You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Fixed
The prompts no longer degrade in silence where Composer never reached. code-review and feature-development cited vendor/scrapkit/engineering-kit/docs/… as their checklist, but Claude Code
materialises only plugins/engineering-kit/ when the plugin is installed:
in a repository that does not require the package those paths do not exist,
so the prompts ran against invented criteria without saying so. Only quality-audit noticed, and reported n/a. Each skill now carries the
guidelines it checks against in its own references/ directory, so the
checklist travels with the prompt on every route.
Added
A SessionStart hook in the plugin, delivering the org-wide rules to the
repositories Composer cannot reach — previously they arrived only through the @vendor/scrapkit/engineering-kit/claude/CLAUDE.md import. It is deliberately
narrow: it runs only when the repository's origin remote is under Scrapkit/, and stays silent when that vendor file exists, because the
Composer import already states the rules there. A fallback, never a second
parallel channel.
composer sync-claude-assets (scripts/sync-claude-assets.php) —
regenerates the copies that have to live inside the plugin subtree. docs/ and claude/ stay the canonical, human-facing sources; the script
writes the skills' references/ and resources/boost/skills/, prunes
orphans, and the test suite fails on a copy that has drifted. Run it after
editing a canonical file and commit what it writes.
Changed
The four skills cite references/<document>.md instead of a vendor/ path,
and quality-audit lost its "standards unreachable → n/a" branch: the
documents ship with the skill, so the degraded path no longer exists. As a
side effect the prompt and its standards are pinned by the same git ref, so
an audit can no longer be scored against guidelines from a different release.
tests.yml now also runs on changes under docs/, claude/, plugins/ and resources/. A docs-only pull request did not run Pest and could ship a
stale generated copy.
composer analyse passes --memory-limit=512M, and PHPStan analyses scripts/ alongside src/.