335 lines
14 KiB
Markdown
335 lines
14 KiB
Markdown
# v2.7.2 Counter 生命周期修复
|
||
|
||
**版本**: v1.1
|
||
**日期**: 2026-05-26
|
||
**作者**: 庞统
|
||
**状态**: 评审修订中(v1.1 采纳部分评审意见)
|
||
|
||
---
|
||
|
||
## 1. 问题背景
|
||
|
||
### 1.1 核心原则
|
||
|
||
> 每次 agent 调用都是独占的。openclaw 无论成功失败都会返回,最差情况是 timeout 返回。谁占用谁持有,进程退出就 release。
|
||
|
||
### 1.2 当前偏差
|
||
|
||
| 偏差 | 当前行为 | 核心原则要求 |
|
||
|------|---------|-------------|
|
||
| counter 生命周期 | 任务级(贯穿 retry 链) | 调用级(spawn acquire,退出 release) |
|
||
| retry 绕过 dispatcher | `_do_retry` 直接调 `spawn_full_agent`,不检查 counter | spawn 只有 `spawn_full_agent` 一个入口,内部统一检查 |
|
||
| A5/A6 fallback 也 retry | fallback 出现说明 agent 被占用时 spawn 了,根因错误 | 不应出现 fallback;出现了说明流程有 bug |
|
||
| A10/A12 也 retry | compact 失败/未知错误都走 retry | 进程退出 = agent 空闲,走正常 dispatch |
|
||
| 429 不推回 pending | release counter 但任务 still working → 30 分钟超时后才标 failed | 应推回 pending,让 ticker 重新调度 |
|
||
| 429 无冷却 | ticker 下次(30秒后)又 dispatch → 又 429 → 循环 | per-agent 冷却期 |
|
||
| 进程意外退出 counter 泄漏 | `_monitor_process` 异常/PM2 重启时 on_complete 不调用 | ticker 层兜底:检测进程存活性 |
|
||
|
||
### 1.3 司马懿事件复盘(2026-05-25 20:58-21:00)
|
||
|
||
```
|
||
20:58:12 dispatch mail → counter acquire → spawn 司马懿 (pid=50512)
|
||
20:59:50 进程退出 (gateway_timeout, exit=0) → release_counter=False → _do_retry
|
||
→ 直接 spawn (pid=50580),不检查 counter → 司马懿被 spawn 第二次
|
||
21:00:08 进程退出 → _do_retry → 直接 spawn (pid=50613) → 第三次
|
||
21:00:13 API 429 → exit=1 → api_error → release_counter=True → 结束
|
||
|
||
问题:counter 占用但 retry 不检查,3 次 spawn 在 ~2 分钟内完成
|
||
每次 retry 没有延迟,叠加 API 调用触发 zhipu 429
|
||
```
|
||
|
||
---
|
||
|
||
## 2. 设计方案
|
||
|
||
### 2.1 核心改动:counter 下沉到 spawn_full_agent
|
||
|
||
**原则:spawn 只有 `spawn_full_agent` 一个入口,acquire/release 统一在内部。**
|
||
|
||
```python
|
||
async def spawn_full_agent(self, agent_id, ...):
|
||
# 1. 检查 counter
|
||
if self.counter and not await self.counter.can_acquire(agent_id):
|
||
raise AgentBusyError(agent_id)
|
||
|
||
# 2. acquire
|
||
if self.counter:
|
||
await self.counter.acquire(agent_id)
|
||
|
||
# 3. 构建 on_complete(release counter)
|
||
original_on_complete = on_complete
|
||
async def _wrapped_on_complete(aid, outcome):
|
||
if self.counter:
|
||
self.counter.release(aid)
|
||
if original_on_complete:
|
||
await original_on_complete(aid, outcome)
|
||
|
||
# 4. spawn 进程
|
||
proc = await asyncio.create_subprocess_exec(...)
