Multi-agent game · local experiment 多 Agent 博弈 · 本地实验
Texas-Poker-Agents Texas-Poker-Agents
A complete, locally runnable multiplayer LLM poker system and a testbed for behavior in imperfect-information games.
一个完整、可本地运行的多人 LLM 德州扑克系统，也可以作为研究不完美信息博弈中模型行为的实验环境。
起点
这不是一个“让 LLM 打扑克”的单点 demo
真正让我感兴趣的是不完美信息：一个座位不应该看到别人的手牌、未来公共牌或服务端调试信息。只要这条边界没守住，后面的策略分析都没有意义。
所以模型不负责发牌和计分。它只收到该座位的可见状态，返回一次动作建议。
牌桌
规则引擎说了算，模型不能改牌局
Node 服务端负责洗牌、发牌、合法动作、盲注、边池、all-in、摊牌和比赛推进。页面支持一名真人和多个 LLM 座位，每个座位可以单独设置模型和牌风 prompt。
模型返回非法 JSON、超出筹码的下注或不可用动作时，服务端会标出 fallback，并选择 check / fold 托管，不让整桌卡死。
复盘
每手牌都要留下能看的记录
所有事件会追加写入 data/sessions/*.jsonl，也能从页面导出 JSON。房主可以在牌后打开 god view，看暗牌、简短 reasoning summary、fallback 和 hand reflection。
引擎测试覆盖牌型、边池、run-it 多次结算、heads-up 行动顺序、盲注翻倍、table-talk 暗牌过滤和筹码守恒。
对 Agent 行为的信任，应该从信息边界、合法性检查和日志开始，而不是从一段看起来聪明的推理开始。
运行
本地开桌
项目只依赖 Node.js 20+ 和内置模块，不需要数据库。没有 API key 时 LLM 座位仍可参与流程，但会使用服务端 fallback。
THE START
More than a one-shot “LLM plays poker” demo
The interesting part is imperfect information. A seat must not see another player's cards, future community cards, or server-only debug state. If that boundary fails, any strategy analysis afterward is meaningless.
The model therefore owns neither the deck nor the score. It receives a seat-visible state and returns one proposed action.
THE TABLE
The rules engine is authoritative
The Node server owns shuffling, dealing, legal actions, blinds, side pots, all-ins, showdown, and match flow. One human can share the table with several LLM seats, each with its own model and style prompt.
If a model returns broken JSON, an impossible bet, or an illegal action, the server marks the fallback and chooses check or fold instead of freezing the table.
THE REVIEW
Every hand leaves a useful trail
Events append to data/sessions/*.jsonl and can also be exported from the page. After a hand, the host can open a god view for hole cards, short reasoning summaries, fallbacks, and hand reflections.
Engine tests cover hand ranking, side pots, run-it-multiple-times settlement, heads-up action order, blind increases, table-talk filtering, and chip conservation.
Trust in an agent should begin with information boundaries, legality checks, and logs, not with a convincing paragraph of reasoning.
RUN IT
Start a local table
The project needs Node.js 20+ and uses only built-in modules, with no database. Without an API key, LLM seats still participate through the server fallback.