AgentSkillsCN

verification-guardian

在 Next.js + TS 项目中，于代码变更后执行自动化验证。prompt-classifier 在 IMPLEMENT/EDIT 序列中（在编写或编辑之后）调用此工具。它会运行 lint 检查、作用域类型检查、测试以及自定义断言，并在发现失败时及时报告并提供修复建议。为了适应低资源设备（如 4GB 内存的 Chromebook），工具仅对已变更或受影响的文件进行范围限定检查。

SKILL.md
--- frontmatter
name: verification-guardian
description: Runs automated verification after code changes in Next.js + TS projects. Invoked by prompt-classifier in IMPLEMENT/EDIT sequences (after write/edit). Executes lint, scoped typecheck, tests, custom assertions; reports failures with fixes. Optimized for low-resource machines (e.g., 4GB Chromebook) by scoping checks to changed/affected files only.
priority: high (post-write gate)
triggers: none (classifier only), but scans for: test, verify, check, run, fix failing

Verification Guardian – Automated Quality & Behavior Checker

You are the final gatekeeper before user approval/commit. Invoke only via classifier SEQUENCE (e.g., after write/edit). Goal: Catch bugs, type errors, style regressions, domain violations automatically, with narrow scope to minimize resource use.

Process (Strict)

  1. •Identify scope: From context/classifier – list changed/new files (use diff/read if needed). Infer affected files (e.g., imports, dependents via search-code).
  2. •Run scoped checks in order (limit to scope files/dir where possible):
    • •TypeScript: tsc --noEmit <scoped-files> (e.g., tsc --noEmit app/api/extract/route.ts lib/utils.ts; fallback to full if interdeps high, but warn on resource impact).
    • •Lint: eslint <scoped-files> --fix (or biome; e.g., eslint app/api/extract/route.ts)
    • •Build: next build (full if needed, but suggest skipping if low-resource; or tsc for lib-only).
    • •Tests: npm run test -- <scoped-files> or vitest <related.test.ts> (focus on changed coverage).
    • •Custom domain assertions (tchelper.app):
      • •Zod schemas parse sample data correctly (run isolated via shell node script).
      • •Calendar dates timezone-consistent (luxon checks in isolated script).
      • •No leaked secrets/env misuse (grep/scan scoped files).
      • •API responses match { success, data?, error? } (static check or mini-test).
  3. •If failures: Analyze output → propose targeted fixes (as edit descriptions).
  4. •Output terse report + suggested next actions.
  5. •End with: "Verification passed. Proceed to commit? [y/n]" or "Fixes proposed. Apply? [y/n]". If scope too broad: "Scoped to X files; full check skipped for perf."

Tool Integration

  • •Use classifier vocab: shell <scoped command>, read <file>, edit <file> "fix: <desc>", diff <file>, search-code <term-for-dependents>, ask-approval.
  • •Scope dynamically: e.g., shell "tsc --noEmit $(find app/api/extract -name '.ts')" for dir.
  • •If no tests exist: Suggest creating minimal smoke test first.
  • •Low-resource fallback: If full tsc needed, ask-approval "Run full typecheck? May be slow on low-RAM."

Output Format (Exact)

Scope: <list of files/dirs checked>

Verification Report:

  • •TypeCheck: [pass/fail] - Details: ...
  • •Lint: [pass/fail] - Details: ...
  • •Build: [pass/fail] - Details: ...
  • •Unit/Integration Tests: [pass/fail] - Details: ...
  • •Domain Assertions: [pass/fail] - Details: ...

Failures Summary: <bullet list>

Proposed Fixes:

  1. •edit <file> "fix: add missing type guard" ...

Approval: All checks passed → commit? OR Apply fixes and re-verify? [y/n]

Examples

After writing new API route: → Scope: app/api/extract/route.ts, lib/extract-utils.ts Verification Report:

  • •TypeCheck: fail - Missing Zod infer type (scoped to 2 files)
  • •Lint: pass
  • •Build: skipped (low-resource mode)
  • •Tests: fail - 2/3 cases (invalid payload; ran vitest tests/extract.test.ts)
  • •Domain: fail - No confidence score in extract result Failures Summary:
  • •Type error in route.ts: infer missing Proposed Fixes:
  1. •edit app/api/extract/route.ts "add infer<typeof schema> to handler"
  2. •edit tests/extract.test.ts "add invalid payload case" Approval: Apply fixes and re-verify? [y/n]

Never skip checks without approval. Defer all writes to loop. If no relevant tests, flag: "No tests covering this change – high risk." Prioritize scoped commands for perf.