council-debates

Run multiple independent LLM roles against the same prompt, optionally rank their responses, and generate a final answer through a chairman model.

council-debates is a small modular Python 3.11+ package for running debates between independent LLM roles through OpenAI-compatible APIs.

The pipeline is deliberately strict:

1. Each role is called separately with only its own system_prompt and the original user prompt.
2. (Optional) A separate ranking LLM receives the original prompt plus all role responses.
3. A final chairman LLM receives the original prompt, role responses, and the optional ranking result, then writes the final Markdown answer.

You can omit the ranking block entirely. In that case the chairman aggregates role responses without an external ranker, which prevents the ranker from biasing the chairman's decision.

The package supports OpenAI, OpenRouter, LM Studio, Ollama OpenAI-compatible endpoint, vLLM, LocalAI, and any /v1/chat/completions-compatible provider.

Pipeline

User Prompt
     │
     ▼
┌───────────────┐
│ Role: Judge   │
├───────────────┤
│ Role: Skeptic │
├───────────────┤
│ Role: Builder │
└───────────────┘
     │
     ▼
┌────────────────────┐
│ Ranking (optional) │
└────────────────────┘
     │
     ▼
┌────────────────────┐
│ Chairman           │
└────────────────────┘
     │
     ▼
Final Markdown Answer

Project layout

council_debates/
├── CHANGELOG.md
├── pyproject.toml
├── README.md
├── config.example.yaml
├── .env.example
├── requirements-dev.txt
├── tests/
└── src/
    └── council_debates/
        ├── __init__.py
        ├── __main__.py
        ├── cli.py
        ├── core/
        ├── pipeline/
        ├── presentation/
        ├── providers/
        └── templates.py

Install

Install the latest released version from PyPI:

pip install council-debates

Upgrade an existing installation:

pip install --upgrade council-debates

PyPI package:

• https://pypi.org/project/council-debates/

Install the latest code directly from GitHub:

pip install "git+https://github.com/ssh-den/council-debates.git"

Install a specific branch, tag, or commit:

pip install "git+https://github.com/ssh-den/council-debates.git@main"

For local development and tests from the repository:

python -m venv .venv
. .venv/bin/activate
pip install -r requirements-dev.txt

Quality checks:

black --check .
ruff check .
mypy src tests
pylint src tests
pytest

Create a config

council-debates init-config
council-debates init-env
cp .env.example .env

Edit config.example.yaml or save it as config.yaml. Set the environment variables referenced by api_key_env.

Run

Prompt from the command line:

council-debates run "Suppose that a judge or magistrate..." --config config.yaml

Prompt from a file:

council-debates run --file prompt.txt --config config.yaml

Save Markdown output:

council-debates run "test prompt" --config config.yaml --output result.md

Show provider/model logging:

council-debates run "test prompt" --config config.yaml --verbose

Configuration

All behavior is configured in YAML:

• providers: OpenAI-compatible base URLs and API keys or api_key_env references.
• defaults: default temperature and max_tokens.
• roles: independent debate roles; each role has its own provider, model, and system prompt.
• ranking: (optional) separate ranking model config. Omit it if you want the chairman to aggregate responses without external ranking.
• chairman: final aggregation model config.
• output: Markdown output options.

Example provider entries:

providers:
  openai:
    base_url: "https://api.openai.com/v1"
    api_key_env: "OPENAI_API_KEY"
  openrouter:
    base_url: "https://openrouter.ai/api/v1"
    api_key_env: "OPENROUTER_API_KEY"
  local:
    base_url: "http://localhost:11434/v1"
    api_key: "ollama"

Local OpenAI-compatible servers usually accept any non-empty API key.

Design Principle

Roles never see each other's responses.

Each role receives only:

• its own system prompt
• the original user prompt

This prevents early consensus, role contamination, and chain-of-thought leakage between participants.

Contributions / Acknowledgments

The original idea of using multiple LLM roles in a council/debate setup comes from Andrej Karpathy's llm-council.