council

TL;DR

A council is for questions where judgment matters more than lookup. The agent states the decision in one sentence, with the options, the code or artifact to inspect, the decision criteria and whether members may edit code. Model diversity comes first: the default roster is the strongest Claude model, the strongest GPT or Codex model and an open-source model such as Kimi or GLM, with any substitution noted; each also gets its own angle, like red-team or implementation cost. On an explicit request it tells you whom it will launch and waits for your approval. Read-only members share the checkout; members who prototype get their own git worktree and branch. Each reports the evidence inspected, current behavior, the alternative, risks, cost, a recommendation and confidence, and a thin report gets a follow-up question to the same member rather than a replacement. The final answer leads with the recommendation, notes consensus and disagreement, and separates what to do now from later hardening.

Summarizes the skill at 8792394.