---
summary: 'Claude Code pre-command hooks for git safety'
read_when:
- Setting up git protection for AI agents
- Debugging blocked git commands
- Understanding hook behavior
---
# Claude Code Git Protection Hooks
This document describes the pre-command hooks that prevent AI agents (Claude Code, etc.) from executing destructive git commands.
## Overview
Claude Code supports `PreToolUse` hooks that intercept tool calls before execution. We use this to enforce git safety policies, preventing agents from accidentally destroying work with commands like `git reset --hard`.
## Architecture
**Single-layer protection:**
1. **Universal block**: `git reset --hard` is ALWAYS blocked for AI agents, regardless of project
## Installation
The hook is already installed in this project. For reference or to reinstall:
```bash
# Create hook directory
mkdir -p .claude/hooks
# Create the pre-command hook
cat > .claude/hooks/pre_bash.py << 'EOF'
#!/usr/bin/env python3
import json
import sys
import re
import os
try:
data = json.load(sys.stdin)
cmd = data.get("tool_input", {}).get("command", "")
# ALWAYS block git reset --hard, regardless of project
if re.search(r'\bgit\s+reset\s+--hard\b', cmd):
print("BLOCKED: git reset --hard is NEVER allowed for AI agents", file=sys.stderr)
print(f"Attempted: {cmd}", file=sys.stderr)
print("Only the user can run this command directly.", file=sys.stderr)
sys.exit(2)
sys.exit(0)
except:
sys.exit(0)
EOF
chmod +x .claude/hooks/pre_bash.py
# Configure Claude Code to use the hook
cat > .claude/settings.local.json << 'EOF'
{
"enableAllProjectMcpServers": false,
"hooks": {
"PreToolUse": [
{
"tool": "Bash",
"command": ["python3", ".claude/hooks/pre_bash.py"]
}
]
}
}
EOF
```
**Activation**: Restart Claude Code after installation.
## What Gets Blocked
### Always Blocked (Universal)
- `git reset --hard` - Destroys uncommitted work
## How It Works
1. **Hook triggers**: When an AI agent tries to use the Bash tool
2. **Hook reads**: Command from stdin as JSON
3. **Hook checks**: Patterns against blocked list
4. **Hook blocks**: Exit code 2 prevents execution
5. **Hook allows**: Exit code 0 lets command through
The hook runs BEFORE the command executes, so blocked commands never reach the shell.
## Testing
```bash
# This should be blocked:
git reset --hard HEAD
# Expected output:
# BLOCKED: git reset --hard is NEVER allowed for AI agents
# Attempted: git reset --hard HEAD
# Only the user can run this command directly.
# This should work:
git status
```
## Troubleshooting
### Hook doesn't trigger
- Restart Claude Code
- Check `.claude/settings.local.json` has the hooks configuration
- Verify `.claude/hooks/pre_bash.py` is executable: `ls -la .claude/hooks/`
### False positives
- Commands containing "git" in arguments (not as a command) might trigger
- Adjust the regex in `pre_bash.py` if needed
### Hook errors
- The hook fails open (exits 0 on errors) to avoid breaking workflows
- Check Python 3 is available: `which python3`
## Files
- `.claude/hooks/pre_bash.py` - The actual hook script
- `.claude/settings.local.json` - Claude Code configuration
## References
- [Claude Code Hooks Documentation](https://docs.claude.com/claude-code/hooks)
- Blog post: [Preventing git commit --amend with Claude Code Hooks](https://kreako.fr/blog/20250920-claude-code-commit-amend/)
- Git hooks: `scripts/git-policy.ts` (lines 28-159)
