Summary-led reports, primer revision stamps, validator self-test
Every decision record now leads with four plain-language blocks (what I found, why it matters, do this first, what I did not check) before the full record; see CLAUDE.md section 14. Consumers merge that section into their root instruction file by hand; the installer writes the new file beside the existing one with a .l-gevity suffix.
Skills
- The three largest skills (
functionality-complexity-tradeoff,ci-cd-reliability-architecture,morphogenetic-architecture) move their catalogs and code patterns intoreferences/*.md, loaded on demand. Section numbers are unchanged; each moved section leaves a stub with the essentials and a link. - The
agents/openai.yamlsidecar files are removed from all skills. A lock v2 installer prunes them on the second run after a v1 lock. - A dated changelog note is removed from
structural-simplification. - The JavaScript architecture-as-code primer no longer recommends running boundary rules at
warnduring migration; it now matches the skill's error-only rule.
Library enforcement
- Every primer carries a stamp of the skill revision it was reviewed against;
npm run validatefails until a changed skill's primer is re-read and restamped (--stamp-primers). npm run validate:mutationbreaks one thing per rule family in a scratch copy and requires the validator to fail on each; it runs in CI.- Unreferenced images fail validation; every
references/*.mdmust be linked from itsSKILL.mdand every link must resolve. - The npm package no longer ships compiled Python caches.
Fixes
- Primer stamps hash skill files in posix order so Windows and Linux agree.
Release commit 4f9400f, followed by ac7fe89 and f40ec28; the tag points at f40ec28.