Provides a 10-second automatic rollback and restart of OpenClaw Gateway if configuration changes cause it to fail, activating only during backup scripts.

Before installing or running this skill: (1) Review bin/watchdog.sh line-by-line — it will run openclaw CLI commands, copy your active config, overwrite openclaw.json with the latest file in ~/.openclaw/backups, and restart the Gateway. (2) Confirm you have the 'openclaw' CLI on PATH and that backups exist at ~/.openclaw/backups; the package does not declare this dependency. (3) Expect possible service disruption when the script restarts the Gateway; test in a safe environment first. (4) If you need the advertised behavior ('only during backups' and '10-second rollback'), modify the script to enforce those constraints (e.g., accept a flag, check for a lockfile created by your backup script, implement a 10s poll-and-wait loop rather than immediate rollback). (5) Make a separate copy of your backups before first use; the script will create failed-config-*.json but still overwrites openclaw.json. (6) Consider adding logging/notification and stricter error handling, and update skill.json to declare the openclaw binary requirement and correct version metadata. These inconsistencies look like sloppy packaging rather than deliberate malice, but verify behavior in a controlled test before trusting it in production.

Type: OpenClaw Skill Name: claw-seatbelt Version: 0.1.2 The skill 'claw-seatbelt' is designed to act as a configuration safety belt for OpenClaw, providing automatic rollback and restart capabilities if the Gateway service becomes unresponsive after a configuration change. The `SKILL.md` clearly outlines this purpose and instructs the agent to run `bin/watchdog.sh`. The `watchdog.sh` script performs expected actions: checking Gateway status, backing up the current configuration, restoring the latest valid backup from `~/.openclaw/backups/`, and restarting the Gateway service using the `openclaw` CLI. All file operations and command executions are directly aligned with its stated purpose and do not show any signs of data exfiltration, malicious execution, persistence mechanisms, or prompt injection attempting to subvert the agent's intended behavior.