dsh-xiangqi

Every time the AI enters a thinking loop, that dozen or so seconds of silence feels like wasted wall time. dsh-xiangqi fills the gap with a small cartoon pet clutching an Elephant piece, floating on your DSH Web GUI and always ready to invite you to a round of Chinese Xiangqi. It is pure entertainment, fully independent from the conversation agent, and never blocks the main loop.

The entire experience works offline out of the box: a hand-written local engine built on negamax with alpha-beta pruning and iterative deepening powers three difficulty tiers, from randomised easy moves up to a depth-4, five-second-budget hard search. Opening lines announce their classic names (Center Cannon, Immortal Opening, Screen Horse, and more), while check and checkmate fire off Sanguosha-style kill phrases, all triggered locally with zero network dependency. If you have an LLM configured in DSH, you can optionally layer on per-move commentary that toggles between a professional chess coach and a variety-streamer voice; any LLM failure degrades silently and the game never stutters. The board panel supports click-to-move, undo, difficulty switching, and a fullscreen stretch (Esc to close). Pet position, game progress, and settings all persist in the browser.

Who is this for? DSH users who want to kill a couple of minutes between prompts without opening a new tab or installing a heavyweight engine like Stockfish. The developer is upfront that the engine is modest, but the whole codebase is pure TypeScript, so you can ask the AI agent in your current DSH session to deepen the search or strengthen the evaluation function, run the build, and have the changes take effect immediately. Zero setup, zero external dependencies, pick it up and play. That is the entire ambition.