Watchdog Watches the AI Jobs You Leave Running

Field notes

The service combines a registered job, configured expectations, and lifecycle events from a Python context manager. Its documented checks cover schedules, runtime limits, progress timeouts, tool fingerprints, costs, and named outcomes, but each depends on signals supplied by the application. A synthetic eight-second failure scenario records a run as still running with critical health after receiving only a run.started event, while explicitly avoiding claims about the underlying cause.