Permissions

Control how much an agent can do on its own — from plan-only to full access — with four per-chat permission levels.

Permissions decide how much an agent is allowed to do without checking with you first. You pick a level in the composer (the message box) for each chat, ranging from “plan but change nothing” to “no prompts at all.”

The four permission levels

Set the level from the composer before or during a chat. It applies to the agent driving that chat.

Level	What the agent does	Good for
Plan only	The agent plans the work and makes no changes, then asks you to proceed.	Understanding scope before anything touches your files; reviewing the approach first.
Supervised	The agent asks before each action (every edit, command, etc.).	Sensitive code, unfamiliar agents, or when you want to watch every step.
Auto-accept edits	The agent auto-applies file edits but still asks for risky things (e.g. running commands).	Steady feature work where you trust the edits but want a check on anything dangerous.
Full access	The agent runs with no prompts.	Throwaway sandboxes, trusted flows, or when you want maximum momentum.

Warning: Full access means the agent can edit files and run commands without asking. Use it only where you’re comfortable with that — for example, in an isolated worktree you can discard.

How approval requests appear

When a level requires your sign-off (Plan only, Supervised, or the risky actions under Auto-accept edits), the agent pauses and surfaces an approval request showing what it wants to do. You approve or decline, and the agent continues or adjusts.

While a chat is waiting on you, Dices shows a “needs input” signal so you can spot it — handy when you have several parallel chats running and one is blocked on your decision.

Plan only, then proceed

Plan only is a two-step flow:

The agent produces a plan and changes nothing.
It asks you to proceed. If you approve, switch to a more permissive level (or approve the actions) to let it carry the plan out.

This is the safest way to start on unfamiliar code: see the plan first, then decide.

Default permission for new chats

You can set the default permission level that new chats start at:

Globally — the default for every new chat.
Per project — override the global default for a specific folder.

For example, keep new chats on Supervised globally, but set a scratch project to Auto-accept edits.

Choosing a level

Start on Plan only when you don’t yet trust the scope.
Use Supervised for careful, watch-every-step work.
Move to Auto-accept edits once you trust the direction but want a guard on risky operations.
Reserve Full access for isolated or disposable contexts.

Agents — the agent each permission level applies to.
Git & worktrees — isolation that makes higher permission levels safer.