claw-guard

A lightweight service that monitors registered events:

1. Sub-agent task PIDs — if PID dies → notify and remove. If log/dir stale → alert and remove.
2. Gateway restarts — if restart fails → revert config backups (newest to oldest) → retry → notify.

ClawGuard only monitors what is explicitly registered. It does not auto-discover.

Install

cd <skill-dir>
bash scripts/install.sh

Installs:

• Daemon: systemd user service (Linux) or launchd agent (macOS) — Restart=always, auto-starts on boot
• CLI: claw-guard in ~/.local/bin/
• Data: ~/.openclaw/workspace/tools/claw-guard/

OpenClaw Integration (Recommended)

1. Auto-register gateway restarts

Add ExecStartPre to your gateway service so every restart (manual, crash, or Restart=always) is automatically registered:

# ~/.config/systemd/user/openclaw-gateway.service
[Service]
ExecStartPre=/home/<user>/.local/bin/claw-guard register-restart
ExecStart=...

Then reload: systemctl --user daemon-reload

Now every gateway restart automatically:

• Snapshots the current config (rotates up to 5 backups)
• Watches for the gateway to come back
• If it fails → reverts config backups newest-to-oldest → notifies default channel

No manual claw-guard register-restart needed — systemd handles it.

2. Add task execution rules to AGENTS.md

Add these rules so the agent always registers its work:

## Task Execution Rules (MANDATORY)

### Sub-agent requirement
- **Any exec/tool call that might take >5s → sub-agent** (`sessions_spawn`).
  Main agent stays responsive.
- **Complex or unpredictable tasks → always sub-agent.** Even if they might
  be fast. If you can't guarantee it won't block, delegate it.
- **Only run in main agent** if certain it won't block I/O (quick file reads,
  short `grep`, `git status`, `claw-guard status`, etc.)

### ClawGuard registration (MANDATORY for all sub-agents)
Every sub-agent and background process **must** be registered:
```bash
claw-guard register --id "<descriptive-id>" --pid <PID> \
  --target "<channel where task was requested>" \
  --log "/path/to/logfile" --timeout 180 \
  --command "<what it does>"

• --target: same channel/room where the user asked for the task
• --log and --timeout: optional but recommended for long tasks
• If PID dies → claw-guard notifies the target channel and removes the entry
• If log goes stale → claw-guard notifies and removes

Gateway restarts

• Never restart the gateway while tasks are running — it kills all sub-agents
• Gateway service has ExecStartPre=claw-guard register-restart — automatic
• No manual registration needed for restarts

### 3. How it works end-to-end

**Sub-agent task flow:**
1. User requests a long-running task
2. Agent spawns sub-agent → gets PID
3. Agent runs: `claw-guard register --id "task-name" --pid $PID --target "room:..." --command "..."`
4. If PID dies → claw-guard does two things:
   - Sends a **Matrix message** to the target channel (user sees the alert)
   - Fires an **`openclaw system event --mode now`** (wakes the agent to investigate)
5. Agent wakes up, sees the `[claw-guard]` event, investigates (check logs, exit code, dmesg for OOM), and reports findings to the target channel
6. If log goes stale → same dual notification → agent investigates

**Gateway restart flow:**
1. Gateway restarts (manual, crash, or auto)
2. `ExecStartPre` runs `claw-guard register-restart` → config backed up
3. Gateway starts successfully → claw-guard logs `✅ Gateway restart succeeded` → watch cleared
4. Gateway fails to start → claw-guard tries config backups → notifies default channel + fires system event

## CLI Reference

### Register a task

```bash
claw-guard register --id "benchmark-q8" --pid 12345 \
  --target "room:!abc:server" \
  --log "/path/to/task.log" --timeout 180 \
  --command "python3 benchmark.py"

# Or watch a directory for new file creation:
claw-guard register --id "export-gguf" --pid 12345 \
  --target "room:!abc:server" \
  --watch-dir "/path/to/output/" --timeout 300 \
  --command "export_gguf.py"

| Flag | Required | Description | |------|----------|-------------| | --id | yes | Unique task identifier | | --pid | yes | Process ID to watch | | --target | yes | Notification target (see formats below) | | --log | no | Log file path — checks mtime only | | --watch-dir | no | Directory — checks newest file mtime | | --timeout | no | Stale threshold in seconds (default: 180) | | --command | no | Description included in notifications |

Register a gateway restart

claw-guard register-restart [--target "room:!abc:server"]

No --target needed — sends to OpenClaw's default channel. Pass --target to override.

Manage

claw-guard status          # Show tasks, restart watch, config backups
claw-guard remove --id X   # Remove a task
claw-guard clear-done      # Remove completed/gone tasks

Behavior

Check cycle (every 15s)

1. Gateway restart: if registered and gateway not active after 30s → revert + retry + notify
2. PID check: if PID gone → notify target → remove entry
3. Log/dir freshness: if mtime exceeds timeout → notify target → remove entry

Deduplication

After notifying, the registered entry is removed from the registry. Once removed, it can't fire again. No dedup tracking needed.

Restart / reboot behavior

On service restart or system reboot:

• All registered tasks are cleared — nothing carries over
• Config backups persist on disk (only thing that survives)

This is by design: after a reboot, all monitored processes are gone anyway. The agent must re-register any new tasks.

Notification Targets

Any format openclaw message send --target accepts:

• room:!roomId:server (Matrix)
• telegram:chatid
• discord:#channel
• slack:#channel

Gateway restart alerts with no --target are sent without a target flag, letting OpenClaw route to the default channel.

Design Principles

• Registration-based, not auto-discovery — only watches what's explicitly registered
• Notify once, then remove — no duplicate alerts, no stale state
• In-memory state — registry clears on service restart (clean slate)
• Disk persistence only for config backups — the only thing worth keeping across restarts
• Cross-platform — Linux (systemd) and macOS (launchd)
• Minimal overhead — ~7MB RAM, negligible CPU