# Seats

> Seats and models, The chat seat, The Assistant, Meetings

## Seats and models

*A project's three roles each pick a model: Design, Dev, Review. The chat seat and scheduled jobs use the Dev model; everyone's Assistant is separate.*

| Role | Default model | Default effort | Runs |
| --- | --- | --- | --- |
| Design | Claude Fable 5.1 | xhigh | The DESIGN phase |
| Dev | Claude Opus 5.5 | xhigh | The engine in DEVELOP / VERIFY / DEPLOY; Auto-Run implement, fix and repair turns; the project conversation in MAINTAIN; scheduled jobs |
| Review | Codex 6.1 Sol | high | The review gate |
| Assistant | Claude Opus 5.5 | high | Everyone's own assistant (fixed per release) |

### Two engines

- **Claude**: A long-lived `claude` CLI process speaking `stream-json` both ways
- **Codex**: `codex app-server`, one persistent thread per project

Effort runs low / medium / high / xhigh / max, and Codex adds ultra. A Space admin keeps the list of models and efforts on offer and can add up to 8 custom Anthropic-compatible endpoints; a project owner picks the model and effort for each of the project's three roles, and an empty choice means the instance default.

A few jobs sit outside the roles: the grader always uses one Codex model, and translation, queue dedupe, live calls and meeting transcription each have their own.

## The chat seat

*Chat and execution are one pipeline: the chat seat reads and edits code. The queue is yours; it adds an item only when you ask.*

- It has full tools while the engine is free. While Auto-Run holds the project it is read-only (read, search, web search and fetch), so it does not compete with Auto-Run for the same working tree.
- When you ask for a small fix right now, it makes the change directly and says exactly what changed. For anything multi-file, risky or worth tracking, it describes the work and leaves the queue to you.
- It never picks requests out of the chat and queues them on its own; it adds an item with `queue_items` only when you ask. It cannot start Auto-Run either: start runs from the queue panel.
- When a decision is yours, it asks with a card.

## The Assistant

*An assistant that is yours alone. Talk an idea through with it, and once the idea is clear it starts the project for you. It remembers what you said.*

- **Room**: A private room only you can see. Every member can create, manage and delete their own Assistants; an owner can also share one with a team, and it then sees only the projects everyone in the room can see
- **Projects**: Reads them through read-only tools (list, read, search, git) and does not edit them directly; it can manage the queue and scheduled jobs itself
- **New projects**: Raises a confirmation card; once you approve, the project is created, the brief is posted into its chat and the engine takes it from DESIGN. If you are not the Space owner, your approval sends a request card to the owner, who creates the project in one tap and it goes into your team
- **Memory**: Three tools, `recall` / `remember` / `forget`; every turn, your words and its reply are appended to L0. Project chats and engines have no memory tools
- **Calls**: You can also talk to it by voice. The call has the same persona and the same memory, and it can hand a task to the text Assistant. A call lasts about 25 minutes, with 60 minutes per person per week

## Meetings

*Tap the meeting button and talk as usual. It recognizes each speaker, writes the transcript and the minutes, and files the items you tick.*

1. Record: up to 90 minutes a meeting.
2. Transcribe: a speaker-aware transcription model labels who said what.
3. Recognize: short clips of each speaker are matched against the team's voiceprint library (2 to 10 s samples, kept per team), so "Speaker 1" becomes a name.
4. Minutes: written up as `## Meeting minutes` and posted into chat.
5. File: you tick items in two rounds, and they go into the Bug / Feature queue through the same dedupe as chat.

Recording in an Assistant room makes a voice message, not a meeting. Meeting transcription is 5 hours per person per week.