What problem does it solve? Tracking many concurrently running ML experiments across local screen sessions and remote SSH servers is tedious and error-prone; this Skill unifies status checks into a single report and can automatically collect finished experiments and advance the research pipeline. ## Core Features & Use Cases - Unified Status Monitoring: Scans all running experiments, checks local screen sessions or remote servers via SSH, and classifies each as running, anomaly, completed-pending-collect, or collected. - Anomaly Detection: Greps recent logs for NaN, OOM, Traceback, and Inf patterns and surfaces the latest metrics (loss, accuracy, step) in a status table. - Auto-Collect & Pipeline Advance: With --collect-ready it triggers /exp-run --collect for finished experiments; with --pipeline and --auto-advance it advances the research pipeline to Stage 4 once all experiments complete. - Use Case: While running a hyperparameter sweep on a remote GPU server, periodically invoke this Skill (or schedule it via cron) to see which runs are alive, which crashed, and automatically move to the verdict stage when everything finishes. ## Quick Start Ask the assistant to check the status of all running experiments and collect any that have finished.
