Aireview

aireview - Enhanced AI Code Review

Pseudo Multi-Agent Protocol (Codex)

Codex does not support native subagents. Simulate role handoffs with explicit sections.

Required sections (in order):

[ANALYST] Summary + constraints
[ARCHITECT] Approach + components
[EXECUTOR] Actions + changes
[REVIEWER] Verification + risks

> Codex invocation: use aireview: ... or $aireview ...

Professional code review combining multi-agent parallel analysis with direct CLI invocations.

Language Configuration

始终使用中文回复: All AI review results will be presented in Chinese.

Quick Usage

# Review changes (most common)
aireview --diff                           # Review unstaged changes
aireview --diff --staged                  # Review staged changes
aireview --diff HEAD~1..HEAD              # Review last commit

# Remote branch/PR review (NEW)
aireview origin/feature-branch            # Review remote branch
aireview origin/feature-branch --quick    # Quick review (small PR)
aireview origin/feature-branch --deep     # Deep review (large PR)

# GitHub/GitLab PR review (if in PR context)
aireview --pr [PR_NUMBER]                 # Review specific PR
aireview --pr --comment                   # Post review as PR comment

# Deep review (Multi-AI collaboration)
aireview --diff --deep                    # Gemini + Codex parallel analysis

# Standard file/directory review
aireview ./src/auth
aireview ./src/auth/service.ts --model codex

Enhanced Architecture

Multi-Agent Parallel Review (Inspired by Official Plugin)

aireview --diff
    ↓
Step 1: Eligibility Check (Haiku Agent)
    - Check if review needed
    - Skip if: closed PR, draft, trivial change, already reviewed
    ↓
Step 2: Gather Context
    Invoke agent `oh-my-antigravity :explore` (model: `haiku`) with prompt: "Get change summary for: {diff_content}"
    - Get change summary
    ↓
Step 3: Parallel Review (5 Specialized Agents)
    ┌─────────────────────────────────────────────┐
    Invoke agent `oh-my-antigravity :architect` (model: `sonnet`) with prompt: "Check Best Practices Compliance from {diff_content}"
    Invoke agent `oh-my-antigravity :qa-tester` (model: `sonnet`) with prompt: "Bug Detection Scan on {diff_content}"
    Invoke agent `oh-my-antigravity :explore` (model: `sonnet`) with prompt: "Git History Context Analysis"
    Invoke agent `oh-my-antigravity :researcher` (model: `sonnet`) with prompt: "Related PR Analysis"
    Invoke agent `oh-my-antigravity :code-reviewer` (model: `sonnet`) with prompt: "Code Comment Compliance"
    ↓
Step 4: Confidence Scoring (Parallel Haiku Agents)
    - Score each issue 0-100
    - Filter issues  /tmp/pr_diff.txt
  Execute: git log origin/master..$REMOTE_BRANCH --oneline
  
  DIFF_CONTENT = read /tmp/pr_diff.txt
  COMMIT_LOG = git log output
  
  # For remote branch, use simplified context gathering

Else:
  Invoke agent `oh-my-antigravity :explore` (model: `haiku`) with prompt: "Get change summary for current diff"
  1. Get change summary:
     - Files modified
     - Lines changed
     - Overall purpose

Return: {summary: "", diff_content: ""}

Step 3: Launch Parallel Review Agents

If MODE == "quick":
  # Quick mode: Skip parallel agents, use single fast Gemini review
  
  Execute quick review via Claude Code CLI:
  
  cat > /tmp/quick_review_prompt.txt &1
  
  Skip to Step 6 (Generate Report)

Else:
  Launch 5 parallel agents using Antigravity delegation:

  # Agent 1: Best Practices Compliance (Architect/Sonnet)
  Invoke agent `oh-my-antigravity :architect-medium` (model: `sonnet`) with prompt:
    你是 INTJ 架构师，专注于规范合规性审查。
    ...

  # Agent 2: Bug Detection (QA-Tester/Sonnet)
  Invoke agent `oh-my-antigravity :qa-tester` (model: `sonnet`) with prompt:
    你是 ISTJ 工程师，专注于 Bug 检测。
    ...

  # Agent 3: Git History (Explore/Sonnet)
  Invoke agent `oh-my-antigravity :explore-medium` (model: `sonnet`) with prompt:
    你是 INTP 性能极客，专注于历史上下文分析。
    ...

  # Agent 4: Related PR Analysis (Researcher/Sonnet)
  Invoke agent `oh-my-antigravity :researcher` (model: `sonnet`) with prompt:
     你是 ENTP 创新者，专注于 PR 关联分析。
     ...

