Auto approval needs a brake

Auto approval is becoming the default. Claude Code now starts in Auto mode, and Codex offers auto-review and full-access modes. That makes sense - research indicates the menial process of approving command after command destroys users' ability to actually catch issues with a protracted coding session dropping to 5% effectiveness.

But approval is not free. A recent Codex user report shows auto-review as a noticeable share of their token usage. The exact cost will vary by account and task, but it is a reminder that provider-managed approval can cause costs you do not see until you check.

Having used these myself, I'm also getting really frustrated by some actions that they won't seek approval for:

• Hijacking your computer, moving your mouse, opening pages in your web browser.
• rm -rf - even when safely in a subdirectory and not mistyped, this gives me nightmares.

Saggar handles routine approval prompts across your providers. You keep one set of rules, rather than learning and maintaining a different approval setup for Claude, Codex, and the next CLI. Turn on Supervised for a session when you want every permission prompt left for you.

Saggar leaves genuinely destructive commands for you by default, including recursive deletes, sudo, force pushes, hard resets, disk writes, infrastructure destruction, and destructive SQL. Add your own rules as case-insensitive text or regular expressions, such as a command that touches production.

Auto approval should remove friction, not remove judgment. Configure approvals in Settings, supervise sessions that need your full attention, and let routine work pass everywhere else.