# Adaptive workspace

The pane beside a conversation follows what the bot is actually doing, not a guess about the task.
No classifier or model decides the layout: `deriveWorkspaceFocus` in `@rakazo/core` reads the thread
state every client already has, so web and mobile agree and the rules stay testable offline.

## Signals

| Focus | Signal | Web | Mobile |
|-------|--------|-----|--------|
| Screen | A fresh `thread.computer` pointer event from a working run, or a takeover request | Live screen opens beside the chat (wide pane), steps back a few seconds after the work ends | Pill above the composer opens the computer |
| Artifact | The latest run attached a file with `present: true` (`attach_file`) | File opens beside the chat until the user closes it or the bot presents another | Unchanged: file cards open the native preview |
| Needs you | An unanswered ask on a waiting run | "Needs you" pill replaces the jump arrow when the question scrolled out of view | Existing push and attention notifications |

Everything else stays a plain conversation.

## Rules

- The pane moves only on focus edges (`applyWorkspaceFocus`). The first focus of a conversation is a
  baseline, so opening a thread never jumps.
- A pane the user opened by hand is never replaced or closed automatically. Auto panes close only
  when their own focus ends; a presented file stays.
- Windows narrower than 1024px keep the conversation full width; file cards fall back to the dialog.
- Per bot, the client learns from the user: opening the screen by hand while the bot works three
  times makes it open on every run; closing an automatically opened screen twice stops it opening
  on its own. A takeover request always shows the screen.
- Pane widths are remembered per device.

## Not in scope

Generated layouts, per-domain dashboards and prompt classification stay out: they guess, move the UI
under the user, and cannot be verified deterministically.