  # Agent 5: Code Comment Compliance (Code-Reviewer/Sonnet)
  Invoke agent `oh-my-antigravity :code-reviewer` (model: `sonnet`) with prompt:
     你是 ISFJ 维护者，专注于代码注释合规性。
     ...

Step 4: Confidence Scoring & Filtering

For each issue from Step 3, launch a Haiku agent:

Invoke agent `oh-my-antigravity :architect-low` (model: `haiku`) with prompt:
  你是专业的代码审查评分员。

  问题描述：{issue}
  代码变更：{diff_content}

  评分标准 (0-100)：
  - 0: 完全不确定，明显的误报
  - 25: 有点怀疑，可能是误报
  - 50: 中等确信，已验证是真实问题，但不是很重要
  - 75: 高度确信，双重检查过，会影响功能
  - 100: 绝对确定，确认是真实问题，经常发生

  False Positive 规则（降低分数）：
  - 预存在的问题
  - 看起来像 bug 但实际不是
  - 吹毛求疵的小问题
  - linter/typechecker 会捕获的问题
  - 缺少测试覆盖率
  - 可能是有意的功能变更
  - 真实问题，但在用户未修改的行上

  输出：最终置信度分数 (0-100)

Filter: 只保留置信度 >= 80 的问题

Step 5: Deep Mode (Optional)

If MODE == "deep":

  Layer 2 - Claude Code Architecture Analysis:
    Use Skill tool with 'code-review' skill (standard mode)
    Or direct CLI:

    cat > /tmp/claude_prompt.txt &1

  Layer 3 - Codex Quality Audit:
    cat > /tmp/codex_prompt.txt &1

  Layer 4 - Synthesis:
    Combine findings from all layers
    Recalculate confidence scores
    Generate comprehensive report

Step 6: Generate Report

Format based on mode:

If filtered_issues.length == 0:
  Output: |
    ## AI 审查报告
    **模式**: {MODE}
    **模型**: {models_used}

    未发现问题。已检查：
    - Bugs
    - 最佳实践合规性
    - 历史上下文
    - 相关 PR
    - 代码注释合规性

    🤖 Generated with Antigravity

Else:
  Output: |
    ## AI 审查报告
    **模式**: {MODE}
    **模型**: {models_used}
    **文件**: {files_count}个文件变更，+{lines_added}/-{lines_removed}行

    发现 {issues_count} 个高置信度问题：

    {for each issue}
    ### {severity} - {issue.title}

    **置信度**: {issue.confidence}/100
    **来源**: {issue.agent} ({issue.persona})
    **文件**: `{issue.file}:{issue.line}`

    {issue.description}

    **修复建议**:
    {issue.fix_suggestion}

    **代码链接**: {github_link_with_full_sha}

    {end for}

    ---

    ## 总结
    - **总体评分**: {overall_score}/10
    - **严重问题**: {critical_count}个 (立即修复)
    - **高优先级问题**: {high_count}个 (本周修复)
    - **中等问题**: {medium_count}个 (本月修复)

    {if deep_mode}
    ### 深度分析
    **Claude Code 架构评分**: {claude_score}/10
    {claude_highlights}

    **Codex 质量评分**: {codex_score}/10
    {codex_highlights}
    {end if}

    🤖 Generated with Antigravity

    {if pr_mode}
    - 如果这个审查有帮助，请回复 👍。否则，回复 👎。
    {end if}

If --comment flag && pr_mode:
  Use gh pr comment to post the report

MBTI Persona Mapping

| Agent Role | MBTI Persona | Focus Area | |-----------|--------------|------------| | Best Practices Compliance | INTJ 架构师 | 规范遵守、系统设计 | | Bug Detection | ISTJ 工程师 | 细节、逻辑错误 | | Git History | INTP 性能极客 | 模式、历史上下文 | | Related PR | ENTP 创新者 | 关联性、创新视角 | | Code Comments | ISFJ 维护者 | 注释、文档一致性 |

Confidence Score Rubric

| Score | Meaning | Example | |-------|---------|---------| | 0 | 完全不确定 | 明显误报、预存在问题 | | 25 | 有点怀疑 | 可能是问题，但未验证 | | 50 | 中等确信 | 真实问题，但不重要 | | 75 | 高度确信 | 影响功能 | | 100 | 绝对确定 | 确认的真实问题，经常发生 |

过滤阈值: 80+ (只展示高度确信的问题)

False Positive Rules

不要标记以下情况 (降低置信度分数):

1. 预存在的问题 - 在变更前就存在
2. 工具可捕获 - linter、typechecker、compiler 会发现
3. 样式问题 - 格式、命名
4. 吹毛求疵 - 资深工程师不会提的小问题
5. 有意变更 - 功能变更可能是有意的
6. 未修改行 - 问题在用户未修改的代码行
7. 缺少测试 - 除非明确要求测试覆盖率

