Metadata Not Passed to Callbacks in Anthropic Messages with Langfuse Integration

[ravitejaseelam]

Description

We have recently integrated Langfuse into our setup, and we’re facing an issue where metadata passed to anthropic_messages is not being propagated to callbacks.
This is impacting our ability to capture observability and tracing data effectively.

Context

The same implementation works perfectly fine when using the /chat/completions endpoint.
However, when switching to the /v1/messages endpoint (for Anthropic models), the metadata is not being forwarded to callbacks.

As a temporary workaround, we tried proxying Anthropic through the OpenAI client, but that introduced several unrelated issues (outside LiteLLM) that are more severe and difficult to manage.

Example Request

curl -X POST http://<proxy-server-ip>:4000/v1/messages \
  -H "Authorization: Bearer sk-************" \
  -d '{
    "model": "claude-3-7-sonnet-latest",
    "messages": [
      {"role": "user", "content": "What is the capital of France?"}
    ],
    "max_tokens": 100,
    "litellm_session_id": "ceb6a9bc-****-****-****-************",
    "metadata": {
      "tags": [
        "[email protected]",
        "entc0754cb"
      ],
      "trace_id": "390c2d53-****-****-****-************",
      "session_id": "ceb6a9bc-****-****-****-************",
      "trace_user_id": "[email protected]"
    }
  }'

Observed Behavior

As seen in the above request, the metadata field is present and correctly structured.
However, it does not get passed to callbacks when using the /v1/messages endpoint with Anthropic models.
In contrast, the same metadata works correctly when using the /chat/completions endpoint.

Expected Behavior

Metadata should be available in callbacks for Anthropic models (via /v1/messages), just as it works with the OpenAI client and /chat/completions endpoint.

Environment

• LiteLLM Docker Image: ghcr.io/berriai/litellm:main-latest
• LiteLLM Version: 1.78.0
• Anthropic Models: Claude-Sonnet 3.7, Claude-Sonnet 4
• Langfuse Version: 3.117.2

Request

We would really appreciate any alternative approach or an update on whether there’s a roadmap to fix this issue.

Reference

A similar issue was reported recently here:
🔗 #15103