|
||
|
||
# 5. schedule monitor(传 wrapped_on_complete)
|
||
asyncio.create_task(self._monitor_process(..., on_complete=_wrapped_on_complete))
|
||
```
|
||
|
||
**变化**:
|
||
- dispatcher 不再 acquire/release counter,只负责路由和构建业务回调
|
||
- spawn_full_agent 内部 acquire + 注册 wrapped on_complete(保证 release)
|
||
- on_complete 不再包含 counter.release,只做业务逻辑(幻觉门控、状态标记等)
|
||
|
||
### 2.2 _classify_outcome 简化
|
||
|
||
去掉 `release_counter` 字段。**进程退出 = release counter**(由 wrapped_on_complete 保证)。
|
||
|
||
只保留两个维度:
|
||
- `should_retry`:是否触发续杯(只有 A2/A3 gateway_timeout)
|
||
- `retry_field`:计入哪个计数器
|
||
|
||
| 情况 | outcome | should_retry | 说明 |
|
||
|------|---------|-------------|------|
|
||
| A1 正常完成 | completed | False | 任务终态 |
|
||
| A4 Agent failed | agent_failed | False | 尊重 Agent 判断 |
|
||
| A2/A3 Gateway timeout | gateway_timeout | True | **唯一续杯场景** |
|
||
| A5/A6 Fallback | fallback_timeout | False | **不应出现**,记录 warning,不 retry |
|
||
| A7 认证失败 | auth_failed | False | 不 retry |
|
||
| A8 Gateway 不可达 | gateway_unreachable | False | 不改任务状态,等 ticker |
|
||
| A9 API 错误/429 | api_error | False | **推回 pending**(见 2.4) |
|
||
| A10 Compact 失败 | compact_failed | False | **不 retry**,推回 pending 等 ticker |
|
||
| A11 Lock 冲突 | lock_conflict | False | 不改任务状态,等 ticker |
|
||
| A12 其他 | agent_error | False | **不 retry**,推回 pending 等 ticker |
|
||
|
||
**A5/A6 不应出现**:如果出现了,说明 spawn 时 agent 被占用(counter 检查失效)。记录 ERROR 级日志,标 failed + escalate,不 retry。
|
||
|
||
**A10/A12 不 retry**:进程退出了 = agent 空闲。release counter → 推回 pending → ticker 重新调度。
|
||
|
||
### 2.3 _do_retry 重构
|
||
|
||
```python
|
||
async def _do_retry(self, session_id, agent_id, task_id, on_complete, db_path, retry_field="retry_count"):
|
||
# 1. 检查任务终态
|
||
if self._is_terminal(db_path, task_id):
|
||
return # wrapped_on_complete 已 release counter
|
||
|
||
# 2. 检查 retry 上限
|
||
count = self._increment_retry(db_path, task_id, retry_field)
|
||
if count >= self.max_retries:
|
||
self._mark_task(db_path, task_id, "failed", {...})
|
||
return # wrapped_on_complete 已 release counter
|
||
|
||
# 3. 构建续杯 message
|
||
message = self._build_retry_message(...)
|
||
|
||
# 4. 通过 spawn_full_agent 重新 spawn(内部会 can_acquire)
|
||
# 此时 counter 已由 wrapped_on_complete release
|
||
try:
|
||
await self.spawn_full_agent(
|
||
agent_id=agent_id,
|
||
message=message,
|
||
task_id=task_id,
|
||
on_complete=on_complete, # 业务回调(不含 counter)
|
||
use_main_session=(session_id is None),
|
||
reuse_session_id=session_id if session_id else None,
|
||
task_db_path=db_path,
|
||
)
|
||
except AgentBusyError:
|
||
# agent 被其他任务占用,release counter,等 ticker
|
||
logger.warning("Retry spawn skipped: %s busy", agent_id)
|
||
# counter 已被 wrapped_on_complete release
|
||
# 任务保持 working,等 ticker 检查
|
||
```
|
||
|
||
**关键变化**:
|
||
- counter 在进程退出时由 wrapped_on_complete release
|
||
- `_do_retry` 调 `spawn_full_agent`,内部 can_acquire 检查
|
||
- 如果 agent 忙(不应该发生,但防御)→ 等 ticker
|
||
|
||
### 2.4 429 处理:推回 pending + 冷却机制
|
||
|
||
#### 推回 pending
|
||
|
||
```python
|
||
# _handle_exit 中 A9 处理
|
||
if outcome == "api_error":
|
||
# wrapped_on_complete 已 release counter
|
||
# 推回 pending,让 ticker 重新调度
|
||
self._transition_task_status(db_path, task_id, "pending", {
|
||
"reason": "api_error_retry",
|
||
"api_retry_count": count,
|
||
})
|
||
```
|
||
|
||
#### 冷却机制
|
||
|
||
```python
|
||
class ActiveAgentCounter:
|
||
def __init__(self):
|
||
self._active: dict[str, int] = {} # agent_id → count
|
||
self._cooldown_until: dict[str, float] = {} # agent_id → timestamp
|
||
|
||
def set_cooldown(self, agent_id: str, seconds: float = 120.0):
|
||
"""设置冷却期"""
|
||
self._cooldown_until[agent_id] = time.time() + seconds
|
||
|
||
def is_cooling_down(self, agent_id: str) -> bool:
|
||
"""检查是否在冷却期"""
|
||
until = self._cooldown_until.get(agent_id)
|
||
if until and time.time() < until:
|
||
return True
|
||
# 冷却期已过,清理
|
||
self._cooldown_until.pop(agent_id, None)
|
||
return False
|
||
|
||
async def can_acquire(self, agent_id: str) -> bool:
|
||
if self.is_cooling_down(agent_id):
|
||
return False
|
||
return self._active.get(agent_id, 0) == 0
|
||
```
|
||
|
||
**冷却时间**:120 秒(zhipu 瞬时限流一般 30-60 秒恢复,留余量)。
|
||
|
||
**冷却触发点**:
|
||
- A9(api_error/429):`counter.set_cooldown(agent_id, 120)`
|
||
- 冷却只阻止新 dispatch,不影响 retry 的 can_acquire(retry 时冷却已过或不存在)
|
||
|
||
### 2.5 进程意外退出兜底
|
||
|
||
在 ticker 的 `_check_timeouts` 中增加**进程存活性检查**:
|
||
|
||
```python
|
||
def _check_timeouts(self, db_path: Path) -> List[str]:
