[Bug]:Blocking SendMessage returns stale task state when the agent ends in an interrupted state (INPUT_REQUIRED / AUTH_REQUIRED)

未关闭
#1,302 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@VISHNU0906 已经在做这个了。

开始于 2026年11月9日。

  • #1314 来自 @VISHNU0906 —— 未关闭
  • #1315 来自 @WMK666666 —— 未关闭

评估

难度
4/5
预计耗时
3-5 天
新手友好度
45/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
java

调研方向

Start with DefaultRequestHandler.onMessageSend, EventQueue.MainQueue.enqueueItemInternal, and MainEventBusProcessor; the report also gives a self-contained reproducer, InputRequiredStaleRepro.java. Trace when queued events are persisted and when the blocking response reads TaskStore, then reproduce with a delayed save and an INPUT_REQUIRED final event. Done means the blocking response returns the persisted interrupted state even when persistence lags.

由索引模型根据 Issue 内容生成。

描述

What happened?

[Bug]: Blocking SendMessage returns stale task state when the agent ends in an interrupted state (INPUT_REQUIRED / AUTH_REQUIRED)

Version: org.a2aproject.sdk:*:1.5.0.Final (also present in main as of 2026-11-09)
Environment: macOS / OpenJDK 17.0.20.1, plain Java (no CDI), InMemoryTaskStore wrapped with a
fixed per-save delay, InMemoryQueueManager, MainEventBusProcessor, handler built via
DefaultRequestHandler.builder()
Affects: non-streaming SendMessage with returnImmediately omitted or false, when the last
event is INPUT_REQUIRED or AUTH_REQUIRED

Summary

For a blocking SendMessage, the returned Task is read from TaskStore
(DefaultRequestHandler.onMessageSend, Step 5, taskStore.get(taskId)), but the blocking path only
waits for the consumer future. If the agent's last event has not been persisted by
MainEventBusProcessor when Step 5 runs, the caller receives a stale state (SUBMITTED / WORKING)
even though the task is already INPUT_REQUIRED in the store a moment later.

Terminal states are protected by the awaitingFinalEvent machinery; interrupted states are not,
because TaskState.INPUT_REQUIRED.isFinal() == false and isInterrupted() == true.

This is the interrupted-state analogue of #1247 (fixed for terminal events by #1227).

Why it happens

  1. EventQueue.MainQueue.enqueueItemInternal only calls ChildQueue.expectFinalEvent() when
    isFinalEvent(event) is true. isFinalEvent accepts TaskState.isFinal() and A2AError —
    INPUT_REQUIRED/AUTH_REQUIRED are excluded.
  2. The blocking path (DefaultRequestHandler.onMessageSend) closes the caller's ChildQueue
    (Step 2) and waits for consumption (Step 3). With isAwaitingFinalEvent() == false, the
    consumer exits as soon as the queue is closed and empty
    (ChildQueue.dequeueEventItem).
  3. Step 5 then reads TaskStore. Events still in MainQueue — acquired on the semaphore in
    enqueueItemInternal, released only after persist + distribute — are invisible at this point.
  4. MainEventBusProcessor is asynchronous and persists before distributing, so the event is
    durable shortly afterwards; the caller already returned.

The consumer's dequeueEventItem(QUEUE_WAIT_MILLISECONDS = 500) gives an accidental grace window:
with a fast store the event lands before the poll times out, which is why this is invisible in
simple examples and only shows up when persistence lags (slow DB, many events, write backlog)
beyond ~500 ms after the agent returns.

Reproduction

Self-contained single file, attached as InputRequiredStaleRepro.java. Only
a2a-java-sdk-server-common + a2a-java-sdk-spec 1.5.0.Final are required.

# pom: the two artifacts above at 1.5.0.Final
mvn -q dependency:build-classpath -Dmdep.outputFile=cp.txt

# interrupted last event, 600ms per persist -> stale
java -DsaveDelay=600 -Dcalls=3 -cp "$(cat cp.txt)" InputRequiredStaleRepro.java blocking

# same store, terminal last event -> correct
java -DsaveDelay=600 -Dcalls=3 -cp "$(cat cp.txt)" InputRequiredStaleRepro.java terminal

Observed:

mode=blocking  expected=TASK_STATE_INPUT_REQUIRED  saveDelay=600ms  artifacts=0  calls=3
  call#1  returned=TASK_STATE_SUBMITTED eventual=TASK_STATE_INPUT_REQUIRED STALE
  call#2  returned=TASK_STATE_SUBMITTED eventual=TASK_STATE_INPUT_REQUIRED STALE
  call#3  returned=TASK_STATE_SUBMITTED eventual=TASK_STATE_INPUT_REQUIRED STALE
  => 0/3 returned the expected state

mode=terminal  expected=TASK_STATE_COMPLETED  saveDelay=600ms  artifacts=0  calls=3
  call#1  returned=TASK_STATE_COMPLETED eventual=TASK_STATE_COMPLETED OK
  call#2  returned=TASK_STATE_COMPLETED eventual=TASK_STATE_COMPLETED OK
  call#3  returned=TASK_STATE_COMPLETED eventual=TASK_STATE_COMPLETED OK
  => 3/3 returned the expected state

A less artificial shape reproduces too — 50 ms per save with 12 artifact events before
requiresInput() (≈650 ms of persistence lag for the tail):

java -DsaveDelay=50 -Dartifacts=12 -Dcalls=3 -cp "$(cat cp.txt)" InputRequiredStaleRepro.java blocking
#   => 0/3, e.g. returned=TASK_STATE_WORKING / SUBMITTED, eventual=INPUT_REQUIRED

GetTask immediately afterwards returns INPUT_REQUIRED — the state is not lost, only the blocking
response is wrong.

Suggested fix

The blocking path should guarantee the same invariant the terminal path gets from
awaitingFinalEvent: no event enqueued by the agent may still be in flight in MainQueue when
Step 5 reads TaskStore.

Option A (general): expose the existing in-flight count and wait for it to reach zero before Step 2,
e.g. QueueManager.awaitDrained(taskId, timeout) implemented on MainQueue
(size() == getQueueSize() - semaphore.availablePermits()), signalled from
releaseSemaphore() — which runs in MainEventBusProcessor's finally block, after persist and
distribute. A callback-based wake-up would miss the last event, so the signal has to be on the
release path.

Option B (narrower): extend the awaitingFinalEvent accounting to
state.isFinal() || state.isInterrupted(). This fixes the status-event case, but an artifact tail
(the last event being an artifact) would still race, so Option A is preferable.

Workaround for applications

Wrap the AgentExecutor so that execute() does not return until the task's queue is drained
(poll QueueManager.get(taskId).size() with a bounded timeout), or use
returnImmediately=true plus client-side GetTask polling.

Related

  • #1247 — blocking SendMessage drops in-flight final events (fixed in 1.5.0 by #1227)
  • #1249 — builder blocking timeouts aligned with documented defaults
  • #990 — API/hook to close streams after INPUT_REQUIRED / AUTH_REQUIRED
Relevant log output

Code of Conduct
  • I agree to follow this project's Code of Conduct
主要语言
Java
星标
514
派生
200
平均合并
1 天 2 小时
30 天内合并 PR
53

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

a2aproject/a2a-java 的其他 Issue

查看 a2aproject/a2a-java 的全部 Issue

相似的 Issue

更多 Java Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。