Skip to content

Spire Agent v0.1.0

Latest
Latest

Choose a tag to compare

View all tags
@Attemory Attemory released this 28 Aug 07:58
· 34 commits to main since this release
v0.1.0
be0491f

Spire Agent v0.1.0

The first public release of Spire Agent, an autonomous agent for complete Slay the Spire runs.

Highlights

  • Full-run support for Ironclad and Defect.
  • A minimal state-action loop with replaceable Map, Build, and Combat agents.
  • Map planning combines LLM judgment with deterministic key and survival constraints.
  • Winning Path card picking searches toward expert-derived deck templates while controlling transitional cards and deck size.
  • Combat uses sts_lightspeed MCTS, including selective potion release when the no-potion search predicts excessive HP loss.
  • Every run records game states, decisions, LLM calls, MCTS searches, card choices, and a strict replay journal.
  • Terminal UI, manual agent takeover, replay-and-continue, and an optional in-game HUD.
  • English and Chinese prompts and HUD labels.

Ironclad has completed multiple Ascension 20 Heart runs with GPT-5.6 Luna. Watch a complete run.

Getting started

Read the installation guide for Linux, macOS, and Windows setup details.

The attached Java mods are platform independent:

  • AgentStateFixes.jar is required for live runs and replay.
  • AgentVisualizer.jar is required only when the in-game HUD is enabled.

Clone the repository with its dependencies:

git clone --recurse-submodules https://github.com/AttemorySystem/spire-agent.git
cd spire-agent
uv sync

Then follow the installation guide to provide the game and Workshop JARs, build the native MCTS tools, configure an OpenAI-compatible model, and run:

uv run spire-agent

Current limitations

  • Linux and macOS are tested; Windows has not yet been tested.
  • A licensed Steam copy of Slay the Spire and the required Workshop mods are not included.
  • Native sts_lightspeed binaries must currently be built locally.
  • Card-picking and route policies remain under active improvement.

Thanks to the maintainers of sts_lightspeed, gym-sts, and to Baalorlord for publishing the expert run archive used by Winning Path.

Assets 4
Loading