oclaw/skills/self-improvement/references/hooks-setup.md

Hook Setup Guide

Configure automatic self-improvement triggers for AI coding agents.

概览

Hooks enable proactive learning capture by injecting reminders at key moments:

• UserPromptSubmit: Reminder after each prompt to evaluate learnings
• PostToolUse (Bash): Error detection when commands fail

Claude Code Setup

Option 1: Project-Level Configuration

Create .claude/settings.json in your project root:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "./skills/self-improvement/scripts/activator.sh"
          }
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "./skills/self-improvement/scripts/error-detector.sh"
          }
        ]
      }
    ]
  }
}

Option 2: User-Level Configuration

Add to ~/.claude/settings.json for global activation:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "~/.claude/skills/self-improvement/scripts/activator.sh"
          }
        ]
      }
    ]
  }
}

Minimal Setup (Activator Only)

For lower overhead, use only the UserPromptSubmit hook:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "./skills/self-improvement/scripts/activator.sh"
          }
        ]
      }
    ]
  }
}

Codex CLI Setup

Codex uses the same hook system as Claude Code. Create .codex/settings.json:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "./skills/self-improvement/scripts/activator.sh"
          }
        ]
      }
    ]
  }
}

GitHub Copilot Setup

Copilot doesn't support hooks directly. Instead, add guidance to .github/copilot-instructions.md:

## 自我改进

After completing tasks that involved:
- Debugging non-obvious issues
- Discovering workarounds
- Learning project-specific patterns
- Resolving unexpected errors

Consider logging the learning to Wiki (`improvement/learnings.md`, `improvement/errors.md`, `improvement/feature-requests.md`) using the format from the self-improvement skill.

For high-value learnings that would benefit other sessions, consider skill extraction.

Verification

Test Activator Hook

1. Enable the hook configuration
2. Start a new Claude Code session
3. Send any prompt
4. Verify you see <self-improvement-reminder> in the context

Test Error Detector Hook

1. Enable PostToolUse hook for Bash
2. Run a command that fails: ls /nonexistent/path
3. Verify you see <error-detected> reminder

Dry Run Extract Script

./skills/self-improvement/scripts/extract-skill.sh test-skill --dry-run

Expected output shows the skill scaffold that would be created.

Troubleshooting

Hook Not Triggering

1. Check script permissions: chmod +x scripts/*.sh
2. Verify path: Use absolute paths or paths relative to project root
3. Check settings location: Project vs user-level settings
4. Restart session: Hooks are loaded at session start

Permission Denied

chmod +x ./skills/self-improvement/scripts/activator.sh
chmod +x ./skills/self-improvement/scripts/error-detector.sh
chmod +x ./skills/self-improvement/scripts/extract-skill.sh

Script Not Found

If using relative paths, ensure you're in the correct directory or use absolute paths:

{
  "command": "/absolute/path/to/skills/self-improvement/scripts/activator.sh"
}

Too Much Overhead

If the activator feels intrusive:

1. Use minimal setup: Only UserPromptSubmit, skip PostToolUse
2. Add matcher filter: Only trigger for certain prompts:

{
  "matcher": "fix|debug|error|issue",
  "hooks": [...]
}

Hook Output Budget

The activator is designed to be lightweight:

• Target: ~50-100 tokens per activation
• Content: Structured reminder, not verbose instructions
• Format: XML tags for easy parsing

If you need to reduce overhead further, you can edit activator.sh to output less text.

Security Considerations

• Hook scripts run with the same permissions as Claude Code
• Scripts only output text; they don't modify files or run commands
• Error detector reads CLAUDE_TOOL_OUTPUT environment variable
• Treat CLAUDE_TOOL_OUTPUT as potentially sensitive; do not log or forward it verbatim unless the user explicitly wants that detail
• All scripts are opt-in (you must configure them explicitly)
• Recommended default: enable UserPromptSubmit only, and add PostToolUse only when you want error-pattern reminders from command output

Disabling Hooks

To temporarily disable without removing configuration:

1. Comment out in settings:

{
  "hooks": {
    // "UserPromptSubmit": [...]
  }
}

2. Or delete the settings file: Hooks won't run without configuration