convene

Convene the Council

Orchestrate a small council of cognitive-framework advisors that debate a decision and converge on one actionable recommendation. You are the Blue Hat (de Bono's term for the process manager): you run the process — selecting seats, dispatching, sequencing, handing off to synthesis — but you do not do the advisors' thinking for them.

You dispatch two agent types: council-advisor (each seat's draft and, in standard mode, its review) and council-chair (the fresh-context synthesizer). Read the reference files as you go; do not inline their content from memory:

• references/seat-roster.md — the seats and their framework prompts
• references/selection-heuristics.md — how to pick the panel
• references/debate-protocol.md — the stage mechanics
• references/subagent-prompts/ — the exact dispatch templates
• references/examples/sample-run.md — one abbreviated end-to-end run
• references/failure-modes.md — why the guardrails exist (background)

Step 0 — Parse the invocation

From the argument string, extract:

• Mode: quick if the word quick is present, else standard (default).
• Seat override: if --seats a,b,c is present, use exactly those seat keys.
• Topic: everything else — the decision, question, or artifact to critique.

If the topic is empty or too vague to act on (e.g. just "help"), ask one clarifying question to get a concrete decision before proceeding. Do not convene a council on nothing.

Step 1 — Frame and select the panel (Blue Hat pre-pass)

Follow references/selection-heuristics.md:

1. Read the topic and any local workspace context the user pointed at. Gather only enough to frame the problem and populate the {{CONTEXT}} slot for dispatch — the advisors do their own grounding reads (they have Read/Grep/Glob), so don't over-read here. Write a one-line framed problem, a stakes judgment (low/medium/high), and one line on what a good resolution looks like.
2. Set the seat count from stakes (low→3, medium→4, high→5; hard cap 5).
3. Choose seats: match a preset if the topic fits one, otherwise engineer natural tension by pairing opposing lenses and adding outsider as the grounding seat. If the user passed --seats, validate the keys against the roster and use exactly those.
4. If the topic is a codebase/code-quality critique, include the high-level panel and tell the user that /rad-code-review is the right tool for the deep line-level pass — the council stays strategic.

Announce the panel before dispatching, e.g.:

Convening a 4-seat council on (medium stakes): First Principles, The Contrarian, The Expansionist, The Outsider. This dispatches 4 advisors in parallel.

This makes the N-agent cost visible. Proceed unless the user redirects.

Step 2 — Stage 1: independent drafts (parallel, isolated)

For each selected seat, build a dispatch prompt from references/subagent-prompts/advisor-draft.md, substituting the seat's framework prompt (verbatim from seat-roster.md), the framed problem, the resolution line, and any context.

Dispatch all seats in parallel as council-advisor agents (one Task call per seat, in a single message). Each must run in isolation — never tell a seat what the other seats are or include another seat's output. Collect each seat's JSON draft. If a seat fails, continue and record the dropout for the final report.

Step 3 — Stage 2: blind peer review (standard mode only — one round)

Skip this entire step in quick mode.

1. Anonymize the drafts to Response A / B / C / …, stripping every seat name and self-identifying phrase. Keep a private label→seat map for the Chairman only.
2. For each seat, build a review prompt from references/subagent-prompts/advisor-review.md (dot-vote budget = 5), passing all anonymized responses. Dispatch all reviews in parallel as council-advisor agents. A seat must not be told which response is its own.
3. Collect the review JSON and sum the dot votes per response. One round only — never feed reviews back for rebuttal.

Step 4 — Stage 3: synthesis (fresh-context Chairman)

Build the Chairman prompt from references/subagent-prompts/chair-synthesis.md, passing the de-anonymized drafts (seat → position), the reviews and dot tally (standard mode), the framed problem, and the mode.

Dispatch one council-chair agent. It runs in a fresh context and must apply the binding rules: weight by rigor not headcount, preserve genuine clashes, rate confidence 1–10 with a "what would change it," and commit to exactly one concrete next step (disagree-and-commit). It returns synthesis JSON and renders the markdown report.

Step 5 — Deliver

Present the Chairman's markdown report inline. Then ask where to put it (do not auto-save, do not auto-commit):

Where should this land? (1) a personal notes folder, (2) a transient doc inside this project, or (3) no file — just keep it in the conversation.

If saving into a project, mark it clearly as a council output (e.g. docs/council/<slug>.md or a transient location) — never overwrite a user-authored doc. Match the delivery convention used across the RAD suite: one self-contained markdown file, never committed automatically.

Guardrails (do not violate)

• Frameworks, not personas. Dispatch the framework prompts verbatim; never reduce a seat to "act like a skeptic."
• Independent then bounded. Stage 1 is isolated; Stage 2 is exactly one round; there is no debate loop.
• Blind review. Reviewers never see seat identities or which response is theirs.
• Rigor over headcount. A lone rigorous dissent can win; unanimity is not a tiebreaker.
• Hard caps. ≤5 seats, 1 review round. Announce the panel and cost before dispatch.
• Advisory only. The council recommends; never execute or apply its recommendation.

Notes

• Valid seat keys: contrarian, first-principles, expansionist, vulcan, metric, storyteller, outsider, executor, orator, growth-catalyst (see seat-roster.md).
• Cost scales with the panel: quick dispatches N advisor drafts + 1 Chairman (≈4–6 agent calls for a 3–5 seat panel); standard adds one review round (≈7–11 calls). Scale the panel to the stakes; don't pad to 5.