What problem does it solve? Long training runs can silently waste GPU hours when loss diverges, NaN values appear, or metrics plateau, and these problems are often only discovered after training finishes. This Skill performs periodic automated health checks on running training jobs so failures are caught early. ## Core Features & Use Cases - WandB Metric Polling: Reads loss trends, eval metrics, gradient norms, and learning rate schedules from the WandB API, with SSH log-file fallback when WandB is unreachable. - Tiered Judgment: Acts directly on clearly good or clearly bad signals, and escalates only ambiguous cases to a Codex model for a STOP/CONTINUE/WAIT verdict. - Adaptive Check Intervals: Runs via CronCreate starting at 10 minutes, increasing to 20, 30, then 60 minutes when consistently healthy, and resetting after anomalies. - Use Case: After launching a multi-day fine-tuning run, schedule this Skill to poll the WandB run every 10 minutes; it kills the job and logs evidence if loss diverges, saving the remaining GPU budget. ## Quick Start Set up a recurring check for my WandB run entity/project/run_id to monitor training health every 10 minutes.
