council

/council — interactive multi-model debate

Drives the council engine: 7 distinct Claude models (Opus 4.8/4.7/4.6/4.5, Sonnet 4.6/4.5, Haiku 4.5) crossed with curated personas → opening positions → cross-examination debate → independent multi-arbiter verdicts → chief synthesizer. Your job: (1) pin down the real question, (2) let the user choose run options, (3) launch the workflow, (4) present results.

Step 1 — Understand the question

• If the user supplied a topic/question with the command, restate it back in one sentence to confirm intent.
• If it is vague, underspecified, or ambiguous, ask 1–3 short clarifying questions first — conversationally, or via AskUserQuestion when the choices are discrete. Pin down: the actual decision/question, any relevant context or constraints, and what a good answer looks like (a recommendation? trade-offs? a forecast? a critique?).
• Do not proceed until you can write one crisp, self-contained question. The engine receives only this single string plus a scale, so fold the gathered context, constraints, and "what a good answer looks like" into the question text itself.

Step 2 — Offer run options (single AskUserQuestion call)

Put the recommended option first and label it "(Recommended)".

• Scale (header "Scale"): Comprehensive (~30 agents, ~600k tokens) (Recommended) · Moderate (~16 agents, faster/cheaper) · Maximum (~50 agents, deepest/most expensive, slow).
• Output (header "Output"): Verdict + key disagreements (Recommended) · Full debate report (positions + rebuttals + verdict) · One-paragraph executive answer.
• Emphasis (header "Emphasis"), only if the topic invites it: Balanced (Recommended) · Decision/recommendation · Explore trade-offs · Stress-test a specific position. Fold the choice into the question framing — do not pass it as a separate arg.

Step 3 — Resolve the engine path

The engine council.js is bundled with this plugin under workflows/council.js. Build its absolute path at runtime:

• If $env:CLAUDE_PLUGIN_ROOT is set, the engine is "$env:CLAUDE_PLUGIN_ROOT\workflows\council.js" (Windows) / "$CLAUDE_PLUGIN_ROOT/workflows/council.js" (POSIX).
• If it is not set (e.g. running from a local checkout), locate it: Get-ChildItem "$env:USERPROFILE\.claude" -Recurse -Filter council.js | Select-Object -First 1 -ExpandProperty FullName

Step 4 — Launch the engine

Invoke by scriptPath (named resolution does NOT find file-based workflows):

Workflow({
  scriptPath: "<resolved absolute path to council.js>",
  args: { question: "<the crisp, enriched question>", scale: "<moderate|comprehensive|max>" }
})

• scale must be exactly moderate, comprehensive, or max.
• The run is backgrounded and takes several minutes. Tell the user it's running and the rough scale/cost; do not block — you'll be notified on completion.
• Optional: run once with args.dryRun: true first for a free preview of the roster (0 agents).

Step 5 — Present results

Render per the chosen Output option:

• Verdict + key disagreements: final.answer, final.confidence, the consensus bullets, and the live disagreements / final.minorityReport.
• Full debate report: lead with final.oneLineTakeaway + final.answer, then consensus, live disagreements / minority report, then a condensed table of panelist stances and the arbiter verdicts.
• Executive answer: just final.oneLineTakeaway and final.answer.

Always surface final.confidence and any notable minority report — the council's value is the divergence, not a false consensus.

Notes

• Models are fixed at the 7 distinct addressable ones; persona variety supplies the rest of the divergence. (claude-opus-4-1/-4-0 silently fall back to 4.8 and are excluded.)
• A full comprehensive run is ~30 agents / ~600k tokens — verify with dryRun first if cost matters.