Observed Behavior

A2AExecutor (agent-framework-a2a) passes only the text of the incoming A2A message to the agent.
Raw (bytes), URL and data parts are dropped.

• A client sends "What is in this picture?" plus a PNG. The agent receives only the string "What is in this picture?" and never sees the image.
• A client sends only a file (for example a PDF). The agent is called with an empty string "".

This also breaks the framework's own round trip: A2AAgent sends Content.from_data(...) / Content.from_uri(...) as raw and URL parts, and an A2AExecutor server throws them away.
The class docstring says the executor supports "text, binary data, and URI-based content", but that is only true for output.

Expected Behavior

Every part of the incoming message reaches the agent: text as text, raw parts as data content, URL parts as URI content, with media type and filename kept.
A text-only message can keep being passed as plain text.

Steps to Reproduce

1. Wrap any agent in A2AExecutor.
2. Call execute() with a RequestContext whose message has a text part and a raw image part.
3. Look at what the agent's run() receives.

Minimal Reproduction

import asyncio

from a2a.server.agent_execution import RequestContext
from a2a.server.context import ServerCallContext
from a2a.server.events import EventQueueLegacy
from a2a.types import Message as A2AMessage
from a2a.types import Part, Role, SendMessageRequest
from agent_framework import AgentResponse, AgentSession, Message
from agent_framework.a2a import A2AExecutor


class RecordingAgent:
    """Stands in for any agent and records what the executor passes to run()."""

    id = "recording"
    name = "recording"
    description = None

    def create_session(self, **kwargs):
        return AgentSession(**kwargs)

    def get_session(self, service_session_id, **kwargs):
        return AgentSession(service_session_id=service_session_id, **kwargs)

    async def run(self, messages=None, **kwargs):
        self.received = messages
        return AgentResponse(messages=[Message(role="assistant", contents=["ok"])])


async def main() -> None:
    agent = RecordingAgent()
    incoming = A2AMessage(
        message_id="m1",
        role=Role.ROLE_USER,
        parts=[
            Part(text="What is in this picture?"),
            Part(raw=b"\x89PNG\r\n\x1a\nfake", media_type="image/png", filename="cat.png"),
            Part(url="https://example.com/report.pdf", media_type="application/pdf"),
        ],
    )
    context = RequestContext(
        ServerCallContext(),
        request=SendMessageRequest(message=incoming),
        task_id="t1",
        context_id="c1",
    )

    await A2AExecutor(agent).execute(context, EventQueueLegacy())

    print("agent.run received:", repr(agent.received))


asyncio.run(main())

Output on main (08851e6):

agent.run received: 'What is in this picture?'

Error Messages and Stack Traces

No error. The image and the PDF link are dropped.

Package Versions

agent-framework-core: 1.22.0, agent-framework-a2a: 1.0.0b261015 (main at 08851e6), a2a-sdk: 1.2.1

Python Version

Python 3.14.5

Operating System

macOS

Regression

No

Additional Context

Root cause: python/packages/a2a/agent_framework_a2a/_a2a_executor.py, A2AExecutor.execute, line 206 on main:

query = context.get_user_input()

RequestContext.get_user_input() in a2a-sdk joins only the text parts ("Extracts text content from the user's message parts"). That string is what _run / _run_stream pass to agent.run(), so all other parts are lost before the agent is called.

The newer agent-framework-hosting-a2a package already converts URL and raw parts into Content (_conversion.py), and A2AAgent._parse_contents_from_a2a does the same on the client side. Only A2AExecutor uses the text-only path.

Suggested fix: when the message has any non-text part, build a user Message from all parts (text, raw → data, url → uri, data → JSON text, keeping media_type and filename). Keep passing the joined text for text-only messages so existing behavior does not change. I have this with four tests using a real a2a-sdk RequestContext (text plus image, file only, data part, and a text-only guard); the first three fail on main and pass with the fix. Happy to open the PR once this is agreed.

AI assistance: found and reproduced with an AI coding assistant; I ran the repro and checked the code myself.

Acknowledgements

• I searched existing issues and did not find a duplicate.
• I personally verified this behavior and the reproduction details are authentic.
• I will wait for explicit maintainer agreement before starting implementation of a non-trivial change.