Rooms
A room is one working directory that more than one agent is in at the same time.
Running two agents on one project normally means two copies of the work and a merge afterwards, or one agent silently overwriting what the other just wrote. A room narrows both. The agents see the same files, each is given a different job, and each is told to read what the others have produced before writing its own. Nothing locks a file: when two of them touch the same one within moments, Vect flags it rather than preventing it.
The three modes
Every session is started in one of three modes. The mode decides how many agents are in the room, not what they are allowed to do.
| Mode | What it is | Reach for it when |
|---|---|---|
| Solo | One agent, on its own. | The task is focused and one pair of hands is enough. This is the default and the cheapest. |
| Manager | A lead agent breaks the work up and hands pieces to workers it creates. | The work splits cleanly and you would rather describe the goal once than assign each piece. |
| Collab | Two or more agents as peers in the same room. | The pieces touch each other, so whoever is working needs to see what everyone else just did. |
Solo and Manager are ordinary sessions. Collab is where the room does the work.
Who can be seated
A room seats agents that speak ACP and are installed on this machine. Today that is Claude Code, Codex and Hermes. OpenCode has no ACP server, so it can hold a session of its own but cannot join a room.
What the room actually gives them
Agents in a room share two channels, and you can pick which ones are open.
- Shared history: every agent reads the others' conversation, so nobody re-derives a decision that was already made out loud.
- The room: a channel the agents talk in. This is where they hand work to each other, ask for a review, and argue about how something should work.
shared_and_room opens both and is the default. shared_only drops the
channel they negotiate in, and room_only lets them talk while hiding the work
they are talking about. Both exist because they are useful to compare, and
neither is a good idea for real work.
Watching a room
Open a room from the Rooms rail and it fills the tab.
- The transcript is one stream, not a column per agent. Each row carries the mark and colour of whoever said it, so four agents stay readable and you can see the order things actually happened in.
- Messages are addressed with explicit chips rather than by typing an
@name, because who a message is for decides which agent processes reach it. - Any message can open a thread beside the transcript, which you can drag wider or narrower.
- Agents can be added to or stopped in a running room, and a room can be renamed inline or from its right-click menu. Deleting asks twice, since the second choice takes the transcript with it.
Adding an agent to a running Solo session upgrades it to a room in place; the session you were already having is kept.
What is written down
A Collab session writes its own record: the session starting, each agent joining and leaving, each file touch, and each conflict between two agents on one file. That record is the audit trail, readable from the palette's Audit view.
Two limits worth knowing. A Solo session writes no audit trail at all, and what an agent runs in a terminal is not recorded anywhere; the trail is about who touched which file, not about every command.
Costs
A room is several agents against one problem, so it spends several agents' worth of tokens. A Collab session's cost meter counts the participants together, not only the one you are reading.