# Module (asset type) A module is a reusable, stateless (or globally-stateful) Luau library that other code pulls in with `require()`. Modules are the unit of factored-out logic in a world: helpers, math libraries, format parsers, registries, shared algorithms. ## When to use one - You have logic you want to share across multiple components, services, scenes, or tools. - You have a registry, lookup table, or shared cache that callers all read from / write to. - You want a stable API surface (`M.fn`) other code can depend on. If the logic is bound to one entity, use a `.component`. If you want a named UI of fields rather than functions, use a `.preset`. ## Where it lives - Source: `/zero/source/.../<name>.module/` - Identity: the source path with `/` written as `.` and each segment's type suffix dropped — `/zero/source/demo/beats.module` is `demo.beats`, and `/zero/source/kit.package/beats.module` is `kit.beats`, the container's name kept and its `.package` suffix gone. `asset.create` returns this string, `asset.resolve(...).identity` reads it back, and the `require` forms below write it as `<identity>`. - Folder shape: - `init.luau` (or `init.lua`) — module body. **Required.** Returns a table (conventionally named `M`). - `README.md` — instance-level documentation. **Required.** - `.metadata` — agent-editable tags + free-form fields. **Required.** ## How to create one ```luau asset.create("module", "<name>") -- Creates: /zero/source/<name>.module/ -- init.luau (canonical `local M = {} … return M`) -- README.md (instance README template) -- `folder` places it in a subfolder of /source, and that subfolder -- leads the identity: `{ folder = "lib" }` produces `lib.<name>`. -- `into` authors it inside a resolved container (a package, a toolbox), -- whose own identity leads the same way: asset.create("module", "<name>", { folder = "lib" }) asset.create("module", "<name>", { into = asset.resolve("myPkg.package") }) ``` ## How it operates 1. **Registration.** Writing `init.luau` into a `.module/` folder indexes the asset and registers the module path with the resolver. 2. **Loading.** The first `require("<identity>")` runs the file's body and caches the returned table. Subsequent requires from anywhere return the **same table** — this is the shared-state lever for "module-local" registries. 3. **Hot reload.** Editing the file re-runs the module body and folds what it returns into the table already loaded, so that table keeps its identity: a caller that captured it with `require` reads the new source's functions and values through the reference it already holds, and a later `require` returns that same table. A key the new source drops goes away; a key something set at runtime stays. The body's own `local`s start again from their initial values, so state that must outlive an edit lives in `modules.state()` (`guides { path: "core/scripting-and-tasks" }`). 4. **Identity resolution.** A module is reached by its `<identity>`, and where the caller lives decides which spelling of it resolves. For a module in a **world**, from a caller in that same world — another module, a component, a scene entrypoint, or an `execute` chunk: - `require("<identity>")` — the identity on its own, resolved against the caller's own root. It carries every folder segment, so a module created with `{ folder = "demo" }` is `require("demo.beats")`. - `require("@root::<identity>")` — the same root-relative resolution, written out. - `require(".sibling")` — a module in the same folder, by its bare name; `..name` steps up one folder, `...name` two. Logs, stack traces and hot-reload notices name that same module `@local.source.<identity>` — its key under the world source root. The `<identity>` tail of that key is the form above, so a trace reading `@local.source.demo.beats` is `require("demo.beats")`. For a module in a **library** — content under `/zero/source/libs/@<lib>/` — `@<lib>::` addresses that library's root from anywhere: `require("@builtin::modules.transform")`. Mounting a world as a library moves its files under such a root and flips the caller root to `@<lib>::`, which is what keeps the root-relative forms above resolving across the move. ## Discovery - `asset.list("module")` — every registered module. - `asset.inspect("<name>")` — public functions, source path, this type README. - `ls /zero/source/<name>.module` lists the files the folder holds. ## Authoring conventions - Return a single `M` table from `init.luau`. Top-level statements with side effects run on first require — useful for one-time initialization, dangerous if they touch the engine before it's ready. - Write exported functions as `typed function`, not `function`: the argument types are enforced at the call and the checker knows the signature, so a caller's mistake is reported where they made it. - Annotate exported functions with `--!desc` / `--!arg` / `--!return` / `--!example` so the LSP, `tools.list`, and `asset.inspect` surface them. A function with no `--!desc` is reachable only by someone already reading this file. - Declare a VALUE with `Field.<kind>(default, mode, description)`, and a table whose members are reached through a metatable with `--!members <TypeName>` — the latter gives the surface its accepted member set, its type, and its documentation from one table. See `guides { path: "core/authoring" }` for all four rules and `modules/api/engine/entity.module/members.module` for a worked member table. - Keep module state in module-local upvalues or in `modules.state()`. Globals leak across reloads. Module-locals start again from their initial values with each hot reload; the table `modules.state()` returns is the same table on both sides of every reload. - Prefer focused modules over kitchen-sink modules. If two halves of a module have no shared state, split them. ## Common pitfalls - **Cyclic requires.** A `require` chain that loops will return the partial table (the half built before the cycle was detected). Design for one-way dependencies. - **Engine-time side effects.** Don't `entity.spawn` at module top-level — the engine may not be ready. Expose an `M.init` and call it from a scene entrypoint. - **Hot-reload + module-local cache.** A module that caches expensive computation in upvalues loses that cache on reload — fine for dev, but be aware in performance work. - **`init.lua` vs `init.luau`.** Either works, `.luau` is canonical. - **Writing one while play runs.** A `/source` write made while play is running lands on the play shadow, and a guarded play-exit discards it unless it is kept. One `.module` write shadows `init.luau`, the `README.md` and the `.metadata` together. `vfs.write(path, bytes, { durable = true })` skips the shadow, and `vfs.promotePlayShadow` keeps what is already on it — the `core/vfs` guide has the model and what each route costs. ## Related types - `.component` — for entity-bound state + lifecycle hooks. - `.service` — generates content you don't have yet (a mesh, sound, texture, …) via a metered provider. - `.tool` — for a single agent-callable function with a YAML schema. - `.package` — to group several modules + components + scenes into one shippable folder. ## In the Inspector Exports: what the module exports (name, kind, signature), parsed from its source without running it, with **Check** (the LSP's diagnostics for its entry file) and **Open**.

# proc Code-authored, graph-backed procedural generation for the Zero engine. A graph is one function of `g`: `proc.define(function(g) ... end)`. Nodes are called by family and name (`g.<family>.<name>({ ... })`), inputs are declared with `g:input(name, default, opts)`, the returned table is the graph's outputs, and `io.output` is how the graph stands in the world. `builder.module/README.md` holds every rule of that form. This module owns the whole procedural model: the graph builder + data model, the operation registry, the type system, the lazy cached evaluator, deterministic scoped hashing, and the node-family op implementations. Authored `.procGraph` graphs are persisted via `procGraph.assetType`; the `procgen.preview` / `procgen.spawn` operations live in `procgen.toolbox` and are invoked as tools: `use_tool` from an agent, or `tools.use("procgen", "preview", args)` from a script. Generated meshes/materials/textures/bundles are persisted through their own `.assetType`. An op family can also come from outside this package: a `.procPack` asset registers its ops in the same registry, called as `g.op.<id>(...)` exactly like the ones shipped here (`proc.registerPack(name, ops)` is what it registers through, and what a test or a tool calls to hold a family for the life of one VM). `terrain.*` is one such pack, shipped in `terrainSystem.package/procPacks/` and covering a whole landscape; `guides { path = "systems/terrainSystem/documentation/terrain" }` is its guide. To find out what the registry holds, use the `procgen` toolbox: `searchNodes` searches it and `describeNode` describes one node or op. `proc.search_ops(query)` and `proc.schema(id)` are the same answers in-process, for a graph builder looking something up while it runs.

# terrain_erode_droplet Droplet hydraulic erosion over one tile, one batch at a time. Each thread carries one droplet. It is placed by a hash of its index, the op's seed, the batch number and the tile's field-wide base cell, then runs downhill: it picks up material where it moves fast over steep ground and drops it where the ground flattens or rises. That is what cuts channels into a ridge and lays the fans out below them, and it is why the result looks like water shaped it rather than like a filter ran over it. ## Reproducibility The pass never writes height. Every change accumulates into `delta` as fixed-point integers, 1/65536 m per unit, through `atomicAdd`. Integer addition commutes; float addition does not. Thousands of droplets landing on one cell therefore sum to the same total whatever order the threads resolve in, and the same seed over the same stack gives the same terrain every run — on any GPU, with any scheduling. A float accumulator would be a few last bits different each run, and everything downstream that assumes a seed names a landscape (storage, streaming, a shared world) would drift. ## What one batch may move A batch works out every droplet's change from the ground as it stood when the batch began and adds them all at the end, so nothing in a batch can see what the rest of it is doing to the cell it is standing on. Two bounds are what keep that from running away. `share` is the part of one cell's rainfall a droplet carries. A batch rains `dropletCount` droplets over the interior and each walks up to `apron` cells, so a cell is stepped on `dropletCount * apron / interiorCells` times a batch and every one of those steps adds into the same accumulator. Scaling each droplet's change by the reciprocal is what makes `rain` a sampling density: doubling it resolves the flow better and moves the same material. Rain too sparse to step on every cell gives a droplet one cell's rainfall and no more, because a droplet carrying several cells' worth would put a whole batch's cut into a single step and reopen the swing below. A run below that density — fewer than `interiorCells / apron` droplets a batch — therefore samples the flow sparsely and moves proportionally less; above it the result stops changing with `rain`. `MAX_SHARE` is the most of the drop below a cell one batch may cut, and the most of the rise above it one batch may fill. Without it a droplet cuts the whole drop, which leaves the cell level with the one it drains into — and every other droplet in the batch cuts that same drop from the same starting ground, so the cell ends the batch far below its neighbour. The next batch finds the pair reversed and cuts back further, and the swing grows every batch until the fixed-point accumulator wraps. `terrain_erode_thermal` names the same bound; a quarter rather than a half, because water gathers where the ground funnels it and a cell along a channel is stepped on several times over what an average cell gets. ## Tiling A droplet takes at most `maxSteps` steps, and the caller keeps that at or below the apron width. Past that it would read cells whose real owner is a neighbouring tile and whose contents have not been exchanged since the batch began, and the two tiles would disagree about the ground at their seam. Droplets spawn in the interior only. A spawn in the apron would place a second droplet on ground the neighbouring tile has already rained on, doubling the rainfall in a band along every seam. Material a droplet carries out of the interior and into the apron is not lost: `terrain_delta_gather` adds it into the neighbour's interior before either tile folds its accumulator into height. What a droplet is still carrying when its walk ends goes with it. A run therefore lowers the ground it cuts: material leaves the uplands, and a field has no basin outside itself for it to arrive in. The fall in the field's mean is a small fraction of the depth the channels reach. ## Parameters Lane 1 carries `seed, batch, dropletCount, maxSteps`; lane 2 `solubility, deposition, evaporation, inertia`; lane 4 `minSlope, capacity, gravity, apron`. Heights are metres and horizontal distances are cells, so `capacity` and `minSlope` read as metres of fall per cell of run. Every op carries a mask, and for a simulation it bounds what a step MOVES. The mask scales both what a step cuts and what it lays down, and the material a step declines to move stays where it was — in the ground, or in the droplet — so a droplet carries its load ACROSS ground the mask holds out rather than losing it there. What a droplet still holds when its walk ends lands where the walk ended, whatever the mask reads there. The ground it walked is already lighter by exactly that much, so scaling that last drop would leave the difference nowhere and the pass would remove material instead of moving it. The mask therefore bounds where a run cuts exactly, and where it deposits everywhere except the cell a walk ends on. An op that names no mask erodes at full strength everywhere.

