Core Workflow
Permission Modes
Ask, Plan, Auto-edit, Auto, and Agent — how much an agent can do without asking.
Permission mode controls how often the agent stops to ask before doing
something. Cycle through modes with Shift+Tab, or set one directly with
/mode. Currently available for Claude Code chats.
| Mode | What it does |
|---|---|
| Ask (default) | Confirms before every gated action — file edits, commands, everything |
| Plan | Read-only: explores freely, but every write still needs confirmation. Ends with a plan you review before anything changes |
| Auto-edit | File edits (create, edit, write) go through automatically; commands and everything else still ask |
| Auto | Everything is automatic, including shell commands |
| Agent | Every tool is allowed, no confirmations at all — full autonomy |
Which one to use
- Start in Ask if you don't know the codebase yet, or the change is risky (migrations, deploys, anything destructive).
- Use Plan when you want to see the approach before committing to it — useful for anything non-trivial.
- Auto-edit is the sweet spot for routine work: you trust file changes, but still want a look before commands run.
- Agent is for when you've reviewed the plan already and just want it done — or for low-stakes, easily reversible work.
Plan → Build handoff
When an agent finishes planning, Quack shows a "Pass the ball" chip in the composer. Clicking it switches the chat to Agent mode and sends the plan straight to implementation — no need to re-explain what was just agreed.