Roundtable

roundtable

A room here is a decision being made, not a debate being staged. Four agents from four vendors research the question independently, argue about what they found, and the person watching walks away able to act. This skill is the craft of running that room. Read it once, then open one.

1. The motion

A motion is not the user's question repeated. It is the question made decidable. Three parts, and all three go in motion.md:

• The decision, in one line. A real fork. "Should the Windows port be native, or WSL2?" is a motion. "Thoughts on Windows?" is a topic, and a topic produces four essays nobody reads.
• The constraints that make it hard. Who is doing the work, what already exists, what it costs to be wrong. This is where a room gets its teeth — without it every seat answers the generic version of the question and they all agree.
• What would settle it. Name the evidence that would end the argument. It tells the seats what to go and look for, and it tells you afterwards whether the room actually did the work.

If the user's first message has only part of this, write the rest yourself and let them correct it. A motion on the page beats an interview.

Point the room at evidence. --evidence <path> gives every seat read access to the real thing — the repository, the folder, the data. It is the difference between four models arguing about your decision and four models arguing about decisions like yours. Use it whenever the subject is something on disk.

Scope the evidence tightly. Hand a seat a whole monorepo and some engines will read it until the timeout and answer nothing at all, while others skim four files and answer well — and you will have learnt about their search strategies instead of about your decision. Name the folder the question turns on. The seats are told to time-box their reading to roughly 70% of --timeout, but that only works if the pile is finite.

2. Seating the room

Rules that matter more than the preset:

• Two vendors minimum. Three or four is the good room. Four seats of the same engine is a mirror with extra steps — if only one vendor is installed, say so plainly and run it anyway, but do not pretend it is a panel.
• Stance ≠ engine. They are separate axes. Keep them separate so that when the room splits you can tell whether the split came from a different model or from a role you assigned. If you assign stances, put that in the chat: an assigned position is not a belief.
• Absent is fine. A vendor that is not installed or not logged in shows as an empty cell with the reason. Never quietly drop it from the seating — the missing vendor is information.

3. The rounds

room run opening → room run cross → room run converge. All seats run at once, so a round costs the slowest seat, not the sum. What each round asks for is already in the prompts; what you control is when to stop.

Openings are sealed. Seats run outside the workspace and cannot read each other's files. This is the load-bearing rule: in any shared-context loop these agents converge on whoever spoke first, politely and fast, and four voices agreeing teaches nothing. If someone asks you to "let them discuss it", explain what that costs and offer another round instead.

Stop after cross-examination. Report the split and ask the chair what to press. A room that runs start to finish without the user is a report, and they could have got a report from one agent.

Pressing a point is a new round on a narrowed motion, not a private chat with one seat: amend motion.md with what the chair wants settled, then run cross again. The pane keeps both passes.