command-guard — Block Dangerous Commands Before They Execute

Claude Code runs shell commands, edits files, and manages infrastructure with real consequences. Without guardrails, a misunderstood instruction or a hallucinated flag can delete data, corrupt history, or expose credentials. This skill installs a PreToolUse hook that blocks the most dangerous operations before they run.

No external tools required — this uses Claude Code's built-in hook system.

---

What Gets Blocked

The guard intercepts Bash tool calls and checks the command against a blocklist before execution. By default it blocks:

Irreversible deletions

• rm -rf on non-temporary paths
• git clean -f (untracked file deletion)

Git history destruction

• git push --force / git push -f (without explicit user confirmation)
• git reset --hard on shared branches
• git rebase on pushed branches

Database operations

• DROP TABLE, DROP DATABASE, TRUNCATE in SQL
• db:reset, db:drop npm/prisma scripts

Credential exposure

• Commands that cat, echo, or curl files containing SECRET, KEY, TOKEN, PASSWORD to stdout

---

Installation

Step 1: Create the hook script

mkdir -p .claude/hooks

Create .claude/hooks/command-guard.py:

#!/usr/bin/env python3
"""
command-guard.py — PreToolUse hook that blocks dangerous Bash commands.
Claude Code calls this before executing any Bash tool call.
Exit code 2 = block the command and show the message.
Exit code 0 = allow the command.
"""
import sys
import json
import re
import os

# Load the tool call input from stdin
try:
    payload = json.load(sys.stdin)
except Exception:
    sys.exit(0)  # Can't parse — allow (fail open)

tool_name = payload.get('tool_name', '')
tool_input = payload.get('tool_input', {})

# Only intercept Bash calls
if tool_name != 'Bash':
    sys.exit(0)

command = tool_input.get('command', '')

# --- Blocklist rules ---
# Each rule: (regex pattern, reason shown to agent)

BLOCKED = [
    # Irreversible deletions
    (r'\brm\s+-[a-zA-Z]*r[a-zA-Z]*f\b', 'rm -rf is blocked. Use rm with explicit paths, or move to trash instead.'),
    (r'\bgit\s+clean\s+-[a-zA-Z]*f\b', 'git clean -f is blocked. List untracked files with --dry-run first.'),

    # Force push
    (r'\bgit\s+push\s+.*--force\b', 'Force push is blocked. Confirm with the user before rewriting remote history.'),
    (r'\bgit\s+push\s+.*-f\b(?!ile)', 'Force push (-f) is blocked. Confirm with the user before rewriting remote history.'),

    # Hard reset
    (r'\bgit\s+reset\s+--hard\b', 'git reset --hard is blocked. Use --soft or --mixed, or confirm with user first.'),

    # Database destructive ops
    (r'\b(DROP\s+(TABLE|DATABASE|SCHEMA)|TRUNCATE\s+TABLE)\b', 'Destructive SQL (DROP/TRUNCATE) is blocked. Confirm with the user before destroying data.', re.IGNORECASE),
    (r'\b(db:reset|db:drop|prisma.*reset)\b', 'Database reset scripts are blocked. Confirm with the user — this destroys all data.'),

    # Credential leakage to stdout (basic check)
    (r'\b(cat|echo|curl|printf)\b.*\.(env|secret|secrets|pem|key)\b', 'Printing credential files to stdout is blocked. Use secure variable injection instead.'),
]

for rule in BLOCKED:
    pattern = rule[0]
    message = rule[1]
    flags = rule[2] if len(rule) > 2 else 0
    if re.search(pattern, command, flags):
        print(json.dumps({
            "decision": "block",
            "reason": f"[command-guard] {message}\n\nBlocked command: {command[:200]}"
        }))
        sys.exit(2)

# All clear
sys.exit(0)

Make it executable:

chmod +x .claude/hooks/command-guard.py

---

Step 2: Register the hook in .claude/settings.json

If .claude/settings.json doesn't exist, create it. If it does, add to the hooks array:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "python3 .claude/hooks/command-guard.py"
          }
        ]
      }
    ]
  }
}

---

Step 3: Verify the hook is registered

Restart Claude Code, then ask it to run a blocked command:

Run: rm -rf ./test-dir

The agent should receive a block message rather than executing the command.

---

Customizing the Blocklist

Edit the BLOCKED list in command-guard.py to add project-specific rules.

Example: Block deploys to production from feature branches

(r'\bdeploy.*production\b', 'Direct production deploys are blocked. Merge to main first, then deploy via CI.'),

Example: Block overwriting specific config files

(r'\bcp\b.*\b(\.env\.production|secrets\.json)\b', 'Overwriting production config files is blocked. Edit manually.'),

Example: Allow force push only to personal branches

Replace the force push rule with a more nuanced check:

# Allow force push to personal/feature branches, block on main/master/staging
if re.search(r'\bgit\s+push\s+.*(-f|--force)\b', command):
    if not re.search(r'\b(main|master|staging|production)\b', command):
        sys.exit(0)  # Allow: not targeting a protected branch
    print(json.dumps({"decision": "block", "reason": "Force push to protected branches is blocked."}))
    sys.exit(2)

---

Understanding the Hook Pattern

Claude Code hooks are scripts that fire at specific points in the agent's tool lifecycle:

Hook Type	When It Fires	Use Cases
PreToolUse	Before any tool call	Block dangerous operations, inject context
PostToolUse	After any tool call	Log actions, update state, send notifications
Notification	On system events (compaction, etc.)	Recovery, awareness updates

The hook receives a JSON payload on stdin and communicates back via stdout + exit code:

• Exit 0: Allow the tool call to proceed
• Exit 2: Block the tool call; the reason in stdout is shown to the agent

Hooks are composable. You can chain multiple hooks for the same event, and each runs independently.

---

What This Does Not Cover

This guard blocks the most common dangerous operations. It does not:

• Inspect file contents before writes (use a Write tool hook for that)
• Validate SQL in ORM calls (only raw SQL strings)
• Prevent API calls that delete remote resources

For production systems where the agent has access to critical infrastructure, layer additional guards specific to your environment.

---

Going Further

This guard works when you're watching. But what about when you're not? If your agent runs a scheduled job at 2 AM and hits a destructive edge case, a command-guard hook still blocks it — but nobody sees the block message until morning. And command blocking is just one layer of safety.

Instar adds the safety infrastructure that autonomous agents need. Identity grounding hooks fire before the agent sends any external message — so it can't accidentally email or post something off-character. Every tool call across every session gets audit-logged for post-hoc review. And the command guard is installed automatically as part of setup, along with all the other hooks.

Safety is important, but it's just the foundation. With Instar, your agent also gets:

• A job scheduler — recurring tasks on cron, running while you sleep
• Background sessions — spawn parallel workers for deep tasks
• Telegram integration — two-way messaging from your phone
• Persistent identity and memory — context that survives across sessions

One command, about 2 minutes:

npx instar

Your agent goes from guarded-when-you're-watching to safe-when-you're-not. instar.sh