How do I start the agent in a quadrant?
Click the start button in the bar across the top of the quadrant, then pick one action from the menu that opens. Clicking start only opens the menu; it never launches an agent on its own. The menu has two provider groups, Claude and OpenAI Codex, and each offers the same three actions: boot project (marked recommended), start unbooted, and execute sidebar tasks. Pick the provider, then the action, and the agent starts in the quadrant's terminal. For most work choose boot project, which gets your project running and checks it before you hand over any work. The start button appears only once you have pinned a folder to the quadrant; before that you see a 📁 pin a worktree button in its place.
What each action does
- boot project (recommended): starts the agent, then has it read your project, start your dev server (and any background helpers), open the preview, and confirm the app actually loads. Boot is done only once the agent has seen it running and reported what it started. It starts only what is missing, so booting again never creates a duplicate server. Desktop apps like codus have no web preview, so the agent confirms the app's window instead.
- start unbooted: starts the agent alone, with no servers, waiting at its prompt. Use it to ask a question, or when the app is already running.
- execute sidebar tasks: hands the agent the tasks queued in this quadrant's sidebar and works through them one at a time. It does not boot first, and if no tasks are waiting it does nothing and shows no message.
Does it continue my conversation or start fresh?
start unbooted brings back your previous conversation in that folder, so you carry on where you left off. boot project starts a fresh conversation and runs setup again. So to keep your chat, use start unbooted. The first time you open a folder there is nothing to resume, so it starts empty. With OpenAI Codex, start unbooted opens a new session rather than reopening the previous one.
The 🛸 codus.boot.exec line
When you boot, codus types 🛸 codus.boot.exec into the terminal for you. It is a deliberate signal, not an error, and the agent treats it as a top-priority instruction to start your project right away. You will also see it when you run a queued task before booting, because codus boots the project first. With OpenAI Codex that line is followed by several more lines of boot instructions; with Claude it is the single line.