What happened?
When the framework itself writes a terminal FAILED state (consumer failure, producer failure), the resulting TaskStatus carries no timestamp. Because TaskManager.save_task_event replaces the status wholesale via task.status.CopyFrom(event.status) (src/a2a/server/tasks/task_manager.py:265), the previously stamped status.timestamp is cleared, and DatabaseTaskStore.save derives last_updated from status.timestamp (src/a2a/server/tasks/database_task_store.py:152-156) — persisting NULL.
Consequences:
ListTasks ordering (which sorts on last_updated) drops the task to the end;
- the
status_timestamp_after filter never matches the task again;
- API responses lose the
status.timestamp field entirely.
The behavior is also internally inconsistent: the cancel-failure path _mark_task_as_failed (src/a2a/server/agent_execution/active_task.py:945-963) copies the existing status and only overwrites state (preserving the timestamp), while the consumer failure path _write_failed_status (active_task.py:161-169) and the producer failure path construct a fresh TaskStatus(state=...) with no timestamp.
Reproduction
Self-contained script (main @ 494a8ec, Python 3.10)
import asyncio
from a2a.auth.user import UnauthenticatedUser
from a2a.helpers.proto_helpers import new_task_from_user_message
from a2a.server.agent_execution import AgentExecutor
from a2a.server.context import ServerCallContext
from a2a.server.events import EventQueue
from a2a.server.request_handlers import DefaultRequestHandlerV2
from a2a.server.tasks import InMemoryTaskStore
from a2a.types.a2a_pb2 import (
AgentCapabilities,
AgentCard,
Message,
Part,
Role,
SendMessageConfiguration,
SendMessageRequest,
TaskState,
)
class ExplodingExecutor(AgentExecutor):
task_id = None
async def execute(self, context, event_queue):
task = new_task_from_user_message(context.message)
self.task_id = task.id
await event_queue.enqueue_event(task)
raise RuntimeError('agent crashed mid-flight')
async def cancel(self, context, event_queue):
pass
async def main():
task_store = InMemoryTaskStore()
executor = ExplodingExecutor()
handler = DefaultRequestHandlerV2(
agent_executor=executor,
task_store=task_store,
agent_card=AgentCard(
name='t',
version='1.0',
capabilities=AgentCapabilities(streaming=True),
),
)
params = SendMessageRequest(
message=Message(
role=Role.ROLE_USER, message_id='m1', parts=[Part(text='Hi')]
),
configuration=SendMessageConfiguration(
accepted_output_modes=['text/plain']
),
)
try:
await handler.on_message_send(
params, ServerCallContext(user=UnauthenticatedUser())
)
except Exception as e:
print(f'on_message_send raised {type(e).__name__} (expected)')
task = await task_store.get(
executor.task_id, ServerCallContext(user=UnauthenticatedUser())
)
print(
f'persisted: state={TaskState.Name(task.status.state)}, '
f'timestamp set={task.status.HasField("timestamp")}'
)
asyncio.run(main())
Observed output:
on_message_send raised RuntimeError (expected)
persisted: state=TASK_STATE_FAILED, timestamp set=False
Expected behavior
Every task state transition the framework writes — including failures — should carry a status.timestamp, matching what TaskUpdater.update_status does for agent-driven transitions.
Suggested fix
Either:
- stamp the current time in the three framework failure paths (
_write_failed_status, the producer failure path, and — for consistency — _mark_task_as_failed), or
- in
TaskManager.save_task_event, preserve the previous status.timestamp when the incoming status event has none.
Happy to submit a PR with tests.
Related
Observed as a secondary effect in the reproduction of #1313.
Code of Conduct
What happened?
When the framework itself writes a terminal
FAILEDstate (consumer failure, producer failure), the resultingTaskStatuscarries no timestamp. BecauseTaskManager.save_task_eventreplaces the status wholesale viatask.status.CopyFrom(event.status)(src/a2a/server/tasks/task_manager.py:265), the previously stampedstatus.timestampis cleared, andDatabaseTaskStore.savederiveslast_updatedfromstatus.timestamp(src/a2a/server/tasks/database_task_store.py:152-156) — persisting NULL.Consequences:
ListTasksordering (which sorts onlast_updated) drops the task to the end;status_timestamp_afterfilter never matches the task again;status.timestampfield entirely.The behavior is also internally inconsistent: the cancel-failure path
_mark_task_as_failed(src/a2a/server/agent_execution/active_task.py:945-963) copies the existing status and only overwritesstate(preserving the timestamp), while the consumer failure path_write_failed_status(active_task.py:161-169) and the producer failure path construct a freshTaskStatus(state=...)with no timestamp.Reproduction
Self-contained script (main @
494a8ec, Python 3.10)Observed output:
Expected behavior
Every task state transition the framework writes — including failures — should carry a
status.timestamp, matching whatTaskUpdater.update_statusdoes for agent-driven transitions.Suggested fix
Either:
_write_failed_status, the producer failure path, and — for consistency —_mark_task_as_failed), orTaskManager.save_task_event, preserve the previousstatus.timestampwhen the incoming status event has none.Happy to submit a PR with tests.
Related
Observed as a secondary effect in the reproduction of #1313.
Code of Conduct