fix(aibox): batch H audit fixes - subagent cancel propagation, scheduler timeout, exec process-group reaping
Fourth audit round (see docs/reports/aibox_audit_20260817.md §11): - H1: shared parent cancel event no longer cleared on subagent first think (N5 regression); parent-already-cancelled skips subagent start - M1: /reload_prompts moved off event loop (run_in_executor, N3 sibling) - M2: skill --help scan prewarmed before serving connections - M3: scheduler task-level timeout (scheduler.task_timeout, default 600s) - M4: get_recent_memories(None) capped instead of full-table load - M5: memo delete now physically removes rows (was complete-only) - M6: PromptManager.reload builds new list then atomically swaps - L1-L12: read_file arg coercion, proxy save failure reported, SSE utf-8 tolerance, persona realpath check, truncation char boundary, compressor div0, skill path quoting, ask_user required, POSIX process-group kill (pgid captured at Popen), web scheme whitelist, registry redact substring match, Bearer redaction in failover logs
This commit is contained in:
qing committed
1 parent
d1f780e381
commit
d3acccd146
14 files changed
+643
-79
No files matched your search
+55
-12
@@ -328,14 +328,35 @@ class TaskScheduler:
|
||||
ai_thread.daemon = True
|
||||
ai_thread.start()
|
||||
|
||||
# 等待 AI 完成，同时处理可能的取消
|
||||
# M3（审计 2026-08-18）：任务级超时兜底。此前 await ai_future
|
||||
# 无任何超时/取消路径——AI think 卡死（LLM 挂起、子进程不可杀）
|
||||
# 时该任务永不返回，调度器 run 循环顺序 await 被阻塞，后续全部
|
||||
# 定时/即时任务不再执行且任务已 mark_as_triggered 不会重试。
|
||||
# 到点 call_later 触发 ai.cancel()（置取消事件 + 关在途 HTTP 流），
|
||||
# worker 线程自然收尾，调度循环继续。
|
||||
task_timeout = float(ConfigManager().get("scheduler.task_timeout", 600))
|
||||
timed_out = threading.Event()
|
||||
|
||||
def _timeout_cancel():
|
||||
timed_out.set()
|
||||
try:
|
||||
ai.cancel()
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
timeout_handle = loop.call_later(task_timeout, _timeout_cancel)
|
||||
try:
|
||||
response = await ai_future
|
||||
result = response or "AI无响应"
|
||||
except Exception as e:
|
||||
self.logger.error(f"AI对话失败: {e}")
|
||||
result = f"任务执行但AI对话失败: {str(e)}"
|
||||
if timed_out.is_set():
|
||||
self.logger.error(f"任务执行超时（{task_timeout}秒），AI 已取消")
|
||||
result = f"任务执行超时（{task_timeout}秒），AI 已取消"
|
||||
else:
|
||||
self.logger.error(f"AI对话失败: {e}")
|
||||
result = f"任务执行但AI对话失败: {str(e)}"
|
||||
finally:
|
||||
timeout_handle.cancel()
|
||||
# 确保线程结束
|
||||
ai_thread.join(timeout=1.0)
|
||||
|
||||
@@ -534,10 +555,11 @@ class ExpressEmotionTool(Tool):
|
||||
if not name:
|
||||
return ""
|
||||
path = os.path.join(constants.SKILLS_DIR, name + ".md")
|
||||
# N12：persona 为 LLM 可控输入，做路径边界校验——normcase 统一大小写与
|
||||
# 分隔符后要求解析结果仍在 SKILLS_DIR 内，防止 ../../ 越界读取并注入
|
||||
# N12 + L4（审计 2026-08-18）：persona 为 LLM 可控输入。abspath 前缀
|
||||
# 检查拦截 ../ 越界，但符号链接可绕过（skills 目录内软链指向任意文件
|
||||
# 即被读取注入）；realpath 解析软链后再做前缀校验。
|
||||
skills_root = os.path.normcase(os.path.abspath(constants.SKILLS_DIR))
|
||||
resolved = os.path.normcase(os.path.abspath(path))
