Improve align from observed mismatches and user feedback. Keep the shared skill small, and put guidance used only during setup or recurring review in the linked references. Personal agreements belong to users, not to a release migration.
- Update the skill and the affected reference. Record the behavior change in CHANGELOG.md. Check realistic scenarios when a change affects judgment.
- Keep VERSION, both plugin manifests, the Claude marketplace entry, and skill metadata on the same version. Claude uses the version to notice updates.
- Run
PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests -p 'test_*.py'. Validate the skill with Skill Creator when available. Test affected install paths in an isolated project; never use personal agreements as fixtures. - Publish the tested source, tag that exact commit, and attach a ZIP of the
complete
skills/alignfolder plus its SHA-256 checksum to the GitHub release. Native plugin installations use the repository manifests. - Refresh directories that snapshot releases, and record actual listing evidence in docs/distribution.md. Do not describe a submission as acceptance.
Follow the client's supported update route in docs/usage.md. Native auto-update settings, or an agreed refresh when alignment starts, bring users the revised method. An update must not alter project ALIGN.md files, confirmation times, client permissions, or unrelated packages. Keep a usable installed version if refresh fails. Report reload requirements instead of claiming immediate activation.
The source repository itself uses the local skill through .agents/skills/align.
Keep AGENTS.md canonical for Codex here; do not auto-update this development
checkout over uncommitted changes or install a duplicate plugin.
Superpowers demonstrates native plugin distribution and a session bootstrap. ALIGN uses that separation between discovery and method, with a static Claude reminder and project instructions for recurring triggers. It does not start an interview on every session or update through a networked hook.