功能描述

4-tier autonomous self-healing and auto-recovery system for OpenClaw Gateway. Monitors gateway health, auto-restarts on crash, detects OAuth token expiry, ki...

使用说明 (SKILL.md)

OpenClaw Self-Healing System

Name: OpenClaw Self-Healing System
Author: ramsbaby

"The system that heals itself — or calls for help when it can't."

A 4-tier autonomous recovery system for OpenClaw Gateway, featuring AI-powered diagnosis via Claude Code. Tested in production on macOS + Linux.

Architecture

Level 1: config-watch        → Config file change detection + instant reload
Level 2: Watchdog v4.4       → OAuth detection, zombie kill, exponential backoff
Level 3: Claude Code Doctor  → AI-powered diagnosis & repair (30 min window) 🧠
Level 4: Discord/Telegram    → Human escalation with full context

What's New in v3.1.0

Complete healing chain fix — config-watch → Watchdog → Emergency Recovery now fully connected
Installer rewrite — single install.sh covers macOS (LaunchAgent) + Linux (systemd)
Watchdog v4.4 — OAuth token expiry detection, zombie process auto-kill, Exponential Backoff
Emergency Recovery v2 — persistent learning repo, reasoning logs, multi-model support (Claude Code + Aider)
Metrics dashboard — success rate, MTTR, trending analysis via tmux

Quick Setup

bash \x3C(curl -fsSL https://raw.githubusercontent.com/Ramsbaby/openclaw-self-healing/main/install.sh)

Or install via ClawHub:

npx clawhub@latest install openclaw-self-healing

The 4 Tiers in Detail

Level	Script	Trigger	Action
L1	`config-watch.sh`	Config file change	Validate + reload gateway
L2	`gateway-watchdog.sh`	Process down / HTTP fail	Kill zombie → restart → backoff
L3	`emergency-recovery-v2.sh`	30min continuous failure	Claude Code PTY diagnosis
L4	`emergency-recovery-monitor.sh`	L3 triggered	Discord + Telegram alert

Configuration

All settings via environment variables in ~/.openclaw/.env:

Variable	Default	Description
`DISCORD_WEBHOOK_URL`	(none)	Discord webhook for L4 alerts
`OPENCLAW_GATEWAY_URL`	`http://localhost:18789/`	Gateway health check URL
`HEALTH_CHECK_MAX_RETRIES`	`3`	Restart attempts before L3 escalation
`EMERGENCY_RECOVERY_TIMEOUT`	`1800`	Claude recovery timeout (30 min)

Verified Recovery Cases

OAuth token expiry — Watchdog v4.4 detects 401 in logs, restarts before agent dies
Zombie process — Preflight detects PID mismatch, SIGKILL + launchctl kickstart
Config schema error — openclaw doctor --fix auto-applied on exit_1 pattern
Level 3 triggered — Claude Code diagnosed and fixed broken config in \x3C 15 min

Links

GitHub: https://github.com/Ramsbaby/openclaw-self-healing
Changelog: https://github.com/Ramsbaby/openclaw-self-healing/blob/main/CHANGELOG.md
Linux setup: https://github.com/Ramsbaby/openclaw-self-healing/blob/main/docs/LINUX_SETUP.md

License

MIT — built by @ramsbaby + Jarvis 🦞

安全使用建议

Install only if you intentionally want a persistent self-healing system with authority to restart OpenClaw and let Claude Code attempt repairs. Inspect the included install.sh and scripts first, prefer a pinned release over curl/bash from main, back up ~/.openclaw configuration, keep webhook channels private, and consider disabling or manually gating the Level 3 AI repair path until you trust the behavior.

功能分析

Type: OpenClaw Skill Name: openclaw-self-healing Version: 3.1.1 The bundle implements a multi-tiered 'self-healing' system that grants an AI agent (Claude Code or Aider) broad shell access to autonomously diagnose and repair the OpenClaw Gateway. It establishes persistence via LaunchAgents (macOS) and systemd (Linux) and exfiltrates system logs and recovery reports to external Discord and Telegram webhooks. While these high-risk capabilities (AI-driven RCE, persistence, and network communication) are aligned with the stated purpose of autonomous recovery, they create a significant attack surface for prompt injection and unauthorized system modification. Key files include emergency-recovery-v2.sh (AI orchestration), gateway-watchdog.sh (process management), and install.sh (persistence setup).

能力评估

⚠ Purpose & Capability

The self-healing purpose is coherent with monitoring and restarts, but the artifacts describe autonomous AI diagnosis and repair of the local OpenClaw environment, including config fixes and process control, without clearly bounded paths or approval steps.

⚠ Instruction Scope

The visible instructions describe automatic escalation to Claude Code for a 30-minute repair window and auto-applied fixes; the reviewed artifacts do not show clear human confirmation, dry-run mode, or modification limits.

ℹ Install Mechanism

Installation uses normal package managers for tmux, Claude Code, and jq, but the quick setup recommends executing an unpinned remote install script from GitHub main via curl/bash.

⚠ Credentials

Killing zombie processes, restarting the gateway, and changing OpenClaw configuration are proportionate to self-healing, but they are high-impact local mutations when performed autonomously by a background system.

⚠ Persistence & Privilege

The skill is designed to run persistently through LaunchAgent/systemd/timers/cron and can continue taking recovery actions after installation; this is disclosed but materially increases the impact of any bad diagnosis or misconfiguration.

版本历史

v3.1.1

v3.1.1: Improved search discoverability. Same features as v3.1.0.

v3.1.0

v3.1.0: Complete healing chain fix. Watchdog v4.4 with OAuth detection + zombie kill + exponential backoff. Emergency Recovery v2 with persistent learning. Installer rewrite for macOS+Linux.

v2.0.2

🐛 Hotfix: Discord webhook HTTP 404 버그 수정 - DISCORD_WEBHOOK_URL 의존성 제거 - stdout으로 알림 출력 → 크론 delivery 자동 전달 - 설정 복잡도 제거 - 실제 장애 시나리오로 검증 완료

v2.0.1

Major update introducing persistent learning, reasoning logs, and expanded alerting. - Added persistent learning with automatic documentation of failures and fixes. - Introduced reasoning logs to explain AI-driven recovery actions. - Expanded alerts to support both Discord and Telegram. - Added metrics dashboard script for visualizing recovery statistics. - New documentation and contributor guides. - Requires new dependency: jq (for metrics dashboard).

v1.3.4

GitHub Actions CI fix

v1.3.3

ShellCheck 100% clean

v1.3.2

Critical bug fixes: file order, lock security, completion detection

v1.3.1

- Added README.md for better documentation and onboarding. - Added demo-script.sh and install.sh example scripts. - Updated CHANGELOG.md and SKILL.md for the new version. - Improved emergency-recovery-monitor.sh and emergency-recovery.sh scripts.

v1.2.1

Add cleanup trap, log permissions, Linux systemd guide

v1.2.0

openclaw-self-healing 1.2.0 - Introduces a 4-tier autonomous self-healing architecture for OpenClaw Gateway. - Adds Claude Code as Level 3 emergency doctor for AI-powered diagnosis and repair. - Integrates Watchdog, Health Check (HTTP 200 with retries), Claude Recovery, and Discord Alert escalation. - Provides macOS LaunchAgent setup and step-by-step installation instructions. - All configuration is now handled via environment variables for easier setup and management.

元数据

Slug openclaw-self-healing

版本 3.1.1

许可证 —

累计安装 16

当前安装数 15

历史版本数 10

常见问题