Plan mode, questions and todos

Three features keep you in control of larger tasks: plan mode lets you review a plan before anything changes, structured questions let the agent ask you to choose between options, and the todo list shows its progress step by step.

Plan mode

In plan mode the agent researches and proposes; it changes nothing. It can read and search files, run read-only commands such as git log, use the language server, look things up on the web, and start explore or plan subagents. Edits and commands that change anything are refused with "plan mode is read-only".

Enter plan mode

• Press Shift+Tab in the composer until the mode chip reads Plan.
• Or choose Plan from the mode chip in the composer's bar, or Switch to Plan in the command palette.
• Or type /mode plan.
• To start every new chat in plan mode, set Settings → Permissions → Permission mode to Plan (permissions.mode = "plan").

Then describe the task:

Add rate limiting to the public API. Plan it first.

Review the plan

When the plan is ready, a Plan ready · Review row appears in the chat and the side panel opens on its Plan tab (Review opens it again), marked Ready for review. It shows the plan in markdown: the goal, the approach, the files to change, the ordered steps, how it will be verified, and risks or open questions; when the agent keeps a todo list, its Checklist and N of M done follow.

Control	Effect
Edit / Preview	Switch to a text editor to change the plan yourself (Esc goes back to the preview). An Edited tag appears when your version differs.
Approve & auto-accept edits	Leave plan mode in Auto-accept edits mode and start implementing.
Approve & ask before edits	Leave plan mode in Ask mode and start implementing.
Keep planning	Opens a feedback box. Type what should change and click Send feedback (or press ⌘Enter / Ctrl+Enter). The agent stays in plan mode, revises, and proposes again. Esc closes the box.

If you edited the plan, the agent is told to implement your version. While the plan waits, the island in the title bar says Plan ready for review in amber (its Review button goes to the plan), and the plan is listed in the Inbox. You can still send messages, which are queued. With no plan waiting, the Plan tab shows the chat's Last plan and its checklist, or No plan yet.

Tip: Plans are only proposed for work that changes files. If you ask a question in plan mode, the agent simply answers it.

Only the main agent proposes plans; the plan subagent returns its plan as a report to the main agent instead.

Questions from the agent

When a decision materially changes the result (two sound designs, a destructive step, a product choice), the agent can ask you one to four multiple-choice questions at once. A Question card takes the composer's place (your draft is kept for later; an approval waiting at the same time comes first):

• With several questions, tabs at the top switch between them; a check mark shows which are answered.
• For a single-choice question, clicking an option selects it and moves to the next question. For a multiple-choice question ("Pick any that apply"), click every option that applies.
• Other… lets you type your own answer. Press Enter in that box to submit once everything is answered.
• Submit sends your answers (it is enabled when every question has an answer).
• Dismiss declines to answer. The agent is told not to ask again and to continue on its best judgment or explain what is blocked.

While a question waits, the island says Question for you, with an Answer button that goes to it. Only the main agent can ask questions; subagents can't.

Todos

For tasks with several steps, the agent keeps a todo list with the TodoWrite tool. Each item is pending, in progress, or completed; the agent keeps exactly one item in progress at a time.

• The island in the title bar shows the current step, such as "Running the parser tests", and a ring around the pet fills as items are done.
• Click the island to open it: its Plan row says how many items are done and which one is current (for example 2 of 5 · Run the tests); click it for the full checklist in the side panel's Plan tab.
• /todos prints the list in the transcript as a checklist.
• A subagent's own todo list is shown in its transcript in the side panel's Agents tab.

The todo list survives compaction, so it also serves as the agent's working memory in long sessions. If the agent has worked for a while on a multi-step task without a list, it gets a reminder to create one.

See also: Permissions and safety · Agents · Everyday use