verification
验证回路,质量保障核心
When & Why to Use This Skill
The Verification Skill implements a robust 'Execute-Verify-Fix' loop designed to stabilize AI-generated code quality. It moves beyond simple code generation by enforcing multi-layered validation—including functional, unit testing, build, and runtime checks—ensuring that tasks are only marked as complete once they pass rigorous quality standards. This skill significantly reduces hallucinations and technical debt by incorporating a self-healing mechanism that automatically analyzes and fixes errors.
Use Cases
- Automated Bug Fixing: Automatically running test suites after a code change, analyzing failure logs, and retrying the fix up to three times before requesting human intervention.
- Continuous Integration Guardrails: Ensuring all generated code passes linting, type-checking, and build processes to maintain a clean and deployable codebase.
- Edge Case Validation: Systematically verifying that new features handle boundary conditions and unexpected inputs through predefined functional checklists.
- Reliable State Management: Maintaining an active verification log in project metadata to track the success status and evidence of every task performed by the AI agent.
| name | verification |
|---|---|
| description | 验证回路,质量保障核心 |
Verification Skill
这是质量提升2-3倍的核心原则 — Boris Cherny 如果AI不能验证自己的工作,质量就不稳定。
核心理念
验证回路是确保代码质量的关键。不是"写完代码就结束",而是"验证通过才算完成"。
验证流程
┌─────────────────────────────────────────┐
│ Verification Loop │
├─────────────────────────────────────────┤
│ │
│ Execute ──→ Verify ──→ Pass? ──→ Done │
│ │ │
│ ↓ No │
│ Analyze │
│ │ │
│ ↓ │
│ Fix │
│ │ │
│ ↓ │
│ Retry (max 3) │
│ │ │
│ ↓ Fail │
│ 寸止: 人工介入 │
│ │
└─────────────────────────────────────────┘
验证类型
1. 功能验证
- [ ] 代码修改是否生效?
- [ ] 功能是否按预期工作?
- [ ] 边界情况是否处理?
2. 测试验证
- [ ] 单元测试通过?
- [ ] 集成测试通过?
- [ ] 相关测试没有回归?
3. 构建验证
- [ ] 编译/构建成功?
- [ ] 无类型错误?
- [ ] 无Lint警告?
4. 运行时验证
- [ ] 应用能正常启动?
- [ ] 关键流程可用?
- [ ] 无控制台错误?
自我修复循环
async function executeWithVerification(task) {
for (let attempt = 1; attempt <= 3; attempt++) {
try {
// 执行任务
const result = await execute(task);
// 验证结果
if (await verify(result)) {
return result; // 成功
}
// 分析失败原因
const analysis = await analyze(result);
// 修复
task = await fix(task, analysis);
} catch (error) {
if (attempt === 3) {
// 请求人工介入
await 寸止.escalate({
issue: "验证失败,已重试3次",
attempts: attempts,
request: "请人工介入"
});
}
}
}
}
验证清单
修改代码后
1. 读取修改后的代码,确认变更正确
2. 运行相关测试
3. 检查没有破坏其他功能
4. 更新任务状态
提交前
1. 所有测试通过
2. 无类型错误
3. 无Lint警告
4. 功能验证通过
5. 更新 .ai_state/active_context.md
验证日志
记录到 .ai_state/active_context.md:
## 📝 验证日志
### [日期] T-001: 用户登录
- **状态**: ✅ 通过
- **测试**: 5/5 通过
- **验证方式**: 单元测试 + 手动验证
- **证据**: [截图/日志]
与Stop Hooks集成
验证失败3次 → 触发 [VERIFICATION_FAILED]
→ 寸止请求人工介入
→ 等待用户决策
使用方式
// 加载验证技能
skill.load("verification");
// 执行带验证的任务
verification.execute(task);
// 验证修改
verification.verify(changes);
// 记录验证结果
verification.log(result);