# terrain_erode_thermal Talus-angle relaxation over one tile, one iteration per dispatch. Loose material resting on a slope steeper than its angle of repose slides down. Repeat that and cliffs grow scree fans at their feet while the crests a droplet pass leaves knife-edged round off — which is why this is usually the finishing pass over hydraulic erosion rather than a substitute for it. The rest angle is a world angle. A cell may stand `tan(talus) * cellSize` metres above a four-neighbour and still rest, so refining the grid resolves the same slope rather than changing which slopes are stable. Each iteration is one dispatch: a cell above the rest angle sheds a fraction of its excess, split across the neighbours below the angle in proportion to how far below each one is. Half the excess is the ceiling per iteration, above which the two cells either side of a step trade places every step instead of settling. Changes accumulate through the same fixed-point integer path droplet erosion uses, so the result is equally independent of the order threads land in. One thread per INTERIOR cell. The apron is read and never written: an apron cell belongs to the neighbouring tile, which relaxes it itself, and relaxing it here as well would move its material twice. Material that slides out of the interior lands in the apron, and `terrain_delta_gather` carries it into the neighbour before either tile folds its accumulator into height. Every op carries a mask, and for a relaxation the mask gates the cell that SHEDS. A cell the mask holds out keeps its own material, and what it still receives from an unmasked neighbour is the scree that slope throws past the mask's edge — material moving downhill, which is what this pass does. An op that names no mask relaxes at full rate everywhere.

# terrain_mask A soft-band rule over one feature of a raster, written into a named layer. | input kind | value | unit | |---|---|---| | 0 height | the surface's own value | world metres | | 1 slope | the angle of its gradient | degrees from flat | | 2 curvature | its Laplacian — positive in a gully, negative on a ridge | 1/metre | | 3 noise | gradient noise over world position | 0..1 | | 4 curve | distance to a polyline's sampled control points | world metres | A band is `min .. max`, feathered by `feather` — a soft half-width in the input's own unit — on either edge; a side left open (`±1e29`) is not a bound at all. `jitter` perturbs the sampled value before the band, by up to `jitter` over a gradient-noise field sized `jitterSize`, so the boundary wanders instead of tracing a contour. `invert` swaps what the band selects. The `height`/`slope`/`curvature` inputs are derived from the raster's own "height" layer, read through `op_source` at the same resolution as the layer this op writes — one indexed read per cell, no resampling. `slope` and `curvature` are central differences over `step` world metres. The `curve` input reads a polyline's sampled control points off `op_payload` — the same array `terrain.project` rides — and measures distance to it in the ground plane; a control point's height plays no part in the distance. `blend` decides how the band's weight meets what the layer already holds: `replace`, `add`, `max`, or `min`. A layer written fresh is usually `replace`; combining two masks into one accumulated layer is `add` or `max`. The written value is a weight in 0..1 — a layer this op writes is what `terrain.mask`'s `name` param names, and any other raster kernel op's own `mask` param can reference it by that name, multiplied in or subtracted out, through the generic `op_mask`/`op_common` contract every raster kernel op already carries.

# terrain_fbm Domain-warped fractal noise written into one scalar channel of one tile. Gradient noise with quintic interpolation. Sample positions come from the field-wide cell index, the field's minimum corner, and the seed, so a tile evaluated on its own is identical to the same tile evaluated as part of the whole field — and a cell reached as one tile's apron lands on exactly the world position it has as its neighbour's interior. `featureSize` is the world size in metres of the largest noise feature. Halving `resolution` therefore refines the same landscape instead of producing a different one. `warp` is a world distance: how far the domain is displaced by two lower-octave fields before the main fractal is sampled. Non-zero warp is what turns isotropic lumps into ridgelines and basins. `ridge` folds the fractal about zero and squares it, giving sharp crests and rounded valleys. `blend` decides how the contribution meets what the channel already holds: `replace`, `add`, `max`, or `min`. Every op carries a mask. The op's weight at a cell is scaled by what the mask reads there — multiplied in, or subtracted out — so the op applies only where the mask lets it and a cell the mask holds out keeps exactly what it already held, at every blend mode. An op that names no mask writes at full strength everywhere.

# terrain_project A curve drives the ground: the points of the curve a geometry holds ride `op_payload`, and every cell's height is banded by its ground-plane distance to the resulting polyline — one dispatch, so the channel it cuts has no seam where a chain of stamps would have met. Inside `width`, falling to zero across `falloff` metres beyond it, a cell blends toward a target height: - `useCurveHeight` set: the curve's own y at the nearest point on the polyline — a road or a riverbed that follows the curve's authored elevation. - otherwise: `fromHeight` .. `toHeight`, interpolated by how far along the polyline (ground-plane arc length, 0 at the first control point, 1 at the last) the nearest point falls — a ramp that runs the whole curve's length in one dispatch instead of one `ramp` stamp per segment. `blend` decides how the target height meets what the layer already holds: `replace`, `add`, `max`, or `min`. The polyline is open — segment `i` to `i+1` for every consecutive pair of the curve's sampled points — and distance and arc length are both measured in the ground plane; a point's height does not distort either.

# water_surfaces The water surfaces a scene holds, gathered from the systems that make them. A system with water — a terrain's lakes, rivers and seas, an ocean — registers a provider here, and whatever stands things on water asks every provider at once. A procedural graph reads them all through `scene.water`. A provider is a table of two functions: - `version()` names what its water is now. A surface that moves every frame answers differently every frame, which is what makes a reader stand its points again. - `stand(request)` puts points on its water on the device. `request` holds: - `run`: the points' GPU positions, in the graph's space; - `count`: how many points there are; - `out`: the `P`, `normal`, `slope` and `wet` GPU columns to write; - `place` / `unplace`: graph to world and back, 16 numbers each; - `offset`: metres above the surface; - `keepY`: whether a point's own height rides on the surface's. It writes the points over its water, leaves the rest as they are, and returns the names of the kernels it dispatched. ```luau local water = require("@builtin::modules.water_surfaces") water.register("myWater", { version = versionNow, stand = standOnMyWater }) for _, p in water.list() do print(p.name, p.provider.version()) end ```

