| name | maintain-verification-skill |
| description | Audit and repair a project verification skill and its feature map against current source and live behavior. Use whenever the user asks to maintain, refresh, or audit a verify skill, its harness, or its user-facing feature coverage. |
Maintain a verification skill
Keep a project-local verification skill honest as the product changes. The unit of rigor is one feature: every mapped feature receives source coverage and a live drive.
Outcomes
- clean: every feature received source and live coverage; no correction is needed.
- changed: proven corrections are ready locally, or published when the user authorized publication.
- blocked: coverage or a safe correction could not finish; name the exact blocker.
Locate and bound the target
Search established project roots, including .agents/skills/verify-*, .claude/skills/verify-*, and .cursor/skills/verify-*. Resolve symlinks and edit only the canonical directory. If several canonical candidates remain, ask which one. If none exists, route to create-verification-skill.
Only edit the verification skill's SKILL.md, features/, and owned harness helpers. Never alter product code during a maintenance pass. Product regressions are findings, not documentation fixes.
| Capability | Preferred route | Fallback |
|---|
| Source wave | One authorized read-only worker per feature, launched within harness limits | Partition or read features sequentially with the same return schema |
| Live drive | The target skill's declared browser, PTY, API, or platform harness | Mark the feature blocked or verified-unreachable with the attempted route and missing capability |
| Publication | Create one PR only when the user's request authorizes external publication | Leave a verified local diff and report it |
Pass
- Index hygiene. Compare the feature index with sibling files. Fix missing, extra, duplicate, and dead entries.
- Source wave. For every feature file, explain current user-visible behavior from source, cite entry points, flag likely drift, and return one concise live recipe. Workers never drive or edit. Only delegate when current policy permits it.
- Reconcile. Account for every feature result, spot-check cited drift, combine recipes into the fewest safe app states, and sweep recent source churn for missing user-facing surfaces. Require a concrete source path before adding one.
- Live pass. Follow the verification skill's launch model: one owned long-lived instance for a server/UI, or a fresh isolated session per short-lived CLI drive. Exercise every feature at least once.
- Triage. Wrong user-facing instructions are doc drift. Working behavior the harness cannot drive is a harness gap. Broken product behavior is a product gap to report outside this edit scope.
- Re-prove corrections. Re-drive any changed harness or recipe. Re-read every changed file.
- Ship or stop. Produce at most one authorized PR. For clean, locally changed, or blocked outcomes, do not create external state without authorization.
Live invariants
- Doctor before the first drive, each fresh session, and after surprising behavior or failed drives.
- Reset or relaunch a wedged user state even when process health looks normal.
- Preserve evidence through every cleanup and verify it still exists.
- Clean residue from failed drives; stop only instances the run owns.
- Report an unreachable feature only with the concrete prerequisite and attempted route.
- Tear down after the final drive and any re-proofs; retain evidence.
Return feature coverage, confirmed drift, changes, product gaps, proof paths, cleanup status, and one outcome label.
