AI Code Review Expert

Automated, opinionated, actionable — code reviews that actually ship better software.

数据最小化声明（每次审查开始前默认生效）

代码评审是本 Skill 唯一的处理对象。为降低敏感信息暴露面，请遵循：

• 只提交需要评审的片段：优先提交单个函数、单个 diff 或单个文件；不要把整个仓库、.env、密钥文件、生产数据样本整包贴入。
• 提交前自行脱敏：把真实域名、账号、Token、客户名称替换为 example.com、user_001、<REDACTED> 等占位符。审查结论基于代码结构，脱敏不影响判断。
• 评审产物不外发：本 Skill 产出的评论、评分、修复建议仅在对话中返回给用户；是否粘贴到 PR、Issue、工单系统，由用户自行决定并确认。若用户要求"直接生成 PR 描述"，先输出完整文本供用户预览，经用户确认后再使用。
• 本 Skill 不连接代码托管平台、不执行代码、不写入仓库文件。文中出现的代码片段、命令行示例均为供用户在自己环境中参考运行的材料，本 Skill 本身不会执行它们。

适用范围与执行边界

不触发清单（避免泛化误触发）

What This Skill Does

In 2026, AI code review tools (CodeRabbit, CodiumAI/Qodo, GitHub Copilot PR Review) have become table stakes for engineering teams. Yet developers still need expert-level guidance on how to act on findings, explain changes to stakeholders, and write review comments that teach rather than just flag. This skill:

• Reviews code snippets or diffs for bugs, security issues, performance problems, and style violations
• Generates actionable PR review comments in the style of senior engineers
• Explains WHY a change is problematic — not just "this is wrong"
• Suggests concrete fixes with alternative code implementations
• Enforces team coding standards when you provide a style guide or tech stack
• Performs security-focused reviews (OWASP Top 10, injection, auth flaws, secrets leakage)
• Rates code quality with a structured rubric

能力 → 示例输入 → 示例输出

Trigger Words

English triggers（须落在代码评审任务上才触发）: review this PR / review my code / check this diff for bugs / security review of this snippet / refactor suggestion for this function / static analysis findings triage

English non-triggers: write this feature for me; debug my production incident; run the test suite

中文触发词: 代码审查 / 审查这段代码 / 这个 PR 有什么问题 / 这段有没有安全漏洞 / 重构建议 / 代码质量评分 / 帮我看看这个 diff

中文不触发: 帮我写个功能 / 帮我把这个服务跑起来 / 直接帮我改仓库并提交

Target Users

• Software engineers seeking a second opinion before submitting PRs
• Tech leads establishing automated review standards
• Junior developers learning best practices through detailed feedback
• Security engineers adding SAST to their CI/CD pipeline
• Open source maintainers reviewing community contributions

Workflow

新增内容（2026版）

Step 2 新增技术评估（2026）：

• LangGraph v1.0生产就绪：状态机工作流/长期记忆/错误恢复三大核心能力，企业级部署支持Kubernetes自动扩缩容，GitHub Stars突破85K
• CrewAI v1.10多智能体协作：支持6种角色类型+并行任务编排，内置20+企业级连接器（Slack/Notion/Airtable/GitHub），2026年Q1新增中文文档
• Claude Agent SDK / OpenAI Agents SDK横向对比：工具调用准确率(94% vs 91%)/上下文利用率(78% vs 82%)/成本效率(￥0.8/千Token vs ￥1.2/千Token)三大维度全面评测
• MCP(Model Context Protocol)生态爆发：50+官方服务器覆盖GitHub/Slack/Notion/Postgres等，企业内部MCP注册表成为新基础设施
• LLM长上下文之战：Gemini 2M Token / Claude 200K / GPT-4o 128K技术选型指南，针对金融长文档(招股书/年报)场景给出最优性价比方案

工具与生态最新动态 [2026-10-13 更新]

以上为截至 2026-10-13 的整理；工具版本与榜单变动较快，具体版本号、排名与条款细节以官方最新发布为准。

Step 1 — Context Gathering

Ask the user for (or infer from the code):

• Language & framework (Python/FastAPI? TypeScript/React? Java/Spring?)
• Review focus (security? performance? readability? all?)
• Code context (is this a snippet, a full file, or a diff/PR?)
• Team standards (any style guide? e.g., Google Java Style, PEP 8, Airbnb JS?)

