[Bug]: A2A Server (Starlette) Fails with HTTP 503/ReadTimeout when Task Artifact Contains Multiple TextParts #112
Comments
|
Hi, from a2a.server.agent_execution import AgentExecutor, RequestContext
from a2a.server.events import EventQueue
from a2a.utils import new_task
from collections.abc import AsyncGenerator
import asyncio
from a2a.server.tasks import TaskUpdater
from a2a.types import Part, TextPart
class HelloWorldAgent:
"""Hello World Agent."""
async def stream(self) -> AsyncGenerator[str, None]:
yield 'Hello '
await asyncio.sleep(2)
yield 'World '
await asyncio.sleep(2)
yield 'again!'
class HelloWorldAgentExecutor(AgentExecutor):
"""Test AgentProxy Implementation."""
def __init__(self):
self.agent = HelloWorldAgent()
async def execute(
self,
context: RequestContext,
event_queue: EventQueue,
) -> None:
task = context.current_task
if not task and context.message:
task = new_task(context.message)
event_queue.enqueue_event(task)
updater = TaskUpdater(event_queue, task.id, task.contextId)
parts: list[Part] = []
async for content in self.agent.stream():
parts.append(Part(TextPart(text=content)))
updater.add_artifact(
parts,
name='result',
)
updater.complete()
async def cancel(
self, context: RequestContext, event_queue: EventQueue
) -> None:
raise Exception('cancel not supported')Here is the response from the server. Just wanted to confirm if this snippet correctly represents the scenario you described? |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
What happened?
When an A2A agent's
AgentExecutorproduces a final result that includes anArtifactcontaining multiplea2a.types.TextPartobjects, the A2A server (usingA2AStarletteApplicationandDefaultRequestHandler) fails to send a successful response to the client. Clients experience anhttpx.ReadTimeoutor an HTTP 503 error.Server-side logs from the
AgentExecutorshow that it correctly processes the agent's output and callsTaskUpdater.add_artifact()with alist[Part](containing multipleTextParts) andTaskUpdater.complete(). No exceptions are observed within theAgentExecutoritself.The issue appears to occur later in the A2A framework, likely during the serialization of the
a2a.types.Taskobject (which includes the multi-part artifact) into a JSON-RPC response, or during the HTTP sending phase by Starlette.The problem is reproducible when an ADK agent using
BuiltInCodeExecutorgenerates code, execution results, and a final text answer, all of which are converted into distinctTextParts for the final artifact.If the
AgentExecutoris modified to ensure the final artifact contains only a singleTextPart, the client receives a successful response. This points to the handling of multipleTextParts within an artifact as the root cause.Steps to Reproduce:
code_interpreter_agentexample) that uses an underlying mechanism (e.g., Google ADK withBuiltInCodeExecutor) which produces multiple distinct pieces of textual information for its final output (e.g., generated code, code output, final summary).AgentExecutor'sconvert_genai_parts_to_a2a(or equivalent) function converts these distinct pieces of information into multiplea2a.types.TextPartobjects.AgentExecutor._process_requestmethod, when the final ADK event is received, create alist[Part]containing these multipleTextParts and pass it toTaskUpdater.add_artifact(). Then callTaskUpdater.complete().AgentExecutorlogs completion of its processing.Expected Behavior: The A2A server should successfully serialize and send a
Taskobject containing an artifact with multipleTextParts.Actual Behavior: Client receives HTTP 503 / ReadTimeout. Server logs from the agent executor are clean, suggesting a failure later in the A2A framework's response pipeline.
Relevant log output
Code of Conduct
The text was updated successfully, but these errors were encountered: