🔍 Smart PR Review

An AI code reviewer with opinions. Not a yes-machine.

If you find this useful, please give it a star — it helps others discover it.

Why Another Code Review Tool?

Most AI code review tools are yes-machines — they surface linting issues, suggest minor refactors, and approve everything else. Smart PR Review is different: it acts like a Staff Engineer who's seen production meltdowns and won't let your code ship until the hard questions are answered.

Feature	GitHub Copilot Code Review	Smart PR Review
Review stance	Suggests	Judges
Says "this approach is wrong"	Rarely	Yes, in the Summary
Review depth	Line-level suggestions	6-layer deep analysis
Devil's Advocate mode	No	Built-in
Architectural review	Limited	Dedicated layer
Output format	Inline comments	Structured report (paste-ready)
Security audit	Basic	Dedicated layer with OWASP checks
Severity discipline	Flat suggestions	MUST FIX / SHOULD FIX / SUGGESTION
Webhook automation	Built-in	Self-hosted via `index.ts`
Language-specific checks	Generic	TS/JS, Python, Go, Rust specific

Overview

Smart PR Review is an opinionated AI code reviewer that performs 6-layer deep analysis with a Devil's Advocate mechanism — it actively challenges your assumptions and stress-tests your code against real-world failure modes.

It doesn't just check if your code works. It asks: what happens when it doesn't?

Core values:

Direct — "This approach is wrong" beats "This is an interesting choice"
Actionable — Every issue comes with replacement code, not vague advice
Prioritized — Clear severity levels that map to merge decisions
Holistic — Reviews architecture, not just syntax

Features

6-layer review — Logic → Boundaries → Performance → Security → Maintainability → Architecture
Devil's Advocate mode — Forces worst-case thinking on every critical change
Structured output — MUST FIX / SHOULD FIX / SUGGESTION / What's Good / Verdict
4 input modes — PR URL, local diff, commit hash, file path
Webhook integration — Auto-review PRs via GitHub webhooks
Language-aware — TypeScript/JS, Python, Go, Rust with language-specific checks
Large PR handling — Automatic diff chunking for PRs > 1000 lines
Strict mode — Raises severity thresholds for critical codebases

Quick Start

As a Claude Code Skill (CLI)

# Install the skill
openclaw install smart-pr-review

# Review a GitHub PR
/review https://github.com/your-org/repo/pull/123

# Review your staged changes
/review --diff

# Review a specific commit
/review --commit=abc1234

# Review with security focus in strict mode
/review https://github.com/your-org/repo/pull/123 --focus=security --strict

As a Webhook Service

# Set environment variables
export GITHUB_TOKEN="ghp_..."
export GITHUB_WEBHOOK_SECRET="your-secret"
export ANTHROPIC_API_KEY="sk-ant-..."

# Install dependencies & start
npm install hono @hono/node-server
npx tsx index.ts

# 🔍 Smart PR Review webhook started: http://localhost:3000

Commands

`/review [PR-URL]`

Review a GitHub Pull Request by URL.

/review https://github.com/acme/api/pull/42
/review https://github.com/acme/api/pull/42 --focus=security
/review https://github.com/acme/api/pull/42 --strict --lang=en

Requires: gh CLI installed and authenticated.

`/review --diff`

Review local uncommitted changes (staged → unstaged → last commit fallback).

/review --diff
/review --diff --focus=performance

`/review --commit=<hash>`

Review a specific commit.

/review --commit=a1b2c3d
/review a1b2c3d              # shorthand

Parameters

Parameter	Values	Default	Description
`--focus`	`security`, `performance`, `logic`, `all`	`all`	Focus on a specific review dimension
`--strict`	flag	off	Lower tolerance thresholds (see Strict Mode)
`--lang`	`zh`, `en`	`zh`	Output language
`--commit`	`<hash>`	—	Review a specific commit

The 6-Layer Review

Every review systematically walks through six dimensions, from correctness to architecture:

flowchart TD
    INPUT["📥 PR / Diff / Commit"] --> PREP["📚 Load Review Knowledge Base"]
    PREP --> L1

    subgraph REVIEW ["6-Layer Deep Review"]
        direction TB
        L1["🧠 Layer 1: Logic Correctness\nControl flow, state mutations,\nconcurrency, type safety, error propagation"]
        L2["🔲 Layer 2: Boundary Conditions\nNull/empty handling, numeric overflow,\nUnicode, timezones, env differences"]
        L3["⚡ Layer 3: Performance\nN+1 queries, memory leaks,\nalgorithm complexity, missing caches"]
        L4["🔒 Layer 4: Security\nInjection, hardcoded secrets,\nCSRF/SSRF, path traversal, auth gaps"]
        L5["🔧 Layer 5: Maintainability\nNaming, function design,\nover/under-engineering, test coverage"]
        L6["🏗️ Layer 6: Architecture\nPattern consistency, API design,\nbackward compatibility, module boundaries"]

