sts_lightspeed

For tree search and simulation of the popular rogue-like deckbuilder game Slay The Spire

Features

• c++ 17 compiled with gcc
• Standalone
• Designed to be 100% RNG accurate*
• Playable in console
• Speed: 1M random playouts in 5s with 16 threads
• Loading from save files (loading into combat currently only supported)
• Tree Search (best result, knowing the state of the game's rng)

Planned Features

• Tree search of possible game outcomes (not given the state of rng)

Implementation Progress

• All enemies
• All relics
• All Ironclad cards
• All Defect cards
• All colorless cards
• Everything outside of combat / all acts

Getting Started

• The project was built with Clion2021 and the mingw64 toolchain on Windows 10
• The main target creates a simulator of the game that can be played in console.
• The test target creates a program with various commands that can be run, including random simulation
• Click the star button at the top of the repo :)

Build tips

• If your build fails with an error about not-return-only constexpr methods, ensure your compiler supports c++17.
• If CLion shows an error about not finding python libs when loading the cmake project, try opening CLion from the msys2 shell.

Card reward counterfactual evaluator

card-reward-eval is a standalone counterfactual executable for comparing skip with one to three permanent card reward candidates. It creates a fresh boss battle from the supplied permanent deck for every candidate and uses the same RNG worlds and MCTS budget for each variant. The executable only reports evidence; the Python agent applies its own quality gate and retains final decision authority.

card-reward-eval <input.json> [simulations_per_world] [rng_worlds] [max_time_ms_per_world]
card-reward-eval --self-test

The input contains:

• game_state: seed, ascension_level, act, floor, HP, gold, class, permanent deck, relics with counters, and potion slots;
• candidates: card id, upgrades, and optional choice_id;
• targets: boss encounter names such as The Guardian or The Heart.

The primary output is the MCTS rollout win rate for each candidate and its delta from skip. Per-world opening hands, monster HP, selected root action, visits, and terminal sample rates are retained for auditing.

This executable deliberately does not predict later cards, upgrades, relics, route damage, or potion use. It rejects state that cannot yet be reconstructed exactly, including unidentified bottled cards and cards upgraded more than once. A zero-win early-game Heart result is saturation, not evidence that every candidate has equal long-term value. Small differences can also change with search budget, so callers must treat unstable or closely grouped results as inconclusive.