# terrainMesh Puts a heightfield on screen as **one draw**. ## What it is A vertex **pool** of `maxNodes` slots, each a fixed `nodeVerts` square grid (33x33 by default). A quadtree **cut** decides which parts of the ground are drawn at which density: near nodes are small and dense, far nodes large and sparse, and every node meshes to the same grid, so a node costs the same wherever it is and the cut stays a few hundred nodes at any camera distance. `terrain_mesh` fills the slots the cut needs; `terrain_index` writes the index run over them; `renderer.mesh.create { vertexBuffer, indexBuffer }` turns the pair into one mesh. One entity carries one `Model` bound to it, so the whole terrain — every node, at every level — is one indexed draw. **Nodes are never entities.** Entity count is the binding constraint on this renderer; a per-chunk entity would put a hundred draws where this puts one. A step that seems to need one is a step that has the design backwards. A cut change calls `renderer.mesh.update` under the **same guid**, so the Model is never re-bound and the scene graph is never touched. ## Using it ```luau local terrainHeightfield = require("~terrainSystem.modules.terrainHeightfield") local terrainMesh = require("~terrainSystem.modules.terrainMesh") local hf = terrainHeightfield.open(groundTexture, { size = { 2048, 2048 } }) local mesh = terrainMesh.create(hf, { camera = terrainMesh.cameraPosition() }) terrainMesh.updateCut(mesh, hf, { x = 900, y = 30, z = 900 }) ``` The `Terrain` component does all of this from a `.texture` on an entity and follows the on-screen camera with the cut every frame, in edit and in play alike. Reach for the module directly only when the ground is not an entity's. ## The vertices `terrain_mesh` writes the engine's 23-word vertex layout, one thread per node vertex, positions in WORLD space — so the entity carrying the mesh stands at the world origin with no transform of its own. It reads the heightfield's texels as one array over the whole rect. The height is bilinear across the four texels around a position, and the normal comes from the gradient those same four texels already carry — four reads a vertex rather than the twenty a central difference of the height would cost. ## Holes `terrain_index` writes one node's triangles, and a quad whose ground the landscape cuts out is written DEGENERATE — six copies of one index — rather than left out. The run is a fixed stride per node, so a quad that vanished would shift every quad after it; a triangle whose three corners are one point covers no pixels and is discarded before it rasterises. A hole therefore costs the index slots it already owned and nothing else. The quad's own centre decides, so the edge of a hole follows the mask at the density the node is meshed at. The run carries the heightfield's generation as well as the slot order, so a mask re-read under a standing terrain moves the hole with it. ## Cracks and popping Two things happen at node edges, and both are in `terrain_mesh`: - **Stitching.** The cut is *balanced* — no leaf borders a neighbour more than one level coarser — and each node carries the mask of which of its four edges do. On such an edge the fine node's odd-indexed vertices snap onto the previous even one, so the fine edge becomes the polyline through the coarse neighbour's own vertices. The two edges are then the same line. - **Geomorphing.** Over the far end of the distance band a node is a leaf across, its vertices slide onto the positions they would hold one level coarser. By the time the cut merges the node into its parent, the node is already drawing the parent's geometry. The morph runs off each vertex's own distance to the camera, so a vertex two nodes share is placed identically by both. ## What re-meshes, and when A node is re-meshed when it is new to the cut, when its coarse-neighbour mask changes, when the camera moves (the geomorph is camera-relative), or when the heightfield's `gen` has moved past the generation that node was last built at. Nothing else touches it — `updateCut` from a viewpoint the cut was already taken from remeshes zero nodes and reports `changed = false`. `terrainMesh.invalidate(state, rect)` marks the nodes overlapping a world rect, for when something changed the ground in one place. ## Nodes that are not drawn `create` takes a `keep` predicate — `(node) -> boolean` — and the cut then covers whatever it keeps rather than the whole rect. That is what lets the terrain END before the grid does: an island's seabed is ground nobody can see, and ground nobody can see is ground nothing has to mesh. The predicate is a parameter of selection rather than a filter applied afterwards, because the node budget is spent on what is **drawn**: a cut that coarsened to fit nodes it then threw away would be spending the budget on seabed. It runs after the tree is balanced, so what it drops cannot unbalance what remains, and a node whose coarser neighbour was dropped has nothing on that side to crack against. ## Ground that is not square The quadtree tiles the rect with square **root** nodes. A square rect is one root; an oblong one is a row or grid of them, sized by `rootSize`, which defaults to the shorter side. A rect whose sides are not a whole number of `rootSize` squares is refused by name — pass a `rootSize` that divides both. ## The bounding box The mesh's box is the heightfield's rect by its declared `heightRange`. It is taken from the declaration rather than reduced from whatever is drawn, because a box that is too small culls the terrain away at grazing camera angles and the symptom is nothing at all on screen. ## Entity binding `terrainMesh.bind(entityId, hf, state)` records that an entity's geometry IS this heightfield drawn through this pool. `forEntity` reads it back, `isBound` asks, and `unbind` releases the mesh, the pool and the heightfield's pixels together. `followCamera(entityId)` is the per-frame entry the `Terrain` component calls.

# terrain_shape Blends an analytic shape into one scalar channel of one tile. One shader covers every placed primitive, selected by a mode lane: | mode | shape | what it draws | |---|---|---| | 0 | disc | a circular pad held at one altitude | | 1 | square | the same, measured in the Chebyshev metric | | 2 | cone | a rounded hill, cosine profile, tangent to flat at the rim | | 3 | ridge | that profile swept along a segment | | 4 | ramp | an altitude interpolated along a segment | | 5 | terrace | the existing height quantized into steps | A negative height on a `ridge` cuts the same profile downward, which is how a valley or a sunken road corridor is drawn. The shape's coverage at a cell is a weight in 0..1 — full inside `radius`, falling to zero across `falloff` metres beyond it — and that weight is what the blend is applied through. A flatten therefore settles into the terrain rather than stamping a cliff-edged plateau into it. Every coordinate is world metres. A shape about a point puts that point in both segment endpoints; a shape along a line puts the two ends in them. `radius` is the half-width for the segment shapes. Because nothing here is in cells, the same op covers the same ground at any resolution. `blend` decides how the shape's value meets what the channel already holds: `replace`, `add`, `max`, or `min`. The rounded shapes read best with `add`, which lays a hill on top of the terrain instead of replacing it; the level shapes read best with `replace`. Every op carries a mask. The op's weight at a cell is scaled by what the mask reads there — multiplied in, or subtracted out — so the op applies only where the mask lets it and a cell the mask holds out keeps exactly what it already held, at every blend mode. An op that names no mask writes at full strength everywhere.

# terrain_stamp Blends a source texture into a scalar channel under a world-space transform. The source is an ordinary `.texture`. That is the point: importing a real-world elevation model and placing an authored shape are the same operation, so there is one path for both and no separate importer. Sampling runs in world space. Each destination cell's world position is carried into the stamp's local square — translated to `at`, rotated by `rotation`, scaled so the square spans `radius * 2` metres — and the texture is read there. A cell outside that square is left untouched, so the op's declared extent is exact and the evaluator dispatches only the tiles the stamp reaches. `channelIndex` picks which of the texture's four channels carries the value; `amplitude` scales it into metres and `bias` offsets it, so an 8-bit greyscale heightmap in 0..1 becomes real elevation. `falloff` feathers the disc inscribed in the square, which is what lets a stamp meet the surrounding terrain without a seam. Compare `terrain_layer`, which blends a raster that already shares the field's grid and therefore needs no sampling at all. Every op carries a mask. The op's weight at a cell is scaled by what the mask reads there — multiplied in, or subtracted out — so the op applies only where the mask lets it and a cell the mask holds out keeps exactly what it already held, at every blend mode. An op that names no mask writes at full strength everywhere.

# terrain_layer Blends a co-registered raster layer into a scalar channel. A layer shares the field's grid exactly — same tile, same cell, same apron — so there is no sampling and no transform involved, only a blend. Compare `terrain_stamp`, which samples a texture through a world-space transform because the source has a grid of its own. That exactness is what lets freehand editing live inside a parametric stack. Brush strokes accumulate into a layer, the layer is one `layer` op among the others, and changing an op above it re-runs the stack with the strokes still in place. Painting is an edit to the layer's raster rather than a bake over the terrain the ops produced. `opacity` scales the layer before the blend, so the same raster can be dialled in and out without repainting it. `blendMode` follows the numbering every terrain op shares: 0 replace, 1 add, 2 max, 3 min. Every op carries a mask. The op's weight at a cell is scaled by what the mask reads there — multiplied in, or subtracted out — so the op applies only where the mask lets it and a cell the mask holds out keeps exactly what it already held, at every blend mode. An op that names no mask writes at full strength everywhere.

# terrain_coast Pulls the ground outside a landmass down to a seabed altitude, by a radial falloff about a point. The land weight is 1 inside `radius`, falls to 0 across `falloff` metres beyond it, and the height is interpolated between the seabed and whatever the stack already built: ``` value = mix(seabed, prev, landWeight) ``` The interior keeps its landscape untouched, the rim slopes off it, and everything past the rim sits at the seabed. Put the seabed below the sea level and the result is an island — the terrain ends before the field's rect does, and the rect goes on carrying the sea. `shapeMode` picks the metric: 0 is a disc, 1 is the Chebyshev square. This op covers the **whole field**, not the disc it is centred on: the ground it changes is the ground outside that disc, and an op that only touched its own radius would leave the far corners standing above the water. It is an ordinary op in the stack, so an erosion below it weathers the coastline it drew, a flatten cuts a harbour into it, and a river carves a mouth through it. `terrain_coast_mask` is the same operation driven by a painted raster instead of a radius, for a coastline whose outline is drawn rather than described. Every op carries a mask. The op's weight at a cell is scaled by what the mask reads there — multiplied in, or subtracted out — so the op applies only where the mask lets it and a cell the mask holds out keeps exactly what it already held, at every blend mode. An op that names no mask writes at full strength everywhere.

# terrain_coast_mask Pulls the ground outside a landmass down to a seabed altitude, with the landmass given by a painted raster rather than a radius. The raster is a layer co-registered with the field — one texel is one cell — so the outline an author paints is the outline the coast follows. The value is the land weight directly: ``` value = mix(seabed, prev, clamp(mask * gain, 0, 1)) ``` 1 is land and keeps the landscape untouched, 0 is open water and sits at the seabed, and the values between are the shoreline slope: the softness of the brush is the softness of the coast. `gain` scales the raster before the clamp, so a stroke painted at a fraction of full strength can still read as solid land. Like the radial form, this covers the **whole field** — the ground it changes is the ground the mask does not cover. `terrain_coast` is the same operation driven by a radius, for a coastline described rather than drawn. Every op carries a mask. The op's weight at a cell is scaled by what the mask reads there — multiplied in, or subtracted out — so the op applies only where the mask lets it and a cell the mask holds out keeps exactly what it already held, at every blend mode. An op that names no mask writes at full strength everywhere.

