一个网页版炸金花游戏，你可以与最多 5 个 AI 对手同桌博弈，每个 AI 由不同的 LLM 驱动。每个 AI 自主决定打法与性格、在牌桌上互相吐槽、从过往牌局中学习，并留下一份你可以在每局结束后翻阅的「思考日记」。
在应用内统一配置 API Key，并选择来自 OpenRouter、Azure OpenAI、GitHub Copilot、SiliconFlow 的模型。
挑选 1–5 个 AI 对手，分别绑定不同的 LLM 模型，自定义昵称、筹码与底注。
实时观看 AI 对手的思考、下注与互相吐槽。聊天面板会显示旁观者的反应与行动解说。
轮到你时，可以选择弃牌、跟注、加注或比牌，AI 对手会对你的每一步做出反应。
窥探每个 AI 的内心：手牌评估、对手分析、风险评估、聊天分析、推理过程、置信度、情绪与桌面发言 — 每一回合都有完整记录。
- 多模型 AI 对手 — 1–5 个 AI 玩家，每个由不同 LLM Provider 驱动（OpenRouter、GitHub Copilot、Azure OpenAI、SiliconFlow），可以在牌桌上自由混搭。
- LLM 驱动策略 — 没有预设性格，没有硬编码规则。每个 AI 的打法、诈唬倾向与风险偏好完全由 LLM 自主推理产生。它们唯一的指令是：「你的目标是赢」。
- 桌面对话 — AI 会吐槽、回应他人的动作、回复你的消息。每个旁观 AI 都会调用 LLM 来决定是否要插话 — 没有概率门控，完全由 LLM 决定。
- 经验学习 — AI 会复盘自己的打法，在连败、巨亏、筹码危机或对手风格变化时调整策略。
- 思考日记 — 结构化决策记录（手牌评估、风险、置信度、情绪）+ 每一局的第一人称叙事 + 含统计与自我反思的整局总结。
- 赛博朋克主题 — 霓虹光效、玻璃拟态、3D 牌桌、全身角色立绘。
标准 52 张牌，每位玩家 3 张。牌型大小（高 → 低）：
豹子 > 同花顺 > 同花 > 顺子 > 对子 > 散牌
未看牌的玩家以一半下注额参与。可用动作：弃牌、跟注、加注、看牌、比牌。
| 层级 | 技术 |
|---|---|
| 前端 | React 19、TypeScript、Vite 7、Tailwind CSS 4、Framer Motion、Zustand |
| 后端 | Python、FastAPI、LiteLLM、SQLAlchemy（async）、SQLite |
| 通信 | WebSocket + REST |
OpenRouter、GitHub Copilot（OAuth Device Flow）、Azure OpenAI、SiliconFlow — 全部可在应用内配置。
cd backend
pip install -e ".[dev]"
uvicorn app.main:app --reload
# → http://localhost:8000cd frontend
npm install
npm run dev
# → http://localhost:5173API Key 在应用内的「模型配置面板」中管理 — 不需要 .env。Key 仅保存在内存中，不会落盘。
浏览器 (React SPA)
↕ WebSocket + REST
FastAPI 后端
├── 游戏引擎 — 牌堆、牌型评估、规则、对局生命周期
├── AI Agent — LLM 决策、聊天、经验学习
├── 思考日记 — 结构化记录、叙事、总结
└── SQLite — 8 张表（异步）
↕ LLM API (OpenRouter, Copilot, Azure, SiliconFlow)
信息隐藏：前端永远看不到其他玩家的牌。容错：非 JSON 的 LLM 响应会触发多层 fallback；非法动作降级为跟注/弃牌；API 超时在重试后自动弃牌。
A web-based Zha Jin Hua (炸金花 / Three-Card Poker) game where you play against up to 5 AI opponents, each powered by a different LLM. Every AI decides its own play style and personality, trash-talks at the table, learns from past rounds, and keeps a detailed "thought journal" you can read after each game.
Configure API keys and select models from OpenRouter, Azure OpenAI, GitHub Copilot, and SiliconFlow, all managed in-app.
Pick 1–5 AI opponents with different LLM models, customize names, and set chip/ante levels.
Watch AI opponents think, bet, and trash-talk each other in real-time. The chat panel shows bystander reactions and action commentary.
When it's your turn, choose from fold, call, raise, or compare. AI opponents react to your every move.
Peek into every AI's mind: hand evaluation, opponent analysis, risk assessment, chat analysis, reasoning process, confidence level, emotion, and table talk — all recorded per turn.
- Multi-Model AI Opponents — 1–5 AI players, each driven by a different LLM provider (OpenRouter, GitHub Copilot, Azure OpenAI, SiliconFlow). Mix and match models at the table.
- LLM-Driven Strategy — No preset personalities or hard-coded rules. Each AI's play style, bluffing tendency, and risk tolerance emerge entirely from the LLM's own reasoning. Their only instruction: "your goal is to win."
- Table Talk — AI trash-talks, reacts to other players' moves, and responds to your messages. Every bystander AI calls the LLM to decide whether to chime in — no probability gating, fully LLM-decided.
- Experience Learning — AI reviews its own play and adjusts strategy on losing streaks, big losses, chip crises, or opponent shifts.
- Thought Journal — Structured decision records (hand eval, risk, confidence, emotion) + first-person narratives per round + full game summary with stats and self-reflection.
- Cyberpunk Theme — Neon glow, glassmorphism, 3D poker table, full-body character illustrations.
Standard 52-card deck, 3 cards per player. Hand rankings (high to low):
豹子 Three of a Kind > 同花顺 Straight Flush > 同花 Flush > 顺子 Straight > 对子 Pair > 散牌 High Card
Unseen players bet at half rate. Actions: Fold, Call, Raise, Peek, Compare.
| Layer | Technology |
|---|---|
| Frontend | React 19, TypeScript, Vite 7, Tailwind CSS 4, Framer Motion, Zustand |
| Backend | Python, FastAPI, LiteLLM, SQLAlchemy (async), SQLite |
| Communication | WebSocket + REST |
OpenRouter, GitHub Copilot (OAuth Device Flow), Azure OpenAI, SiliconFlow — all configurable in-app.
cd backend
pip install -e ".[dev]"
uvicorn app.main:app --reload
# → http://localhost:8000cd frontend
npm install
npm run dev
# → http://localhost:5173API keys are managed in the in-app Model Config Panel — no .env needed. Keys are memory-only, never persisted to disk.
Browser (React SPA)
↕ WebSocket + REST
FastAPI Backend
├── Game Engine — deck, evaluator, rules, game lifecycle
├── AI Agents — LLM decision, chat, experience learning
├── Thought Journal — structured records, narratives, summaries
└── SQLite — 8 tables (async)
↕ LLM APIs (OpenRouter, Copilot, Azure, SiliconFlow)
Information hiding: frontend never sees other players' cards. Fault tolerance: non-JSON LLM responses trigger multi-layer fallback; illegal actions degrade to call/fold; API timeouts auto-fold after retries.
- PRD — Requirements, game rules, feature specs
- Technical Design — Architecture, data models, API specs
- Tasks — 30 tasks across 8 phases