nightcall

Deterministic fixed-point RTS simulation with a PPO-trained policy: C++ engine, Python training, ONNX export for in-game inference.

Docs

Document	Contents
TRAINING.md	Env, observation/action shapes, sweep, PPO, export, rewards
ARCHITECTURE.md	Engine layout, determinism, planned game stack

Quick links

• Build: cmake --preset default then cmake --build build (produces build/python/nightcall_sim.pyd).
• Train / sweep: pip install -r python/requirements.txt and ffmpeg on PATH for sweep videos; e.g. python python/sweep.py --quick --configs 8v8 (uses CUDA when available).
• Tests: build/tests/test_engine.exe (Catch2).

Repo layout (high level)

engine/     simulation (collision, combat, observation, …)
python/     pybind11 module + training (env, model, train, export) + sweep.py
tests/      C++ engine tests
tools/      replay_logger, visualize.py