RSI-Jev-Slay-the-Spire-2
Chinese Slay the Spire 2 agent research: code handles rules, cheap Jev makes most choices, a stronger model takes key moments.
Overview
RSI Jev for Slay the Spire 2 is a public, evidence-driven experiment log (in Chinese) for building a game agent. The aim is to use deterministic computation for rules and risk, let the cheap Jev model make the large number of choices, and call a stronger model (Astra) only for critical, uncertain situations, with the goal of raising full-run win rates across characters and ascension levels.
The system has three layers: code controllers enforce legal actions, combat budgets and boundary checks; Jev (typesafe/jev-1.13) makes structured choices among candidates; Astra can give single-step decisions or room plans bound to fresh state. Experiments are numbered and documented with traces, costs and failures, for example one where Jev picks one plan per fight from an experience pack and replays seen positions 12/12, while noting an exact lookup table does equally well and generalisation is unproven. Full unseen-seed clears have not yet been achieved.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which legal action or card should be chosen from the candidates in this combat state? | — |
| Choice | Which plan from the experience pack should be used for this fight? | — |
How it works
- Code enumerates legal actions and enforces budgets
- Jev chooses among candidates for routine decisions
- Escalate critical uncertain moments to Astra
- Record traces, costs and outcomes per experiment
Strengths
- Cheap Jev requests (e.g. about $0.02 for 88 requests in one experiment)
- Every experiment documented with failures
Caveats
- Documentation in Chinese
- No full win on unseen seeds yet
- Expert input partly supplied manually via Codex tasks
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Game CLI simulator and MCP
- Has code
- yes
Use cases and integrations
How curated lists describe it
This project integrates Jev to provide structured decisions for its workflow. See the repository for implementation details.
logicrw/awesome-jev-projectsHttps://v.douyin.com/iJtUdfsywHU/.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Structured choices among legal candidate actions and fight plans.
When is the stronger model used?
Only for critical and uncertain positions, by design.
Does it win full runs?
Not yet on unseen seeds, per the README.
Related projects
AI life-and-civilization simulation where Jev makes every person's next action as an auditable Choice and LLMs only plan.
ProjectprototypeJavaScriptFactorio agent on FLE where Jev picks goals, next actions and stuck detection while code owns rules and actuation.
ProjectresearchPythonSource for browser chess and Chess960 where code computes each legal move's facts and Jev picks one per turn as a typed Choice.
ProjectproductionTypeScriptBrowser 1v1 sniper arena where Jev reads match state about 9 times a second and chooses movement, aim, ADS, firing and jumping.
ServicebetaTypeScriptInstrumented 2048 web lab where every move is a Jev Choice over four directions, with probabilities, latency and cost shown live.
ProjectbetaTypeScriptFifteen live arcade games where Jev and LLMs play each other, led by WikiRace in which Jev picks only real links.
ProjectbetaTypeScript