SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillscouncil

council

Agents & models

View repository

Updated 10-16 09:49

About this skill

council — lensed deliberation, one synthesized answer

Purpose

An advisory council over subscription CLI harnesses: each member answers the same question independently through a forced lens (Direct & Practical / Skeptic & Risks / Context & Tradeoffs), then a boss model synthesizes the anonymized answers into one decisive verdict that names its confidence, how aligned the members were, and the strongest unresolved dissent. No API keys.

When to use / when NOT to use

Use for decisions worth a second opinion: architecture calls, adopt-or-not questions, risk assessments, "is this a good idea". NOT for: seeing each model's raw answer (that is /advise), sourced/cited research (that is /explore), or trivial factual lookups (a council is 6 harness calls — 5 members + the boss).

Workflow

  1. Run from the researchkit repo root (or --directory a checkout):

    uv run researchkit council "QUESTION"
uv run researchkit council "QUESTION" --context-file notes.md
uv run researchkit -v council "QUESTION"        # -v appends full member answers

  2. Read the output top-down:

    • # Council answer (confidence: …, convergence: …) — the synthesized verdict. Low convergence means the members genuinely split; weigh it accordingly.
    • ## Dissent — the strongest unresolved disagreement. Treat it as signal, not noise: relay it to the user, never smooth it over.
    • ## Members — one line per member with lens, confidence, rationale (and failed: … for members that errored).

  3. If stderr warns "boss synthesis unavailable", the shown answer is the first valid member's (deterministic fallback) — say so when relaying.

Members/boss default to the harness preset in models.yaml (claude:claude-opus-5@xhigh, codex:gpt-6-astra@xhigh, agy:gemini-3.8-flash-high, grokcli:grok-4.6, kimicli:kimi-code/k3; boss claude:claude-opus-5@xhigh); override with --harnesses <specs…> and --boss <spec> (specs accept @<effort>). uv run researchkit doctor --preset harness verifies every pin before spending anything.

Failure modes

  • Individual member failures are isolated; the run errors only when every member fails (the message lists each member's error — usually a CLI not installed, not logged in, or out of subscription quota).
  • Expect ~2–10 min (members in parallel, then one boss call). Each harness call gets a 30-minute budget: a dense question at @xhigh can take a member past 10 minutes, and a tighter clock only costs you that member's answer. RESEARCHKIT_CLI_TIMEOUT=<seconds> overrides it.
  • Members answer from model knowledge — no web tools in deliberation.