A fantasy post-apocalyptic sandbox RPG on a custom, numpy-first engine that uses Panda3D as a rendering SDK only. Deterministic voxel terrain, GPU radiance-cascade global illumination, a physically-based sky, volumetric weather, a living wind field, and procedurally grown forests — every texture, tree, cloud, and blade of grass generated from a single world seed. No hand-made environment art.
Built almost entirely by AI coding agents in eight days (June 9–16, 2026, ~130 commits), directed by a human. The visual build log — with the development screenshots the agents took to check their own work — is in docs/progression/. Start there if you want the story; stay here if you want to run it.
Everything in this frame is procedural: trees grown per-species from Python scripts, leaves placed by cellular automaton, instanced grass and flowers in tagged zones, ray-marched clouds — all lit by GPU radiance cascades and swaying in a simulated wind field.
- Deterministic voxel world — 0.5 m voxels in 32³ chunks, streamed and meshed around the camera; brush edits (left-click explosions) remesh and relight within a frame or two.
- Delta saves — a save is the world seed plus per-system diffs. No pickles, ever.
- GPU volumetric lighting — three nested radiance cascades, ray-marched voxel shadows, bounced global illumination, froxel fog, dynamic point/spot lights (torches, flashlight, explosion flashes).
- Physical HDR sky — Rayleigh/Mie atmosphere feeding the light grid, volumetric ray-marched clouds, bloom, lens flare, god rays, FXAA.
- Spatial weather — storm cells that drift across the map, humidity-driven fog, volumetric rain that respects roofs and canopy, procedural lightning and thunder.
- Wind field — a travelling, spatially-varying gust field driving grass, tree sway, dust motes, and leaf litter. Costs zero bytes in a save.
- Procedural vegetation — GPU-instanced grass and wildflowers in tagged zone volumes; 3-D trees and bushes grown per-species from Python scripts with impostor LODs and per-meter canopy light extinction.
- Buildings — free-form floorplan authoring (walls, arcs, openings, auto-detected rooms, slabs), meshed headlessly in numpy.
- Unity-style object model — GameObjects, components, and lifecycle in snake_case; quaternions only.
- Fire Editor — a VS Code extension driving the live engine: scene tree, component inspector, gizmos, scene save, screenshot RPC.
- Frame profiler — headless core, in-game F3 overlay, scripted benchmark harness.
The deep simulation layers (AI, economy, politics) are still stubs — that's the next arc.
Python 3.11+ required.
python -m venv .venv
.venv\Scripts\activate # Windows (source .venv/bin/activate on POSIX)
pip install -r requirements.txt
python main.py| Input | Action |
|---|---|
W A S D + mouse |
Free-fly camera (Shift sprints, Space/Ctrl up/down) |
ESC |
Toggle mouse capture |
| Left-click | Fire an explosion — carves a crater and flashes a point light |
F |
Toggle camera flashlight |
L / K |
Drop a torch / clear dynamic lights |
G |
Build a Cornell-style GI test room ahead of the camera |
F1 |
Developer overlay |
F5 / F9 |
Quick save / quick load (delta saves to saves/quick.ta) |
F6 |
Cycle forced weather (clear → cloudy → overcast → fog → rain → storm → natural) |
F7 / F8 |
Fast-forward time scale / jump the clock +6 hours |
pytest -q # full headless suite (no window / GPU required)
pytest -m window # tests that need a real Panda3D window
pytest -q tests/standards/ # the machine-enforced standards gateThe headless suite (236 test files) runs without a GPU because only render/ and
lighting/ may import panda3d — everything else is pure Python/numpy. The standards gate
enforces structure limits, lint, strict types, docs-link resolution, and per-module test
presence; a violation fails the build like any failing test
(docs/systems/standards.md).
python tools/screenshot.py --out shot.png # offscreen screenshot of the demo world
python tools/preview_texture.py wasteland_ground # render a ProceduralTextureDef → PNG
python tools/preview_tree.py # preview a tree species' variants
python tools/verify_weather.py # summon a thunderstorm overhead
python tools/profile_run.py # scripted performance benchmark
python tools/dump_save.py saves/quick.ta # inspect a save's per-system deltasThese tools are how the AI agents verified their own visual work — the screenshots they saved along the way became the progression docs.
The engine is layered: downward calls are direct imports, upward/sideways notifications go through an event bus, and only the render bridge touches the rendering SDK. Full design authority: docs/ARCHITECTURE.md.
simulation/ (player, ai*, economy*, politics*) * = stubs
│
world/ (terrain, weather, wind, sky) buildings/ zones/ scene/
│
lighting/ ── render/ ← the only packages that may import panda3d
│
core/ procedural/ resources/ save/ ← pure Python/numpy foundation
docs/is grep-first — it's the AI search index. Every package has a doc indocs/systems/with identical headings, so agents grep docs before reading code. Stale docs fail the build.- Determinism: the world is a pure function of
world_seed(config.toml). All randomness flows throughcore.rng.for_domain(*keys). Same seed → byte-identical world; that's what makes delta saves and bug repro possible.
main.py # demo entry point (python main.py [--load save.ta])
config.toml # world_seed + all sizes/distances (no magic numbers in code)
fire_engine/ # the engine: core/ procedural/ resources/ save/ lighting/ render/
# world/ simulation/ buildings/ zones/ scene/ devtools/
editor/ # Fire Editor VS Code extension + Python daemon
docs/ # grep-first knowledge base: ARCHITECTURE, systems/, sessions/,
# content/, progression/ ← the visual build log
tests/ # headless suite + standards gate (tests/standards/)
tools/ # screenshot / preview / profiling / save-inspection utilities
assets/ # hand-crafted files only (env textures are procedural — never here)
See DECISIONS.md for the dated decision log, docs/sessions/ for per-session handoff notes, and docs/progression/ for the illustrated eight-day build story.