litellm - 💡(How to fix) Fix [Bug]: Bedrock pass-through /converse returns 200 with empty body on v1.85.0

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

Summary

After upgrading LiteLLM proxy from v1.83.10-stable to 1.85.0, native Bedrock pass-through calls to the non-streaming Converse endpoint return HTTP 200 with Content-Length: 0 and an empty body.

LiteLLM observability/logging records the upstream Bedrock call as successful and shows output tokens, but the HTTP client receives no body. AWS boto3 then parses only ResponseMetadata, with no output, causing downstream callers to fail with KeyError: 'output'.

Version

• LiteLLM Docker image: docker.io/litellm/litellm-database:1.85.0
• Previous working image: docker.io/litellm/litellm-database:v1.83.10-stable
• Route: /bedrock/model/{modelId}/converse
• Model tested: us.anthropic.claude-haiku-4-5-20251001-v1:0

Minimal repro

Call native Bedrock pass-through Converse:

curl -i "$LITELLM_PROXY_URL/bedrock/model/us.anthropic.claude-haiku-4-5-20251001-v1%3A0/converse" \
  -H "Authorization: Bearer $LITELLM_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  --data '{
    "messages": [
      {
        "role": "user",
        "content": [{"text": "Reply with exactly: ok"}]
      }
    ],
    "inferenceConfig": {
      "maxTokens": 16,
      "temperature": 0
    }
  }'

Actual behavior

The response is successful but empty:

HTTP/1.1 200 OK
content-type: application/json
content-length: 0
x-litellm-version: 1.85.0

Body length is 0.

When using boto3 with endpoint_url=$LITELLM_PROXY_URL/bedrock, the parsed response contains only ResponseMetadata and does not contain output.

Example downstream failure:

KeyError: 'output'

Expected behavior

The /converse response should include the Bedrock Converse JSON body, including output.message.content.

The response should not send Content-Length: 0 when LiteLLM has already read a non-empty upstream response body.

Possible related change

This may be related to a response-header merge introduced between v1.83.10-stable and 1.85.0.

In v1.83.10-stable, the non-streaming pass-through response path used:

headers=HttpPassThroughEndpointHelpers.get_response_headers(
    headers=result.headers,
    custom_headers=None,
)

In 1.85.0, this became:

headers=HttpPassThroughEndpointHelpers.get_response_headers(
    headers=result.headers,
    custom_headers=dict(fastapi_response.headers),
)

Possibly related PR: https://github.com/BerriAI/litellm/pull/27412

Notes

This appears specific to native Bedrock pass-through non-streaming /converse.

These paths did not show the same failure in our testing:

• /bedrock/model/{modelId}/converse-stream
• OpenAI-compatible /v1/chat/completions

Steps to Reproduce

Call native Bedrock pass-through Converse:

curl -i "$LITELLM_PROXY_URL/bedrock/model/us.anthropic.claude-haiku-4-5-20251001-v1%3A0/converse" \
  -H "Authorization: Bearer $LITELLM_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  --data '{
    "messages": [
      {
        "role": "user",
        "content": [{"text": "Reply with exactly: ok"}]
      }
    ],
    "inferenceConfig": {
      "maxTokens": 16,
      "temperature": 0
    }
  }'

Relevant log output

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

1.85.0

Twitter / LinkedIn details

No response