# terrain_water_kernel Fills a water surface over a region of one tile, into the `water` and `waterMask` layers of a `proc.raster` `Raster` — `terrainOps.water`'s `lake`, `river` and `sea` share this one dispatch. One shader covers every water body, selected by a region-mode lane: | mode | region | what fills | |---|---|---| | 0 | disc | a round body about a point — a lake | | 1 | square | the same, measured in the Chebyshev metric | | 2 | path | a body swept along a segment, its surface running from one end's level to the other's — a river | | 3 | everywhere | the whole raster — a sea | A cell takes the surface only where the ground stands **below** it, so a body fills the basins its region reaches and leaves ground that rises out of it dry. That is the whole difference between a lake and a sea: the region an op is applied to. The ground it reads is `op_source`, bound to the raster's `height` layer. ## The two layers are one thing | layer | what it holds | |---|---| | `water` | the world-space Y of the surface covering that cell | | `waterMask` | 1 where a surface covers the cell, 0 where none does | A mask of 0 is what "no water here" IS — there is no height in the `water` layer standing in for it, so no authored water level can ever be mistaken for an absence. Everything that reads water reads the pair. ## Bodies meet through `blend` Where two bodies reach one cell, the second meets the first through `blend` — `replace`, `add`, `max` or `min`. The default `max` makes the meeting independent of authoring order: the higher surface wins whichever op ran first, so a lake in a coastal basin reads the same whether it was placed before or after the sea. Every coordinate and every level is world metres, so a body keeps its size, its place and its altitude when the raster's resolution changes. Every op carries a mask. The op's weight at a cell is scaled by what the mask reads there, and here the water mask layer is what carries that coverage: a cell the mask holds out keeps both layers as they stand, and a partly covered cell holds part of the surface it already had. An op that names no mask fills at full strength everywhere.

