consilium

Convene a decision/opinion council on the cli-council engine (council decide).

Input. If this ran as /consilium <decision>, that argument is the decision. If it auto-triggered from the conversation, extract the decision to weigh yourself — the forks, the constraints, and what a good answer must address. If the decision is empty or too vague to frame, ask ONE clarifying question before firing (a run is $0 via the user's subscriptions but burns a few minutes of wall time). If the request is a trivial or quick-fact question, do NOT convene the council — just answer.

What this is

A decision/opinion council on the cli-council engine (council decide, pipeline.run_decide_pipeline) — the SAME engine as consreview, pointed at a DECISION instead of a diff. One council decide call runs the whole shape: every enrolled voice gives a recommendation → anonymized peer-ranking → a chairman synthesizes ONE recommendation with the action tiers → a mandatory audit panel gates it (worst-wins — a rule, not a chairman, so no model can soften the gate). Decision mode = recommendations + trade-offs + action tiers, NOT code-review SHIP/FIX verdicts. $0 (subprocess to the subscription CLIs you authed), stdlib-only.

Why audit but no redteam by default: a redteam refutes a claim — high-signal for code (a bug is true or false), but a decision has no ground truth to refute. The load-bearing guard for a recommendation is the audit — is the synthesis faithful to the raw voices (no invented convergence, misattribution, or action-tier inflation). Opt into adversarial pressure on the recommendation with --redteam.

Voices: whatever you enrolled in council.toml ([council].voices, and the [decide] audit/redteam panels). A decision needs ≥3 model families — a family quorum, so no single vendor's models decide alone; the shipped council.example.toml enrols a diverse keyless CLI roster that clears it. Two voices of one house (e.g. two Claude models) count as one family.

Step 0 — Exa grounding (before any voice runs)

Run an Exa search on the decision's factual premises BEFORE launching the council and put the findings, with their URLs, into question.md: facts, not guesses. If Exa fails, label the run "ungrounded" in the question and in the report. Consilium is not a review loop:

• the prompt asks each voice for its perspective, not for bugs;
• the output is opinions with arguments, one round, then a synthesis and the user decides;
• Claude is the synthesiser, not a vote counter: one unique minority insight can outweigh five agreeing voices; resolve by the project concept and the evidence.

Procedure

1. Preflight. council must be on PATH (run the repo's install.sh, or put bin/ on PATH) and council.toml must enrol your voices — the blessed path is python3 installer/doctor.py enroll agy grok muse chatgpt (the roster below; a re-enroll keeps the chairman already in council.toml): it re-smokes each and writes a GATED config with the [decide] audit panel already filled in from your voices. Hand-copying council.example.toml works too, but its [decide] block ships commented out — leave it commented and the run is unverified (no audit gate, no AUDIT_VERDICT.md). The engine auto-discovers council.toml (cwd or the repo root); an explicit --config <path> to a file that doesn't exist fails loud (it won't silently fall back to one voice).

2. Frame the decision → write it to a file so a long / multi-line decision pipes cleanly on stdin (the engine reads a positional arg OR stdin):

   DIR="$(mktemp -d)/consilium"; mkdir -p "$DIR"
   # write the framed decision to "$DIR/question.md": state the forks, the constraints,
   # and what a good answer must address — a decision, not a vague prompt.

3. Run the decision council (background it — a few minutes with the slow reasoning voices and the audit panel):

   council decide --out "$DIR" < "$DIR/question.md" 2>"$DIR/engine.log"
   echo "exit=$? · dir=$DIR"; tail -12 "$DIR/engine.log"

   • A short one-liner instead of the file: council decide "<the question>" --out "$DIR".
   • --redteam requested → append --redteam <v1,v2> (a lean adversarial panel that stress-tests the recommendation; off by default because a decision has no ground-truth claim to refute).
   • --no-verify → append --no-verify — skips the audit panel → status unverified, no gate. Use only for a quick unguarded opinion sweep; an unverified run is NEVER reported as a clean recommendation.
   • --voices a,b,c narrows the roster; --chairman X changes the synthesizer. Peer-ranking is intrinsic — there is no --no-rank.
   • Non-zero exit = the council itself failed — all voices dead, or fewer than 3 model families answered (the family-quorum abort) — read engine.log. A degraded status is exit 0: a gate flagged the synthesis (steps 5–6).
4. Artifacts in $DIR: SYNTHESIS.md (the recommendation) · v-<voice>.md (raw recommendations) · a-<voice>.md (audit panelists) · AUDIT_VERDICT.md · RANKINGS.md · MECHANICAL.md · pipeline-status.json (machine-readable gate) · PIPELINE_DEGRADED.md (only when not clean) · decide_prompt.md (the exact prompt). With --redteam: also REDTEAM_VERDICT.md + r-<voice>.md. Read pipeline-status.json first for the verdict, then the prose in SYNTHESIS.md.