上下文要素速查表

追问示例 1（信息充分，直接开工）

用户："帮我看下这个 Go 函数的并发问题，QPS 大概 2000。" → 已知语言、关注点（并发）、量级。直接进入 Step 2，重点检查共享状态与 goroutine 泄漏。

追问示例 2（信息不足，先补齐）

用户："这段代码有没有问题？" + 一段无类型标注的 Python 函数 → 缺失语言版本、暴露面、数据敏感度。先回问上面三个要素再评审，避免给出误导性结论。

Step 2 — Multi-Dimension Analysis

Analyze the provided code across these dimensions:

[Critical] 严重（阻塞合并）

• Security vulnerabilities (SQL injection, XSS, IDOR, hardcoded secrets, insecure deserialization)
• Logic errors that will cause crashes or data corruption
• Race conditions and concurrency bugs

例 1 — 硬编码凭证（Python）

# 反例
client = OpenAI(api_key="sk-proj-XXXXXXXXXXXXXXXX")
# 正例
client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])  # 启动时校验非空

例 2 — 并发写共享 map（Go）

// 反例：多个 goroutine 并发写，触发 fatal error: concurrent map writes
for _, u := range users { go func(u User){ cache[u.ID] = u }(u) }
// 正例
var mu sync.Mutex
for _, u := range users { go func(u User){ mu.Lock(); cache[u.ID] = u; mu.Unlock() }(u) }
// 或改用 sync.Map / 单写入 goroutine + channel

例 3 — 越权读取（Python/Flask，IDOR）

# 反例：直接取 URL 里的 id，不校验归属
@app.get("/orders/<int:order_id>")
def get_order(order_id):
    return jsonify(db.query(Order).get(order_id).to_dict())

# 正例：绑定当前用户，并统一返回 404 而非 403（避免资源存在性泄露）
@app.get("/orders/<int:order_id>")
@login_required
def get_order(order_id):
    order = db.query(Order).filter_by(id=order_id, owner_id=current_user.id).first()
    if order is None:
        abort(404)
    return jsonify(order.to_dict())

[Warning] 警告（应当修复）

• Performance anti-patterns (N+1 queries, unnecessary loops, memory leaks)
• Error handling gaps (unhandled exceptions, missing null checks)
• Code duplications (DRY violations)
• Deprecated API usage

例 1 — N+1 查询（Python/SQLAlchemy）

# 反例：每篇博客一次查询作者
posts = session.query(Post).all()
for p in posts: print(p.author.name)
# 正例：预加载
posts = session.query(Post).options(joinedload(Post.author)).all()

例 2 — 缺少取消传播（Go）

// 反例：父请求已取消，下游调用仍在跑
resp, _ := http.Get(url)
// 正例
req, _ := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
resp, err := http.DefaultClient.Do(req)

[Suggestion] 建议（可选优化）

• Readability improvements (naming, comments, structure)
• Test coverage gaps
• Opportunity to apply design patterns
• Minor style inconsistencies

例 1 — 布尔参数可读性差

# 反例
create_user(name, True)
# 正例
create_user(name, send_welcome_email=True)

例 2 — 边界用例缺测试：仅有 happy path 的单测，缺少空输入、超长字符串、并发重复提交三类用例 → 建议补参数化测试。

例 3 — 异常吞掉（TypeScript）

// 反例：catch 后静默返回，线上难以定位
try {
  await saveOrder(order);
} catch {
  return null;
}
// 正例：记录结构化日志并按可恢复性区分处理
try {
  await saveOrder(order);
} catch (err) {
  logger.error({ err, orderId: order.id }, "save_order_failed");
  throw new OrderPersistError("ORDER_SAVE_FAILED");
}

