Approving Commands

Learn how to safely approve AI actions and configure auto-approve settings.

Why Approvals Exist

AI coding assistants are powerful, but they can make mistakes. The approval system gives you:

●
Safety: Review destructive operations before they run (deleting files, running scripts, git operations)
●
Control: You decide what the AI can do autonomously and what needs your approval
●
Awareness: See exactly what's happening on your system, even when you're not at your desk

What Needs Approval?

By default, these actions require your approval:

Action	Why It Needs Approval	Risk Level
Bash commands	Can run any shell command on your system	High
File edits	Modifies existing code (configurable)	Medium
File creation	Creates new files in your project	Low
Git operations	Commits, pushes, or modifies git history	High
Package installs	npm install, pip install, etc.	Medium
Destructive ops	rm, unlink, truncate, etc.	Critical

Read-only operations are auto-approved

Tools like Read, Search, and Glob don't modify anything, so they run automatically without asking. This keeps the flow smooth while you stay safe.

The Approval Card

When the AI wants to do something that needs your approval, you'll see a card like this:

Permission Required

The AI wants to run a shell command

git add . && git commit -m "Fix login bug"

Card Elements

Element	Description
Warning icon	Indicates this action needs your review
Title	"Permission Required" or similar
Description	What the AI wants to do
Details	The exact command, diff preview, or file path
Deny button	Reject the action (AI will be told you denied it)
Approve button	Allow the action to proceed

Diff Preview for File Edits

For file edits, the approval card shows a diff preview so you can see exactly what will change:

Edit: src/utils.ts

export function formatDate(date: Date) {

- return date.toString()

+ return date.toISOString().split('T')[0]

}

How to Respond

Review the request

Read what the AI wants to do. For commands, check the exact syntax. For edits, review the diff.

Make your decision

If it looks correct and safe, tap Approve. If something seems wrong, tap Deny.

AI continues or adapts

After approval, the action runs. After denial, the AI is notified and may suggest an alternative.

When to deny

• Command looks suspicious or unfamiliar
• Diff changes code you didn't expect
• Operation seems too broad (e.g., rm -rf)
• You want to do it differently

Always Allow (Auto-Approve)

For trusted tools you use frequently, you can enable auto-approve:

Permission Modes

Mode	Description	Best For
default	Ask for all potentially dangerous operations	New users, careful workflows
plan	Auto-approve planning/analysis tools	When you trust the AI to explore
auto-edit	Auto-approve file edits (still asks for bash)	Faster coding, you review via git
allowAll	Auto-approve everything (use with caution!)	Trusted, isolated environments

Configuring Permissions

In the app:

Go to Settings → AI Backend → Permissions
Choose your permission mode
Optionally, enable/disable specific tools

Per-Tool Settings

You can also configure individual tools:

Tool	Recommendation
Read, Glob, Grep	Always auto-approve (read-only)
Edit	Auto-approve if you review via git, otherwise ask
Write	Auto-approve for new files, review for overwrites
Bash	Always ask (can run any command)
Task	Ask if you want control over sub-agents

Security Best Practices

Do:

✓ Review bash commands before approving
✓ Check diffs to understand what's changing
✓ Use git to track changes (easy to revert)
✓ Start with default mode, loosen as you gain trust
✓ Deny anything that looks suspicious

Don't:

✗ Use "allowAll" mode on production systems
✗ Approve commands you don't understand
✗ Ignore the details in approval cards
✗ Give the AI access to sensitive directories
✗ Approve broad delete operations without review

Pro tip: Use git

Keep your project in a git repository. Even if you approve something by mistake, you can always git checkout or git reset to undo changes.

What Happens When You Deny

When you deny a request:

The action is not executed
The AI receives a "permission denied" response
The AI typically acknowledges and may suggest an alternative
You can explain why you denied (helps the AI understand)

Denying is safe

Denying never breaks anything. The AI is designed to handle denials gracefully. If you're unsure, it's always safer to deny and ask for clarification.

Permission Notifications

When a permission request comes in while you're not looking at the app:

• Push notification: Your phone notifies you (if enabled)
• Badge count: App icon shows pending requests
• Session status: Session shows "Waiting for approval"

Enable notifications in Settings → Notifications to never miss a permission request.

Your First Session

Complete walkthrough for beginners

Privacy & Security

How Broski protects your data