# assetType (asset type) The `assetType` is the **type of types** — the meta-type that every `<typename>.assetType/` folder is an instance of, including itself. A `<typename>.assetType/` folder declares a new asset type called `<typename>`: its on-disk shape, its identity rule, and the structural contract instances of the type must satisfy. This type is self-hosting. `assetType.assetType/` conforms to its own schema — a `type.yaml` plus a `README.md` — so the type system is defined in terms of itself rather than a hidden engine special-case (the same way a C compiler is written in C). Every asset in the engine, including each `.assetType` definition, now resolves to a registered assetType: a `.material` resolves to `material.assetType`, and `material.assetType` resolves to `assetType.assetType`, which resolves to itself. ## Why it exists Assets used to be mapped to their type *implicitly*, by matching the folder suffix against a name-keyed registry. That match carried no stored link: nothing on a `Foo.material` recorded *which* `material.assetType` it was written against, so two worlds shipping a same-named type could silently disagree. Making `assetType` a real, resolvable asset closes that gap — the link from an asset to its type is now an explicit reference pinned in the asset's `.refs` sidecar (`via: "asset_type"`), exactly like every other dependency. ## Where it lives - Source: `/zero/source/.../<typename>.assetType/` - Identity: `<typename>` (the `.assetType` suffix strips from the identity; the folder retains the suffix on disk). - Folder shape: - `type.yaml` — the structural spec the validator reads. **Required.** - `README.md` — type-level documentation. **Required.** - `about.yaml` — three short plain-language texts saying what the type is for (see "Say what your type is for" below). Every built-in type ships one. - `behavior.luau` — optional behavior/scaffolding module. - `template/` — optional canonical placeholder body the templater copies into a new asset of the type. - `<name>.module/` — optional shared code the type ships to **every instance** of it, conventionally `shared.module/` (see below). ## Shared code its instances reach (`asset.containing` + `.modules`) A type ships code every instance of it uses by declaring the modules in its own `behavior.luau`, as a map of tracked `require`s: ```luau -- inside <typename>.assetType/behavior.luau M.modules = { shared = require(".shared") } -- the sibling shared.module/ ``` An instance reaches it with one location-independent line: ```luau -- inside any foo.<typename>/init.luau local api = asset.containing(__FILE__).modules.shared ``` `__FILE__` is the chunk's own VFS path; `asset.containing` resolves the calling asset, and `.modules.<name>` follows the instance's pinned `typeRef` (the type's **guid**, recorded in the instance's `.refs` as `via = "asset_type"`) to the module its type declares. Because resolution follows that link rather than a category name, the same line in an instance of a different type resolves to that type's module, and identically-named modules in two types never collide. `asset.typeRef(target)` reads that guid for any target; `asset.resolve(asset.typeRef(target))` gives the full `AssetRef<assetType>`. See `assetTypes/README.md` § "Shared code: type-owned modules" for the full mechanism. ## How to create one Authoring a new asset type is just creating a `<typename>.assetType/` folder under `/zero/source/`. The engine registers `<typename>` as a category **the moment the folder is written** — no engine restart, no Rust change, and it works for built-in (`@builtin/assetTypes/`) and user-defined types identically. Once `<typename>` is registered, `<name>.<typename>/` folders are recognised as assets of that type and resolve their `asset_type` link back to this definition. Creating a `<name>.<typename>/` folder *before* its `<typename>.assetType/` exists fails fast: the suffix has no registered type, so the folder can't be a valid asset. Author the type first. ## Say what your type is for (`about.yaml`) Every type carries a small `about.yaml` beside its `type.yaml`, and the README you write for the type documents how to author an asset of it. Both are required, and both are meant to be written as you make the type, while you know what it is for. `about.yaml` answers three questions in plain words, one key each: ```yaml intent: "Sets the typeface of the text in your game: UI, labels, signs and 3D lettering." function: "A font, from a TTF, OTF or WOFF file, made available to all of Zero's text, including 3D text." usage: "You choose it by name for your UI or your text, and the text appears in that typeface." ``` - **`intent`** is what the type is for in someone's game. Name the whole range of things it covers, in the words a game maker would use for them. - **`function`** is how it works today: what an asset of the type is made of and what the engine does with it. - **`usage`** is what a person or an agent does with one: how they pick it, place it, call it, or hand it to something else. A few habits keep it useful. Write about the type as a whole, so the text holds for every asset of it, and leave out any one asset's name. Say what the type does right now. Stay with words a game maker would say, and mention a file name or a call only where a game maker would know it. The limits are in bytes: `intent` 800, `function` 1500, `usage` 1200. Each is generous, and a few sentences usually say it. Why bother: agents working in a world read these three texts to learn what your type is for, so they reach for it when it fits and use it the way you meant. People searching find the type by what they want to make, and the three texts let them tell at a glance whether it is the one they need. `asset.create("assetType", name)` scaffolds the file with all three keys, and a README whose first paragraph is a `[[describe]]` note. Each value starts with `[[describe]]` followed by a note on what to write; replace the whole value with your own words. A value or a README line that still starts with `[[describe]]` counts as not written yet, which the push's check of each asset it carries against its type reports, naming the file. A type you leave as it was is not asked, so older types stand. Many types at once, or one you would rather not write by hand: `services describe` lists the types that lack these texts and what filling them costs, and `{ mode = "all" }` writes them into the same files, where you edit them like anything else you wrote. Where the texts show up: `asset.inspect("<typename>", "assetType")` returns them as `detail.about`, `assetTypeRef:getAbout()` returns them as a table with the keys `intent`, `function` and `usage`, the type's `types/<typename>` guide page ends with an "About this type" section holding them, and the Inspector's Type section lists them at the top. The guide page is this README plus that section, so a README that explains the type well and an `about.yaml` that says what it is for read well together. ## Ask your assets to say what they are A type can ask each of its assets to carry a description: a file, written by the author or generated, that says what that asset is and what it is for. The people and agents who find an asset read it to decide whether it fits, and the author knows what it is for at the moment of making it. **1. Declare the file with a `described` check.** This is the whole requirement: ```yaml files: required: - path: "README.md" # any file name works kind: "file" content: - { check: "described", message: "Say what this widget is for and where it goes." } ``` With that in place: - An asset whose file is missing, empty, or still has a line starting with `[[describe]]` does not meet the type's requirements. `asset.validate` names it, `zm status` and `world.publishBlockers()` list it, and a push that carries it is refused with `message`. - `services describe` finds it (`mode = "find"` lists it and the price) and can write the description for it (`mode = "all"`). It writes only this file. For a YAML file that holds several texts, name the keys and their limits in bytes, the way this type does for `about.yaml`: ```yaml content: - check: "described" keys: - { key: "intent", max_bytes: 800 } - { key: "usage", max_bytes: 1200 } message: "Write intent and usage." ``` `headings` names sections a Markdown file must have. `assetTypes/README.md` has the full check. **2. Optionally, steer what gets generated with a `describe:` block:** ```yaml describe: guidance: "Say what the widget shows, what input it reacts to, and where it fits in a UI." examples: - "<identity of an asset of this type whose description reads well>" ``` `guidance` says what a description of your type covers, and `examples` names assets whose descriptions show the shape you want. Without the block, generation still works from the general describing guide and the check's `message`. An example is an asset of your type in the library whose description is a true, written one, never a scaffold's placeholder, and never the asset being described. When the library has none, leave `examples` empty rather than name one that is not. A type whose assets the model cannot perceive (a sound it cannot hear, a motion it cannot see) says so in `guidance`: the description comes only from what the asset's own files state, such as its name and the `.metadata` where a generation service records the prompt that made it, and when they state nothing the model answers `cannot` with that reason. `services describe` then fails that one asset with the reason and leaves its file as it was. **3. Start the file unwritten in your template.** Put a heading and one line that starts with `[[describe]]` and a short note on what to write, in `template/<file>`, so a new asset starts unwritten rather than with text that only looks written. Do not put example content in it: every new asset carries the template verbatim, and the same text on many assets makes them impossible to tell apart in search. Types whose assets say what they are through their own content (a mesh, a texture, a sound) leave this out: declare nothing and nothing is asked. ## The `behavior.luau` contract `behavior.luau` is the type's code. It returns ONE table, and the keys below are the ones the framework reads off it for EVERY type — the surface you get by declaring them. Anything else on the table is your own: reachable by name through `require("@builtin::assetTypes.assetType.shared.ref").loadTypeModule("<typename>")` for a system that knows your type (`computeShader` publishes `dispatchKey` that way), and never called by the framework. ```luau local M = {} M.ref = { ... } -- methods on every AssetRef of this type M.modules = { ... } -- shared code the instances reach M.global = {} -- reserved key; ship the empty table function M.onCreate(name, opts) ... end return M ``` ### The tables | Key | Shape | What the engine does with it | |-----|-------|------------------------------| | `M.ref` | `{ [name] = function(self, ...) }` | Every function becomes a method on every `AssetRef` of this type: `ref:name(...)`, with `self` the ref. This is how a type gives its instances an API. Two names in here are contracted — see below. | | `M.modules` | `{ [name] = require(".sibling") }` | Shared code the type's instances reach through `asset.containing(__FILE__).modules.<name>`, resolved by the instance's pinned `typeRef` guid. See the section above. | | `M.refShapes` | `{ [method] = "TypeName" }` | Declares the result type of a `M.ref` method so the LSP can check what a call site does with it. See the `M.refShapes` section below. | | `M.events` | an events schema | The events an ASSET of this type fires, declared the way a component declares its own. Readers subscribe through `ref.events.<name>:connect(...)`; firing authority stays with the type. Keyed by the asset's guid, so every resolver of the same asset shares the signals and a re-resolved ref re-attaches to subscriptions already there. | | `M.namePattern` | a Lua pattern string | The name shape `asset.create` enforces for instances of this type, replacing the default `^[A-Za-z][A-Za-z0-9_]*$`. | | `M.global` | `{}` | Reserved. Ship the empty table. | ### The lifecycle hooks Each is optional; a type that omits one costs nothing. The engine calls them: | Hook | Signature | When it fires | |------|-----------|---------------| | `M.onCreate` | `(name, opts) -> { [filename] = contents }` | On `asset.create("<type>", name, opts)`. Returns the files that scaffold the new instance, as a map of relative path to contents; the framework writes them. The `opts` type annotation is the schema `asset.create` validates the caller's arguments against, so annotate it. | | `M.onReplace` | `(dest, files) -> { [filename] = contents }` | On an `asset.create` that writes over an asset already at `dest`, after `onCreate`. Receives the files `onCreate` produced and returns the files to write, so a type that re-derives its content (a rig re-imported from its source file) can carry what an author added to the asset it replaces. Optional. | | `M.onRegister` | `(self)` | Once per instance, the first time it registers: on `engine.onWorldLoaded` for instances already in the world, immediately on a live `asset.create`, and when an instance arrives after that sweep (installed, pulled, or synced from a peer). Guarded by the instance's path, so a double trigger never double-registers. An instance that leaves the runtime releases the guard, so one arriving at the same path again registers. This is what makes "write an asset into the world and it takes effect live" work for content that registers into a runtime registry. The `@builtin` library is excluded from the sweep. | | `M.onChange` | `(ref, change)` | On every VFS write to a file INSIDE one of this type's instances. `change` is `{ path, asset, type, kind, origin }`: `kind` is `"edited"` or `"seeded"`, `origin` is `"local"` or `"remote"` (an importer runs on the originator only). **Filter on `change.path`**: the hook fires for any file under the folder, so act on the one you care about. It runs as its own task and may wait on what it reads: a `vfs.read` of bytes only the blob store holds (a binary file on the web build) fetches them. The writes to one asset reach the hook in the order they were made, each once the hook before it has finished. A re-entrancy guard suppresses writes back into the same asset routed inline from the hook, and a write the hook makes is routed on a later frame like any other, so the hook must be convergent on its own: diff the meaningful state and short-circuit while your own regeneration is in flight. | | `M.onDelete` | `(ref)` | When an instance of this type leaves the runtime: it is deleted, or the engine swaps away from the world binding that held it. Its files are gone from `/zero/source` by then. A removal routed while the instance's `onChange` is waiting reaches `onDelete` once that hook, and every write queued behind it, has finished. On a swap the instance still exists in the world it belongs to, and a VFS write the hook makes reaches the world now bound. | | `M.validate` | `(assetRef) -> { { code, message, severity? } }` | On `asset.validate`, after the structural `type.yaml` check, for the type's own SEMANTIC validation. `severity` defaults to `"error"`; error-severity problems flip `ok` to false, warnings do not. A hook that raises or returns a non-table is itself reported as a `validate.hook_failed` error. `world.push` runs this per user asset, so declaring it enforces your type's rules at publish time with no further wiring. | ### The two contracted `M.ref` names Most `M.ref` methods are your type's own surface: name them what you like, return what you like. Two are read by the engine and mean the same thing for every type, so their shape is fixed: - **`instantiate(self, target?, opts?) -> (root, idMap)`** — makes an instance part of the scene. Defining it is the whole opt-in: `ref:canInstantiate()` is true exactly when it exists. Its **return is a contract** — see the next section. - **`inspect(self) -> detail`** — the type-specific half of `asset.inspect`. See its section below. ### Names a type cannot shadow Some keys resolve on every `AssetRef` before per-type dispatch, so an `M.ref` entry of the same name is never reached: `canInstantiate`, `getSource` / `getBytes` / `getText`, `exists`, `deps`, `meta`, `runtime`, `events`, `modules`, `typeRef`, and the residency flags `has_backing_asset`, `has_runtime_changes`, `cpu_resident`, `gpu_resident`. These are the behaviours every asset must expose identically, which is why they win. ### Reloading A `behavior.luau` edit ripples to existing refs with no restart: dispatch reads `mod.ref` fresh each time and `require`'s cache is re-run in place. A type whose folder appears at runtime is picked up on the next dispatch — negative results are not cached either. ## The names your type's assets hold (`references:`) An asset that names another asset by a string depends on it: the named asset travels with it into a pack or a pull, and a name nothing answers to is an unresolved dependency `worldValidation` reports and the push gate refuses. The engine reads `require(...)