A Codex Skill that replaces repeated "check it again" requests with proportionate research, cross-file checks, and self-review. It preserves existing approvals and asks about consequential unresolved decisions.
SKILL.md is the canonical workflow. Ordinary changes do not require an approval record, candidate freeze, lock, or independent evaluator.
Ask Codex to install this repository's skill-change-loop Skill, or use the available skill-installer.
Example requests:
- "Use $skill-change-loop to review this Skill. Report findings without editing."
- "The proposed change is approved. Use $skill-change-loop to implement it, check affected callers and docs, and verify the result."
Automatic invocation remains enabled for substantive Skill work. Mechanical edits and unrelated work do not need this workflow.
- convergence-protocol.md: focused behavioral evaluation for complex changes or releases.
- self-update.md: installed-file replacement and recovery.
- legacy-self-update.md: compatibility for transactions already bound to the former protocol.
scripts/candidate_guard.py and its tests are retained unchanged for legacy consumers. They are not the default path for a new ordinary change. Task evidence, local backups, caches, and repository metadata are not release payloads.
Run the available skill-creator structure validator. For legacy-helper maintenance, run:
python -B scripts/test_candidate_guard.pyThese tests check the helper, not the quality of agent decisions. Use isolated realistic tasks to evaluate workflow changes.
MIT