Core Advantages

From Official Plugin

• ✅ 多 agent 并行审查 (5 agents)
• ✅ 置信度评分系统 (0-100)
• ✅ False positive 过滤 (阈值 80)
• ✅ 一般最佳实践检查
• ✅ GitHub PR 集成

From Original aireview

• ✅ 直接 CLI 调用 (无超时)
• ✅ claude -p 优先
• ✅ MBTI 人格系统
• ✅ Deep mode (多模型协作)
• ✅ 中文输出

Enhanced Features

• ✅ Git history 上下文分析
• ✅ 相关 PR 分析
• ✅ 代码注释合规性检查
• ✅ 分层分析 (multi-agent → Claude Code → Codex → synthesis)
• ✅ 详���的置信度评分标准

CLI Implementation Templates

Claude Code CLI (优先)

cat > /tmp/claude_prompt.txt &1

Claude Code CLI (备用)

cat /tmp/claude_prompt.txt | claude -p --output-format text 2>&1

Codex CLI

cat > /tmp/codex_prompt.txt &1

Requirements

• Claude Code CLI (claude -p) configured
• Codex CLI configured
• GitHub CLI (gh) for PR mode

Usage Examples

Example 1: Standard Review

$ aireview --diff

🔍 正在执行多 agent 并行审查...

✅ Step 1: 资格检查通过
✅ Step 3: 启动 5 个并行审查 agents
✅ Step 4: 置信度评分完成，过滤后保留 3 个问题

## AI 审查报告
...

Example 2: Remote Branch Quick Review (NEW)

$ aireview origin/feature-cr

🔍 正在获取远程分支信息...
✅ 已获取远程更新
📊 变更统计: 1 file, +1/-1 lines
🎯 自动选择: quick mode (小型 PR)

🔍 正在执行快速审查...

## ⚡ 快速审查报告

**PR**: `feature-cr` → `master`
**提交**: 2338116 - "docs(BDD_FOR_ELECTRON): 更新测试案例说明"
**变更**: 1 file, +1/-1 lines

---

### ❌ 阻止合并

**问题: 非专业用语**
- 文件: `BDD_FOR_ELECTRON.md:2`
- 问题: "这是给王大爷的测试案例" 不符合技术文档规范
- 修复: 移除此行或改为专业描述

**建议修复**:
```markdown
# Electron 项目的 BDD 完整实施方案

> 本文档提供 Electron 桌面应用的 BDD 完整实施指南

---

✅ 合并建议

• [ ] 可以合并
• [x] 修复后合并 (移除非专业用语)
• [ ] 不建议合并

---

审查模型: Claude Code (claude -p) 审查时间: 2025-12-29 14:00 耗时: < 30 秒

🤖 Generated with Antigravity

### Example 3: Remote Branch Deep Review
```bash
$ aireview origin/feature-payment --deep

🔍 正在获取远程分支信息...
✅ 已获取远程更新
📊 变更统计: 15 files, +450/-120 lines
🎯 自动选择: deep mode (大型 PR)

🔍 正在执行深度多模型审查...

✅ Layer 1: 多 agent 并行审查 (5 agents)
✅ Layer 2: Claude Code 架构分析 (INTJ)
✅ Layer 3: Codex 质量审计 (ISTJ)
✅ Layer 4: 综合分析

## 深度 AI 审查报告
...

Example 4: Deep Review (Original)

$ aireview --diff --deep

🔍 正在执行深度多模型审查...

✅ Layer 1: 多 agent 并行审查 (5 agents)
✅ Layer 2: Claude Code 架构分析 (INTJ)
✅ Layer 3: Codex 质量审计 (ISTJ)
✅ Layer 4: 综合分析

## 深度 AI 审查报告
...

Example 5: PR Review with Comment

$ aireview --pr 123 --comment

🔍 正在审查 PR #123...

✅ 审查完成，发布评论到 PR

---

核心创新:

1. 结合官方的多 agent 并行架构
2. 保留原有的 CLI 直接调用优势
3. 引入置信度评分和 false positive 过滤
4. 增强的 MBTI 人格专业化
5. 分层深度分析 (multi-agent → Gemini → Codex)

Output

• Produce a concrete deliverable in markdown aligned with the workflow/skill goal.
• Include key decisions, actions taken, and final status for Antigravity IDE visibility.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: materialofair
• Source: materialofair/oh-my-antigravity
• License: MIT
• Homepage: https://github.com/materialofair/oh-my-antigravity

Install and usage instructions live in the source repository linked above.