The AI assistant

Open the assistant panel and describe the change — “add a payments container behind the gateway, wire it to Stripe”, “split the monolith into checkout and catalog”, “draw our stack: Next.js on Vercel, Postgres, Redis, Stripe”. It draws on the canvas while it answers — nodes pop in, parents expand, edges connect.

• Everything lands in a draft. Review it, then Save to model (a normal version, labelled AI · your prompt in history) or Discard. The assistant never saves on its own.
• It can just talk. Ask a question, think out loud, or ask it to interview you before it draws — it answers in chat and draws only when you ask for a change. The Talk only switch under the composer pins a reply to chat, whatever you type.
• Follow-ups compose. Keep talking — it appends, adjusts and removes without starting over; the reply says what changed, and the canvas keeps your framing, panning only if something new landed off-screen. Large deletions ask before they land.
• It authors flows. “Show me the checkout flow step by step” creates a playable flow.
• It navigates. Pure “show me X” requests move the camera and expand the right nodes without touching the model.
• Diagram source. Every reply carries its Mermaid source — open it, copy it, keep it in a repo. Pasting Mermaid back to the assistant rebuilds it on canvas (that's the import path).
• Sending. On a keyboard, Enter sends and Shift+Enter adds a line. On a phone, Return adds a line and Send sends.
• Limits. To keep the preview fair: 8 requests per minute, 200 per day, per account.