- Review unresolved criteria when a loop or goal is configured, stops, or is marked complete against user expectations.
- Separate judgment mismatches from known errors, usage limits and requested pauses. Preserve restart permissions and completion evidence.
- Clarify that discovery requires an active agent turn; the skill cannot wake a stopped process.
- Retain loop/goal discovery evidence, including a rejected description that loaded the method for unnecessary cases.
- Ask for an actual criterion before creating an agreement from a vague request to learn preferences. Keep unrequested drafts in the conversation.
- Continue saving explicit lasting preferences without another confirmation.
- Add reproducible multilingual discovery checks with retained requests, execution evidence and offline replay.
- Check existing context before asking about a brief request.
- Group questions about consequential gaps and proceed within delegated scope.
- Use repeated context-related misunderstandings to review when the user wants questions.
- Keep task answers separate from confirmed lasting preferences.
- Rename the project, agreement file, repository, and site to ALIGNS.md.
- Keep
$alignas the skill invocation. - Preserve existing criteria when migrating a legacy
ALIGN.md; never replace it with a new agreement. - Retain original benchmark transcripts and source snapshots under their recorded names for verification.
- Add optional recent/full PR and project conversation review with disclosed time estimates.
- Collect paginated PR feedback with partial coverage reports and resumption.
- Parse selected Codex and Claude JSONL text while retaining project boundaries.
- Use scoped, revisable taste invariants to propose criteria from evidence, with counterexamples and confirmation before inferred criteria become agreements.
- Make align a maintained practice with focused review after consequential drift, recurring missing criteria, uncertain direction, and before long autonomous work.
- Add Codex and Claude Code plugin marketplaces, explicit update policies, and a static Claude SessionStart reminder.
- Connect the primary CLI's instructions while preserving distinct client files. Provide a helper for safe relative links and an idempotent trigger block.
- Replace the standalone README interview with an installation request.
- Record invocation and update mode in new agreements; preserve confirmation dates during method updates and narrow reviews.
- Initial project alignment skill and public distribution.