        L1 --> L2 --> L3 --> L4 --> L5 --> L6
    end

    L6 --> DA["😈 Devil's Advocate"]
    DA --> OUT["📊 Structured Verdict"]

Layer 1: Logic Correctness 🧠

Control flow completeness, state mutation consistency, race conditions, type safety, error propagation chains.

Layer 2: Boundary Conditions 🔲

Null/undefined/nil handling, empty collections, integer overflow, floating-point precision, Unicode edge cases, timezone & daylight saving, cross-platform path differences.

Layer 3: Performance ⚡

N+1 queries, unnecessary re-renders (React), memory leaks (listeners, timers, closures), O(n²) on large datasets, missing database indexes, redundant network requests.

Layer 4: Security 🔒

Hardcoded secrets, SQL/XSS/command injection, CSRF/SSRF, unsafe deserialization, path traversal, missing auth checks, sensitive data in logs, vulnerable dependencies.

Layer 5: Maintainability 🔧

Naming clarity, single responsibility, over/under-abstraction, magic numbers, meaningful error handling, test coverage for new logic.

Layer 6: Architecture 🏗️

Consistency with existing patterns, dependency coherence, API design conventions, backward compatibility, module boundary violations, circular dependencies.

Devil's Advocate Mode

This is the core differentiator. Even when code looks fine, the reviewer forces itself through five stress tests:

Question	What it catches
What if traffic is 100x current?	Scaling bottlenecks, connection pool exhaustion
What if input is maliciously crafted?	Injection attacks, DoS vectors
What if this needs to change in 6 months?	Rigid coupling, poor extensibility
What if a dependency goes down?	Missing fallbacks, cascading failures
What if a junior dev maintains this?	Implicit knowledge, unclear control flow

Only when all five questions have satisfactory answers does the reviewer give APPROVE.

Output Format

Every review produces a structured, GitHub-pasteable report:

## 🔍 Code Review: PR #247 — Add user search API endpoint

### Summary
New user search API with name/email fuzzy matching. **The approach has security risks**:
the search endpoint has no auth and contains a SQL injection vulnerability. Must fix before merge.

---

### 🚨 MUST FIX (2 issues)

**[MF-1] SQL Injection Vulnerability**
📍 `src/routes/users.ts:45`
```typescript
const results = await db.query(
  `SELECT * FROM users WHERE name LIKE '%${req.query.q}%'`
);

Problem: User input directly concatenated into SQL. An attacker can craft q=%'; DROP TABLE users; -- to destroy the database. Impact: P0 security vulnerability — arbitrary database read/write. Suggested fix:

const results = await db.query(
  "SELECT id, name, email FROM users WHERE name LIKE $1",
  [`%${req.query.q}%`]
);

[MF-2] Search endpoint missing authentication 📍 src/routes/users.ts:38 Problem: No authMiddleware — anyone can search user data including emails. Impact: Privacy violation, potential GDPR non-compliance.

⚠️ SHOULD FIX (2 issues)

[SF-1] Returns unnecessary user fields 📍 src/routes/users.ts:45 Problem: SELECT * exposes password_hash, reset_token. Suggestion: Explicitly select id, name, email, avatar_url.

[SF-2] No pagination — OOM risk at scale 📍 src/routes/users.ts:45-48 Suggestion: Add LIMIT $2 OFFSET $3, default 20 results per page.

💡 SUGGESTION (1 issue)

[SG-1] Add minimum search length 📍 src/routes/users.ts:40 Suggestion: if (q.length < 2) return res.status(400)... to prevent single-char queries.

✅ What's Good

Clean route organization following existing src/routes/ patterns
Proper async/await usage, good readability

📊 Verdict

[x] REQUEST CHANGES — Must fix critical issues

Two P0 security issues (SQL injection + missing auth) must be resolved before merge.


### Severity Rules

| Tag | Meaning | Merge Impact |
|---|---|---|
| 🚨 **MUST FIX** | Bugs, security holes, data loss risk | **Blocks merge** |
| ⚠️ **SHOULD FIX** | Performance, maintainability, missing tests | Strongly recommended |
| 💡 **SUGGESTION** | Style, naming, better practices | Non-blocking |

### Strict Mode (`--strict`)

When `--strict` is enabled:
- Missing tests → **MUST FIX** (normally SHOULD FIX)
- Any `any` type usage → **SHOULD FIX**
- Missing error handling → **MUST FIX**
- Complex logic without comments → **SHOULD FIX**

---

## Webhook Integration (OpenClaw)

The `index.ts` webhook server enables **automatic PR review** — every PR opened or updated gets reviewed without manual invocation.

### Architecture

