Agent Self Healing

Agent Self-Healing

Patterns

Health Check Loop

#!/bin/bash
# Check if a service is healthy, restart if not
check_and_heal() {
  local name=$1 port=$2
  if ! curl -sf "http://localhost:$port/health" > /dev/null; then
    echo "[$name] UNHEALTHY — restarting..."
    pm2 restart "$name"
    sleep 5
    curl -sf "http://localhost:$port/health" && echo "[$name] ✅ Recovered" || echo "[$name] ❌ Still down"
  fi
}

check_and_heal frontend 3002
check_and_heal bridge 3001
check_and_heal cowork-mcp 3003

PM2 Restart With Backoff

// ecosystem.config.js
{
  restart_delay: 3000,
  exp_backoff_restart_delay: 100,
  max_restarts: 10  // stop trying after 10 crashes
}

Port Conflict Resolution

kill_port() {
  local port=$1
  local pid=$(lsof -ti:$port 2>/dev/null)
  [ -n "$pid" ] && kill "$pid" && echo "Killed PID $pid on port $port"
}

Self-Healing Agent Checklist

• Service reachable (HTTP health check)
• PM2 process in online state (not errored)
• No port conflicts (EADDRINUSE)
• Log file not growing unboundedly
• API keys still valid (test a lightweight call)

Julia's Health Checker

• Runs every 15 min via PM2 cron
• Checks ports: 3000, 3001, 3002, 3003
• Checks OpenClaw: openclaw health
• Auto-restarts via pm2 restart <name>
• Escalates to Telegram if a service fails 3x in a row