System Watchdog

Monitor system resources and flag wasteful or suspicious processes. Works standalone as a bash script — see openclaw.md for scheduled cron setup.

Standalone Usage

Run the check script directly — no OpenClaw required:

bash check.sh

Outputs a JSON object to stdout. Parse it however you like — pipe to jq, feed to an agent, integrate into your own monitoring stack.

Output Format

{
  "suspicious": true,
  "summary": {
    "ram": "12.3/31.2 GB (39%)",
    "swap": "0.5/8.0 GB (6%)",
    "load": "1.2/0.8/0.6",
    "cores": 8,
    "disk": "120/256 GB (45%)"
  },
  "issues": [
    {
      "type": "high_ram",
      "description": "claude (PID 1234) 4650MB RAM",
      "details": { "pid": 1234, "name": "claude", "cpu_pct": 2.1, "mem_mb": 4650, "elapsed": "3d" }
    }
  ],
  "top_processes": [
    { "pid": 1234, "name": "claude", "cpu_pct": 2.1, "mem_mb": 4650, "elapsed": "3d" }
  ]
}

• suspicious: true → at least one issue exceeded a threshold
• suspicious: false → system looks healthy

Thresholds

Common Offenders

• claude / codex — AI coding agents left running for days
• whisper / whisper-server — speech-to-text servers consuming GPU/RAM
• python / python3 — runaway scripts or leaked processes
• node — dev servers or builds that never stopped

Agent Workflow (for AI agents)

1. Run check.sh
2. Parse the JSON output
3. If suspicious is false → do nothing (no report needed)
4. If suspicious is true → format a concise report and notify the user

Suggested Report Format

⚠️ System Watchdog Report

📊 System: RAM 12.3/31.2 GB (39%) | Swap 0.5/8.0 GB (6%) | Load 1.2/0.8/0.6
💾 Disk: / 45% (120/256 GB)

🔴 Issues Found:

HIGH RAM — claude (PID 1234)
  CPU: 2.1% | RAM: 4650 MB | Running: 3 days
  → Likely stale, safe to kill

💡 Suggested: kill 1234

Platform Notes

• macOS only — check.sh uses sysctl, vm_stat, and macOS ps flags. It won't work on Linux without adaptation (replace with free, /proc/meminfo, etc.).
• Intermittent jq-style parse error — the script occasionally fails on first run due to a race in process scanning. Retry once before reporting failure.