Hacktoberfest 2026: những issue maintainer đã đánh dấu cho tháng Mười, đang mở và phù hợp người mới. Xem issue Hacktoberfest

[Bug]: DefaultRequestHandler runs follow-up messages on a task with the first request's contextvars

Đã đóng
  • Có gì không ổn?
#1,316 0 bình luận 1 reaction 2 người được giao Xem trên GitHub

Maintainer thường phản hồi trong vòng 2 ngày

@rohityan đang làm issue này rồi.

Từ ngày 5/10/2026.

  • #1317 của @tejaskash — đang mở

Đánh giá

Issue này chưa được đánh giá.

Mô tả

What happened?

With DefaultRequestHandler (v2), a follow-up message on an existing task runs AgentExecutor.execute() with the contextvars of the request that created the task, not the request that sent the follow-up. A common case is a reply after the task went input-required.

Any per-request contextvar set by ASGI middleware or by the ServerCallContextBuilder comes through stale: auth tokens, tenant or caller identity, request-scoped tracing. ServerCallContext.state is correct per message. Only contextvars are wrong.

Reproduced on 1.1.2 and 1.2.2, for both SendMessage and SendStreamingMessage. 0.3.x is not affected because its handler started a new producer per message.

Cause

ActiveTask.start() creates the producer with asyncio.create_task(self._run_producer()) during the first request, so the producer copies that request's context once. Later requests call ActiveTask.enqueue_request(), which only puts the RequestContext on _request_queue. The producer then calls execute() from the first request's context. A task parked in input-required keeps its ActiveTask (#1296), so a reply hours later still sees the first request's values, which may by then be an expired token.

Reproduction
import asyncio, contextvars, uuid

import httpx
from a2a.helpers import new_task_from_user_message
from a2a.server.agent_execution import AgentExecutor
from a2a.server.request_handlers import DefaultRequestHandler
from a2a.server.routes import create_jsonrpc_routes
from a2a.server.tasks import InMemoryTaskStore, TaskUpdater
from a2a.types import AgentCapabilities, AgentCard
from starlette.applications import Starlette

REQUEST_TAG = contextvars.ContextVar("request_tag", default="unset")


class TagMiddleware:
    def __init__(self, app):
        self.app = app

    async def __call__(self, scope, receive, send):
        token = REQUEST_TAG.set(dict(scope.get("headers", [])).get(b"x-tag", b"unset").decode())
        try:
            await self.app(scope, receive, send)
        finally:
            REQUEST_TAG.reset(token)


class Executor(AgentExecutor):
    async def execute(self, context, event_queue):
        print(f"execute({context.get_user_input()!r}) sees request_tag={REQUEST_TAG.get()!r}")
        task = context.current_task or new_task_from_user_message(context.message)
        updater = TaskUpdater(event_queue, task.id, task.context_id)
        if context.current_task:
            await updater.complete()
        else:
            await event_queue.enqueue_event(task)
            await updater.requires_input()

    async def cancel(self, context, event_queue):
        pass


card = AgentCard(name="a", description="a", version="1", capabilities=AgentCapabilities())
handler = DefaultRequestHandler(agent_executor=Executor(), task_store=InMemoryTaskStore(), agent_card=card)
app = Starlette(routes=create_jsonrpc_routes(request_handler=handler, rpc_url="/"))
app.add_middleware(TagMiddleware)


async def send(client, tag, text, task=None):
    message = {"messageId": str(uuid.uuid4()), "role": "ROLE_USER", "parts": [{"text": text}]}
    if task:
        message |= {"taskId": task["id"], "contextId": task["contextId"]}
    body = {"jsonrpc": "2.0", "id": 1, "method": "SendMessage", "params": {"message": message}}
    resp = await client.post("/", json=body, headers={"x-tag": tag, "A2A-Version": "1.0"})
    return resp.json()["result"]["task"]


async def main():
    async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://t") as client:
        task = await send(client, "request-1", "book a flight")
        await send(client, "request-2", "Friday", task)


asyncio.run(main())

Actual:

execute('book a flight') sees request_tag='request-1'
execute('Friday') sees request_tag='request-1'

Expected:

execute('book a flight') sees request_tag='request-1'
execute('Friday') sees request_tag='request-2'
Proposed fix

Capture the sender's context in enqueue_request() and run execute() in it:

async def enqueue_request(self, request_context):
    request_id = uuid.uuid4()
    await self._request_queue.put((request_context, request_id, contextvars.copy_context()))
    return request_id

# in _run_producer()
request_context, request_id, sender_context = await self._request_queue.get()
...
await sender_context.run(
    asyncio.create_task,
    self._agent_executor.execute(request_context, self._event_queue_agent),
)

I applied this patch to 1.2.2 locally and the reproduction prints request-2 for the follow-up, for both SendMessage and SendStreamingMessage. Context.run(asyncio.create_task, ...) works on Python 3.10. Awaiting the child task keeps cancellation working, because cancelling the producer cancels the child. The MCP Python SDK fixed the same problem this way in modelcontextprotocol/python-sdk#2298.

Downstream workaround

bedrock-agentcore works around this today without touching a2a internals. It wraps SimpleRequestContextBuilder to store contextvars.copy_context() in call_context.state, and wraps the executor to run execute() in that copy (aws/bedrock-agentcore-sdk-python#690). A fix here would let downstream projects delete that code, and it would also cover users who build DefaultRequestHandler directly.

Relevant log output

No response

Code of Conduct
  • I agree to follow this project's Code of Conduct
Ngôn ngữ chính
Python
Star
2.2k
Fork
509
Merge trung bình
3 ngày 18 giờ
Pull request đã merge (30 ngày)
44

Chuẩn bị môi trường

Bắt đầu từ đâu

  1. Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
  2. Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
  3. Fork repository và làm thay đổi trên một nhánh.
  4. Mở pull request có tham chiếu số hiệu của issue.

Issue khác của a2aproject/a2a-python

Tất cả issue của a2aproject/a2a-python

Issue tương tự

Thêm issue về Python

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.