Slay the Spire 2 AI Copilot & Card Evaluator

MEDIA PREVIEW

Overview & Result

Paulwei built a real-time autonomous copilot for Slay the Spire 2. Rogue-like deck builders require deep situational calculations where mistakes are fatal. Traditional LLMs take 3-5 seconds per decision and frequently miscount energy costs. By representing game state as structured choices and using Jev for ranking, the agent evaluates the entire hand against enemy intent in 70ms, executing superhuman 0.7s total turns.

How Jev fits in the loop

1. Hooks game state (draw pile, player HP, energy, relics, enemy incoming damage and debuffs).
2. Generates candidate play permutations and feeds them to Jev Score API to evaluate survival and damage maximization.
3. Jev selects the winning action sequence within 70ms.
4. Directly sends virtual input commands to execute the turn seamlessly.

How to reproduce

1. Clone coolish/spire-jev and install python requirements.
2. Launch game and attach python bot hooks to track player hand and enemy intent.
3. Execute run_bot.py to enable real-time Jev card evaluation.
4. Watch automated high-speed turn execution with calculated health preservation and energy optimization.

Why this build matters

Showcases Jev's capability in complex game-theoretic state spaces where latency, deterministic logic, and zero hallucination are absolute prerequisites.

Reported performance

Cost: $0.0003 / combat encounter

Latency: 70ms / card decision

Limitations

• Requires game memory hook or clean 1080p screen capture to parse current energy, relics, and enemy block.
• Deep meta strategy for specific ascension modifiers requires calibrated heuristic prompts.

Patterns