|
||||
resolved = os.path.normcase(os.path.realpath(path))
|
||||
if not resolved.startswith(skills_root + os.sep):
|
||||
self.logger.warning(f"人格技能名非法（越界路径）: {name}")
|
||||
return ""
|
||||
@@ -673,6 +695,14 @@ def run_full_subagent(role, task, config, logger=None, model=None,
|
||||
# 深度中断（sub_chat.cancel() 强关在途 HTTP 流并 kill 其子进程）。
|
||||
if cancel_event is not None:
|
||||
sub_chat._cancel_event = cancel_event
|
||||
# H1（审计 2026-08-18）：共享父事件后禁止子代理首轮 clear——
|
||||
# _think_and_respond_impl 首轮无条件 clear() 会吞掉父级已置位的
|
||||
# 取消信号（用户中断 / --once 超时恰落在子代理生成窗口时被清掉，
|
||||
# 观察者线程随之失效，子代理跑完全程）。置位后立即检查一次，
|
||||
# 父级已取消则不再启动子代理。
|
||||
sub_chat._shared_cancel_event = True
|
||||
if cancel_event.is_set():
|
||||
return "❌ 子代理已因父级取消而终止"
|
||||
|
||||
def _watch_parent_cancel():
|
||||
while not _watch_done.wait(timeout=0.2):
|
||||
@@ -1456,8 +1486,9 @@ class DeepSeekChat:
|
||||
|
||||
def _think_and_respond_impl(self, input_text: str, output_callback: Callable = None, is_continuation: bool = False) -> Optional[str]:
|
||||
"""think_and_respond 的实际实现，在 _msg_lock 保护下执行。"""
|
||||
# 首轮开始时清除取消标志
|
||||
if not is_continuation:
|
||||
# 首轮开始时清除取消标志（H1：共享父事件的子代理不清除——清除会
|
||||
# 吞掉父级已置位的取消信号，使中断/--once 超时失效）
|
||||
if not is_continuation and not getattr(self, "_shared_cancel_event", False):
|
||||
self._cancel_event.clear()
|
||||
|
||||
# 立即检查是否已被取消
|
||||
@@ -2492,10 +2523,17 @@ class WebSocketHandler:
|
||||
elif command == "/reload_prompts":
|
||||
session = self._get_session(websocket)
|
||||
ai = session.get("ai") if session else None
|
||||
if ai:
|
||||
new_count = ai.reload_prompts()
|
||||
else:
|
||||
new_count = self.ai_manager.prompt_manager.reload()
|
||||
# M1（审计 2026-08-18）：reload_prompts 内部同步获取 _msg_lock
|
||||
# （繁忙对话持锁可达数十秒），直接调用会冻结整个事件循环——
|
||||
# N3 只修了 /new，这里同类遗漏。移入线程池执行。
|
||||
loop = asyncio.get_running_loop()
|
||||
|
||||
def _do_reload_prompts():
|
||||
if ai:
|
||||
return ai.reload_prompts()
|
||||
return self.ai_manager.prompt_manager.reload()
|
||||
|
||||
new_count = await loop.run_in_executor(None, _do_reload_prompts)
|
||||
await self.send_message(websocket, {
|
||||
"type": "command_result",
|
||||
"content": i18n.get('prompts_reloaded', new_count)
|
||||
@@ -2883,12 +2921,17 @@ class EnhancedAIManager:
|
||||
port = constants.WEBSOCKET_PORT
|
||||
|
||||
self.ws_handler = WebSocketHandler(self)
|
||||
# M2（审计 2026-08-18）：技能扫描对每个可执行技能串行跑 --help
|
||||
# 子进程（各 ≤10s），懒加载于首条消息/首个任务时在事件循环线程
|
||||
# 执行会冻结整个服务。启动前预热，把扫描成本移到连接建立之前。
|
||||
self.ws_handler._get_shared_skill_manager()
|
||||
|
||||
# 如果启用了任务调度器，创建并启动
|
||||
if constants.SCHEDULER_ENABLED:
|
||||
self.scheduler = TaskScheduler(asyncio.get_running_loop(), self.logger)
|
||||
self.scheduler.set_websocket_handler(self.ws_handler)
|
||||
self.scheduler.set_ai_manager(self) # 设置AI管理器引用
|
||||
self.scheduler._get_shared_skill_manager()
|
||||
|
||||
async def handler(websocket):
|
||||
await self.handle_connection(websocket, None)
|
||||
|
||||
Reference in new issue
Block a user