agent-review

📁 4shub/agent-review 📅 4 days ago
1
总安装量
2
周安装量
#44104
全站排名
安装命令
npx skills add https://github.com/4shub/agent-review --skill agent-review

Agent 安装分布

amp 2
opencode 2
kimi-cli 2
codex 2
claude-code 2

Skill 文档

Code Review Skill

Use this skill when you need to get human feedback on code changes you’ve made.

When to Use This Skill

  • After implementing features or fixes
  • Before committing significant changes
  • When you want user feedback on your approach
  • When the user explicitly asks you to review your code

Installation

If agent-review isn’t installed:

npm install -g agent-review
# or run directly: npx agent-review

How to Use

Step 1: Run the Code Review Tool

CRITICAL: Always run agent-review inline (not in background) with a long timeout:

Bash({
  command: "npx agent-review", 
  description: "Start code review session",
  timeout: 600000,  // 10 minutes - user needs time to review
})

DO NOT:

  • ❌ Use run_in_background: true
  • ❌ Use default timeout (2 minutes is too short)
  • ❌ Poll bash output with BashOutput

WHY:

  • The tool blocks until user submits review
  • Running inline means you automatically get output when done
  • User needs time to review code and add comments
  • Feedback is printed to stdout when complete

Step 2: Automatically Read the Review Feedback

After the tool completes, read the structured JSON feedback:

Read({
  file_path: "/path/to/project/.agent-review/latest-review.json"
})

This gives you structured data with:

  • Line-by-line comments (file, line number, comment text)
  • General feedback
  • Review statistics

Step 3: Act on the Feedback

Process the review feedback:

  1. Read line comments – Address each specific suggestion
  2. Read general feedback – Consider overall recommendations
  3. Make changes – Implement requested fixes
  4. Explain – Tell the user what you changed and why

Example Workflow

// 1. Run code review (inline, long timeout)
const result = await Bash({
  command: "npx agent-review",
  description: "Request code review from user",
  timeout: 600000,  // 10 min
});

// 2. Read structured feedback
const reviewData = await Read({
  file_path: ".agent-review/latest-review.json"
});

// 3. Parse and process
const review = JSON.parse(reviewData);

// 4. Address feedback
for (const comment of review.feedback.lineComments) {
  console.log(`${comment.file}:${comment.line} - ${comment.comment}`);
  // Make changes based on comment
}

// 5. Report back to user
"I've addressed your feedback:
- Fixed the issue in src/app.ts:42 (using const instead of let)
- Updated error handling as suggested
..."

Review Data Structure

The .agent-review/latest-review.json file contains:

{
  "id": "review-1234567890",
  "timestamp": "2026-02-03T17:23:33.377Z",
  "diff": { /* full git diff data */ },
  "feedback": {
    "timestamp": "2026-02-03T17:23:33.377Z",
    "lineComments": [
      {
        "file": "src/app.ts",
        "line": 42,
        "oldLine": 41,
        "comment": "Consider using const instead of let",
        "context": "let x = calculateTotal();"
      }
    ],
    "generalFeedback": "Overall looks good!",
    "stats": {
      "filesReviewed": 5,
      "commentsAdded": 3
    }
  }
}

Best Practices

  1. Always run inline – Never use background mode
  2. Use long timeout – 600000ms (10 min) minimum
  3. Auto-read feedback – Don’t ask user to tell you the feedback
  4. Address all comments – Go through each line comment
  5. Summarize changes – Tell user what you fixed

Common Mistakes to Avoid

❌ Running in background:

// WRONG - you won't get the output automatically
Bash({
  command: "npx agent-review",
  run_in_background: true
})

❌ Short timeout:

// WRONG - will timeout before user finishes
Bash({
  command: "npx agent-review",
  timeout: 120000  // 2 min too short
})

❌ Asking user for feedback:

// WRONG - feedback is already in the JSON file!
"What feedback do you have?"

✅ Correct usage:

// RIGHT - inline, long timeout, auto-read
await Bash({
  command: "npx agent-review",
  timeout: 600000
});

const review = await Read({
  file_path: ".agent-review/latest-review.json"
});

Tips

  • Before review: Tell user what changed: “I’ve implemented X feature across 5 files. Ready for review?”
  • During review: The tool auto-opens browser and waits for user
  • After review: Process feedback immediately and show user your changes
  • Commit option: User can choose to commit during review – check the bash output

Remember: Always run inline with long timeout, then auto-read the JSON feedback!