```mermaid
sequenceDiagram
    participant GH as GitHub
    participant WH as index.ts (Webhook)
    participant AI as Anthropic API
    participant PR as PR Comments

    GH->>WH: PR opened/synchronize/reopened
    WH->>WH: Verify webhook signature (HMAC-SHA256)
    WH-->>GH: 202 Accepted
    WH->>GH: Fetch PR files + diff
    WH->>WH: Filter reviewable files
    WH->>WH: Chunk large diffs (500KB limit)
    loop Each chunk
        WH->>AI: Send review prompt
        AI->>WH: Structured review
    end
    WH->>WH: Merge chunk results + extract verdict
    WH->>PR: Post review (APPROVE / REQUEST_CHANGES / COMMENT)

Setup

1. Start the webhook server:

export GITHUB_TOKEN="ghp_..."
export GITHUB_WEBHOOK_SECRET="your-webhook-secret"
export ANTHROPIC_API_KEY="sk-ant-..."
export REVIEW_MODEL="claude-sonnet-4-20250514"   # optional
export PORT=3000                                  # optional

npm install hono @hono/node-server
npx tsx index.ts

2. Configure GitHub repository:

Go to Settings → Webhooks → Add webhook
Payload URL: https://your-server:3000/webhook/github
Content type: application/json
Secret: same as GITHUB_WEBHOOK_SECRET
Events: select Pull requests

Environment Variables

Variable	Required	Default	Description
`GITHUB_TOKEN`	Yes	—	GitHub token with `repo` scope
`GITHUB_WEBHOOK_SECRET`	Yes	—	Webhook signature secret
`ANTHROPIC_API_KEY`	Yes	—	Anthropic API key
`REVIEW_MODEL`	No	`claude-sonnet-4-20250514`	Model for review
`PORT`	No	`3000`	Server port
`MAX_DIFF_SIZE`	No	`512000` (500KB)	Max diff chunk size in bytes
`REVIEW_LANGUAGE`	No	`zh`	Output language (`zh`/`en`)
`REVIEW_MAX_TOKENS`	No	`4096`	Max tokens per review chunk

Endpoints

Method	Path	Description
`GET`	`/health`	Health check
`POST`	`/webhook/github`	GitHub webhook receiver

Supported Languages

Language	Specific Checks
TypeScript / JavaScript	`any` abuse, unhandled Promise rejections, React `useEffect` dependency arrays, stale closures, ESM/CJS mixing
Python	Mutable default arguments, bare `except:`, missing context managers, GIL concurrency traps, type annotation consistency
Go	Unchecked errors, goroutine leaks, interface pollution, concurrent slice/map access, `defer` in loops
Rust	Unnecessary `.clone()`, `unwrap()`/`expect()` in non-test code, lifetime annotations, unnecessary `unsafe`, Error type design
All languages	Hardcoded config, missing observability, inconsistent error handling, stale comments

vs GitHub Copilot Code Review

Dimension	Copilot Code Review	Smart PR Review
Personality	Neutral, suggestive	Opinionated, decisive
Will say "this is wrong"	No	Yes
Review depth	Line-level	6-layer (logic → architecture)
Direction check	No	Flags wrong approaches in Summary
Devil's Advocate	No	5-question stress test
Output format	Inline suggestions	Structured report with severity
Severity discipline	Flat	MUST FIX / SHOULD FIX / SUGGESTION
Merge verdict	No explicit verdict	APPROVE / REQUEST_CHANGES / COMMENT
Replacement code	Sometimes	Always for MUST FIX
Large PR handling	Per-file	Automatic chunking with merge
Anti-pattern library	Built-in rules	Extensible `references/` knowledge base
Self-hosted webhook	No (GitHub native)	Yes (`index.ts`)
Customizable focus	No	`--focus=security\|performance\|logic`
Strict mode	No	`--strict` raises severity thresholds

