@jaspergrant/constellation

Constellation

Secure parallel agent swarm for confidential work. The whole mission runs inside Flower:

ServerApp = trusted coordinator. It decomposes the mission, sends each worker only its own task and context slice, quarantines and replaces an attacked worker, and synthesizes the final result (services/swarm.py).
ClientApp = untrusted worker, one per SuperNode (services/flower_app.py).

A compromised worker cannot leak information it never received.

Run with Flower

uv sync
uv run flwr run .                                    # built-in AI-research demo mission
uv run flwr run . --run-config 'attack="literature"' # inject an attack mid-mission
uv run flwr run . --run-config 'prompt="Confidential: ..."'

Set options.num-supernodes (in ~/.flwr/config.toml) to at least the number of workers plus one spare node for replacements, e.g. 10.

The FastAPI backend (uvicorn main:app) runs the same coordinator as one long-lived ServerApp; POST /mission and POST /workers/{id}/attack are handed to it. Set CONSTELLATION_RUNTIME=local to run the coordinator over local threads instead of Flower.

Optional model calls

Set FLWR_MODEL_API_ENDPOINT, FLWR_MODEL_ID and FLWR_MODEL_API_KEY (OpenAI Responses-compatible endpoint) on the nodes, or put them in a local .env (never committed or bundled). Without them, workers return deterministic seeded output.

Tests

uv run pytest