360 Secure SkillHub
360 安全技能中心

Delegated Code Review

该 Skill 定义了一次委派式代码评审流程:启动只读子代理评审代码,按问题、文件或行号、理由、建议记录发现;关闭子代理后逐条核对,采纳的问题交由实现子代理或自行修复,并在每次修复后重跑相关测试,最后汇总修复、验证结果与剩余风险。

wufei-png
wufei-png
热度 21.45
最近更新 2026-08-08 00:20:51
分类:开发依赖:内部调用大模型 / 代码执行 / 读写文件应用场景:代码审查 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add delegated-code-review
Metadata
name
delegated-code-review
description
Perform a single review with a fresh, read-only subagent, then verify findings and apply accepted fixes.
license
MIT-0

Delegated Code Review

Review

  1. Start one fresh subagent with fork_turns: "none" and $review-agent. The reviewer may inspect code, tests, and call sites, but must only review. Wait patiently for its conclusion; unless the reviewer is confirmed to be in an abnormal state, do not send prompts to rush or wrap up the review.
  2. Ask the reviewer to focus on Critical, Important, and clearly worthwhile issues. Record Minor issues as non-blocking. Every finding must include:
    • Issue
    • File/line number
    • Reason
    • Recommendation
  3. Close the subagent after its report. If there are no findings, only Minor findings, or all findings are not accepted, finalize. Otherwise, verify each finding yourself:
    • Record invalid or unaccepted findings as Not accepted with the reason.
    • If an implementation subagent owns the change, delegate accepted fixes to it; otherwise fix them yourself. In either case, rerun the relevant tests after each fix.

Finalize

Summarize the implementation, findings and decisions, fixes, verification results, unverified items, and remaining risks.