CodeMind — AI Codenames Spymaster

CodeMind is a solo-playable version of Codenames, the popular word-guessing party game — except one of the two hardest roles, the Spymaster, is played by a real AI teammate instead of a second human. Codenames normally needs four or more people split across two teams. CodeMind removes that barrier: the AI Spymaster looks at a hidden 5x5 grid of words, knows exactly which belong to your team, the opponent, neutral space, and the game-ending Assassin, and has to generate a single one-word clue plus a number that safely connects multiple of your team's words without pointing toward danger. This isn't a chatbot wrapped around a game — it's a real constraint-solving problem. The AI has to hold every hidden word in context simultaneously, validate its own clue against every danger word before it's ever shown to the player, and retry if a generated clue fails that safety check. Every clue ships with the AI's own plain-language reasoning, displayed live in the "Spymaster Transmission" panel, so players can actually see why a move was made instead of trusting a black box. The full game loop is implemented end to end: grid generation, guessing, word reveals, opponent auto-play, a running Mission Log, and win/loss detection with a full replay option. Built solo with FastAPI and vanilla JavaScript, fully containerized with Docker, and using OpenAI-compatible inference routing that mirrors AMD's Fireworks-hosted model pattern — with Fireworks AI, Groq, and Gemini all supported as swappable backends via a single environment variable.