An autonomous One Night Ultimate Werewolf simulation where 6 AI players — each powered by a different LLM — make every decision themselves: who to inspect at night, what to say in debate, who to vote out, and how they reflect afterwards.
This is not a scripted demo. Every move, every speech, every vote is generated by the AI player's own model in real-time, based on its persona, private role knowledge, and the live conversation.
Live Showcase: https://shawt.im/endless-werewolf/
| Distinct persona | Each player has a unique character — a relentless prosecutor, a chaos agent, a nervous underdog — that shapes their speech and strategy |
| Different LLM | Each player runs on a different model (GLM, Kimi, Qwen, Grok, etc.). You see real differences in reasoning style |
| Private information | Players only know what their role allows. The Seer knows one card, the Werewolf knows their teammate — nobody sees the full picture |
| Live awareness | Players hear each other's speeches in real-time and can respond, accuse, or bluff accordingly |
| Cross-game memory | Players remember the previous game. A player who was betrayed last round may hold a grudge |
Night → AI players take secret actions (inspect, rob, swap cards)
Day → 6 AI players debate concurrently — speeches, accusations, defenses
Vote → Each player casts a vote
Resolve → Winner determined by One Night Ultimate Werewolf rules
Postgame → In-character interviews: winners celebrate, losers explain
| Phase | Who decides |
|---|---|
| Card dealing & role assignment | Random shuffle (rule-based) |
| Night actions | AI — each player decides based on role + persona |
| Daytime debate | AI — 6 concurrent subagents generate speeches |
| Voting | AI |
| Outcome resolution | Rule-based (game rules) |
| Postgame interviews | AI — in-character reflection |
| Chinese translation | LLM bridge agent (正體中文書面語) |
Game Engine — Python. Handles card dealing, night action resolution, debate orchestration, voting, and outcome determination.
AI Bridge — Each decision request is routed through an OpenClaw bridge agent that spawns a short-lived subagent using the player's assigned LLM. The subagent thinks, returns a JSON decision, and is destroyed.
Frontend — Three.js 3D scene on GitHub Pages. Candlelit round table, 6 stylized characters, procedural night-sky shader with stars and moon, AI-generated day sky and texture set, a village with lit windows, ground mist, a guided full-trace replay, phase navigation, speech bubbles, vote arrows, and confetti.
Bilingual — All game content is generated in English, then translated to 正體中文書面語 via an LLM bridge agent for a parallel Chinese track.
gm_night.py deal cards, build night action plan
bridge_agent.py LLM bridge: request → spawn subagent → JSON response
night_phase.py night actions: seer, robber, troublemaker, werewolf peer
day_phase.py concurrent AI debate loop
resolve_phase.py rule-based outcome resolution
postgame_phase.py AI postgame interviews
tag_phase.py post-process: add <Role> [Player] markup
translate_zh_phase.py LLM translation (EN → 正體中文書面語)
cross_game_memory.py inject previous game data into player context
state_manager.py game directories, manifests, counters
run_full_game.py orchestrates the full pipeline
This project runs on the OpenClaw platform and cannot be run standalone.
data/
players.json player roster (personas, models, avatars)
roles_pool.json role definitions
games/game_XXXXXX/ per-game results
night_result.json night phase (English)
day_result.json day phase (English)
vote_result.json votes + tally (English)
resolve_result.json outcome + final roles (English)
postgame_result.json interviews (English)
chat_history.md full dialogue log (English)
*_zh.json Chinese translations
docs/ GitHub Pages frontend
app.js Three.js 3D village scene + replay engine (source)
styles.css UI styling (source)
app.min.js / styles.min.css generated bundles actually served by index.html
index.html entry point
data/games/ published game archive
index.html loads the minified app.min.js and styles.min.css, but the
readable app.js and styles.css remain the source of truth. After editing
either, regenerate the bundles:
bash scripts/build_assets.shscripts/prepublish_check.sh rebuilds both files and fails if the committed
copies differ, so a stale bundle cannot reach the published site.
Game text uses a lightweight tagging convention:
<Werewolf><Seer>— role names (highlighted purple in UI)[The Prosecutor]— player names (highlighted gold in UI)
Tags are applied by tag_phase.py (deterministic post-processing, not by AI agents) to ensure:
- Consistent visual highlighting
- Accurate translation (
<Seer>→<預言家>, never<先知>)
6 個 AI 玩家，各自使用不同的 LLM 模型，自主完成一整局狼人殺。
這不是劇本演示 —— 每個決策（夜晚行動、白天發言、投票、賽後感言）都是 AI 玩家根據自己的角色記憶與場上局勢，在執行時自主判斷的結果。
- 🔮 全自主決策 — 夜晚行動、白天辯論、投票、賽後訪問，全部由 AI 自主完成
- 🧠 跨局記憶 — 每個玩家記得上一局的完整過程，並讓這些經驗影響下一局的判斷
- 🤖 多模型協作 — 6 個玩家使用不同的 LLM，呈現各異的推理風格與性格
- 🌐 雙語支援 — 英文 + 正體中文書面語，角色名稱與玩家名稱翻譯一致
- 🎮 3D 互動介面 — Three.js 場景、完整軌跡重播、階段導覽，以及 AI 生成的貼圖與燈光