|
||
# ... 现有超时检查 ...
|
||
|
||
# 新增:检查 counter 占用但进程已死的情况
|
||
if self.counter and self.spawner:
|
||
for agent_id in self.counter.active_agents:
|
||
session_info = self.spawner.get_session_by_agent(agent_id)
|
||
if not session_info:
|
||
continue
|
||
pid = session_info.get("pid")
|
||
if pid and not self._is_pid_alive(pid):
|
||
# 进程已死但 counter 还占着
|
||
crash_count = self._get_crash_count(db_path, task_id) + 1
|
||
self._set_crash_count(db_path, task_id, crash_count)
|
||
|
||
if crash_count >= 3:
|
||
# 3 次连续崩溃,标 failed
|
||
logger.error("Agent %s crashed %d times, marking failed", agent_id, crash_count)
|
||
self._mark_task(db_path, task_id, "failed", {
|
||
"reason": "process_crash",
|
||
"crash_count": crash_count,
|
||
})
|
||
self.counter.release(agent_id)
|
||
else:
|
||
# 释放 counter,让 ticker 下次重新 dispatch
|
||
logger.warning("Agent %s process dead (crash %d/3), releasing counter",
|
||
agent_id, crash_count)
|
||
self.counter.release(agent_id)
|
||
# 任务保持 working,下次 tick 会看到 counter 空闲 → 重新 dispatch
|
||
```
|
||
|
||
**crash_count 存储**:`task_attempts.metadata` 的 `crash_count` 字段。
|
||
|
||
**重置时机**:agent 成功完成一次任务后重置为 0。
|
||
|
||
### 2.6 dispatcher 简化
|
||
|
||
dispatcher 不再管 counter:
|
||
|
||
```python
|
||
async def dispatch(self, task, ...):
|
||
# 不再 acquire counter(spawn_full_agent 内部处理)
|
||
# 不再构建含 counter.release 的 on_complete
|
||
|
||
# 只做路由和业务回调
|
||
on_complete = None
|
||
if is_mail:
|
||
on_complete = lambda aid, outcome: self._mail_auto_complete(task_id, aid, db_path, must_haves)
|
||
# Task 不需要业务回调(counter release 由 spawn_full_agent 的 wrapped_on_complete 处理)
|
||
|
||
try:
|
||
session_id = await self.spawner.spawn_full_agent(
|
||
agent_id=agent_id,
|
||
message=message,
|
||
task_id=task.id,
|
||
on_complete=on_complete, # 只含业务逻辑
|
||
use_main_session=is_mail,
|
||
task_db_path=db_path,
|
||
)
|
||
except AgentBusyError:
|
||
return {"status": "skipped", "reason": "Agent busy"}
|
||
```
|
||
|
||
---
|
||
|
||
## 3. 改动范围
|
||
|
||
| 文件 | 改动 | 行数 |
|
||
|------|------|------|
|
||
| `spawner.py` | spawn_full_agent 加 counter acquire/release + wrapped_on_complete | ~30 |
|
||
| `spawner.py` | _classify_outcome 去掉 release_counter 字段 | ~10 |
|
||
| `spawner.py` | _handle_exit 简化(A9 推回 pending,A10/A12 不 retry) | ~20 |
|
||
| `spawner.py` | _do_retry 通过 spawn_full_agent 重试 | ~15 |
|
||
| `spawner.py` | 新增 AgentBusyError | ~5 |
|
||
| `dispatcher.py` | 去掉 counter acquire/release/on_complete 逻辑 | ~-40 |
|
||
