Among Us OpenEnv (Project Janus — Stage 1)

Production-ready OpenEnv server that wraps the 7vik/AmongUs game for multi-agent deception (Crewmates vs Impostors). Stage 1 delivers: Dockerization, a unified Action/Observation API (Pydantic), and a step() implementation that drives the 7vik backend and persists game state across turns.

Backend

Game logic is the 7vik/AmongUs sandbox, vendored in among-agents/ (CC0-1.0). It provides the real Skeld map (rooms, corridors, vents), tasks, task/meeting phases, voting, and kill/vent rules. One OpenEnv-controlled player per session; reset(seed=...) uses 7vik’s config (e.g. FIVE_MEMBER_GAME). The bridge in server/sevenvik_bridge.py maps our GameAction to 7vik Action types (MoveTo, Kill, Vote, Speak, CompleteTask/CompleteFakeTask) and builds observations from the game state.

Stage 1 Delivered

• 1.1 Dockerization & API

  • Container: server/Dockerfile — run the environment in isolation (build from repo root so among-agents is included).
  • Action space (Pydantic): GameAction in models.py: type: Literal["move", "task", "kill", "talk", "vote"], target: Optional[str]. For move target is a room name (e.g. "Medbay"); for kill / vote it is a player color or name (e.g. "red"); for talk it is the message text.
  • Observation space (Pydantic): GameObservation: location, visible_players, internal_monologue, chat_history, game_status, plus OpenEnv done/reward.

• 1.2 OpenEnv step() implementation

  • Logic bridge: server/amongus_environment.py — step(action) delegates to SevenVikBridge, which turns the LLM’s JSON into 7vik actions and updates the game (map, activity log, phases).
  • State persistence: Each WebSocket session has its own AmongUsEnvironment and thus its own 7vik AmongUs instance; game state is tracked across turns.

Quick start

Prerequisites

• uv — install with: curl -LsSf https://astral.sh/uv/install.sh | sh (or pip install uv)

Local server (no Docker)

From the repo root:

uv sync
export PYTHONPATH="${PWD}"
uv run uvicorn server.app:app --host 0.0.0.0 --port 8000

• uv sync creates a .venv (if needed), installs the Python version from .python-version, and installs dependencies from pyproject.toml (and locks with uv.lock).
• uv run runs the command inside the project’s virtual environment.

Docker

Build from the repo root (so among-agents is in the build context):

docker build -f server/Dockerfile -t amongus-openenv .
docker run -p 8000:8000 amongus-openenv

Client usage

from client import AmongUsEnv
from models import GameAction

with AmongUsEnv(base_url="http://localhost:8000") as env:
    result = env.reset(seed=42)
    print(result.observation.location, result.observation.visible_players)
    result = env.step(GameAction(type="move", target="Medbay"))
    print(result.observation.location)

reset() accepts optional seed, episode_id, and kwargs (e.g. current_player_index, game_config for the bridge).

Layout

├── models.py              # GameAction, GameObservation (Pydantic)
├── client.py              # AmongUsEnv (EnvClient)
├── among-agents/          # 7vik/AmongUs backend (vendored)
│   └── amongagents/       # envs (map, game, action, player, task), agent, configs
├── server/
│   ├── app.py             # FastAPI create_app(AmongUsEnvironment, ...)
│   ├── amongus_environment.py  # reset/step/state, delegates to 7vik bridge
│   ├── sevenvik_bridge.py # GameAction → 7vik Action, observation, get_episode_outcome()
│   ├── game_sandbox.py    # Minimal sandbox (unused when 7vik backend is active)
│   ├── Dockerfile
│   └── requirements.txt
├── training/              # Stage 2: GRPO fine-tuning with real RL env
│   ├── README.md          # Full pipeline, Docker, Northflank, reward details
│   ├── grpo_train.py      # TRL GRPOTrainer entrypoint (env reward by default)
│   ├── env_reward.py      # Real env reward: replay game → step action → verifier
│   ├── verifier.py        # Rule-based reward from episode outcome
│   ├── reward_proxy.py    # Proxy fallback (format + reasoning)
│   ├── collect_episodes.py # Run random games, emit JSONL snapshots
│   ├── data/              # JSONL (generated by collect_episodes.py)
│   ├── Dockerfile         # Northflank image (includes full Among Us env)
│   └── requirements.txt
├── openenv.yaml
├── pyproject.toml
├── uv.lock              # lockfile (commit this); use `uv sync` to install
├── .python-version      # Python version for uv
└── project-plan.md