Projects / Simulator
Slay the Spire
conquer-the-spire
C++17 Slay the Spire simulator covering Ironclad, Silent and Defect cards, intended for reinforcement learning with the PyTorch C++ API.
Related
- decapitate-the-spire — Headless Python clone of Slay the Spire made for reinforcement learning and published on PyPI. The README calls it very much a work in progress, focused on Silent in Exordium.
- STS-AI-Master — Reinforcement-learning decision system for Slay the Spire with headless multi-process training and protobuf messaging between processes.
- sts-ironclad-agent — A20 Ironclad bot: one frozen network makes every out-of-combat choice and simulator search plays combat. The README reports a 50.1% Heart win rate over 501 runs in the original game (95% interval 45.7–54.5%).
- sts_lightspeed — Standalone C++17 reimplementation of Slay the Spire built for tree search, designed to be RNG-accurate. The README reports 1M random playouts in 5 seconds on 16 threads.
- sts-rl-agent — A0 Ironclad agent: a learned policy handles every out-of-combat choice and sts_lightspeed MCTS plays combat. On 50 held-out seeds it reaches floor 42.5 and wins 14%, against 31.2 and 6% for heuristics.
- sts2-cli — Runs the real Slay the Spire 2 engine headless in a terminal, so damage, card effects, enemy AI, relics and RNG match the game exactly. It needs a Steam copy of the game, and the README says it was tested on v0.111.0.