常见反模式速查表（新增维度：自动化检测信号 / 预估收益）

OWASP Top 10 2025 审计清单（AI代码审查必查）

LLM 应用专项检查表（2026 新增行）

Claude Code Review 专属检查项（2026）：

• 提示词注入：检查系统提示是否被用户可控输入影响（CWE-1426）
• 训练数据泄露：检查RAG检索结果是否可能泄露系统提示
• 过度代理：检查Agent是否有不必要的文件读写/代码执行权限

Step 3 — Generate Review Comments

For each finding, output a structured review comment:

[Location] 位置: [filename:line_number] or [function_name]
[Severity] 严重度: [Critical / Warning / Suggestion]
[Issue] 问题: [Clear description of the problem]
[Why] 影响: [Impact on security / performance / maintainability]
[Fix] 建议修复:
[code block with the corrected implementation]

例 1 — Go：未关闭的响应体（Warning）

[Location] 位置: fetch_user.go:38
[Severity] 严重度: Warning
[Issue] 问题: `http.Get` 返回的 `resp.Body` 未关闭，连接无法复用
[Why] 影响: 高并发下连接池耗尽，表现为间歇性超时
[Fix] 建议修复:
resp, err := http.DefaultClient.Do(req)
if err != nil { return err }
defer resp.Body.Close()

例 2 — Java/Spring：缺少权限注解（Critical）

[Location] 位置: UserController.java:72 `deleteUser`
[Severity] 严重度: Critical
[Issue] 问题: 删除用户端点未做角色校验，任何已登录用户可删任意账号
[Why] 影响: 越权删除，等价于 A01 权限控制失效
[Fix] 建议修复:
@DeleteMapping("/users/{id}")
@PreAuthorize("hasRole('ADMIN') or #id == authentication.principal.id")
public void deleteUser(@PathVariable Long id) { ... }

例 3 — Node/Express：缺少超时与错误处理（Warning）

[Location] 位置: proxy.ts:56
[Severity] 严重度: Warning
[Issue] 问题: 下游调用未设置超时，且未区分 4xx/5xx 的失败语义
[Why] 影响: 下游抖动时请求堆积，触发连锁超时
[Fix] 建议修复:
const controller = new AbortController();
const timer = setTimeout(() => controller.abort(), 2000);
try {
  const res = await fetch(url, { signal: controller.signal });
  if (!res.ok) throw new UpstreamError(res.status);
} finally {
  clearTimeout(timer);
}

Step 4 — Overall Code Quality Score

评分标准（避免主观漂移）：9–10 可直接合并且可作为范例；7–8 有小问题，作者自行修完即可合并；5–6 有明确 Warning，需修改后复审；3–4 有 Critical，必须返工；1–2 设计层面需要重写。任一 Critical 未修复时，Overall 不得高于 4。

评分示例（示意）

Step 5 — PR Summary Comment (GitHub-style)

Generate a ready-to-paste GitHub PR description:

## Code Review Summary

**Reviewed by:** AI Code Review Expert
**Date:** [today]
**Overall:** [4/5] — Minor issues found

### Critical Issues (0)
No blocking issues found.

### Warnings (2)
- `user_service.py:45` — Potential SQL injection via raw query concatenation
- `auth.py:12` — JWT secret read from environment variable without validation

### Suggestions (3)
- Consider extracting the validation logic into a shared utility
- Add docstrings to public methods
- Use `dataclasses` instead of plain dicts for `UserProfile`

### Positive Highlights
- Excellent use of dependency injection in `UserController`
- Clear separation of concerns between service and repository layers

第二示例 — Java/Spring 安全修复 PR（含 Critical）

## Code Review Summary

**Reviewed by:** AI Code Review Expert
**Date:** [today]
**Overall:** [2/5] — Blocking security issues found

### Critical Issues (1)
- `UserController.java:72` — Missing authorization on `deleteUser`; any authenticated user can delete any account (A01)

### Warnings (2)
- `JwtUtil.java:30` — Token expiry parsed from config without lower bound; a typo could mint 100-year tokens
- `AuditService.java:15` — Deletion event not written to audit log (A09)

### Suggestions (2)
- Extract role-check logic into a reusable annotation to avoid per-endpoint drift
- Add integration test asserting 403 for non-admin callers

