feat(a2a_client): expose extension metadata in a2a_send_message

[prashant1rana]

Summary

Pass A2A request metadata to sub-agents via tool_context instead of exposing it as an LLM parameter.

Motivation

The A2A protocol's MessageSendParams.metadata field enables passing metadata data between agents.

The original approach exposed this as an LLM-facing parameter, allowing the model to fabricate values that could cause:

• Wrong session routing
• Failed authentication
• Misleading trace data
• Billing issues on remote servers

While hallucinated metadata won't break the protocol (it's optional JSON), it creates semantic risks when remote servers interpret fabricated values.

Solution

Use @tool(context=True) to hide tool_context from the LLM schema and extract metadata transparently:

python
@tool(context=True)
async def a2a_send_message(
    self,
    message_text: str,
    target_agent_url: str,
    message_id: str | None = None,
    tool_context: ToolContext | None = None,  # Framework-injected, not LLM-visible
) -> dict[str, Any]:
    metadata = None
    if tool_context is not None:
        metadata = tool_context.get("invocation_state", {}).get("metadata")

    return await self._send_message(message_text, target_agent_url, message_id, metadata)

How it works:

1. Parent A2A request arrives with MessageSendParams.metadata = {"session_id": "abc"}
2. Strands framework stores in tool_context.invocation_state.metadata
3. LLM calls a2a_send_message (only sees message_text, target_agent_url, message_id)
4. Framework injects real tool_context at runtime
5. Tool extracts and forwards metadata to sub-agent
6. LLM never sees or controls metadata—transparent passthrough

Testing

• ✅ 36/36 tests passing
• ✅ Validates extraction from tool_context
• ✅ Handles None and empty tool_context cases
• ✅ Proper ToolContext type annotations

[mkmeral]

Hi, can I ask what the use case is? The way you are adding the metadata right now, it's another parameter for LLM to add. Is that the desired behavior?

Would it be better to get it through tool_context.invocation_state["a2a_metadata"]?

[prashant1rana]

Hi, can I ask what the use case is? The way you are adding the metadata right now, it's another parameter for LLM to add. Is that the desired behavior?

I was considering an approach like this on the client side to have greater control over the metadata.

class MetadataAwareA2AClient(A2AClientToolProvider):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._original_send_message = self._send_message

    @tool
    async def a2a_send_message(
        self,
        message_text: str,
        target_agent_url: str,
        message_id: str | None = None,
        metadata: dict[str, Any] | None = None,
        tool_context: dict[str, Any] | None = None,
        timestamp: str | None = None,
    ) -> dict[str, Any]:
        if metadata is None and tool_context:
            invocation_state = tool_context.get("invocation_state", {})
            metadata = invocation_state.get("metadata", {})

        updated_metadata = {**(metadata or {}), "timestamp": timestamp}

        return await self._original_send_message(message_text, target_agent_url, message_id, updated_metadata)

[prashant1rana]

Would it be better to get it through tool_context.invocation_state["a2a_metadata"]?

If no metadata is provided, I think we could auto-detect it as a fallback.

async def a2a_send_message(
    self,
    message_text: str,
    target_agent_url: str,
    message_id: str | None = None,
    metadata: dict[str, Any] | None = None,
    tool_context: ToolContext | None = None, 
) -> dict[str, Any]:
    # Auto-extract metadata from tool_context
    if metadata is None and tool_context is not None:
        invocation_state = tool_context.invocation_state
        metadata = invocation_state.get("metadata")
    
    return await self._send_message(message_text, target_agent_url, message_id, metadata)

[prashant1rana]

Hi, can I ask what the use case is?

My use case involves passing a portion of the metadata to the A2A sub-agents, which can be solved directly using tool_context.

[mkmeral]

So, I have done a bit more research (see below). The concern is now that we are adding this tool, it can break anyone who uses it, because now LLM can hallucinate metadata, and it will be sent to the A2A servers. It can have unintended consequences.

I think we should to move to tool context method only, unless there is an actual need to generate metadata from LLM, but even then, I'd consider it risky for our core SDK, because it might impact other developers who already use the tool

I was considering an approach like this on the client side to have greater control over the metadata.

I think we are aligned in the sense that we want devs to have greater control over metadata, but this approach gives the control to LLM by default, and it's risky

Agent investigation: A2A metadata hallucination risk analysis

What the PR does: Adds an optional metadata: dict[str, Any] | None parameter to a2a_send_message, which gets passed through to the A2A SDK's client.send_message(message, request_metadata=metadata). This ends up in MessageSendParams.metadata in the JSON-RPC request.

What A2A metadata actually is: In the A2A protocol spec, metadata on MessageSendParams is defined as an optional google.protobuf.Struct (basically a free-form JSON dict). It's described as "Extensions Metadata attached to the request" in the a2a-python SDK. It's meant for passing extension-specific data alongside the message, things like session IDs, trace IDs, or custom extension payloads. The receiving A2A server can read it or ignore it.

Your core question: if the LLM hallucinates metadata, would it break stuff?

Short answer: it won't break the protocol, but it's a real concern.

Here's why:

1. Protocol-level safety: The metadata field is optional and free-form (dict[str, Any]). The A2A spec and SDK treat it as opaque extension data. Unknown keys are simply ignored by servers that don't care about them. So hallucinated metadata like {"priority": "urgent", "session_id": "abc123"} won't cause protocol errors or crashes.

2. The real risk is semantic, not structural. If the LLM fabricates metadata that a receiving server does interpret (for example, a session_id that doesn't exist, or a trace_id that's garbage), the remote agent could behave unexpectedly. It won't crash, but it might route to the wrong session, log misleading trace data, or trigger extension logic with nonsensical inputs.

3. Your comment on the PR is spot on. Exposing metadata as a direct tool parameter means the LLM decides what to put in it. That's the fundamental issue. The LLM has no way to know what metadata keys the remote A2A server expects or cares about. It'll just make something up based on the parameter description. Using tool_context.invocation_state["a2a_metadata"] instead would let the developer control the metadata programmatically, keeping the LLM out of the loop for this field.

4. Extension-specific metadata can have real consequences. The A2A protocol's extension system means servers can define custom behaviors keyed on metadata. If an extension expects {"auth_token": "..."} and the LLM hallucinates a fake token, that's not just noise, it's a failed auth attempt. If it expects {"billing_account": "..."}, hallucinated values could cause billing issues on the remote side.

Bottom line: Hallucinated metadata won't break the wire protocol (it's just JSON in an optional field), but it can cause incorrect behavior on servers that actually consume that metadata. The safer pattern is what you suggested: source metadata from tool_context.invocation_state so developers control it, rather than letting the LLM populate a free-form dict it doesn't understand.

[mkmeral]

We should not allow LLM to send metadata by default, as it can impact. I am leaving the review for now to make sure we do not unintentionally merge this

[prashant1rana]

We should not allow LLM to send metadata by default, as it can impact. I am leaving the review for now to make sure we do not unintentionally merge this

Thanks for calling out, hallucinating metadata could cause real issues. Let me revise the PR.