Show HN: Self-Healing AI Agents with Claude Code as Doctor

3 points

4 months ago

I built a 4-tier self-healing system for OpenClaw (AI agent platform running on my Mac Mini 24/7). The interesting part is Level 3: when health checks fail repeatedly, the system spawns Claude Code in a tmux PTY session to autonomously diagnose and repair issues.

Recovery escalation: - Level 0-1: LaunchAgent KeepAlive + Watchdog - Level 2: Automated "doctor --fix" (config validation, port checks) - Level 3: Claude Code spawns in tmux, reads logs, attempts repairs - Level 4: Discord alert if all automation fails

Production-tested in my homelab over 3 months: 99% recovery rate, recovery time reduced from 45min → 3min avg. Handled 17 consecutive crashes, config corruption, port conflicts.

Built for macOS (stable) with Linux systemd support (beta). MIT licensed.

Curious what others think about AI-powered infrastructure self-healing.