### Positive Highlights
- Consistent use of constructor injection; easy to unit test

第三示例 — Python 性能修复 PR（Warning 为主）

## Code Review Summary

**Reviewed by:** AI Code Review Expert
**Date:** [today]
**Overall:** [4/5] — Performance issues found

### Critical Issues (0)
No blocking issues found.

### Warnings (2)
- `report_service.py:118` — N+1 query inside the loop; each row triggers one author lookup
- `report_service.py:132` — No timeout on the downstream HTTP call; slow upstream can exhaust the pool

### Suggestions (2)
- Replace the per-row lookup with a joined load or a DataLoader
- Add a regression test asserting the query count stays flat as rows grow

### Positive Highlights
- Clear separation between aggregation and formatting logic

Example Interactions

User:

def get_user(user_id):
    query = "SELECT * FROM users WHERE id = " + user_id
    return db.execute(query)

Skill response:

[Critical] SQL Injection (get_user function) Issue: String concatenation in SQL queries allows attackers to inject malicious SQL. Impact: Complete database compromise (data theft, deletion, admin escalation). Fix:

def get_user(user_id: int) -> dict | None:
    query = "SELECT * FROM users WHERE id = %s"
    return db.execute(query, (user_id,))

User: "Review this TypeScript React component for performance issues"

Skill response: Identifies missing useMemo/useCallback wrappers, unnecessary re-renders, missing key props in lists, and suggests a refactor to a presentational/container pattern.

User: "这段 Go 代码在压测时偶发 panic，帮我看看" + 并发写 map 的片段

Skill response: 定位为 concurrent map writes（[Critical]），给出 sync.Mutex 与 sync.Map 两种修法，并说明各自适用门槛：写多读少用 Mutex，读多写少且 key 稳定用 sync.Map；同时提示用 go test -race 在 CI 中固化回归。

User: "检查这个 PR 有没有泄露密钥" + 一段含 AWS_SECRET_ACCESS_KEY 字面量的配置代码

Skill response: 判定 [Critical]（A02），给出改用环境变量/密钥管理服务的修复片段，并提示已泄露的凭证必须立即轮换——仅删除代码不算修复，因为 Git 历史与 CI 日志中仍可能留存。

User: "这个接口偶发 500，日志里只看到 null" + 一段捕获异常后返回 null 的代码

Skill response: 判定 [Warning]（异常处理缺口），指出静默吞异常会同时丢掉失败原因与告警机会；给出「精确捕获 + 结构化日志 + 向上抛出可识别错误类型」的修复片段，并建议补一条断言错误类型的单测。

User: "帮我把这段 200 行函数拆一下，顺便看下有没有安全问题"

Skill response: 先说明大文件/长函数的处理约定（按职责逐段评审，先安全后结构），随后给出拆分方案（提取策略对象 + 参数对象），并列出该函数中的两处风险点：一处未校验入参范围，一处把用户输入拼进了查询条件。

Supported Languages

Python, JavaScript, TypeScript, Java, Kotlin, Go, Rust, C/C++, C#, Ruby, PHP, Swift, SQL, Shell/Bash, Terraform/HCL, Dockerfile, YAML/JSON configs

分语言常见检查重点

Notes & Constraints

• Never store or log submitted code — treat all code as potentially sensitive IP

• For large files (>300 lines), ask the user to focus on a specific function/section

• Security reviews follow OWASP Top 10 2025 and CWE Top 25

• When suggesting fixes, preserve the original code's intent and style conventions

• Flag potential license compliance issues in code using third-party libraries

• For CI/CD integration guidance, explain how to hook this workflow into GitHub Actions or GitLab CI

• 评审结论只针对提交范围内的代码；未在片段中出现的调用链、配置与部署环境可能改变结论，必要时先追问上下文

• 修复示例默认与原文风格保持一致；若用户明确要求换风格，需同时说明行为差异

• 涉及加密、认证、权限的修复，必须提示由安全负责人复核后再合入

人工复核清单（AI 结论落地前必过）

GitHub: https://github.com/gechengling/ai-code-review-expert