Council

/council — Parallel Critics, Then a Separate Synthesis

Dispatch 3–4 critic agents in parallel on the same input, collect their raw critiques, then run ONE synthesis pass in a fresh context. Hard cap: 4 critics. Single round only — critics speak once, the synthesizer reads once. Never turn the critiques into a majority vote.

Why this shape:

• Parallel, not sequential — independent critics with different lenses catch different problems, and nobody anchors on whoever spoke first.
• One round only — no debate or rebuttals. Multi-round AI debate drifts toward consensus, not accuracy.
• Separate synthesis — a fresh-context pass reads the critiques as data instead of being swayed by the conversation that produced them.

The four shipped critics

Persona files live at ${CLAUDE_PLUGIN_ROOT}/agents/<name>.md.

Default panels

Paper and grant-proposal panels need additional personas — future release. If the user asks for one, say so and offer the plan panel as the closest fit.

Inferring the type when --type isn't given: "plan / design / architecture / workflow" → plan; "should I / decide / go or no-go / accept or reject" → decision; unclear → ask one line: "Is this a plan review or a decision?"

When NOT to run a council

• Quick lookups, single-file edits, casual brainstorming — just answer.
• Inputs under ~50 words of substance — there isn't enough for three critics to disagree about. Offer a quick single review instead.
• If it's unclear whether the user wants a council run or is just discussing councils as a topic, ask one line before dispatching.

Steps

1. Parse and resolve

1. Identify the input. If file:<path> is given, Read that file — its content is what the critics review. Otherwise the topic is the text or recent conversation content the user pointed at.
2. Resolve the panel: explicit --panel wins; else --type; else infer from keywords; else ask.
3. Cap at 4 critics. If the user asks for more: "Hard cap is 4 — more voices adds noise, not signal. Pick a tighter panel."

2. Check the persona files exist

For each persona in the panel, Glob ${CLAUDE_PLUGIN_ROOT}/agents/<name>.md. If any file is missing, stop with a clear message:

"Council can't run: the persona file for <name> is missing from the starter-kit plugin. Reinstalling the plugin usually fixes this, or pass --panel with personas that exist."

No silent substitutions for missing personas.

3. Optional peer swap (--peer codex or --peer gemini)

By default all critics are Claude agents — that works with zero extra setup. --peer swaps ONE critic for a different AI engine, so the panel isn't all one model family.

0. Privacy check and explicit confirmation (required). A peer critic sends the full input to another AI service. If the input contains confidential records, personal contact details, research-participant or human-subjects data, or anything the user wouldn't paste into a public website — skip the peer swap, say why in one sentence, and run the all-Claude panel. If the privacy check passes, ask: "Send this plan to Codex/Gemini as one council critic? (yes/no)" Proceed with the peer critic only if the user answers yes; otherwise run the all-Claude panel.

1. Read ~/.claude/starter-kit/engines.json (written by /kit-setup). The engine counts as available only if ALL of these hold:

   • the file exists and parses as JSON (missing or broken file = every engine unavailable — that's normal, not an error),
   • the engine's top-level entry (codex or gemini) has "available": true — the flat schema in ${CLAUDE_PLUGIN_ROOT}/references/engines-contract.md is the only one,
   • command -v codex (or command -v gemini) still finds the CLI — the config can go stale; the live check wins.

   If any check fails, print exactly:

   "Running on Claude only — that's the normal setup and everything here works. (Have ChatGPT or Gemini? Run /kit-setup to plug them in.)"

   and continue with the all-Claude panel.

2. If the engine is available, pick the critic it replaces: codex replaces skeptic (closest overlap — rigor and failure-finding); gemini replaces completeness-checker (broad coverage). If the replaced persona isn't in the panel, replace the last critic in the panel.

3. Build the peer prompt: the replaced persona's role and output shape (from its persona file), the same input the Claude critics get, and the line "End with VERDICT: APPROVE or REVISE + brief rationale. Do not reconcile with other critics — a separate synthesizer does that."

4. Create two temp files in one Bash call — echo "$(mktemp /tmp/council-peer-prompt.XXXXXX) $(mktemp /tmp/council-peer-out.XXXXXX)" — note both paths, then use the Write tool to put the peer prompt in the first one.

5. In Step 4's parallel message, include ONE Bash call alongside the Task calls:

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/codex-run.sh" <out-file> < <prompt-file>
   # or, for gemini:
   bash "${CLAUDE_PLUGIN_ROOT}/scripts/gemini-run.sh" <out-file> < <prompt-file>

6. If the script fails or reports the engine unavailable, don't fail the council — note "peer critic unavailable, continuing with Claude critics" and synthesize from the critics that did return.