Developer builds AI agent to play Slay the Spire 2 using rejection-feedback, not LLM calls

A developer built an autonomous agent that completed a full solo combat session in Slay the Spire 2 on August 9, 2026, executing 35 actions across 7 turns without any human input. Notably, none of the decisions were made by a large language model; instead, a local score-based greedy policy running inside a Python bridge handled all in-combat choices. The project was enabled by reverse-engineering the game's internal C# codebase, where the developer discovered an existing AutoSlay system and a public card-selection interface called ICardSelector. A key engineering insight was the rejection-feedback pattern, where game errors became persistent state used to filter the agent's next actions — requiring no retraining or re-prompting. The developer also navigated two major technical failures: a main-thread deadlock in Godot and brittle patches that broke with each Early Access update.