AI Collaboration System

🔄 BIDIRECTIONAL autonomous AI collaboration - Claude Code + GitHub Copilot working together

A file-based task queue system that enables Claude Code and GitHub Copilot to work together autonomously on your codebase with mutual code review and iterative improvements. Both AIs can execute tasks AND review each other's work!

Features

• 🔄 Bidirectional Collaboration - Both AIs execute AND review each other's work
• 🤖 Autonomous Execution - Copilot executes tasks automatically without manual intervention
• 🔍 Dual Code Review - Claude reviews Copilot's work, Copilot reviews Claude's work
• ♻️ Iterative Improvement - Auto-creates follow-up tasks until BOTH AIs approve the code
• 📋 Task Queue System - Simple JSON-based task queue for coordination
• 🛡️ Quality Assurance - Catches syntax errors, missing imports, and code quality issues
• ⚙️ Configurable - Adjust review strictness and workflow to your needs
• 🚫 Loop Prevention - Built-in safeguards against infinite task loops

How It Works

Copilot Flow:

You create task → Copilot executes → Claude reviews →
  ✅ Approved → Done!
  ⚠️ Issues → Creates fix tasks → Copilot fixes → Review again

Claude Flow:

You create task → Claude executes → Copilot reviews →
  ✅ Approved → Done!
  ⚠️ Issues → Creates fix tasks → Claude fixes → Review again

Both AIs collaborate and iterate until the code meets quality standards! ✅

Requirements

• Node.js 16+
• GitHub Copilot CLI installed and authenticated
• Claude Code (for code review)
• Git repository (for change detection)

Installation

# Install globally
npm install -g @ai-collab/task-system

# Or install in your project
npm install --save-dev @ai-collab/task-system

Quick Start

1. Initialize in your project:

cd your-project
ai-collab init

This creates:

• .ai-workspace/ directory
• tasks.json - task queue file
• Configuration files
• Example tasks

2. Start the watcher:

ai-collab start
# or
npm run ai-collab

3. Create tasks for Copilot:

Edit .ai-workspace/tasks.json or use the helper:

ai-collab create-task --title "Add login page" --assignedTo copilot

4. Watch the magic happen!

Copilot executes → Claude reviews → Iterates until approved

Usage

Creating Tasks

Via JSON file:

Edit .ai-workspace/tasks.json:

{
  "tasks": [
    {
      "id": "task-001",
      "title": "Build user dashboard",
      "description": "Create a modern dashboard page at app/dashboard/page.tsx...",
      "assignedTo": "copilot",
      "status": "pending",
      "priority": "high",
      "files": ["app/dashboard/page.tsx"]
    }
  ]
}

Via CLI:

ai-collab create-task \
  --title "Add authentication" \
  --description "Implement JWT auth with login page" \
  --assignedTo copilot \
  --files "app/lib/auth.ts,app/login/page.tsx"

Task Assignment (Bidirectional)

• assignedTo: "copilot" - Copilot executes automatically → Claude reviews
• assignedTo: "claude" - You execute via Claude Code CLI → Copilot reviews
• assignedTo: "any" - Either AI can handle it

Key Point: Regardless of who executes, the OTHER AI will review! This creates a bidirectional quality assurance system.

Task Statuses

• pending - Not started
• in-progress - Currently being worked on
• completed - Done, awaiting review
• verified - Approved by Claude review
• blocked - Has errors, needs fixing
• cancelled - No longer needed

Configuration

Review Strictness

Edit .ai-workspace/config.json:

{
  "review": {
    "checkSyntax": true,
    "checkQuality": true,
    "failOnConsoleLog": false,
    "requireExports": true,
    "maxFollowUpTasks": 5,
    "delayBeforeReview": 2000
  },
  "claudeReviewer": {
    "enabled": true,
    "reviewCopilotTasks": true,
    "createFollowUpTasks": true
  },
  "copilotReviewer": {
    "enabled": true,
    "reviewClaudeTasks": true,
    "createFollowUpTasks": true,
    "detailedReview": true
  },
  "watcher": {
    "interval": 2000,
    "enableAutoReview": true,
    "enableBidirectionalReview": true
  }
}

Custom Review Rules

Create .ai-workspace/review-rules.js:

module.exports = {
  customChecks: [
    {
      name: "No TODO comments",
      check: (content) => !content.includes("TODO"),
      message: "Remove TODO comments before completion"
    }
  ]
};

🔄 Bidirectional Workflow

How Bidirectional Review Works

1. Copilot executes → Claude reviews → Copilot fixes → Repeat until approved
2. Claude executes → Copilot reviews → Claude fixes → Repeat until approved

Example: Claude Task with Copilot Review

Step 1: Create a task for Claude:

{
  "id": "task-001",
  "title": "Implement user authentication",
  "description": "Create JWT-based authentication with login/logout",
  "assignedTo": "claude",
  "status": "pending",
  "files": ["lib/auth.ts"]
}

Step 2: Execute the task using Claude Code CLI

Step 3: Mark the task as completed in tasks.json:

{
  "id": "task-001",
  "status": "completed",
  "result": "Authentication implemented with JWT tokens"
}

Step 4: Copilot automatically reviews Claude's work!

