# hermes-swarm

Multi-agent discussion swarm for Hermes Agent. Replace one expensive reasoning model with several cheap models that argue, refine, and converge.

## What it does

- Runs a round-robin discussion between Hermes profiles via Redis Pub/Sub.
- Each agent sees the full transcript before replying.
- First agent synthesizes the final answer.
- Saves transcript and final answer as Markdown.

## Install

```bash
git clone https://forgejo.redtask.ru/youruser/hermes-swarm.git
cd hermes-swarm
```

Install dependency:

```bash
pip install redis
```

## Quick start

1. Start Redis:

```bash
docker compose -f templates/docker-compose.redis.yml up -d
```

2. Add Hermes profiles (see `templates/hermes-profiles.yaml`).

3. Run:

```bash
python3 scripts/swarm_chat.py \
  --topic "Refactor this function to use async SQLAlchemy" \
  --agents swarm-architect swarm-critic swarm-coder \
  --rounds 3
```

## Environment variables

| Variable | Default |
|----------|---------|
| `REDIS_URL` | `redis://localhost:***@dataclass | `SWARM_OUTPUT_DIR` | `./swarm_outputs` |
| `HERMES_WORKDIR` | `.` |
| `HERMES_CMD` | `hermes` |

## Example use cases

- **Code review**: feed a diff, let critic + reviewer + coder discuss.
- **Design decisions**: compare two approaches with architect + critic.
- **Test planning**: let tester + coder + architect write a test strategy.

## License

MIT