Auto-Approval

Auto-approval lets CACD automatically respond to agent permission prompts when the requested action appears safe. Instead of you clicking “approve” for every read-only operation, CACD can handle the obvious ones.

Caution

This feature is experimental. It uses AI to judge safety, which isn’t foolproof. Use it in low-risk environments and keep an eye on what’s being approved.

How It Works

When an agent enters a “waiting for input” state, CACD can analyze the terminal output and decide:

1. Needs permission - Something risky is happening. Wait for you.
2. Safe to approve - It’s a read-only or low-risk operation. Approve automatically.

The analysis uses Claude Haiku to quickly evaluate the situation.

What Gets Blocked (Requires Your Approval)

• File modifications - Writing, deleting, moving files
• Package managers - npm install, pip install, apt install, etc.
• Git changes - Commits, pushes, history rewrites
• System operations - sudo, service restarts, permission changes
• Network operations - curl/wget to unknown hosts, ssh connections
• Sensitive areas - SSH keys, credentials, dotfiles, /etc configs

What Gets Approved Automatically

• Read-only operations - Viewing files, listing directories
• Tests and linting - Running test suites, lint checks
• Help and info - Showing help text, version info
• Dry runs - Formatting checks, build previews

Enabling Auto-Approval

1. Open Settings in the WebUI
2. Go to the General section
3. Toggle Auto-Approval on
4. Optionally adjust the timeout (default: 30 seconds)

Timeout

If the safety check takes too long, auto-approval falls back to waiting for your input. The default timeout is 30 seconds, but you can adjust this in settings.

Custom Verification Command

You can replace the built-in Claude Haiku check with your own script. Your command receives the terminal output and should return JSON indicating whether permission is needed:

{"needsPermission": true, "reason": "Attempting to delete files"}

or

{"needsPermission": false}

Configure this in settings under Auto-Approval Command.

Tips

• Start with auto-approval disabled until you’re comfortable with how your agents behave
• Use it on isolated worktrees where mistakes are easy to undo
• The Haiku model is fast but not perfect - review your session history occasionally