27 lines
1.1 KiB
Markdown
27 lines
1.1 KiB
Markdown
# Hacker News "Show HN" 게시문 초안
|
||
|
||
## Title (80자 이내)
|
||
Show HN: Self-Healing System for AI Agents – Claude Code as Emergency Doctor
|
||
|
||
## URL
|
||
https://github.com/Ramsbaby/openclaw-self-healing
|
||
|
||
## Text (선택적 - URL만 있어도 됨)
|
||
I built a 4-tier autonomous recovery system for OpenClaw Gateway (AI agent framework).
|
||
|
||
The interesting part: Level 3 uses Claude Code as an "emergency doctor" - it launches in a tmux PTY, reads logs, diagnoses issues, and autonomously fixes them.
|
||
|
||
Architecture:
|
||
- Level 1: Process watchdog (180s)
|
||
- Level 2: HTTP health check (300s)
|
||
- Level 3: Claude Code diagnosis + repair (30min timeout)
|
||
- Level 4: Discord alert (human escalation)
|
||
|
||
Key insight: If we trust AI to write code, why not trust it to fix infrastructure? The AI already knows how to diagnose - we just gave it permission to act.
|
||
|
||
All actions are logged and auditable. 30-minute timeout ensures humans can intervene.
|
||
|
||
MIT licensed, bash scripts only, works on macOS (Linux guide included).
|
||
|
||
Would love feedback on the security model - how do you prevent the "doctor" from weakening its own checks?
|