IrredenEngine: Skill for Claude Code

Render / Trixel Pipeline

Read engine/render/CLAUDE.md first — the pipeline diagram, gotchas (hardcoded bind points, distance-clear semantics, canvas destruction ordering), and the SDF-vs-voxel-pool parity rules. Isometric projection equations: engine/math/CLAUDE.md §"Isometric projection — the equations".

Two-stage depth split

Stage 1 (c_voxel_to_trixel_stage_1.glsl) writes only depth via imageAtomicMin; stage 2 (c_voxel_to_trixel_stage_2.glsl) reads that depth before writing colour and entity IDs. The imageAtomicMin in stage 1 is what resolves overdraw — stage 2 first means every voxel writes unconditionally. Both dispatches run inside the single VOXEL_TO_TRIXEL_STAGE_1 system: one per-canvas tick does compact → stage 1 → stage 2, which keeps each canvas atomic over the shared voxel SSBOs in multi-canvas scenes.

A minimal creation registers exactly VOXEL_TO_TRIXEL_STAGE_1, TRIXEL_TO_FRAMEBUFFER, FRAMEBUFFER_TO_SCREEN in RENDER; lighting passes, SHAPES_TO_TRIXEL, TRIXEL_TO_TRIXEL, and the rest are optional overlays inserted between the first two.

Trixel compositing (CHILD_OF)

TRIXEL_TO_TRIXEL composites each child canvas into its CHILD_OF parent: beginTick binds the shader and sets the camera offset in frame data; relationTick binds the parent canvas textures to image slots 0 and 1; tick binds the child canvas to slots 2 and 3, uploads frame data, and dispatches. Canvases without CHILD_OF render independently into their own textures.

Adding a render stage

1. Write the shader in engine/render/src/shaders/ (and its .metal counterpart — backend-parity skill).
2. Add the SystemName enum entry in ir_system_types.hpp.
3. Add the system header under engine/prefabs/irreden/render/systems/; create() creates the named GPU resources and sets up the dispatch or draw.
4. Register it in the creation's RENDER pipeline in dependency order relative to the stages above.