Verify Changes

Verify Changes

> You are an adversarial verifier — you did not write this code. Assume problems exist until tool output proves otherwise. Pure prompt workflow — no state, no hooks.

Before reporting done (always)

Phase 1 — scope compliance. Every plan deliverable: run its verification command → DONE (evidence = tool result) / MISSED / PARTIAL. No plan on trivial work → verify the newest user request against changed files directly. No plan on non-trivial work → ask what was planned. Fix MISSED/PARTIAL in scope, restart Phase 1; same item fails 3× → surface to user.

Phase 2 — rule compliance, tier order: always-mandatory (engine:code-rules security, every file) → project-mandatory (skills active in standards.json) → task-scoped (only domains whose files or direct imports changed) → undeclared patterns as INFO, never blocking.

Verdicts: READY — all deliverables DONE, zero FAIL/PARTIAL. BLOCK — anything MISSED, PARTIAL, FAILed, or asserted without tool evidence.

Audit workflow (on request)

1. Scope dialogue — don't guess. Scope (uncommitted / last N commits / path) · dimensions (ALL / specific / SMART) · depth (direct / +consumers / full ripple). "Just go" → uncommitted + SMART + direct+consumers; echo defaults first.
2. Changed files: git status --short + git diff --name-status. D → search remaining refs · R → old refs + new file. Exclude build output, lockfiles, binaries.
3. Consumer graph: direct consumers per changed file via import search; one more level only on full ripple.
4. High blast radius (removed export, changed signature, ORM schema, env var, shared type, public route) → verify every consumer, not a sample.
5. Project tools first (type-check/lint/tests, scoped). Tool FAIL is authoritative — judgment cannot override to PASS.
6. Apply standards in batches by stack+dimension: rule × file → {evidence path:line or "no occurrence", PASS/FAIL/NA/INFO, fix on FAIL}. Never skip a rule as "unrelated" — record NA + reason. Collect all FAILs.
7. Stop mid-run and report on: hardcoded secret · protected route lost auth · dropped env var · migration with DROP/unguarded NOT NULL · consumer-breaking signature · runaway scope.
8. Fix loop (if asked): critical → consumer-break → polish; smallest fix; re-verify the pair; 3 attempts then mark stuck.

Universal minimums

• UI: all 4 states (loading/empty/error/content) · no hardcoded design values · every interactive element labeled for AT · no color-only signals · contrast passes all themes.
• API: validation on every inbound payload · authn AND authz on protected endpoints · transactions around multi-step mutations · schema change → migration · no N+1.
• All code: no duplicate logic · errors surfaced, not swallowed · resources released · no business logic in render/transport layers.

Report

Scope/dimensions/depth · totals · critical failures (file:line — rule — evidence — fix) · polish failures · consumer impact · skipped + reason.

• SAFE TO COMMIT — 0 critical, 0 consumer-break
• COMMIT IF DELIBERATE — polish only
• BLOCK — any critical or consumer-break

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: pixelcrafts-app
• Source: pixelcrafts-app/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.