Documentation

Coding & Cloud Tasks

Switch a chat to coding mode and it becomes a real development workspace: the assistant edits files, runs commands and reads diagnostics on a connected machine — while you watch in an editor and terminal right beside the conversation.

Coding mode

Pick Coding from the chat-mode switcher (or when starting a new chat). The layout gains a Workspace region beside the conversation - on the right on desktop, or a Workspace tab in the bottom nav on mobile. The Workspace combines:

  • a Monaco editor — the same editor that powers VS Code — with a file tree, syntax highlighting, and code intelligence (LSP navigation, diagnostics and formatters);
  • a real terminal showing the commands the agent runs, live;
  • save straight from the editor (Ctrl/⌘+S), with a dot on any file that has unsaved edits.

The agent’s file edits show up in the conversation as diffs, and you can revert an answer’s changes (and rewind the chat to before it) from the message actions — see Chatting. Read tools run in parallel for speed, while writes to the workspace are serialized so they can’t race each other.

Picking where to work

The first time you send a message in a coding chat, the workspace asks where to work:

  • My device — a connected computer running the Computer Service; or
  • Cloud sandbox — a disposable, zero-setup environment in the cloud.

…and which project: an existing folder, a new one, or a clone from Git (a repo URL plus a branch). Your first message is queued and runs as soon as the workspace is ready. Chats branched off the main chat start “clarify-first” — the agent plans and asks questions before building, rather than inventing requirements.

Built-in coding commands

In a coding chat the slash palette adds /review, /explain, /fix, /tests, /terminal and /init (generate an AGENTS.md). See slash commands. An AGENTS.md in the repo — and your own custom commands — guide how the agent works in that project.

!!! tip “Hooks for your repo” Want the agent to auto-format after every edit, or to refuse writes to a read-only repo? That’s what hooks are for — rules that run a command on the workspace device (or block a tool) around tool calls.

The cloud sandbox

No machine to connect? Start a cloud sandbox and code entirely in the browser. It spins up on demand, the editor and terminal work exactly the same, and it’s disposable — ideal for a quick experiment or when you’re away from your own device. You can also spawn a cloud browser for web-automation tasks (see Devices & apps).

Cloud Tasks

For work you’d rather not babysit, Cloud Tasks (in the account menu) runs an autonomous coding task end-to-end on its own cloud sandbox and hands you a reviewable result.

  1. Start a task — give it a title, a prompt (“what should the agent do?”), a repository (owner/repo or a GitHub URL) and a base branch (defaults to main). You need your GitHub integration connected first, since a task clones from and pushes to your repo.
  2. It runs in the background; the status moves through Starting → Running → Ready to review (or Failed).
  3. When it’s ready, open it to read the agent’s summary and view the diff.
  4. Apply pushes the work to a pa/task-… branch and opens a pull request — or Discard throws it away.

Nothing touches your repository until you click Apply, so you’re always reviewing before anything ships. Run several at once and review them as they finish.