About this project

This is a vertical-screen AI turtle soup (situational reasoning puzzle) interactive game for Douyin livestreams. The system automatically generates eerie puzzles; viewers send "#question" danmaku to probe, and the AI instantly issues rulings such as "yes/no/irrelevant/getting close/reveal" for each question. When someone guesses the core truth or a set time is reached, the answer is automatically revealed and the next puzzle begins, making it suitable for long unattended operation. The project offers multiple run modes including offline simulation, state machine testing, real livestream room connection, and manual danmaku input, and includes a built-in 1080x1920 vertical-screen web interface that can be used with livestream companion window capture or browser source. The interface includes puzzle display, scrolling Q&A stream, ruling badge color coding, folding of irrelevant messages, countdown prompts, and a cumulative soup-guessing leaderboard. Technically, the underlying layer only reads Douyin danmaku and does not send messages or control the livestream room. The core finding is to use forced tool calling (tool_choice) to make the model return structured rulings validated by schema, significantly reducing parsing failure rates and stably triggering correct-guess determination; a lenient parser is also retained as a fallback. The project supports multiple LLM interfaces and multi-model alias routing, and each stage (puzzle generation, puzzle review, Q&A, judging, hints, reveal) can be configured with separate models. Data is persisted as JSONL, the leaderboard persists across livestreams, and a migration tool for older leaderboard versions is provided.