Maintain Verification Skill

Maintain a verification workflow

Treat the request text that activated this skill as the verification surface to audit.

This method repairs verifier drift. When evidence exposes a product defect, report it and route authorized product repairs to pkstack-verified-goal with the unchanged expectation. A verifier-maintenance request alone does not authorize product edits.

Do not assume a passing old check still covers current behavior. Preserve this sequence:

1. Locate the surface. Find the authoritative verification skill, its feature-map index, every feature record, executable commands, implementation entrypoints, and existing evidence. Use .pkstack/bin/projectctl feature list --output json and feature show <slug> --output json when PKStack is set up. Feature records use schema 2; unsupported records need a fresh reviewed contract grounded in the current implementation and public surface.
2. Check index hygiene. Reconcile the index, feature files, and user-visible surface. Record missing, stale, duplicate, or orphaned entries before editing. Also find native .kiro/specs/*/pkstack-verification.json bridges. Treat their requirements.md or bugfix.md, design.md, and tasks.md as Kiro-owned planning artifacts; reconcile each bridge to its exact published feature or reviewed command without rewriting the native plan.
3. Run independent source audits. Keep one live coordinator responsible for the shared runtime and evidence surface. Assign one read-only native Kiro sub-agent per feature when the records are independent. Each compares the claimed behavior and path with current code and real entrypoints and returns file pointers, drift, and risks. Do not let auditors edit shared files.
4. Reconcile first. Combine the audits into one coverage ledger. Classify each feature as clean, changed, or blocked and distinguish product drift from verifier drift.
5. Run the mandatory live pass. The single coordinator exercises every feature's launch, doctor/preflight, drive, evidence capture, failure recovery, evidence survival, and cleanup/residue checks on its real surface. Static inspection or a unit-only shortcut cannot replace this pass. After an initial failed drive, allow exactly one recovery retry. If the real surface remains unreachable, record verified-unreachable with the failed commands and environment evidence rather than retrying indefinitely or substituting inspection. Verify that failed attempts leave no unsafe residue and that bounded evidence remains available after cleanup.
6. Triage and repair verifier scope. Modify only the project-local, user-owned verification skill and feature records unless the user separately requested product repairs. Never relax an expectation to preserve green output. Keep blocked checks explicit.
7. Prove or stop. For every remaining initial-map draft, review its complete contract and run .pkstack/bin/projectctl feature publish <slug> --output json; publication executes the stored verifier and changes draft state only after it passes. For an already published verifier changed during this pass, run feature verify <slug> --output json. Never hand-edit draft: false or use another feature's evidence. Then run .pkstack/bin/projectctl feature validate --output json and nearby tests. Report a clean, changed-and-proved, verified-unreachable, or blocked outcome. When verifier assets changed and the user separately authorized a pull request, group the maintenance result into one reviewed pull request rather than one per feature. Without that authorization, leave the local result and report the exact next action.
8. Refresh spec bindings last. After a changed feature passes its required proof, rerun .pkstack/bin/projectctl goal bind-spec <spec-name> --feature <slug> --output json. An identical bridge is a no-op. A changed bridge requires explicit --overwrite only after comparing the old and new contracts. Never preserve green output by pointing a spec at a weaker verifier.

Keep scratch audit notes uncommitted. Return the index reconciliation, per-feature outcome, changed contracts, exact live runs, recovery and residue results, observed side effects, and unverified gaps.