fix(bedrock): strip output_config from Converse requests; fix spend tracking redaction test

[Sameerlite]

Summary

• Strip output_config (snake_case) from Bedrock Converse requests — Nova and other Converse models reject this Anthropic-only param
• Fix test_spend_logs_redacts_request_and_response_when_turn_off_message_logging_enabled to match actual perform_redaction behavior (in-place content redaction)

Changes

• converse_transformation.py: pop output_config before building request
• test_converse_transformation.py: add test_output_config_snake_case_stripped_from_bedrock_converse_request
• test_spend_tracking_utils.py: update assertion for dict response redaction

Made with Cursor

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Mar 10, 2026 6:25am

[greptile-apps]

Greptile Summary

This PR contains two focused bug fixes:

1. output_config leaking into Bedrock Converse requests – When the Anthropic-style snake_case output_config parameter is passed (either by users or internal param-mapping code), it was not being stripped and ended up in additionalModelRequestFields, causing Nova and other non-Claude Converse models to reject the request. The one-line fix pops output_config unconditionally before the supported/unsupported parameter split, mirroring how the camelCase outputConfig is already handled just above it.

2. Incorrect spend-log redaction test assertion – The pre-existing test asserted that the redacted response would equal {"text": "redacted-by-litellm"}, but perform_redaction actually mutates the response in-place and leaves the full choices structure intact, with only message.content replaced. The updated assertions match the actual behaviour.

Key observations:

• The production change is a single, unconditional pop — it is applied to all Bedrock Converse requests, which is correct because output_config (snake_case) is Anthropic-specific and has no Bedrock Converse equivalent; the Bedrock-native outputConfig path is unaffected.
• All new tests are mock-only, consistent with repository conventions.
• No model-specific flags are hardcoded in production code; the strip is applied universally.

Confidence Score: 5/5

• This PR is safe to merge; both changes are minimal, well-scoped bug fixes with accompanying tests.
• The production change is a single unconditional pop that removes a parameter known to break non-Claude Bedrock models. The test fix aligns an incorrect assertion with observed behaviour. No backwards-incompatible API surface changes, no new dependencies, no real network calls in tests.
• No files require special attention.

_{Last reviewed commit: 200b001}