| `counter.py` | 新增 cooldown 机制 | ~25 |
|
||
| `ticker.py` | _check_timeouts 加进程存活性检查 | ~30 |
|
||
| **合计** | | ~95 行净增 |
|
||
|
||
## 4. 续杯语义变化
|
||
|
||
| | 当前 | 修复后 |
|
||
|---|------|--------|
|
||
| counter 生命周期 | 任务级(贯穿 retry 链) | 调用级(spawn acquire,退出 release) |
|
||
| retry 路径 | 直接 spawn,不检查 counter | 通过 spawn_full_agent,内部检查 |
|
||
| retry 失败 | 只看 retry_count 上限 | can_acquire 失败也会停止(agent 忙) |
|
||
| A9/429 | release + 任务 working → 30 分钟超时 | release + 推回 pending + 冷却 120s |
|
||
| A10 compact | retry(counter 不 release) | 不 retry,推回 pending |
|
||
| 意外退出 | counter 泄漏 | ticker 检测 → crash_count → 最多 3 次 |
|
||
|
||
## 5. 续杯时间计算
|
||
|
||
Gateway timeout = 600s(10 分钟),3 次续杯 = ~30 分钟。
|
||
|
||
每次续杯的流程:
|
||
```
|
||
进程退出 → release counter → _do_retry → spawn_full_agent(can_acquire)
|
||
→ acquire → spawn → 新进程执行(最多 10 分钟)
|
||
```
|
||
|
||
30 分钟内 agent 无法完成任务 → 第 3 次续杯后标 failed + escalate。
|
||
|
||
## 6. 测试计划
|
||
|
||
| 用例 | 验证 |
|
||
|------|------|
|
||
| 正常完成 | counter acquire → release,任务 done |
|
||
| Gateway timeout 续杯 | release → re-acquire → spawn,复用 session |
|
||
| 续杯时 agent 被占用 | can_acquire 失败 → 不 spawn,等 ticker |
|
||
| 429 → 冷却 | 推回 pending,120s 内不 dispatch |
|
||
| 进程崩溃 | ticker 检测,crash < 3 → 重新 dispatch |
|
||
| 3 次崩溃 | 标 failed + escalate |
|
||
| PM2 重启 | ticker 检测进程死 → release counter |
|
||
| Mail 幻觉门控 | on_complete 业务回调正常执行 |
|
||
|
||
---
|
||
|
||
## 7. 司马懿评审意见处理(mail-1779726169654)
|
||
|
||
| 评审意见 | 结论 | 理由 |
|
||
|---------|------|------|
|
||
| wrapped_on_complete 加 try/finally | ✅ 采纳 | 防御性编程,确保 counter release 和业务回调都执行 |
|
||
| A5/A6 加 context 日志 | ✅ 采纳 | 排查方便 |
|
||
| per-provider 冷却 | ⏭ 延后 | 低优先级,先做 per-agent |
|
||
| crash_count per-agent 累计,禁用 agent | ❌ 不采纳 | 崩溃可能是任务问题不是 agent 问题。保持 per-task 3 次标 failed,通过 escalate 通知用户自行判断 |
|
||
| can_acquire 失败推回 claimed | ❌ 不采纳 | retry 路径下 can_acquire 不会失败(asyncio 单线程无竞态)。release → can_acquire → acquire 是内存同步操作,中间无 await |
|
||
| release 和 acquire 之间有竞态窗口 | ❌ 不存在 | Python asyncio 单线程,三步都是内存同步操作,无竞态 |
|
||
|
||
## 8. 实现检查清单
|
||
|
||
- [ ] counter.py:新增 cooldown 机制
|
||
- [ ] spawner.py:spawn_full_agent 加 counter acquire/release + wrapped_on_complete(try/finally)
|
||
- [ ] spawner.py:_classify_outcome 去掉 release_counter,只有 A2/A3 触发 retry
|
||
- [ ] spawner.py:_do_retry 通过 spawn_full_agent 重试
|
||
- [ ] spawner.py:_handle_exit 简化(A9 推回 pending,A5/A6 标 failed + context 日志)
|
||
- [ ] dispatcher.py:去掉 counter acquire/release 逻辑
|
||
- [ ] ticker.py:_check_timeouts 加进程存活性检查(per-task crash_count)
|