` and the asset lookups on its own; a name your type's format or API carries anywhere else is one you declare, in `type.yaml`: ```yaml references: # A top-level key of this type's own files that holds a name (or a table of them). keys: - { key: "shader", type: "shader" } # A call's argument, counting from 1, wherever the call is written. args: - { call: "g:node", position: 2, type: "procNode" } # A field of the options table a call takes. fields: - { call: "fx.beam", field: "material", type: "material" } ``` `type` is the category the name resolves in, one of `asset.categories()`. An argument declared once per category, at the same `call` and `position`, takes the name from whichever category answers. A key or a field holds one category: declared again with another `type`, it takes the later one. The block is declared as the type registers, when the library or the world holding it is laid into the VFS, so every reference table reads it whatever script has or has not run. `asset.declareReferenceKey`, `asset.declareReferenceArg` and `asset.declareReferenceField` declare the same things from a script, for a surface that is no type's. ## What search does with your type (`indexing:`) A type declares what search embeds for its instances. **This is not optional and there is no useful default**: ZeroMind runs your declaration and adds nothing of its own, so anything you do not point at is absent from the index — silently, and for every instance of the type forever. Twenty types once shipped with no `indexing:` block and every instance of them was unfindable. A type whose instances are never searchable (a run record, a log) says so with `indexing: none`; every other type declares a recipe. A type without an `about.yaml` beside its `type.yaml` refuses every push that carries it, whether or not the push edited the type. Texts in it that are not written yet follow the rule every `described` file follows: they refuse the push that edits, stages or commits the type. Two words carry the whole thing, and they are the same two the search tool exposes: - **`identity`** — what an instance IS. For anything you can look at, that is the picture (`content:` + `modality: image`, matched directly by a text query). For everything else it is the authored text that says what it is. - **`capability`** — what it DOES and how it is made. Code, settings, the model-written summary of them. A type with no behaviour has no `capability` entry. A type with nothing to look at has no image. Leaving a slot empty is a statement; filling it with whatever happens to be lying around is not. The one you have to decide when authoring a type is: *for an instance of this, what is the thing a person would recognise it by, and what is the thing it does?* A texture answers "the image" and "its compression settings". A module answers "its README" and "its code". Write those two answers into `indexing:` and the rest follows. Three specifics worth knowing before you write one: - **`derive` is instructed by this type's own `README.md`.** A model reads an instance's source and writes a description; what it embeds is the description, not the source. It follows your README to know what it is looking at, so a type whose README says what its instances are gets good derivations for free. There is no prompt to name. - **`derive` loses detail.** It keeps the main technique and drops secondary ones. If the source is code you want searchable by its own vocabulary, declare the same files a second time as `capability` with `extractor: verbatim`. Two entries, same role, different failure modes. - **Never `source: { field: name }`.** A filename is one or two words with no usable embedding, and a corpus indexed that way answers "language runtime" with `say_runtime_2.soundClip`. `facets:` is the second stage: a `file:` embedded whole (`.metadata`, so keys nobody declared still get indexed) plus computed keys that state a value for every instance including `false` — the raw file can imply "not rigged" and can never say it. The full schema, every extractor and the facet sources are in [`assetTypes/README.md`](../README.md); the scaffolded block with inline guidance is in [`template/type.yaml`](template/type.yaml). ## A container whose contents work together (`system: true`) A type that is a container can say its contents are one working whole by declaring `system: true` at the top level of its `type.yaml`. `package` declares it: the components, modules, scenes and tools a package ships are parts of one system. The declaration is plain data, read from the decoded `type.yaml` (`require("@builtin::assetTypes.assetType.shared.declaration").of("package").system`), and a type that leaves it out is an ordinary type. ## What your type looks like (`preview:`) A type declares how its instances are pictured, and the engine makes the picture when something needs it: staging an instance whose `preview.png` is missing or pictures content it no longer holds queues a render, a view that shows the instance asks for one, and `ref:preview()` renders one on demand. An edit renders nothing on its own. Each `preview.png` records the content it pictured, which is how a stale one is told from a current one. The preview studio owns the light, the backdrop, the framing and the camera, and nothing in the world reaches it, so an instance pictures the same in every world. Declare one of four recipes: | `kind` | The picture is | Your type provides | |---|---|---| | `instantiate` | the asset stood in the studio through its own `instantiate` | `ref.instantiate` | | `image` | a file the instance holds, fitted to the frame, nothing rendered | `source.files`, the candidate files in order | | `surface` | the asset painted on a builtin mesh | `mesh`, and `ref.previewMaterial(self) -> { material }` | | `custom` | whatever `ref.preview(self, opts)` returns | `ref.preview` | ```yaml preview: kind: surface mesh: sphere ``` A scene-instantiable type that declares nothing still answers `ref:preview()` through its `instantiate`, but persists no picture: the declaration is what makes the engine keep one. List `preview.png` among the type's files so an instance may hold it. The picture is findable only when `indexing.content` names it. The two are separate statements, and neither is derived from the other: ```yaml indexing: content: - role: identity modality: image source: { file: "preview.png" } ``` `@builtin::assetTypes.assetType.shared.previewRecipe` reads the declaration; the scaffolded block with inline guidance is in [`template/type.yaml`](template/type.yaml). ## Typing a result from the instance (`M.refShapes`) Every method on `M.ref` is checked against the type its `--!return` names, and every asset of the type gets the same one. That is right for most methods and wrong for the ones whose result is shaped by the ASSET: a method answering one entry per child folder, per row of a config, per binding an input map declares. The widest true annotation for those is `{ [string]: Thing }`, and an indexer accepts every key — so a caller's typo reads as valid and nothing reports. `M.refShapes` lets the type state the result per instance. Each entry is `function(self) -> (typeExpression, source?)`: a Luau type expression for THIS asset, and the module whose type vocabulary the expression uses. Read the asset; never run it — stating a type must not take effect. ```lua -- inputMap.assetType/behavior.luau — the shipped example. -- `activate()` hands back one handle per control the map declares. M.refShapes = { activate = function(self): (string, string) local names = {} for _, control in ipairs(M.ref.controls(self)) do table.insert(names, control.name .. ": Handle") end if #names == 0 then return "", "" end return "{ " .. table.concat(names, ", ") .. " }", "@builtin::modules.zinput.scheme" end, } ``` With it, `map.jump:onPressed(fn)` resolves and `map.noexisting` is reported with the map's real controls. Without it, both pass silently. ### What the call site has to say A per-instance type is matched by the asset's IDENTITY, so it applies only where the reference the method is called on says WHICH asset. Two spellings do: ```lua -- a component field, whose declared default names the asset public = { map = Field.assetRef("inputMap", "@builtin::inputMaps.default", Sync) } local controls = public.map:activate() -- typed for THAT map -- an annotation, naming category and identity local m: AssetRef<"inputMap", "@builtin::inputMaps.default"> ``` `asset.resolve("<identity>", "<category>")` carries the CATEGORY, so the methods the category defines are checked on the result — but not which asset it found, so a per-instance result keeps the category's declared return. A reference that names no asset (`asset.resolve(someVariable)`, a value passed in as a parameter) carries neither, and calls on it are unchecked. That is the same rule everywhere: the checker states what the code states, and a name computed at runtime states nothing. ### The result is a value, not a binding The type belongs to the expression, so it travels the way any inferred type travels. Indexing the call's result is checked; stashing it in a module local and indexing it from another function is not, because the local's declared type is what carries across — and `any` (or no annotation) carries nothing: ```lua -- checked: `d.greting` is reported function awake() local d = public.dialogue:lines() d.greeting:say() end -- NOT checked: `lines` is a module local typed `any` local lines: any = nil function awake() lines = public.dialogue:lines() end function start() lines.greting:say() end ``` Bind what you need at the call site and the names stay checked in both places — the asset-derived type at the boundary, ordinary scoping after it: ```lua local lineGreeting = nil function awake() local d = public.dialogue:lines() lineGreeting = d.greeting -- `d.greting` is reported here end function start() lineGreeting:say() end ``` ### When nothing is reported and you expected something From a call site, a type that is correct and a type that was never published look the same: no diagnostic either way. `require("@builtin::assetTypes.assetType.shared.refShapes").published()` answers which it is. It returns two maps — the category surfaces, and the per-asset results keyed by identity and then by method: ```lua local shapes = require("@builtin::assetTypes.assetType.shared.refShapes").published() shapes.categories.dialogue --> "{ lines: () -> any, lineNames: () -> { string } }" shapes.returns.Shopkeeper.lines --> "{ browsing: Line, farewell: Line, greeting: Line, wares: Line }" ``` An identity absent from `returns` was never published for, which is a different thing from published-and-correct — and the reason to look here rather than at the call site. The answers are recomputed when an instance is written, so a control added to a map reaches the checker with no restart. Only types that declare `refShapes` are read per instance, and at most 64 assets of one type are — past that the type keeps its declared return and the engine log names what was dropped. ## The instantiate hook (`M.ref.instantiate`) — a contracted return A type opts into becoming part of the scene by defining `instantiate` on its `ref` table. That is the whole opt-in: `ref:canInstantiate()` is true exactly when the hook exists, so consumers offer a scene path — an `Asset.source` field, a viewport drop, a tool argument — by CAPABILITY rather than by a list of type names. Unlike `inspect`, this hook's **return is a contract**. A caller writes one piece of code against every instantiable type, so what comes back cannot vary by type: ```lua -- behavior.luau local Instantiable = require("@builtin::assetTypes.assetType.shared.instantiable") local M = {} M.ref = { instantiate = function(self, target, opts) local root = Instantiable.root(self, target, opts) -- IN: the base opts -- ... compose whatever this type is, under `root`, NOW ... return Instantiable.result(self, root, idMap) -- OUT: the contract end, } return M ``` **IN.** `target` is an owning entity ref: the instance lands under (or, for a hierarchy type, onto) that owner. With no target the type spawns a fresh root. `position`, `rotation`, `scale`, `name` and `temporary` mean the same for every type — `Instantiable.root` applies them to a root you mint, `Instantiable.place` to one you adopt, so you never re-read the spec. Honour more opts of your own if your type needs them, and document them in the hook's `--!arg` docs. **OUT.** Return `(root, idMap)` through `Instantiable.result`: - **`root`** — the composed root, as an `EntityRef`. Composition is **synchronous**: everything your type builds is live when you return, so the caller can parent to it and read its components in the same statement. Do not defer the work to a component's `awake` and return an empty shell — a root that is not live is refused. - **`idMap`** — the `originalId -> runtimeId` map naming what you spawned, or nil for a type with no addressable children (`result` normalises it to `{}`). **A caller never receives nil.** A component that re-composes your asset on every load keeps this map and passes it back in, which is how a cross-entity reference into the composition survives a reload. `AssetRef` runs every `instantiate` through `result` on the way out whether or not your type called it, so returning something else fails at your own call rather than handing a caller a nil root. Calling it yourself is still how you say what you return, and running it twice changes nothing. `@builtin::assetTypes.assetType.shared.instantiable`'s README is the full reference for both halves. ## The inspect hook (`M.ref.inspect`) `asset.inspect(ref)` resolves `ref`, builds a common envelope shared by every asset (`identity`, `name`, `guid`, `source`, `typeName`, `typeDefinitionPath`, `scope`, `origin`, `description`, `tags`), then calls the resolved type's own inspect hook — `M.ref.inspect(self)` on `behavior.luau` — for the type-specific `detail`. The hook returns *only* `detail`; the framework fills in everything else: ```lua -- behavior.luau local M = {} M.ref = { inspect = function(self) -- self.path is the asset's own folder. Parse the asset's OWN -- files — source text, a native payload header, its `.metadata` -- — never runtime state (a live component instance, a GPU -- upload, a compiled schema): inspect must work on an asset that -- hasn't been compiled, uploaded, or registered yet. return { -- type-specific fields, whatever this type wants to surface } end, } return M ``` A type that omits `M.ref.inspect` yields the generic record: the envelope alone, `detail = nil`. That's a valid, working state — not an error — but it means agents calling `asset.inspect` / `tools.use("assets","describe", ...)` on an instance of the type see only identity and scope, none of what makes an instance of it meaningful. Declaring the hook is what makes a type's instances inspectable for what they actually are. A hook that raises leaves `detail = nil` and sets `record.warning` to the error text — `asset.inspect` itself never raises for a resolved ref. ## Discovery - `asset.list("assetType")` — every registered asset type. - `asset.inspect("<typename>")` — the type's identity, source path, and this README, plus its `about.yaml` texts as `detail.about`. - `asset.resolve("<typename>", "assetType")` — the `<typename>.assetType` asset itself (used by the validator to find the type's `type.yaml`). ## Related - `assetTypes/README.md` — the full `type.yaml` schema and validator rules. - `@builtin::assetTypes.assetType.shared.instantiable` — the instantiation contract's shared implementation, both halves. - Every other `*.assetType/` here — the concrete types built on this meta-type. ## The Inspector a type ships Every type names the file its Inspector sections come from with `inspector: { type: "inspector.luau" }` in its type.yaml. `asset.create("assetType", name)` scaffolds the block and an `inspector.luau` whose Contents section lists an asset's files, built from `ctx.parts`. `asset.validate` on a type that names none reports `assetType.inspector_required` (an error for a `@builtin` type the engine's library authors; a warning for a world's own, and for one the library carries by reference, which the world that published it answers for), and one whose file does not load `assetType.inspector_invalid`. ## In the Inspector Type: what the type.yaml declares (suffix, container, primary files, the settings and inspector blocks), whether the type ships a behavior, how many instances it has, and **Open** for its files. Schema: the settings every instance carries, as read-only typed controls at their defaults (a type whose schema is per instance says where it is declared). Instances: a link to each of the first 20 assets of the type, and how many there are when there are more.

