Council

The Council — five seats, two models, one verdict

One agent gives you one answer, and you can't tell if it's the good one. The council fixes that structurally: five seats think about the question independently, each seat is a real, separate agent — not one model wearing five hats — the benches review each other's work blind, and the chair rules.

What makes this council different from a persona exercise: the seats run on different models. Codex and Claude disagree for real — different training, different instincts — so the tension between seats is genuine, not theatrical. (Methodology descends from Karpathy's LLM Council by way of Ole Lehmann's skill — see THIRD-PARTY-NOTICES.md; this implementation is Omnia Vault's own.)

Where it sits in the vault: sparring is depth — one rival attacking one locked plan. The council is breadth — five angles on an unlocked decision. Convene the council before locking a direction; spar the design after. A roadmap restructure or a big /intel adoption is exactly a council question.

The five seats (and who fills them)

The tensions are the design: Skeptic vs Maximalist (downside/upside), Rebuilder vs Operator (rethink it/ship it), Outsider keeping both benches honest. The driver (whichever agent runs the session) is the Chair — it frames, convenes, and rules, but never sits a seat.

Running a session

1. Frame (Chair, ~1 minute)

Pull the vault's context the cheap way — _relay/STATE.md, python scripts/plan_tool.py context if a plan exists, search-catalog on the topic — and write ONE neutral framed question: the decision, the options, the constraints and numbers that matter, and what it costs to get it wrong. No steering, no chair opinion. If the ask is too vague, ask the user exactly one clarifying question first. Check _relay/council/ for a prior session on the same ground before re-convening.

The Outsider gets a stripped version: the decision and options ONLY — no project name, no history, no vault context. What survives translation to a stranger is the signal.

2. Convene (all five seats in parallel, blind)

Independence rules — the whole value lives here:

• Seats never see each other's opinions.
• Each Codex seat is its OWN fresh codex exec thread (never resume one seat's thread for another).
• Each Claude seat is its own subagent with only the framed question.
• Launch everything in parallel: Claude seats via the Agent tool in one message; Codex seats as background bash calls.

Every seat gets: its seat identity + thinking style, the framed question (stripped version for the Outsider), and the instruction to argue its angle at full strength in 150–300 words — no hedging, no balance, the other seats cover the rest. A seat must end with one line: POSITION: <one-sentence stance>.

Codex seat mechanics (per the sparring skill's rules — stdin feeding, -o capture, 10-minute timeout; resolve $CODEX the same way):

# Skeptic / Operator — read-only, repo visible
"$CODEX" exec -s read-only --json -o "$OUT" - < "$PROMPT" 2>/dev/null >/dev/null
# Outsider — same, but from an EMPTY directory so the repo does not exist to it.
# --skip-git-repo-check is REQUIRED here: codex refuses untrusted non-git dirs
# without it (verified 2026-08-27 — the isolated call then reports it can see
# no project and no repo, which is exactly the point).
cd "$(mktemp -d)" && "$CODEX" exec -s read-only --skip-git-repo-check --json -o "$OUT" - < "$PROMPT" 2>/dev/null >/dev/null

An empty output file = a dead seat; relaunch it once, and if it dies again, report the council as four seats — never ghost-write a missing opinion.

3. Cross-bench review (two reviewers, anonymized)

Shuffle the five opinions into Opinion A..E (note the mapping privately; randomize order so no bench can be identified by position). Then two blind reviewers — each bench reviews the mixed set, without knowing which opinions are its own model's:

• One fresh Codex thread and one fresh Claude subagent each receive the framed question + all five anonymized opinions and answer exactly:
  1. Strongest opinion and why (pick one letter).
  2. Biggest blind spot in any single opinion.
  3. What ALL five missed.

Cross-model review is the point: each model grades work partly written by the other, and neither knows which is which.

4. The Chair rules

De-anonymize. The Chair reads everything and writes the verdict — in chat, in this exact shape:

## Council Verdict: <topic>
**The bench:** Skeptic ⚙ codex · Rebuilder ✳ claude · Maximalist ✳ claude · Outsider ⚙ codex (blind) · Operator ⚙ codex

### Consensus
<what multiple seats reached independently — high-confidence signal>
### Dissent
<the real clashes, by seat name, both sides stated fairly>
### Caught in review
<what only surfaced when the benches graded each other>
### The Chair's ruling
<one committed answer with reasoning — the Chair may side with a lone
dissenter against the majority, and says so when it does>
### First move
<exactly one concrete next step>

Rules of the ruling: no "it depends," no smoothing dissent into mush, and the minority position is preserved in writing even when overruled.

5. Keep the record

Write the full session to _relay/council/<date>-<slug>.md: framed question (+ the Outsider's stripped version), seat-by-seat opinions with their agent, both reviews, the verdict. Committed — a council is convened precisely because the decision is expensive, and expensive decisions deserve receipts. If the ruling changes the project: update the plan/roadmap or spar the resulting design, note it in the relay baton, and compile the decision into a Wiki note citing the council record. Gate, commit.

When Codex drives (the mirror)

Same session, benches swapped: Codex fills its seats with its own fresh threads and fills the Claude seats headlessly — claude -p "<seat prompt>" --permission-mode plan --output-format json, one call per seat, never reusing a session between seats. The Outsider trick inverts too: run the headless call from an empty directory. The driving agent chairs; the Chair never sits.

Hard rules

• One real agent per seat. The Chair sits zero seats. No seat sees another seat's opinion before the review round.
• The Outsider's isolation is mechanical (empty directory), not honor-system.
• Dead seats are reported, never impersonated.
• Anonymize before review, always — including from yourself as far as practical: shuffle first, map later.
• Don't convene for trivia, validation-seeking, or anything with a testable right answer. Five agents agreeing that Paris is the capital of France is not insight.
• The record is part of the verdict — no transcript, no council.