Fixed
- The
docx,pdf,pptx, andxlsxskills are now discoverable — they were nested atskills/document-skills/<name>/, one level deeper than any client scans. Agent Plugins §7.1 states that "Clients MUST NOT recursively search deeper descendants for additional skills", so all four were invisible to every Agent Plugins client and to Claude Code.skills.lock.jsonnow maps them to top-level destinations,.claude-plugin/marketplace.jsonregisters each individually, and all 26 skills load.scripts/validate_agent_plugin.pyreports zero warnings across all three plugin roots. - Correction to the v2.20.0 notes — that entry claimed the fix belonged upstream in
scientific-agent-skillsbecauseskills/is a hash-locked snapshot. That was wrong on both counts: upstream has always published these four as top-level skills and has nodocument-skillsdirectory at all, and the nesting came from this repository's owndestinationmapping inskills.lock.json. Re-vendoring the same pinned commit with flat destinations fixes it here, andsnapshot_sha256was regenerated accordingly.
Changed
scripts/check_consistency.pyno longer special-cases skill bundles — theSKILL_BUNDLE_DIRSescape hatch existed only to letdocument-skills/pass frontmatter checks while hiding its sub-skills from discovery. Every directory underskills/is now checked as one skill, which is what the specification requires.
Full Changelog: v2.20.0...v2.21.0