# proc.builder What a graph builder hands the author, in the call form: `proc.define(function(g) ... end)`. - `g.<name>(args)` adds a node of the node asset the name resolves to, by asset identity: the nodebox and the node's folder name (`g.points.disc`), a path (`g["/zero/source/fx/ring.procNode"]`), `~` (the nearest package, effect or world root: `g["~/ring"]`), a guid, or a short name (`g.disc`). A name that matches two nodes, and a name that matches none, are errors that name the candidates. Each call is recorded in the graph by the node's guid. Arguments are a table naming inputs or positional values in the node's input order. - What a call answers: a node with one output answers that output's socket, a node with several answers a record of sockets by output name (`local f = g.core.normals.face(geo); f.geometry, f.area`), a node with none answers nothing. The record does not depend on declaration order. - `g.op.<name>(args)` (and `proc.op.<name>`) adds a node of an op, for library nodes built on the ops. - `g:input(name, default, opts)` is typed by the default: `Float`, `String`, `Bool`, `Vec3`, `Color`, `List<T>`, `Record` with its fields. `opts` takes `min`, `max`, `options`, `description`, `optional` and `type` for a default that does not state its type or no default at all. An empty table default with a stated `type` is that type's empty value (`g:input("entities", {}, { type = "List<Entity>" })`). - An input with `options`, or of type Bool or String, is a plain value while the graph is built: `g:input` answers the value itself, `if` over it is allowed, and it passes into calls as a constant. Number and vector inputs are sockets. The graph records the plain values it was built with in `graph.plain`, and each such declaration in `graph.inputs[name].plain = true`. - Sockets take `+ - * / % ^ // ..` and unary minus, `math.*` and `vector.*` (each a `math.<name>` node; `vector.create` is `math.vec2`, `math.vec3` or `math.vec4` by its component count), `x.name`, `x.name = v` and `x[name] = v` for a computed name (which rebind `x`), `x[class][name]`, `x[n]`, and `for index, item in list do ... end`. A source loaded through `builder.prepareSource` reads plain `math`, `vector` and `rand` over sockets unchanged; `define` loads a node or graph entry that way when it names them. - `g.compare(a, op, b)`, `g.where(cond, a, b)`, `g.pick(choice, alternatives)`, `g:fold(n, init, fn, opts)`, `g:state(name, columns)`, `g:step(dt, substeps, fn)`, `g:check(cond, message)`, `g.has(geometry, class, name)` and `g.copy(socket)` record the constructs the evaluator knows. - `g:check(cond, message)`: a `cond` known while the graph is built refuses the build with the message (a true one records nothing); a `cond` that reads a socket is recorded for the plan or the cook to answer. The build of a node's own defaults, which is what indexes the node, holds such a refusal instead of raising it (`graph.refusals`), and `forInputs` states it where the node is called or cooked as it stands, so a check over a plain input fires when the call gives it the refused value. - A plain Luau table filled inside a `for` body over sockets keeps one element (`rings[#rings + 1] = v`, `table.insert(rings, v)`): the body is one pass for every item. `builder.check` refuses it with the code `socket-loop-append`, naming the form to write: a local declared before the loop and assigned inside it, read after the loop, where it is the list of every pass's value. - `g.copy(value)` of a record or a list written in the graph (`g.copy({ total = 0 })`) is a value of its own that member writes rebind, so a `for` body that fills it carries it from pass to pass. - A class view made with a socket key (`names[i]`) and read as a value is a node of the scope it was made in, whichever loop body first reads it. - A `for` body is one pass that stands for every item. `index` is the item's place (a socket, from 1). A value built outside the body that the body reads or assigns (`geometry[class][name] = v`, `record.name = v`, `record[name] = v`) is carried: each pass starts from what the pass before left, the first from the value as it stood before the loop, and after the loop the value is what the last pass left. The loop records the carries in `scope.carries[id] = { init, result }` and the refs `loop_index`, `loop_carry` (the value at the start of a pass) and `loop_result` (the value after the loop). A value the body only reads reads as it was. The emitter writes a carrying loop as it was authored: a local that starts as `g:copy(value)`, the body's assignments to it, and the loop reading that local after it (`proc.emit`). - `g.clock.placement` (16 numbers, column after column), `g.clock.inverse`, `g.clock.rotation` (`{ x, y, z, w }`) and `g.clock.entity` answer where the graph stands in the scene it is cooked in: each is a `construct.scene` node that reads the scene the cook is handed (a graph standing on no entity has the identity placement and no entity). `g.transform.point(p)`, `g.transform.direction(d)` and `g.transform.rotation(q)` carry a world position, direction or quaternion into the graph's space through the inverse placement; each takes plain values, sockets or Fields. - `g:fold(n, init, fn, { stop = "changed" })` or `{ stop = function(before, after) return <Bool> end }`; `fn` returns the state only. The fold runs `n` passes and ends early at the first pass whose stop holds. - `g.element` is the element a Field evaluates over: `g.element.id` (Int), `g.element.P` (the position leaf that `fieldc.withPosition` substitutes, also written `g.element.point.P`) and `g.element.<class>.<column>` for a column of the domain the field stands in. Each read is a socket, so `g.element.id * 1.0` is a math node. - `rand.unit(seed, ...scope, id)` (a global in a node source, also `builder.rand.unit`) is the Float field of unit draws keyed by the seed, the scope (`graphId, nodeId, channel`, or `graphId, compositePath, nodeId, channel`) and the element `id`, last (`g.element.id`). It adds an `io.luau` node `{ module = "@builtin::systems.procgen.proc.fieldc.leaves", fn = "rand", args = { seed, { ...scope }, id }, returns = { "Any" } }` and answers its `value`. - `geometry.names[class]` lists the attribute names a class holds (a list socket the plan answers); `geometry[class].count = n` assigns the domain count; `geometry.material = m` assigns the material. `names` is read by that name on any socket: what is read off it by a class (`geometry.names.point`) lists the names, and any other read (`record.names[2]`, `record.names`) is the member a record holds under that name, as `count` and `keys` are. `material` is assigned by that name on any socket. - A member, an item, a column or the names of a value built outside a fold, step or stop body that the body reads first is a node of the scope the value was built in: the body and every later fold read the one value, which the plan writes once. A read keyed by a socket of the body (a fold's index) stays in the body. - `g.clock.time`, `.frame`, `.delta` and `.duration` are sockets; `.fps`, `.frames`, `.loop` and `.realTime` are the timeline's values. - Luau's `< <= > >= == ~=`, `and`, `or`, `not` and `if` over a socket have no value while the graph is built. `builder.check(source)` names each one with the `g.*` form to write, and `define` holds a node or graph entry to it. Plain inputs are not sockets to the check. A plain Luau table indexed by a socket (`LANES[values.type]`) answers nil while the graph is built, so the check refuses it with the code `socket-table-index`, which names `g.pick`. - A list of values holds a nil as a value: `{ a, nil, b }` is three, and the second is a constant of nothing. ## Building a graph for other plain values A graph whose plain inputs change structure is built again for the values a cook was given. - `builder.forInputs(graph, inputs)` answers `graph` while its plain inputs hold what the build used, else the graph built again from `graph.source` with the values `inputs` states (an input it leaves out takes its default); the result is cached by the source text and the values, so a node written again is built again. It answers `nil, reason` when the source cannot be built again or a value is outside the input's options. - An optional input with no default (`g:input(name, nil, { optional = true })`) is nil while the graph is built for a call that leaves it unwired, and its socket otherwise, so `if input == nil then ... end` in a node source is decided by what the call wired. The declaration states `wiring = true` and the graph records the wiring it was built for in `graph.wiring[name]`; `forInputs` builds the graph again for a cook given no value for such an input (`values[name] == false` states none in `rebuild` and `withInputs`). - `builder.rebuild(graph, values)` builds again unconditionally; `builder.withInputs(values, fn)` runs a function that calls `proc.define` and gives that graph the values. A node asset called with a plain value records it as a constant param; the evaluator reads a node's graph through `forInputs` with those params. ## What a graph records A call-form graph states `surface = "call"`. Each node it makes carries `call = true`, `seq` (the order it was made in, which the emitter writes it back in) and `scope`. A node asset is `op = { __ref, type = "procNode", name }`; an op is its id string; a construct is `op = "construct.<kind>"` with `construct = <kind>`. `graph.scopes` holds each `for`, `fold`, `stop` and `step` body, `graph.states` each `g:state`, `graph.plain` the plain input values. The `inputs` of a construct node are refs (a constant is `{ kind = "const", value }`); `params` are constants. | construct | inputs | params | answers | |---|---|---|---| | `member` | `value`, `name` | | the member of `value` | | `member_set` | `value`, `name`, `to` | | `value` with the member written | | `item` | `list`, `index` | | the item | | `column` | `geometry`, `class`, `name` | | the column read (a read of `count` is the domain count) | | `column_set` | `geometry`, `class`, `name`, `value` | | the geometry with the column written | | `column_remove` | `geometry`, `class`, `name` | | the geometry without the column | | `count_set` | `geometry`, `class`, `count` | | the geometry with that class's domain count set to `count` | | `material_set` | `geometry`, `material` | | the geometry with its material set | | `names` | `geometry`, `class` | | the List of attribute names the class holds, answered at plan | | `has` | `geometry`, `class`, `name` | | a Bool answered at plan: whether the class holds the column | | `check` | `cond` | `message` | nothing; a `cond` that is false when cooked is an error stating `message` | | `compare` | `a`, `b` | `op` (`eq neq lt le gt ge`) | a Bool | | `where` | `cond`, `a`, `b` | | `a` where `cond` holds, else `b` | | `pick` | `choice`, `alternatives` | | the alternative the choice names | | `step_sample` | `step`, `value` | | the value of an animated `value` at the step's time | | `step_between` | `step`, `index` | | the fraction (0 to 1) of the step's movement that source `index` of the step has covered; the evaluator defines it | A call that has several outputs names them in refs: `{ kind = "node_output", node, output = "<name>" }`; a node with one output has `output = nil`. ### Refs The v1 refs (`const`, `graph_input`, `node_output`, `list`, `record`, `animated`) and: | ref | stands for | |---|---| | `loop_item { loop }`, `loop_collect { loop, ref }` | the item of a `for` scope, and the list of `ref` over every item | | `fold_value { scope }`, `fold_index { scope }`, `fold_result { scope }` | the value and the index (from 1) of a fold pass, and the value after the last pass | | `fold_before { scope }`, `fold_after { scope }` | the state before and after a pass, inside a fold's `stop` function; `scope` is the `stop` scope | | `step_state { scope }`, `step_index { scope }`, `step_result { scope }` | the state record, the step record (`dt time graphTime index first start`, with `step_sample` and `step_between`), and the state the step reached | | `state { name }`, `clock { field }`, `engine` | declared state, the clock's field, the engine API namespace | | `element { field, class? }` | the element a Field evaluates over: `field = "id"` its index, `field = "P"` its position leaf (no class), else the column `field` of the domain `class` | ### Scopes `graph.scopes[id]` is `{ kind, parent, seq, ... }`: - `for`: `list`, `item`. - `fold`: `count`, `init`, `result`, and `stop` when it states one: `"changed"` (the pass ends the fold when the state it produced equals the one it read) or `{ scope = <id of a stop scope> }`. - `stop`: a scope whose `parent` is its fold. Its nodes read `fold_before` and `fold_after`; `result` is the Bool ref the fold stops on. It is evaluated after each pass. - `step`: `dt`, `substeps`, `result`. The `plain` table of a graph and the `plain = true` flag of an input declaration state which inputs are build-time values. Part of `@builtin::systems.procgen.proc`.

