council-of-llms（法学硕士理事会）

Council of LLMs

Overview

A real council spawns 3 parallel subagents, each with a different model and perspective, then synthesizes their outputs into a unified verdict. This is NOT one model roleplaying 3 experts — it's genuinely different models providing independent analysis.

Models

Read from ~/.openclaw/council-config.json:

{
  "council_models": [
    "ollama/kimi-k2.6:cloud",
    "ollama/deepseek-v3.2:cloud",
    "ollama/gemma4:31b-cloud"
  ],
  "default_timeout": 900,
  "max_tokens": 8192
}

Perspectives

Each model gets a different lens:

How to Run a Council

Step 1: Prepare the Context

Gather all relevant data BEFORE spawning. Council agents cannot browse the web or access your conversation history. Paste everything they need inline.

Step 2: Spawn 3 Parallel Subagents

sessions_spawn(
  runtime: "subagent",
  mode: "run",
  model: "ollama/kimi-k2.6:cloud",
  label: "Council-Strategos",
  lightContext: true,
  runTimeoutSeconds: 900,
  task: "You are Strategos, a strategic analyst. [PASTE CONTEXT HERE]
  
  Analyze from a STRATEGIC perspective:
  - Business impact and feasibility
  - Market positioning and competitive advantage
  - Resource requirements and ROI
  - Strategic risks and opportunities
  
  Return your analysis as a structured review with: verdict, conditions, risks, recommendations."
)

sessions_spawn(
  runtime: "subagent",
  mode: "run",
  model: "ollama/deepseek-v4-pro:cloud",
  label: "Council-Analyticos",
  lightContext: true,
  runTimeoutSeconds: 900,
  task: "You are Analyticos, a data and logic analyst. [PASTE CONTEXT HERE]
  
  Analyze from an ANALYTICAL perspective:
  - Data quality and completeness
  - Technical correctness and edge cases
  - Statistical validity and sample sizes
  - Logical consistency and contradictions
  
  Return your analysis as a structured review with: verdict, conditions, risks, recommendations."
)

sessions_spawn(
  runtime: "subagent",
  mode: "run",
  model: "ollama/gemma4:31b-cloud",
  label: "Council-Creativos",
  lightContext: true,
  runTimeoutSeconds: 900,
  task: "You are Creativos, a creative and UX thinker. [PASTE CONTEXT HERE]
  
  Analyze from a CREATIVE perspective:
  - User experience and usability
  - Novel alternatives and unconventional approaches
  - Design and presentation improvements
  - What's missing that no one else would think of
  
  Return your analysis as a structured review with: verdict, conditions, risks, recommendations."
)

Step 3: Synthesize

When all 3 return, merge their verdicts:

1. Consensus points — where all 3 agree
2. Disagreements — where they differ and why
3. Blind spots — what none of them caught
4. Final verdict — weighted synthesis with conditions
5. Action items — concrete next steps

Write the synthesis to council-review-[topic].md.

Critical Rules

1. Paste ALL context inline — agents have no conversation history
2. Keep task descriptions under 2000 words — longer = context overflow = failure
3. Use lightContext: true — always, to prevent context bloat
4. Set runTimeoutSeconds: 900 — councils need time
5. Don't spawn with too much data — if pasting 10k+ words, summarize first
6. Wait for ALL 3 to complete — don't synthesize with 2/3 results
7. Never re-spawn — if one model times out, note it in the synthesis

Common Failure Modes

Anti-Patterns

• ❌ Spawning one subagent and asking it to "be 3 experts" — that's roleplay, not a council
• ❌ Pasting 10k+ words of raw data — summarize first
• ❌ Using the same model for all 3 perspectives — defeats the purpose
• ❌ Synthesizing before all 3 complete — wait for everyone
• ❌ Ignoring disagreements — disagreements are the most valuable output