skill-connectivity

Point the audit at a skills/<repoKey>/ directory and it extracts every reference from Markdown, JSON, YAML, Python, TypeScript, Go, and Shell files, then resolves each against the filesystem. Each edge is classified as hit, dangling, or external, and results are delivered as a structured review report.

Beyond path resolution, it validates code imports against package manifests, cross-checks meta.ts targets and SYNC.json hashes with on-disk skills, searches README.md for catalog coverage, verifies evals ground-truth chains, and classifies backtick-quoted skill identifiers.

No, the audit is strictly read-only. It only produces a review opinion report listing dangling references, reference graphs, problems, and optional improvement suggestions, leaving all files untouched.

Relative imports that resolve within the collection become internal edges. Modules listed in requirements.txt, package.json, go.mod, or the language standard library are marked external, while imports that look local but fail to resolve are flagged as dangling.

Zero cross-skill edges are recorded as an observation of complete isolation, not an automatic defect. The audit then checks whether isolated skills share a domain and may suggest adding see-also lines, while self-contained skills remain acceptable.