Pairs with: autonomous-loops, agentic-engineering, deployment-patterns

Safety Guard

When to Activate

• Agent running in autonomous/full-auto mode
• Working on production systems with irreversible operations
• Restricting edits to a specific directory during focused work
• Sensitive operations: migrations, deploys, permission changes, data transformations

Three Protection Modes

Mode 1: Careful Mode

Intercepts destructive commands before execution and warns, then asks for confirmation:

Watched patterns:

rm -rf (especially /, ~, or project root)
git push --force
git reset --hard
git checkout . (discard all changes)
DROP TABLE / DROP DATABASE
docker system prune
kubectl delete
chmod 777
sudo rm
npm publish (accidental publishes)
Any command with --no-verify

On detection: explains what the command does → suggests safer alternative → requires explicit confirmation.

Mode 2: Freeze Mode

Locks all file writes to a specific directory tree:

activate: safety-guard freeze src/components/

Any Write/Edit outside src/components/ is blocked with an explanation. Use when you want an agent focused on one area without touching unrelated code.

Mode 3: Guard Mode (Careful + Freeze combined)

Maximum safety for autonomous agents — reads anywhere, writes only to allowed directory, destructive commands blocked everywhere:

activate: safety-guard guard --dir src/api/ --allow-read-all

Disable

safety-guard off

Implementation

Uses PreToolUse hooks to intercept Bash, Write, Edit, and MultiEdit tool calls. Checks command/path against active rules before allowing execution.

Hook entry (hooks.json):

{
  "matcher": "Bash|Write|Edit",
  "hooks": [{
    "type": "command",
    "command": "node hooks/scripts/safety-guard.js"
  }]
}

Activation Guide

Logs

All blocked actions are logged to ~/.claude/safety-guard.log for audit trail.