Architecture

Smart PR Review operates in two complementary modes:

┌─────────────────────────────────────────────────────┐
│                  Smart PR Review                     │
├──────────────────────┬──────────────────────────────┤
│   CLI Mode           │   Webhook Mode               │
│   (SKILL.md)         │   (index.ts)                 │
├──────────────────────┼──────────────────────────────┤
│ • Manual invocation  │ • Auto on PR events          │
│ • /review command    │ • Hono HTTP server            │
│ • Full project       │ • Diff-only context          │
│   context access     │ • Anthropic API direct       │
│ • Claude Code tools  │ • GitHub Review API          │
│ • Terminal output    │ • Async processing           │
├──────────────────────┴──────────────────────────────┤
│              Shared Knowledge Base                   │
│  references/review-checklist.md — per-language checks│
│  references/anti-patterns.md   — pattern library     │
│  references/review-examples.md — output templates    │
└─────────────────────────────────────────────────────┘

Recommended workflow: Use webhook mode for automatic first-pass review on every PR, then use CLI mode for deep-dive reviews on critical changes.

Contributing

Fork this repository
Add checks to references/review-checklist.md for new patterns
Add anti-patterns to references/anti-patterns.md
Submit a PR (and yes, it will be reviewed by Smart PR Review 😈)

License

MIT

中文说明

简介

Smart PR Review 是一个有立场的 AI 代码审查工具 — 它不是无脑点头的橡皮图章，而是像一个有 10 年经验的 Staff Engineer 那样审查你的代码。

核心差异

直言不讳：发现问题直接说"这个方案有问题"，不说"这也是一种方式"
有判断力：能区分"必须修"和"建议改"，不把所有问题都列为 nit
给方案：每个 MUST FIX 都附带可执行的替代代码
主动反对：即使代码看起来没问题，也会强制进行 5 个维度的压力测试

6 层审查维度

🧠 逻辑正确性 — 控制流、状态变更、并发竞态、类型安全
🔲 边界条件 — 空值、空集合、数值溢出、Unicode、时区
⚡ 性能影响 — N+1 查询、内存泄漏、算法复杂度、缺少缓存
🔒 安全风险 — 注入攻击、硬编码密钥、CSRF/SSRF、路径遍历
🔧 可维护性 — 命名、函数设计、过度工程化、测试覆盖
🏗️ 架构一致性 — 模式一致性、API 设计、向后兼容、模块边界

使用方式

# 审查 GitHub PR
/review https://github.com/owner/repo/pull/123

# 审查本地变更
/review --diff

# 审查特定 commit
/review --commit=abc1234

# 聚焦安全审查 + 严格模式
/review https://github.com/owner/repo/pull/123 --focus=security --strict

# 英文输出
/review --diff --lang=en

双模式架构

CLI 模式（SKILL.md）— 手动调用 /review，可访问完整项目上下文，适合深度审查
Webhook 模式（index.ts）— GitHub PR 事件自动触发，仅基于 diff 审查，适合 CI/CD 集成

推荐搭配使用：Webhook 自动捕获每个 PR 做初步审查，开发者对重要 PR 再用 CLI 做深度审查。

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
references		references
LICENSE		LICENSE
README.md		README.md
SKILL.md		SKILL.md
index.ts		index.ts

Folders and files

Latest commit

History

Repository files navigation

🔍 Smart PR Review

Why Another Code Review Tool?

Overview

Features

Quick Start

As a Claude Code Skill (CLI)

As a Webhook Service

Commands

/review [PR-URL]

/review --diff

/review --commit=<hash>

Parameters

The 6-Layer Review

Layer 1: Logic Correctness 🧠

Layer 2: Boundary Conditions 🔲

Layer 3: Performance ⚡

Layer 4: Security 🔒

Layer 5: Maintainability 🔧

Layer 6: Architecture 🏗️

Devil's Advocate Mode

Output Format

⚠️ SHOULD FIX (2 issues)

💡 SUGGESTION (1 issue)

✅ What's Good

📊 Verdict

Setup

Environment Variables

Endpoints

Supported Languages

vs GitHub Copilot Code Review

Architecture

Contributing

License

中文说明

简介

核心差异

6 层审查维度

使用方式

双模式架构

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

`/review [PR-URL]`

`/review --diff`

`/review --commit=<hash>`

Packages