Council: the debate framework (bring your own seats)

A council is a structured way to get MORE than one model's single opinion: several independent seats, each carrying a distinct lens, that reason separately and then argue to a consensus. It exists because one model's one pass has one set of blind spots. Diverse seats that must defend their claims against each other catch both false positives (a claim that does not survive scrutiny) and surviving bugs (a real problem the happy-path pass missed).

This skill is the pattern, not a fixed panel. You define the seats for your domain. A code team defines correctness / security / simplicity seats; a research team defines different ones. The protocol below is the same regardless.

When to use

• Any decision or artifact important enough that one model's single review is not enough: a design, a diff, a plan, a risky change.
• As a gate inside a larger pipeline (e.g. a conductor's review phase).
• Whenever you catch yourself trusting a lone "looks good to me."

Define your council (this is the only thing you supply)

A council is a list of seats. Each seat is config the agent READS:

seat:
  name: <who/what this lens is>
  lens: <the ONE thing this seat interrogates, and how it judges>
  output: severity (BLOCKER/HIGH/MEDIUM/LOW) + confidence (0-1) + file:line or section evidence

Seats can be role lenses ("the security reviewer") or embodied personas ("a veteran operator who has run this in production"). Both work; the honesty rail below governs personas.

Isolation ladder (Codex)

Blind seats require separate model contexts, not three headings written by the driver. Before Round 1, probe the live host once:

1. Use native Codex subagents when one test dispatch succeeds and returns a real task identifier.
2. If the dispatch errors, returns no task identifier or reports that the parent thread does not exist, do not wait on it and do not retry it. Immediately run one fresh codex exec --ephemeral process per seat against the same frozen commit. Give each process only the common evidence packet plus that seat's lens, write each raw output to a different file and wait for all process exit codes before cross-critique.
3. Verify codex --version and codex login status before using the CLI fallback. If either fails, mark the independent council unavailable and stop. Never manufacture blind agreement inside the driver context.

The fallback is still Codex-first. It changes process isolation, not the reviewer vendor. Never recursively ask a fallback reviewer to convene another council.

A neutral worked example (a generic code-review council) ships in references/example-code-review-council.md - copy it and swap the seats for yours.

Mode A - roundtable (default, fast)

Round 1 only. Each seat, as an INDEPENDENT and BLIND pass (it has not seen the other seats' output), gives a thesis (a committed ruling, not a hedge) plus a confidence 0 to 1. Synthesize the theses. Use when you want diverse coverage quickly and the stakes do not warrant a full debate.

Mode B - debate protocol (merge-blocking, three rounds)

1. Round 1, independent and BLIND. Each seat is a SEPARATE pass that has NOT seen any other seat's output. It writes a thesis: a load-bearing argument, a ruling, and a confidence 0 to 1. A seat that hedges without ruling has failed its round. Independence here is non-negotiable; it is what preserves the blind-spot coverage the whole method exists for.
2. Round 2, cross-critique. Every seat now receives the others' theses and attacks or defends them. Claims get strengthened with new argument, softened, or killed. This round kills false positives AND surfaces bugs the first round missed (the seats review each other's reviews).
3. Round 3, consensus + dissent register. ONE unified verdict from what survived, recording what changed between rounds, per-finding confidence, and a dissent register of any disagreement that did NOT resolve. Bounded to 3 rounds; if it does not converge, surface the deadlock honestly rather than faking consensus.

Cross-cutting doctrine (applies to every council)

• Diversity charter. Consensus among diverse seats is signal; consensus among similar seats is a shared blind spot wearing a badge. Make the seats genuinely different lenses, and where possible run at least one seat on a different model/vendor (different training, different blind spots).
• Shared-blind-spot rule. Answer the upstream question before the downstream one (e.g. "is the data / premise sound?" before "is the method good?"). A council that all assumes a bad premise agrees its way off a cliff.
• Honesty framing. Embodied personas are an ANALYTICAL DEVICE, not real people. Never fabricate a quote or a credential; any quoted text must be a real, cited quote. A persona's value is the lens, not a fake biography.
• Per-seat known-limits. Each seat states what it is NOT positioned to judge, so its silence on something is not mistaken for a pass.
• Findings are hypotheses. A council finding is a claim to validate against the actual artifact/code before it gates anything. Cite evidence (file:line or section). An unvalidated finding never blocks on its own.

Output schema (what a council returns)

A ranked findings list, each with: severity (BLOCKER / HIGH / MEDIUM / LOW), confidence (0 to 1), evidence (file:line or section), and a one-line statement of the defect. Then a verdict (e.g. GO / BLOCKED, or adopt / adopt-with-fixes / reject) and the dissent register. Nothing else. Severity is "how bad if real"; confidence is "how sure it is real" - keep them separate, and give a high-severity / low-confidence finding extra validation before it gates.

Hard rails

• Config the agent READS, never a runtime it executes. A council is a lens the agent applies, not an engine. Do not build a parser that "runs" councils; that forfeits portability across harnesses.
• Independence before debate. Never let seats see each other's Round 1 output. If they do, you have one opinion in three costumes.
• Dissent is preserved, never averaged away. A minority finding that survives the debate is logged with its rationale, not erased by "consensus."
• No fabricated quotes or credentials (see honesty framing).