Models Consensus

Produce a decision report for a user who wants more than one model's opinion. The next consumer is the user. Done means the report names the recommendation, dissent, evidence limits, one next step, configured and observed seats, and receipt limits.

This is a user-controlled, read-only workflow. A council is justified only when the user decides that independent opinions are worth its time and cost.

Entry and approval

Run only after the user explicitly invokes models-consensus or explicitly asks for extra model opinions. Another skill may suggest this workflow as a manual next step, but must not invoke it.

Before dispatching any model:

Choose poll by default, debate for a contested direction, or personas when the user wants five angles from one chosen model.
Read references/role-routing.md, resolve the current seat and requested-model mapping from shared/references/model-roster.md, and make read-only transport checks. If .rar-skills/config.local.yaml exists, read its advisory seats and models only while forming this preview, as shared/references/local-config.md defines. Validate every value against the roster and runner, let direct user instructions win, and never reread it after approval.
Show the approval preview below and ask: "Approve this council plan, or specify changes."
Wait for a clear approval. Silence, a timeout, a prior general approval, an auto request, or another skill's instruction is not approval.

The preview must show every planned and conditional call: