PR #28311: fix(a2a): forward agent_extra_headers through completion bridge for custom_llm_provider agents

• Repository: BerriAI/litellm
• Author: VANDRANKI
• State: open | merged: False
• Link: https://github.com/BerriAI/litellm/pull/28311

Description (problem / solution / changelog)

Root Cause

When an A2A agent is configured with custom_llm_provider (e.g. LangGraph, Bedrock AgentCore), asend_message and asend_message_streaming route through _send_message_via_completion_bridge instead of the native A2A client path.

The native path correctly merges agent_extra_headers into the HTTP client:

if agent_extra_headers:
    extra_headers.update(agent_extra_headers)
a2a_client = await create_a2a_client(base_url=api_base, extra_headers=extra_headers)

The completion-bridge path silently drops them:

return await _send_message_via_completion_bridge(
    request=request,
    custom_llm_provider=custom_llm_provider,
    api_base=api_base,
    litellm_params=litellm_params,
    # agent_extra_headers not passed!
)

The same problem exists in asend_message_streaming before its handle_streaming call.

Impact

• Method 3 convention-based headers (x-a2a-{alias}-{target}) are dropped for all LangGraph/completion-bridge agents.
• Admin-allowlisted extra_headers configured on the agent are never forwarded.

Fix

1. Add agent_extra_headers: Optional[Dict[str, str]] = None to _send_message_via_completion_bridge.
2. Before calling the bridge handler, shallow-copy litellm_params and merge agent_extra_headers into its extra_headers key — this flows through to litellm.acompletion(**completion_params) because the handler already does completion_params.update(litellm_params_to_add).
3. Pass agent_extra_headers from asend_message to the bridge call.
4. Apply the same merge in the custom_llm_provider branch of asend_message_streaming before handle_streaming.

Test Plan

• Configure a LangGraph agent with custom_llm_provider: langgraph
• Send a request with x-a2a-<agent-name>-authorization: Bearer token header
• Verify the Authorization header reaches the downstream LangGraph backend
• Same test for admin-allowlisted extra_headers on the agent config
• Verify standard A2A agents (no custom_llm_provider) are unaffected

Fixes #28267

Changed files

• litellm/a2a_protocol/main.py (modified, +16/-0)