Watchdog — Health Check and Mechanical Recovery

The watchdog exists because agent sessions can crash, time out, or simply disappear. Without cleanup, work items would stay in implementing forever with expired leases, blocking progress. The watchdog detects these situations and takes safe, mechanical recovery actions.

When This Skill Runs

• Routine health check (recommended: at start of each orchestrator session)
• After a session crash or unexpected termination
• When the user suspects something is stuck
• Periodically during long-running orchestration loops

Prerequisites

• Orchestration must be initialized

Allowed Tools

• Read (state files, HDR files), Bash (git root, state-commit), Glob