What problem does it solve? W&B experiment directories and local server databases accumulate empty or failed runs that contain no useful step metrics, wasting disk space and cluttering the experiment UI. This Skill identifies and removes only those runs that lack any successful step evidence, while preserving runs with real training or validation signals. ## Core Features & Use Cases - Filesystem Cleanup: Scans wandb and offline-run-* directories for success signals such as wandb-summary.json metrics, step:N log lines, and val-core//val-aux/ validation output, then dry-runs or deletes no-signal units. - Local Server Cleanup: Soft-deletes empty runs in the Dockerized local W&B MySQL database (runs, runs_flat, files tables) and optionally removes matching MinIO object directories. - Safe Deletion Workflow: Always supports dry-run reporting first, deletes only the highest nested candidate to avoid double counting, and can prune leftover empty directories. - Use Case: After a series of failed smoke tests left dozens of empty offline runs under ~/autoResearchData/autoresearch-log/wandb/offline-runs, run a dry-run to review candidates, then delete them and rebuild the local W&B UI. ## Quick Start Ask the assistant to dry-run the W&B empty run cleanup over the default AutoResearch roots and show which runs would be deleted before confirming deletion.
