safety-guard

To prevent destructive shell commands during production deployment, apply PreToolUse hooks to intercept Bash, Write, and Edit tool calls. The safety guard detects high-risk patterns like rm -rf and force pushes, blocking them and suggesting safer alternatives.

To restrict autonomous agent file edits to a specific directory tree, enable Freeze Mode. This enforces directory write locking, preventing modifications to unrelated areas while optionally allowing configurable read access to the rest of the repository.

You can block dangerous SQL drops and force pushes while retaining write access by enabling Combined Guard Mode. This simultaneously enforces destructive-command blocking and directory write restrictions, allowing configurable read access while logging blocked actions.

To enable directory write locking for a specific path during a migration, run the guard command with your target directory and read permissions. For example, start a session with the guard command targeting your specific directory tree and allowing global read access.

Blocked destructive operations and high-risk command attempts are logged to the safety-guard.log file within the .gemini directory. This captures all intercepted Bash, Write, Edit, and MultiEdit tool calls that triggered the command and path pattern checks.

To restore normal agent behavior after a risky production maintenance window, use the global unlock or off switch. This operational control disables the destructive command blocking and directory write restrictions, returning the agent to its standard execution state.