• ✅ If approved: Task marked as verified
• ⚠️ If issues found: Copilot creates follow-up tasks for Claude to fix

Step 5: Fix any issues and repeat until Copilot approves

Example: Copilot Task with Claude Review

Step 1: Create a task for Copilot:

{
  "id": "task-002",
  "title": "Create login form component",
  "description": "Build a React login form with email and password",
  "assignedTo": "copilot",
  "status": "pending",
  "files": ["components/LoginForm.tsx"]
}

Step 2: Copilot automatically executes the task

Step 3: Claude automatically reviews Copilot's work!

• ✅ If approved: Task marked as verified
• ⚠️ If issues found: Claude creates follow-up tasks for Copilot to fix

Step 4: Copilot automatically fixes issues until Claude approves

Benefits of Bidirectional Review

• 📊 Double Quality Check - Two different AI perspectives on every piece of code
• 🔍 Complementary Strengths - Claude catches what Copilot misses and vice versa
• 🎯 Reduced Errors - Higher confidence in code quality
• 🔄 Continuous Improvement - Both AIs learn from each other's feedback
• ⚡ Flexible Workflow - Choose the best AI for each specific task

Examples

Example 1: Build a Feature

{
  "id": "task-001",
  "title": "Create user profile page",
  "description": "Build a user profile page with avatar, bio, and settings",
  "assignedTo": "copilot",
  "status": "pending",
  "files": ["app/profile/page.tsx"]
}

Result:

• Copilot creates the page
• Claude reviews for errors
• If issues found, creates fix tasks
• Copilot fixes issues
• Repeats until approved ✅

Example 2: Add API Endpoint

{
  "id": "task-002",
  "title": "Add user API endpoint",
  "description": "Create GET /api/users endpoint that returns user list",
  "assignedTo": "copilot",
  "status": "pending",
  "files": ["app/api/users/route.ts"]
}

Example 3: Refactoring

{
  "id": "task-003",
  "title": "Extract auth logic to utility",
  "description": "Move authentication logic from pages to app/lib/auth.ts",
  "assignedTo": "copilot",
  "status": "pending",
  "files": ["app/lib/auth.ts", "app/login/page.tsx"]
}

Advanced Usage

Pausing the Watcher

Create .ai-workspace/STOP file:

touch .ai-workspace/STOP

Remove to resume:

rm .ai-workspace/STOP

Monitoring

Check status:

ai-collab status

View logs:

ai-collab logs

Task Dependencies

Tasks can depend on others:

{
  "id": "task-002",
  "title": "Add styling to dashboard",
  "dependencies": ["task-001"],
  "status": "pending"
}

Task-002 waits until task-001 is completed.

Troubleshooting

Copilot Not Executing Tasks

1. Check Copilot is installed: copilot --version
2. Check authentication: copilot -p "test"
3. Verify permissions in .ai-workspace/copilot-agent.js

Review False Positives

If Claude flags valid code:

1. Check review notes in tasks.json
2. Adjust strictness in config
3. Manually mark as verified if incorrect

Infinite Task Loops

The system has built-in loop prevention (max 5 attempts). If it triggers:

1. Check Copilot permissions
2. Review the error pattern in task history
3. Manually fix the issue

File Structure

your-project/
├── .ai-workspace/
│   ├── tasks.json              # Task queue
│   ├── auto-watcher.js         # Main orchestrator (bidirectional)
│   ├── copilot-agent.js        # Copilot execution wrapper
│   ├── claude-reviewer.js      # Claude reviews Copilot's work
│   ├── copilot-reviewer.js     # Copilot reviews Claude's work (NEW!)
│   ├── config.json             # Configuration
│   ├── claude-output.json      # Execution logs
│   └── STOP                    # Pause file (optional)

API

CopilotAgent

const CopilotAgent = require('.ai-workspace/copilot-agent');
const agent = new CopilotAgent({ workingDir: process.cwd() });

await agent.executeTask('Create a hello world function');

ClaudeReviewer

Reviews Copilot's completed tasks:

const ClaudeReviewer = require('.ai-workspace/claude-reviewer');
const reviewer = new ClaudeReviewer(workingDir, tasksFile);

const review = await reviewer.reviewTask(task);
console.log(review.approved); // true/false
console.log(review.issues);   // Array of issues found
console.log(review.suggestions); // Array of suggestions

CopilotReviewer (NEW!)

Reviews Claude's completed tasks:

const CopilotReviewer = require('.ai-workspace/copilot-reviewer');
const reviewer = new CopilotReviewer(workingDir, tasksFile);

const review = await reviewer.reviewTask(task);
console.log(review.approved); // true/false
console.log(review.issues);   // Array of issues found
console.log(review.suggestions); // Array of suggestions

Best Practices

1. Be Specific - Write detailed task descriptions
2. Small Tasks - Break large features into smaller tasks
3. Review Suggestions - Even approved tasks may have suggestions
4. Monitor First - Watch a few cycles before leaving it unattended
5. Git Commits - Review changes with git diff before committing

Contributing

Contributions welcome! Please open an issue or PR.

License

MIT

Links

• GitHub Repository
• Documentation
• GitHub Copilot
• Claude Code

Support

• GitHub Issues: github.com/yourusername/ai-collab-system/issues
• Discussions: github.com/yourusername/ai-collab-system/discussions