sue-cleanup

What problem does it solve?

This Skill solves the critical issue of inode exhaustion on shared parallel filesystems (such as LUMI Lustre and Snellius GPFS) caused by tens of thousands of small output files generated during ML/HPC scale-up experiments, which can block new runs and disrupt team workflows.

Core Features & Use Cases

• Safe Bulk Archiving: Archives only high-inode, already-summarized bulk output directories (images, benchmarks, legacy outputs) after verifying all metrics are extracted into reports and ledgers, never touching critical artifacts like checkpoints, logs, or readiness stamps.
• Sandbox-Scoped Operations: Operates exclusively on a single selected sandbox backend's workspace output tree, avoiding unintended project-wide or multi-workspace deletions.
• Use Case: After completing a full training run on LUMI that generated 18,000 per-variant image files, use this Skill to archive the image directory into a single tar file, freeing up nearly 18,000 inodes while preserving all benchmark reports and progress CSVs needed for result summarization.

Quick Start

Invoke the sue-cleanup skill after a dryrun or fullrun completes to archive high-inode bulk output directories from your sandbox-scoped experiment workspace, ensuring all benchmark reports, ledgers, and checkpoints are preserved for downstream processing.