| name | repo-rule-authoring |
| description | Propose, implement, and calibrate repo-standards lint rules when deterministic evidence, false-positive review, or pending-rule governance matters. |
Repo rule authoring
Use this skill for repository-policy rules, not ordinary application lint fixes.
Modes
- Propose: read only. Return a RuleProblem, upstream/catalog overlap, labeled cases,
severity, limitations, and the authorization needed for implementation. Do not write files.
- Implement: read rule-anatomy.md, write labeled cases first,
then add implementation, catalog metadata, governance, versions, and tests atomically.
- Calibrate: read calibration.md, verify immutable corpora, run
the helper, inspect and classify findings, compare performance, and refine until clean.
- Review or promote: require a separate immutable review commit. Never approve a rule in its
authoring commit or activate consumers without explicit rollout authorization.
New judgment-heavy rules start as pending warnings. Prefer syntax-aware evidence, emit one
diagnostic per semantic defect, and preserve false negatives when avoiding them would require
guessing. Never infer AI authorship or intent.
The calibration helper is developer-only and is not a shipped CLI:
uv run python .agents/skills/repo-rule-authoring/scripts/calibrate_rules.py verify \
--corpus "$PRIVATE_CORPUS"
uv run python .agents/skills/repo-rule-authoring/scripts/calibrate_rules.py evaluate \
--corpus "$PRIVATE_CORPUS" --rule RULE_ID --fixtures-passed \
--private-output /private/tmp/rule-private.json \
--public-output /private/tmp/rule-public.json
Do not clone, fetch, install, publish, commit, or mutate consumers unless the user separately
authorizes that action.