What problem does it solve? When an A2MC calibration ensemble runs on a workstation with A2MC_EXEC_MODE=local, there is no scheduler to query, so the HPC three-layer watcher contract does not apply and a crashed dispatcher looks identical to a long quiet run. This Skill defines how to watch such a run correctly: reading the dispatch log, checking the dispatcher process, and verifying completion through the filesystem. ## Core Features & Use Cases - Dispatcher liveness checks: Extracts the dispatcher PID from each case's job_id.txt and verifies it with ps, distinguishing a dead dispatcher from a slow case. - Monitor filter design: Arms a Claude Monitor on local_dispatch.log with both progress signals (e.g., EcoSIM start/end) and failure signatures (Killed, Cannot allocate, Traceback), since silence on a crash is identical to silence on a long run. - Completion verification: Uses the backend's check_case_status, which requires the final restart file rather than an output tape, and never treats dispatcher exit as ensemble completion. - Use Case: After submitting an EcoSIM ensemble locally on a laptop, arm a Monitor on the dispatch log, detect an OOM-killed case via the Killed signature, and propose lowering A2MC_LOCAL_WORKERS and relaunching only the affected cases. ## Quick Start Ask the AI to monitor the local A2MC ensemble by checking the dispatcher process and arming a Monitor on local_dispatch.log with progress and error filters.
