071906df59
- Core engine: simulator, game mechanics, triggers (138 tests) - Fog-of-war per-player state tracking - Meeting flow: interrupt, discussion, voting, consolidation - Prompt assembler with strategy injection tiers - LLM client with fallbacks for models without JSON/system support - Prompt templates: action, discussion, voting, reflection - Full integration in main.py orchestrator - Verified working with free OpenRouter models (Gemma)
The Glass Box League
Among Us as a headless discrete event simulation. LLMs are players. Classic rules, no GUI.
Quick Start
# Setup
python3 -m venv .venv
source .venv/bin/activate
pip install requests
# Set API key
export OPENROUTER_API_KEY="your-key-here"
# Run tests
python3 -m unittest discover -v tests/
# Run simulation (coming soon)
python3 -m src.main
Project Structure
among-us-agents/
├── config/
│ ├── game_settings.yaml # Game rules & parameters
│ └── prompts/ # LLM prompt templates
├── data/
│ ├── maps/skeld.json # The Skeld map
│ └── agents/{agent_id}/ # Per-agent scratchpads
├── docs/ # Design docs & API reference
├── src/
│ ├── engine/
│ │ ├── simulator.py # Discrete event simulator
│ │ ├── game.py # Game mechanics
│ │ ├── triggers.py # Trigger registry
│ │ ├── fog_of_war.py # Per-player knowledge
│ │ ├── available_actions.py # Dynamic actions
│ │ ├── trigger_messages.py # Trigger JSON schemas
│ │ ├── meeting_flow.py # Meeting lifecycle
│ │ ├── discussion.py # Discussion orchestrator
│ │ └── types.py # Core data types
│ ├── map/graph.py # Graph-based map
│ ├── agents/
│ │ ├── agent.py # LLM agent wrapper
│ │ ├── scratchpads.py # File-based memory
│ │ └── prompt_assembler.py # Prompt builder
│ ├── llm/client.py # OpenRouter client
│ └── main.py # Game orchestrator
└── tests/ # 138 tests
Architecture
Engine
- Continuous time simulation (seconds, not discrete ticks)
- Event-driven: triggers fire → time freezes → agent responds → time resumes
- Fog of war: each agent only knows what they've observed
Agents
- Stateless LLM calls: each trigger = fresh invocation
- JSON scratchpads: persistent memory across ticks
- Toggleable settings: personas, strategy injection, meta-awareness
Discussion
- Priority bidding: agents bid to speak, highest priority wins
- Vote-when-ready: discussion ends when all vote
- Ghost chat: dead players observe
Documentation
| Document | Description |
|---|---|
| design_main_game.md | Game loop, triggers, tools, state |
| design_discussion.md | Discussion, voting, personas |
| openrouter_api.md | LLM integration |
Tests
python3 -m unittest discover -v tests/
# 118 tests across map, simulator, triggers, game, discussion
Status
- ✅ Phase 1: Core Engine
- ✅ Phase 2: Agent Scaffolding
- ✅ Phase 3: Prompt Engineering Design
- 🔄 Phase 4: Implementation
- ⏳ Phase 5: Prompt Templates
- ⏳ Phase 6: Testing & Polish
License
MIT