typed-decision-slay-the-spire

Related

• AgenticSTS — Research testbed for long-horizon language-model agents on Slay the Spire 2, built around bounded, typed memory instead of a growing transcript. Reports 3/10 wins at Ascension 0 without learned skills and 6/10 with them.
• ai-spire — Slay the Spire 2 mod that asks an OpenAI-compatible model (DeepSeek by default) for decisions, adds Spire Codex card data to its prompts, and falls back to a rule engine when the API call fails.
• AutoClad — Imitation-learning pipeline on sts_lightspeed: it generates scenarios, records a scripted agent's choices, trains a network to copy them, and exports it for C++ inference. Early work aimed at A20 Ironclad.
• LearnTheSpire — Early experiment (2018) training a feed-forward network on local run files to predict which card a player picks, with cards encoded as 1×234 vectors.
• MCPTheSpire — Slay the Spire mod that runs an MCP server inside the game (Streamable HTTP on port 8080) with tools to read state, play cards, choose options and start runs, so language-model agents can play.
• Orak — Benchmark for training and evaluating language-model agents across 12 video games, Slay the Spire among them, with a public leaderboard.