据 AlphaMoat 统计,截至 2026-09,Error Guard — Control‑Plane Safety 累计下载 4,444 次、获 8 个星标,在全部 5,078 个 AI 技能中排名第 2,187 位,在「运维安全」分类(264 个)中排名第 122 位。

累计安装

375

累计下载

4,444

星标

全站排名

#2,187

全站前 44% · 「运维安全」前 47%

平台	累计安装	累计下载	星标	评论
ClawHub ↗	141	4,248	8	2
SkillHub ↗	375	4,444	8	—

各平台的计数口径不同,统计时间不一致。

这个技能是做什么的

设计涵盖状态查询、取消全部活动任务与进程会话并清空待发队列的紧急 flush，以及重置控制状态并可选重载的恢复流程。命令目标是不调用模型、不等待 worker，且只保留任务元数据。看门狗和子代理运行辅助仍列为未来扩展，不能据此认定已实现。

资料发布于 2026-10-27

可以用来做什么

为停滞代理任务规划恢复控制

什么时候用

长任务运行需要明确状态与取消语义时

示例提示词

以下是根据技能文档生成的示例,不代表实际运行结果。

解释状态、flush 和恢复的行为，说明 flush 会取消什么，以及哪些扩展尚未实现。

月度数据趋势

累计下载4,444

星标8

2026-09 起统计方法调整:同名技能按作者分别统计,并覆盖全部技能。图中虚线为调整前的月份,与之后的数据不直接比较,也不计算跨越调整的增长。

月份	累计下载	环比	星标
2026-09	4,444	—	8
2026-08	4,124	+3.6%	8
2026-07	3,980	—	8
2026-05	3,566	—	8
2026-03	3,104	—	8

SKILL.md 原文

以下是作者发布的原始文件,按纯文本显示,本站未做修改或审核;其中的链接不可点击,命令请核实后再运行。抓取于 2026-10-23。

2.7 KB · ClawHub v1.0.0

---
name: error-guard
description: >
  System safety and control-plane skill that prevents agent deadlocks and freezes.
  Provides non-LLM control commands to inspect task state, flush message queues,
  cancel long-running work, and recover safely without restarting the container.
  Use when implementing or operating long-running tasks, sub-agents, benchmarks,
  background monitors (e.g., Moltbook, PNR checks), or when the system becomes
  unresponsive and needs immediate recovery controls.
---

展开全部(共 90 行)

# error-guard

⚠️ **System‑level skill (Advanced users)**

This skill defines the **control‑plane safety primitives** for OpenClaw.
It is intentionally minimal, non‑blocking, and designed to prevent agent freezes, deadlocks, and unrecoverable states when running long‑lived or high‑risk workloads.

## Design Principles

> **Warning:** This skill operates at the agent control‑plane level.
> It should be installed only by users who understand OpenClaw’s execution model and are running workloads that can block, hang, or run for extended periods.

- **Main agent never blocks**: no long exec, no external I/O, no LLM calls.
- **Event-driven**: workers emit events; the control plane listens.
- **Fail-safe first**: recovery commands must always respond.
- **Minimal state**: track only task metadata (never payloads).

## Command Surface (Phase 1)

### /status

Report current system health and task registry state.

Returns:
- Active tasks (taskId, type, state)
- Start time and last heartbeat
- Flags for stalled or overdue tasks

Constraints:
- Must run in constant time
- Must not call any model or external API

### /flush

Emergency stop.

Immediately:
- Cancel all active tasks
- Kill active exec/process sessions
- Clear pending message queue
- Reset in-memory task registry

Constraints:
- Must always respond
- No waiting on workers
- No model calls

### /recover

Safe recovery sequence.

Steps:
1. Execute `/flush`
2. Reset control-plane state
3. Optionally reload skills/state (no container restart)

## Future Extensions (Not Implemented Yet)

- Sub-agent runner helper (event-driven)
- Task watchdogs with TTL and silence detection
- Structured event protocol (task.started, task.heartbeat, task.completed, ...)
- Back-pressure and task classes (interactive / batch / background)

## Security & Privacy

- This skill **does not** store payloads, prompts, messages, or model outputs
- Only minimal task metadata is persisted (taskId, timestamps, state)
- No API keys, credentials, or user data are read or written
- Safe to publish and share publicly

## Non-Goals

- No business logic
- No background polling loops
- No user-facing features
- No LLM reasoning paths

This skill is the **last line of defense**. Keep it small, fast, and reliable.