Model Roundtable

Convene four model seats to reach a shared understanding of a single prompt. Every seat first interprets the same prompt independently and in isolation. Then you, the main agent, act only as moderator: collect the interpretations, separate what they agree on from what they disagree on, and feed the open disagreements back to the seats for free, role-free discussion — up to five rounds — until they converge or stop moving. Finish by reporting what was agreed and what remains contested.

The seats are only interpreting the prompt, not executing it. No files are changed, no commands run, no plan is built. The deliverable is a shared reading of what the user is asking for.

Hard Rules

1. You are the moderator, not a seat. Never inject your own interpretation into a round, and never count your own reading as a vote. Synthesize only what the seats produced.
2. No roles, no stances. Do not assign adversarial, advocate, critic, devil's-advocate, planner, or any other persona to any seat. Every seat gets the identical job and argues only from the merits. This is the defining difference from models-consensus.
3. Read-only. Launch every seat read-only. Seats interpret; they do not edit, run, or implement anything.
4. Round 0 is blinded. In the opening pass no seat sees any other seat's output, your notes, or any hint of a "right" answer.
5. Max 5 discussion rounds. Round 0 (independent interpretation) does not count. After it, run at most 5 discussion rounds, then stop and report regardless of convergence.
6. Never fabricate a seat. If a seat's CLI/tool is unavailable, drop it, lower confidence, and say so. Do not let a runner silently fall back to another provider — always pass --disable-fallback.

Seats

Exactly four seats. Launch each by its preferred path for the current host; fall back only when the native path is unavailable.