Doc

You don't open Doc to write a page; you tell the doc assistant what you want to see ("my tasks this week", "the Q3 pipeline by stage", "everything in the brain about Acme"). The assistant emits a renderView tool call, which creates a draft page server-side and streams a deep-link pill back into the chat. Click the pill to open the full page in Doc; the chat dock stays open in the bottom-right so you can refine without leaving.

Every data block on a Doc page is bound to a workspace primitive (tasks, CRM contacts, files, deals, entities, memories), not a snapshot. Open a saved page tomorrow and the data block re-resolves against current state. Editing the underlying primitive in chat changes the page; editing the page nudges the primitive. There is no "sync". The page is a view, not a copy.

Every renderView creates a draft row in saved_views with state='draft'. Drafts auto-prune 30 days after their last touch (any read or edit bumps the deadline). Click "Save" on a page to flip state='saved' and clear the prune timestamp. Saved pages live forever. Click "Unsave" to drop back to draft. The sidebar lists both groups side by side.

Pages are composed of block-shells (text, heading, divider, data, chart). Data blocks render through the A2UI v0.8 renderer with the same property registry as chat-inline views: Table, Board, KPI, BarChart, LineChart, PieChart. The renderer is share-safe (no eval, no model output reaches the DOM directly) and dark-mode aware via Tailwind theme tokens.

Doc editing is a context-injected skill, not a dedicated assistant. The default interlocutor is your workspace primary, and the chat dock lets you switch to any other accessible assistant; whichever one runs on the Doc surface gets the page tools (tasks, CRM, views, files) injected automatically. The skill steers it to "render, don't narrate": visibility requests ("show me", "list", "everything about") begin with a renderView call, while analytical questions still answer in prose.