5. Read pipeline-status.json + SYNTHESIS.md + AUDIT_VERDICT.md (+ RANKINGS.md) and report:

   • Recommendation (the one-line lead of SYNTHESIS.md) + per-voice stances (which voices converged / which split) + status (clean / degraded [finding|infra| mixed] / unverified) + opinion_errors.
   • Peer leaderboard (RANKINGS.md): top / bottom voices. A weighting SIGNAL only — never present rank as proof, and never drop a voice because of rank.
   • Key convergences (2+ voices) with verbatim quotes · disagreements.
   • Audit verdict — AUDIT_VERDICT.md first line # CLEAN | ISSUES | INVALID (worst-wins across panelists) + which panelist(s) set it. (redteam = SKIPPED on a default run — expected; a REDTEAM_VERDICT.md appears only with --redteam.)
   • Action classification — SYNTHESIS.md headings `## BLOCKER / ## IMPORTANT /

     CHECK / ## ACCEPT / ## NOISE`. Present in that order.

   For CHECK items, verify before presenting as actionable. For BLOCKER/IMPORTANT items with specific numbers/paths, spot-check against the actual source/data first.

   status: degraded is NOT a failure — a gate flagged the synthesis. degraded_kind: finding = a verifier caught a real problem (hand-verify before applying); infra = a verifier or the chairman synthesis couldn't complete (re-run); mixed = both.

6. 🚨 Re-synth gate — MANDATORY when AUDIT_VERDICT is ISSUES / INVALID: do NOT apply raw SYNTHESIS.md. The engine gates but does not auto-re-synth (an auto-loop would let the panel out-vote a correct minority). Re-synth + primary-source verify: a. Read each v-*.md directly (don't trust SYNTHESIS.md counts) + the flagging a-*.md. b. Rebuild by hand: recount convergences, drop inventions (phrases not literally in any voice), fix attributions, add missed convergences. c. Verify each claim with a concrete identifier (number, %, date) — 2+ independent sources. d. Apply only verified; consilium-only = tentative; disproven → drop.

Auto-Consilium (auto-decide)

• Activation: the user says "auto-decide"; it stays on until "stop auto-decide" or the session ends.
• Scope: the implementation phase only. Discovery, research and plan decisions always go to the user.
• Rule: where Claude would ask the user an implementation question, it runs this skill's council instead and acts on the synthesis. At most 5 auto-consiliums per task. A wrong answer that costs under 1 h of rework → decide; one that costs over a day → ask the user.
• Never auto-decided (always ask the user): permission-gated actions, spending over $1, scope expansion, security decisions, anything contradicting the concept or charter.
• Self-bias safeguard: write the initial lean BEFORE launching the voices. Confidence high (3 of 4 voices or more, plus evidence) → decide; medium (2 of 4) → decide with caution; low (split or weak) → ask the user. A council that merely confirms the lean with less than high confidence → ask the user.
• Log: append to the project's consilium-log.md: question, initial lean, votes, audit verdict, confidence, decision.

Fallback chain (optional)

Voices can declare fallbacks = ["name1", "name2"] in council.toml. If the primary fails (usage limit, version error, timeout), fallbacks are tried in order; the first success wins. A dead-provider cache ensures a failed primary is skipped instantly on subsequent stages (stage 2/3/panels pay zero retry cost). Codex voices can add auth_check for a quick usage-limit preflight (~20s instead of 30min hang). SSH-based fallback providers (bin = "ssh") get automatic ControlMaster warm-up at pipeline start.

All fallback features are opt-in via council.toml — a fresh install with no fallbacks configured behaves identically to before.

Dependencies

• cli-council on PATH (bin/council → python3 -m council; Python 3.11+, stdlib only)
  • a council.toml enrolling your voices (per-machine, git-ignored — copy council.example.toml).
• The subscription CLIs you enrolled, each authed (python3 installer/doctor.py smoke <voice> PASSES).

Related

• consreview — code-review council on the same engine (council review; it adds a redteam panel — a bug has ground truth to refute).
• commands/consilium.md — the explicit slash-command form of this exact procedure.
• Engine internals: council/{decide,pipeline,audit,stages}.py.

Required roster (2026-10-01)

Every consilium runs all four outside voices: Gemini (agy), Grok (grok-p), Muse (muse-batch) and the OpenAI voice: ChatGPT Chat, falling back to dot, then Sol (gpt-6.1-sol), for an ordinary council (owner rule 2026-10-21), Astra (gpt-6-astra at "extra" = xhigh) for a heavy one (skill codex-cli, section "Sol or Astra — which one"). The roster, routes and fallbacks live once, in the global contract (~/.claude/rules/team-rules.md, from claude-setup contract.md, section "Review Cycle", "The four outside voices"); this section only says how this skill meets it.

• council.toml enrols agy, grok, muse, chatgpt by default (chairman chatgpt; both use the chain chatgpt → dot → sol; sol on a Mac without a Chat route; decide audit agy, grok, muse), so the procedure's bare council decide runs them; a heavy decision (costly trade-off, or stuck after a Sol council) runs --voices agy,grok,muse,astra --chairman astra. Never pass a --voices list that drops a family. The Claude voices, the terra/sol codex voices and agy38 remain opt-in.
• After the run, check pipeline-status.json: every voice in reviewers, none in opinion_errors. Since 2026-10-01 a missing voice makes the run degraded (required voice <v> missing). Four houses leave a family-quorum margin of one: two missing voices abort the decide.
• A missing voice is re-probed and re-run, never replaced by another model (no Qwen, DeepSeek, Mistral, default Codex, Claude or OpenRouter stand-in); if it still cannot run after both routes and retries, report it as missing.