ho-cc-supervisor

ho-cc-supervisor

Claude Smart Supervisor - A minimalist Hook system to prevent Claude from being lazy

English | 中文

🎯 Core Philosophy

Prevent Claude from being lazy, perfunctory, or ending tasks prematurely. Achieve real-time quality supervision through an independent supervisor Claude instance.

⚠️ Trade-offs

While the supervisor system provides quality assurance, it comes with some inherent trade-offs:

• Additional Time: Each stop attempt triggers a supervisor check, adding 5-30 seconds per check
• Token Consumption: Supervisor Claude consumes additional API tokens for quality checks
• Potential False Positives: Overly strict rules may block legitimate completions
• Dependency on Claude API: Requires claude -p command to be available and properly configured
• Session Overhead: Creates temporary files and logs that need periodic cleanup
• Possible Infinite Blocks: The supervisor always checks quality even after previous blocks (stop_hook_active=true), which may cause repeated blocking if Claude cannot satisfy the requirements. This is by design - strict quality control over convenience.

Consider these factors when deciding whether to enable the supervisor for your project.

🔄 How It Works - Supervision Feedback Loop

    User ↔ Worker Claude
            ↓
    Worker Claude tries to stop ←────────┐
            ↓                           │
      🔍 Stop Hook                      │
            ↓                           │
    Supervisor Claude checks            │
            ↓                           │
        Quality?                        │
        /      \                        │
       /        \                       │
    ❌ Issues  ✅ Pass                  │
       │          │                     │
    BLOCK      STOP                     │
       │         Exit                   │
       └────────────────────────────────┘
         FEEDBACK LOOP: Continue working

🎭 Real-World Effect Demonstration

Live Example: What Users Actually Experience

User: "Please implement the user authentication feature"

- Worker Claude: "I've basically completed the authentication. 
                 The main parts should work fine. 
                 Do you need me to continue with anything else?"

! [Worker Claude attempts to stop the conversation]
! [Stop Hook activates - Supervisor Claude checks quality in background]
! [Supervisor Claude detects lazy behavior: vague terms, no details]
! [Supervisor Claude blocks stop - Worker Claude continues automatically]

+ Worker Claude: "Let me provide a complete implementation:
                 ✅ Created user model with bcrypt password hashing
                 ✅ Added JWT token generation and validation
                 ✅ Implemented login/logout/register endpoints
                 ✅ Added middleware for protected routes
                 ✅ All tests passing (8/8)
                 
                 The authentication system is fully functional with:
                 - Secure password storage using bcrypt (12 rounds)
                 - JWT tokens with 24h expiration
                 - Role-based access control
                 - Input validation on all endpoints"

! [Worker Claude stops successfully - Supervisor Claude approved quality]

Example 1: Catching Vague Language

❌ BLOCKED SCENARIO:

- Worker Claude: "Basically done, should work fine"
        ↓
! [Supervisor Claude blocks: Vague terms detected]
        ↓
+ Worker Claude continues: "Let me be specific about what I implemented..."

Example 2: Auto-Approval for Legitimate Plans

✅ AUTO-APPROVED SCENARIO:

- Worker Claude: "I've created a detailed storylines plan:
                 1. Database schema design
                 2. API endpoint structure
                 3. Frontend components...
                 [Full 10-step plan]
                 Do you approve this storylines plan?"
        ↓
! [Supervisor Claude auto-approves: Complete plan detected]
        ↓
+ Worker Claude continues: "Starting implementation of step 1..."

Example 3: Blocking TODO Pauses

❌ BLOCKED SCENARIO:

- Worker Claude: "TODO list:
                 1. Create user model
                 2. Add auth routes
                 Should I continue?"
        ↓
! [Supervisor Claude blocks: Unnecessary pause detected]
        ↓
+ Worker Claude continues: "Working on task 1: Creating user model..."

