Bedrock (from --detailed_debug, drop_params: true active and not helping):

-d '{"messages": [{"role": "user", "content": [{"text": "say hi"}]}], "inferenceConfig": {"maxTokens": 16}, "additionalModelRequestFields": {"client_metadata": {"originator": "codex_cli_rs"}}}'

After, at cdac5fd, same curl, no passthrough field left for Bedrock to reject:

-d '{"messages": [{"role": "user", "content": [{"text": "say hi"}]}], "inferenceConfig": {"maxTokens": 16}}'

Both runs stopped at Bedrock auth: the AWS credentials I have access to are no longer authorized for bedrock:InvokeModel on this account, so I could not get a paid 200 back. The before/after difference is in the request LiteLLM signs and sends, which is exactly the field Bedrock was rejecting, so someone with working Bedrock credentials can confirm the 200 in a minute

Until a release ships, additional_drop_params: ["client_metadata"] on the model works as a config-only workaround, and it stays a valid escape hatch afterwards

Type

🐛 Bug Fix

Changes

AmazonConverseConfig._prepare_request_params drops client_metadata next to where it already drops parallel_tool_calls, so it never reaches additionalModelRequestFields. AnthropicConfig.transform_request does the same thing for the direct Anthropic, Vertex, and Bedrock invoke paths, this closes the one remaining hole

Note that drop_params is the wrong lever here and always was: it only governs known OpenAI params a provider does not support, while client_metadata is not an OpenAI param, so it flows through the provider-specific passthrough instead

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

Link to Devin session: https://app.devin.ai/sessions/e57cd0ce775749e595f7da41d0795378

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

The PR prevents Codex’s unsupported client_metadata field from being forwarded through Bedrock Converse while preserving legitimate model-specific passthrough fields.

• Removes client_metadata before constructing additionalModelRequestFields.
• Adds regression coverage confirming removal while retaining anthropic_beta.

Confidence Score: 5/5

The PR appears safe to merge, with the unsupported metadata removed from every Bedrock Converse request-construction path.

The exact field is removed before additional model request fields are serialized, both synchronous and asynchronous paths use that transformation, and the regression test confirms legitimate passthrough data remains intact.

Important Files Changed

Filename	Overview
litellm/llms/bedrock/chat/converse_transformation.py	Safely removes the unsupported field from the common synchronous and asynchronous Converse request-building path.
tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py	Adds focused regression coverage that verifies the offending field is removed without dropping a valid passthrough field.

_{Reviews (1): Last reviewed commit: "fix(bedrock): strip client_metadata from..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_bedrock_strip_client_metadata (cdac5fd) with litellm_internal_staging (54fb717)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (0659738) during the generation of this report, so 54fb717 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[pjahanlou-godaddy]

Do we have plans on reviewing and getting this into main?