Safety Guard

Safety Guard Skill

Three-mode safety guardrail system for autonomous agent sessions. Prevents destructive operations and logs all interventions.

What Claude Gets Wrong Without This Skill

Without safety guardrails, autonomous agents:

1. Execute destructive commands (rm -rf, DROP TABLE) without confirmation
2. Force push to protected branches, losing commit history
3. Run unrestricted operations in production environments
4. Bypass safety hooks (--no-verify flags) without justification
5. Provide no audit trail of potentially dangerous actions

Safety guard ensures autonomous operations remain safe and auditable.

Three Modes

Careful Mode (Warn)

Behavior: Warns on destructive commands but allows execution after user confirmation.

Use When: Standard development, trusted environments, human oversight available.

Protected Operations: rm -rf, git push --force, git reset --hard, DROP TABLE/DATABASE, docker system prune, kubectl delete, chmod 777, sudo rm, npm/cargo publish, --no-verify flags.

Freeze Mode (Lock Writes)

Behavior: Blocks write operations to specific paths. Config: FREEZE_PATHS="/etc/,~/.ssh/". Blocks Write, Edit, and Bash ops to frozen paths.

Guard Mode (Careful + Freeze)

Behavior: Warns on destructive ops AND locks write paths. Use for: autonomous agents, docker-sandbox, permissionMode: dontAsk sessions.

Hook Implementation Pattern

Safety guard extends existing hooks with additional checks.

PreToolUse Bash Hook Enhancement:

# Add to existing block-destructive.sh or create safety-guard-bash.sh
SAFETY_MODE="${SAFETY_GUARD_MODE:-off}"  # off | careful | freeze | guard
LOG_FILE="${HOME}/.claude/safety-guard.log"

if [ "$SAFETY_MODE" = "off" ]; then
  exit 0  # Safety guard disabled
fi

COMMAND="$1"  # Tool input from stdin

# Careful mode: destructive operation detection
if [[ "$SAFETY_MODE" =~ (careful|guard) ]]; then
  if echo "$COMMAND" | grep -qE "rm\s+-rf|git push.*--force|git reset --hard|DROP (TABLE|DATABASE)|docker system prune|kubectl delete|chmod 777|sudo rm|--no-verify"; then
    echo "$(date -Iseconds) WARN: $COMMAND" >> "$LOG_FILE"
    echo "⚠️  SAFETY GUARD: Destructive operation detected"
    echo "Command: $COMMAND"
    echo "Continue? (y/N):"
    # Exit 1 to prompt user, or exit 0 if user confirms
  fi
fi

# Freeze mode: write path locking
if [[ "$SAFETY_MODE" =~ (freeze|guard) ]]; then
  FREEZE_PATHS="${FREEZE_PATHS:-}"
  if [ -n "$FREEZE_PATHS" ]; then
    # Check if command affects frozen paths
    # Implementation depends on command parser
  fi
fi

PreToolUse Write/Edit Hook Enhancement:

# Check if write target matches frozen paths
SAFETY_MODE="${SAFETY_GUARD_MODE:-off}"
TARGET_FILE="$1"

if [[ "$SAFETY_MODE" =~ (freeze|guard) ]]; then
  FREEZE_PATHS="${FREEZE_PATHS:-}"
  for frozen_path in ${FREEZE_PATHS//,/ }; do
    if [[ "$TARGET_FILE" == "$frozen_path"* ]]; then
      echo "$(date -Iseconds) BLOCK: Write to $TARGET_FILE (frozen path)" >> ~/.claude/safety-guard.log
      echo "❌ SAFETY GUARD: Write blocked"
      echo "Path: $TARGET_FILE is frozen"
      exit 2  # Block operation
    fi
  done
fi

Action Logging

All interventions logged to ~/.claude/safety-guard.log with ISO timestamp, action (WARN/BLOCK), command, reason, outcome. Rotate at session start if >10KB.

Mode Activation

Via env vars: SAFETY_GUARD_MODE=guard and FREEZE_PATHS="/etc/,~/.ssh/". Or in agent frontmatter. Unset to disable.

Integration with Existing Hooks

Complements:

• block-destructive.sh: Safety guard extends this with logging and mode flexibility
• block-secrets-write.sh: Freeze mode adds pattern-based path protection

Does NOT replace:

• Existing hooks remain active
• Safety guard adds layer of user-controlled protection
• Hook order: existing blocks first, then safety guard warnings

Anti-Patterns

Always-on Guard mode: Guard mode for routine development slows workflow. Use Careful mode for normal work.

Freezing entire project: Freeze mode for / blocks all writes. Be specific with frozen paths.

Ignoring warnings repeatedly: If you're clicking through 10 safety warnings, the guard is misconfigured for your task.

No log review: Safety log accumulates warnings. Review monthly to identify risky patterns.

Mandatory Checklist

1. Verify mode selected matches risk profile (Careful for dev, Guard for autonomous)
2. Verify if Freeze or Guard: FREEZE_PATHS specified and tested
3. Verify hook implementation added to PreToolUse: Bash, Write, Edit
4. Verify log file location created at ~/.claude/safety-guard.log
5. Verify warning format includes command, risk, and mode
6. Verify blocked operations exit with code 2 (block)
7. Verify environment variables documented in activation instructions