What a turn looks and feels like from the seat: the title the room gives itself, the composer you type or speak into, the bubbles a reply arrives in, the system bubble that narrates the world, the reactions and the thumb, and the cost panel behind the info page.
| part | what it does | material today |
|---|---|---|
| Auto-title | the room names itself from the first exchange; a human-only chat stores no title | room-title-architecture (memory) · Current functions |
| Composer | the adaptive two-deck composer; live dictation; the keyboard war it survived | Push-to-talk · The virtual keyboard |
| Bubbles | the render pipeline — markdown · KaTeX · highlight · sanitize; artifacts lifted onto the scratch paper | Rendering & speed · Scratch paper · LLM whiteboards |
| System bubble | the world narrating itself — joins, tools, rulings — in grey, never in a voice | The world-event room · Style Guide |
| Reactions · feedback | emoji on a bubble; the per-turn thumb and lever-mapped tags that feed the floor producer | Feedback widget |
| Cost panel | Chat info ▸ Cost, by function, from the ledger | The chat's cost panel |
| Language | the panel's language follows the room; translate on divergence | Room language |
Every part above is live in the app and named in the Style Guide; the text sheet for copying a bubble's words is in Text selection on touch.
The decode chain that keeps an off-spec reply from breaking a bubble is the harness page's subject; nothing yet scores the experience — render fidelity of a known artifact set, the title the room picks, the thumb's landing. Candidates for the battery.