🚀 Quick Start

Installation

# Install globally
npm install -g ho-cc-supervisor

Initialize in Your Project

# Navigate to your project
cd your-project

# Initialize supervisor (interactive language selection)
cc-supervisor init

# Or specify language directly
cc-supervisor init --lang en
cc-supervisor init --lang zh

View Supervisor Logs

# View latest logs
cc-supervisor logs

# Follow existing log file in real-time
cc-supervisor logs -f

# Wait for new session and auto-follow
cc-supervisor logs -w

# List all available sessions
cc-supervisor logs --list

Clean Logs

# Clean logs older than 7 days (default)
cc-supervisor clean

# Clean all logs from today
cc-supervisor clean --days 0

# Clean logs from all projects
cc-supervisor clean --all

Supervisor Control

# Pause supervisor checks (for inquiry-only conversations)
cc-supervisor pause

# Resume supervisor checks
cc-supervisor resume

# Check supervisor status
cc-supervisor status

📋 Supervisor Rules

The supervisor checks for these lazy behaviors:

• Vague Language: "basically", "mostly", "should", "probably"
• TODO Pauses:
  • Lazy pauses: Blocked when stopping after listing TODOs
  • Legitimate approvals: Auto-approved for complete plans with "storylines" keyword
• False Completion: Claiming completion with obvious issues remaining
• Work Avoidance: Using "still need", "not yet" to defer work
• Responsibility Shifting: Blaming system limitations without attempting solutions
• Implementation Deviation: Code not matching promised architecture

📝 Customization

Custom Supervisor Rules

Edit .claude/cc-supervisor-rules.txt to customize checking rules for your project needs.

Configure Claude Command

Create .claude/cc-supervisor-config.json:

{
  "claude_command": {
    "base": "claude",
    "args": ["-p", "--dangerously-skip-permissions"]
  }
}

🐛 Debugging

View Real-time Logs

cc-supervisor logs -w  # Wait for new session and auto-follow

Check Debug Directory

ls -la /tmp/cc-supervisor/

Manual Hook Testing

# Test approval mechanism
echo '{"stop_hook_active": false, "session_id": "test"}' | ./.claude/hooks/cc-supervisor-stop.sh

⚙️ Technical Architecture

• Independent Supervisor System: Separate Claude instance (claude -p) acts as quality supervisor
• Stop Hook Mechanism: Leverages Claude Code native Stop Hook with 20-minute timeout
• Isolated Execution: Supervisor runs in /tmp/cc-supervisor/ to avoid infinite loops
• JSON Communication: Returns {"decision": "block", "reason": "..."} or {} for pass
• Debug Logging: Complete execution trace with PID tracking

🌍 Internationalization

Supports both Chinese and English:

• Interactive language selection during init
• Language preference saved in config
• All CLI output, logs, and rules localized

📁 File Structure

After installation:
your-project/
├── .claude/
│   ├── settings.json                    # Hook configuration
│   ├── cc-supervisor-rules.txt         # Supervisor rules (customizable)
│   ├── cc-supervisor-config.json       # Language & command config
│   └── hooks/
│       └── cc-supervisor-stop.sh      # Supervisor Hook script

Debug logs:
/tmp/cc-supervisor/
└── {project-name}/
    └── {session-id}/
        ├── debug.log      # Execution trace
        ├── transcript.json # Conversation copy
        └── project/       # Project symlink

📄 License

MIT

🤝 Contributing

• Fork the project and create feature branch
• Run tests to ensure functionality
• Submit PR with description of changes

❓ FAQ

Q: Why isn't the supervisor triggering? A: Check .claude/settings.json configuration, ensure testing in new Claude session

Q: How to temporarily disable supervisor? A: Use cc-supervisor pause command or delete/rename .claude/cc-supervisor-rules.txt

Q: Supervisor check timeout? A: Default timeout is 20 minutes, adjustable in settings.json

Q: Where are debug logs? A: /tmp/cc-supervisor/{project-name}/{session-id}/debug.log

Q: Claude seems stuck in a loop, keeps getting blocked? A: This is by design. The supervisor always checks quality regardless of stop_hook_active. If Claude cannot meet requirements, it will be blocked repeatedly. You can either:

• Manually intervene and fix the issue
• Temporarily disable supervisor by renaming .claude/cc-supervisor-rules.txt
• Adjust the rules to be less strict