Chatting
The chat is the same in VS Code and in the app’s window.
Sending messages
Section titled “Sending messages”- Enter sends, Shift+Enter starts a new line. On a touch screen, Enter starts a new line and the Send button sends.
- While the agent works, Enter queues your message: it is sent when the agent is done, one per turn. Queued messages show above the box, where you can edit one (click it), remove it, or send it now.
- Ctrl/⌘+Enter sends now. With Claude Code and Codex, the agent reads it right away and keeps going (steering); with the others, its turn stops first.
- Stop ends the agent’s turn. Queued messages go back into the box.
Slash commands
Section titled “Slash commands”Type / for the agent’s own commands (e.g. /compact). Enter runs the highlighted one,
Tab completes it.
Shell commands
Section titled “Shell commands”Start a message with ! to run a shell command in the sandbox, in the agent’s workspace, e.g.
!ls /data/input. Its output streams into the chat, and the agent sees it with your next message. Commands stop
after 15 minutes.
Mode, model and effort
Section titled “Mode, model and effort”The bar under the box shows what the agent offers: its mode (e.g. for Claude Code: ask before edits, accept edits, plan, auto), model, reasoning effort, and more under More settings. Shift+Tab cycles through the modes. What each agent offers is its own: code flower shows its choices as they are.
Permissions and questions
Section titled “Permissions and questions”When the agent asks before doing something, a card shows what it wants to do, with its options (allow once, always, …) and Cancel. With DeepSeek Harness, a switch in the bar allows all of its requests in that chat (on by default). Questions the agent asks you (e.g. multiple choice) appear as forms. While the agent waits for you, the chat and the sandbox are marked, and VS Code’s status bar says An agent needs you.
Actions
Section titled “Actions”The Tools menu in the chat’s header sends ready-made requests: write or update a report, answer comments on reports, write a methods section, save a checkpoint, or wrap up. See all chat actions.
The meter shows how full the agent’s context is, the chat’s tokens (and estimated cost, for Claude Code and OpenCode), and your login’s quota windows with when they reset. A sandbox nearly out of quota turns orange in VS Code’s worker list.
Several windows at once
Section titled “Several windows at once”You can open the same sandbox, and the same chat, in VS Code and in the app’s window (or on your phone) at the same time. Messages from either take turns; a permission request shows in all of them, and the first answer counts. The agent keeps working when you close every window: reopen the chat later to see what it did.
Sessions
Section titled “Sessions”Each chat is saved in the project, named after its first message. Rename or delete it from the sandbox’s list (a chat that is working cannot be deleted). A chat open in another sandbox of the same project shows read-only.