# Work with the assistant

Use persistent conversations, clear approvals and discoverable commands.

## A conversation in your project

Write in the bottom composer. Enter sends; Shift+Enter adds a line. New chat starts fresh conversation authority in the same project; it does not create a new workcell. Subsequent turns should retain the underlying conversation and selected model.

## Models, thinking and commands

Type `/` to see commands supported by your installed release. `/model` chooses a model and its supported thinking levels. `/permissions` controls approval policy; `/status` shows workspace, runtime and usage information. Available models depend on your route and entitlement, not a permanent marketing list.

## Approvals and safety

Read the action, its target and its expected effect before approving. Grant only the access needed. Full access is not a guarantee of correctness and should carry a warning. Workspace text and instructions cannot grant themselves permission to your credentials or other projects.

## Follow-ups and Stop

A follow-up can wait in the queue; explicit steering addresses an active task. Stop should cancel the intended run, retire its children and pause queued work. If the UI reconnects, check the retained run status before resending so you do not accidentally repeat an action.