Chulinuwu/skillloom

Curate Skills

Improve a managed skill from observed evidence while preserving its provenance and the user's ability to reject or reverse the change.

Workflow

1. Identify the installed skill, its active path, the behavior that failed or drifted, and the evidence supporting a change.
2. Inspect the full package and its promotion history. Separate confirmed defects from stylistic preferences.
3. Copy the skill to a temporary directory and make the smallest coherent revision. Keep SKILL.md harness-neutral, remove stale resources, and do not add credentials, transcript bodies, symlinks, binaries, or machine-specific tool instructions.
4. Run skillloom capture <revised-skill-directory> --created-by agent --base <installed-skill-path> --evidence <reference>. The recorded base hash protects concurrent human edits.
5. Run skillloom validate <candidate-id>. Compare the candidate with the installed base and review every trust finding. The scan is preflight evidence, not a malware-proof sandbox.
6. Present the behavioral change, diff summary, provenance, base and candidate hashes, findings, intended targets, and reload impact. Stop on danger findings and ask before accepting warnings or promoting.
7. Only after explicit approval, run skillloom promote <candidate-id> --target <targets> --scope <scope> --yes. Add --accept-warnings only for warnings the user explicitly accepts.
8. Verify destination hashes and the promotion record. If the base changed after capture, do not force an overwrite; recapture from the new base.
9. If verification fails, inspect skillloom status. Resume only when staged evidence still agrees. Otherwise run skillloom rollback <promotion-id> --yes and report the rollback record and restored hashes.

Never edit an active destination directly and never promote autonomously.