Slay the Spire Watcher Solver

An RL agent for Slay the Spire with an ambitious goal: beat every single seed at the highest difficulty (Ascension 20). Top human players achieve ~90% win rates, but no unlosable seeds have been proven - the question is whether ML can find a path to 100%.

Thesis

The game has no proven unlosable seeds. If a model can learn to always win, it either proves perfect play exists or the model found strategies humans haven't discovered.

Technical Approach

Strategic Features

Engineering high-level strategic features rather than raw card data:

• Kill probability calculations per enemy
• Stance-specific mechanics (Wrath doubles damage)
• Block efficiency and damage predictions

Learning Architecture

• Behavioral cloning on official game run data
• Neural network for decision evaluation
• Modular system for training and inference

Status

Currently paused - will resume when time permits.