# proc.demos Procedural-generation demo recipes — graph builders that showcase the proc system end to end. Part of `@builtin::systems.procgen.proc`. - **`cashPile(proc, opts?)`** — scatters a billboard source over a disc of points and merges it into a single mesh (the scatter → merge showcase).

# proc.eval Lazy topological evaluator of a graph target, with per-eval memoization and cycle detection. Part of `@builtin::systems.procgen.proc`. ## Graphs built in the call form A graph built with `proc.define(function(g) ... end)` has `surface = "call"`. It records nodes (ops and node assets), scopes (`for`, `fold`, `step`), declared state and the refs between them. It has no output the evaluator is asked for: every node nothing reads is cooked from. ```lua local play = eval.play(graph, { inputs = { count = 8 } }) local report = play:present({ time = 0, frame = 0, fps = 30, duration = 0 }) play:release() ``` `eval.play(graph, opts)` plans the graph, cooks from every node nothing reads and keeps what those nodes return. `play:present(clock)` cooks at a clock (`{ time, frame, fps, duration }`; a `duration` above 0 is a timeline with an end, where a seek replays from the start), `play:setInputs(inputs)` gives the graph other inputs, `play:handle(id)` answers what a node cooked from answered last, `play:report()` says what happened and `play:release()` lets go of everything the play holds, the Handles its output nodes answered included. Every call-form graph cooks through the plan, whoever asks: `proc.eval(graph, target, opts)` of a graph with `surface = "call"` plans it (`opts.inputs`, `opts.clock`), cooks what the plan waits on and then the target through the flat graph, and a node asset used as an op cooks its graph the same way. `output:<name>` names an output of the graph (`flat.outputs` holds each one lowered), and the nodes the eval cooked into its own cache are released apart from what the answer reaches. `eval.plan(graph, opts)` is the plan alone, and `eval.references(graph)` the assets and ops the graph's calls reach, by guid and id, without cooking. ### The document ``` graph.surface "call" graph.nodes[id] { op, inputs = { socket = ref }, params = { name = ref | value }, scope = scopeId? } graph.scopes { [id] = { kind = "for", list = ref } | { kind = "fold", count, init, result } | { kind = "step", dt, substeps, result }, parent = scopeId? } graph.states { [name] = { columns = ref } } graph.inputs { [name] = { type, default, ... } } graph.outputs { [name] = { ref } } -- what a graph used as a node gives ``` `op` is an op id (`"device.kernel"`, or `"op.device.kernel"`), a node asset (`{ __ref = guid, type = "procNode", name }`) or `"construct.<kind>"`. The refs are the v1 refs (`const`, `graph_input`, `node_output`, `list`, `record`, `animated`) and: | ref | stands for | |---|---| | `loop_item { loop }` | the item of the `for` scope in the copy being written out | | `loop_collect { loop, ref }` | the list of `ref` in every copy of the `for` scope | | `loop_index { loop }` | the place (from 1) of the item in the copy being written out | | `loop_carry { loop, carry }` | a value built outside the `for` scope as a pass begins: the value before the loop for the first copy, what the copy before left for every other | | `loop_result { loop, carry }` | the same value after the last copy | | `fold_value { scope }`, `fold_index { scope }` | the value and the index (from 1) of the pass of a fold | | `fold_result { scope }` | the value after the last pass | | `step_state { scope }`, `step_index { scope }` | the state record and the `step` record of a step body | | `step_result { scope }`, `state { name }` | the state the step has reached | | `clock { field }` | `time`, `frame`, `fps`, `duration` or `delta` of the graph's clock | | `engine` | the engine API namespace | ### The plan - A node asset built in the call form is written out where it is called, its inputs bound to what the call was given. A call nothing reads is cooked from through every output it declares. - A `for` body is written out once per item of its list, a `fold` body n times in a chain, a `pick` takes the one alternative its choice names. A copy is keyed by its content, so reordering, inserting or removing items cooks only what is new. - A count or a list a scope needs is a plan value: a constant, an input, or the answer of a node, which the plan cooks and lists in `report.waits`. A column left on the device is not a plan value; reading it back with `transfer.readback` is, and the wait names the readback. The length a scope was written out for is watched: a cook that answers another length plans the graph again. - A `where` whose condition the plan knows takes one alternative. What only the other alternative stands on is not written out, and a call nothing reads that stands on any of it is not written out either (the nodes a node asset leaves unread are written after every alternative of the plan is decided), so nothing reachable only through the untaken alternative is cooked while the plan is made. A node of the graph itself that nothing reads is always cooked. - A `for` scope's `carries` (`{ [id] = { init, result } }`) chain the copies: copy 1 starts a carried value at `init` and copy n+1 at what copy n's `result` lowered to; a `loop_result` read after the loop is what the last copy left. - A step becomes one solver node. Its state is one record of device Columns per play (`g:state(name, { column = { type, count } })` allocates them zeroed). The columns a state allocates are stated through plan values: a literal record, or a record a fold fills or constructs read off a list or an input (per-source columns, counts from `list.count`, names from `geometry.names`), whose counts the plan cooks and lists in `report.waits`. A `member` or `item` read inside the body of a value the plan knows (a count or a list a fold inside the step is written out for) is answered while the plan is made. The step record carries `dt`, `time`, `graphTime`, `index` and `first`. The body runs `substeps` times in each fixed step of `dt` seconds. Pausing holds the state, a seek back replays from the start on a timeline with an end, and a changed value the step reads begins it again. ### A plan is made once A graph is planned once for the values its plan stands on, and every later cook of it reads that plan. - A plan stands on the registry (`registry.revision()`), on the graph its plain inputs build, on the values of the inputs it read (as keys), on the animated values it read at the clock, on the answers it waited on and the nodes it cooked for a count or a list, and on the modules it required (by what their files hold). A cook reads a kept plan after asking each of those again (`expand.stillStands`); a plan that no longer holds is made again, and a plan that does is read. - Each answer and node the plan waited on is held by the key of its cone, without a cook. The leaves a cone stands on are named once, when the plan is kept (`expand.keyWatches`): the graph inputs it reads, the animated values, the constants that hold a device value (a Geometry, a Column) or an asset reference, and the nodes that read live state; the ops, their versions, the plain constants and the wiring are fixed by the plan and the registry. A read asks the evaluator for the key of each leaf as it stands now (`opts._leafKeys`: a device value is keyed by the run it stands on and the writes the run has had, an asset reference by what the asset's files hold) and compares them with the keys the watch stood at. A watch whose leaves are unmoved holds and cooks nothing; a watch whose leaf moved is cooked again, only its cone, and holds when it answers what it answered, and the keys it stands at are renewed. A plan that cooked a node reading the clock, a play's input or a live run, or whose cone stands on a solver or an op that always cooks, is not keyed: each of its watches is cooked again on every read. - A kept plan also holds the answer of each target it was cooked for (up to four a plan, and sixteen plans at a time), filed by the keys of the target's cone: a cook of `proc.eval` that shares no node cache with its caller, finds the cone unmoved and passes no `stats` or `cookRecord` is answered with a copy of the held answer (a Geometry or a Column is cloned, a list or a record of them rebuilt), and cooks nothing. An answer holding anything else (a Handle, a Field) is not held. The held answers are let go of with the plan. - A plan is kept unless it registered a step solver or wrote a run of the device into itself (a `for` over geometries). A plan that stands on a run of the device holds by the key of its cone, and is made again when the key moves. `eval.planCounts()` answers how many plans were made and read, the milliseconds each took, `notKept`, the plans that could not be kept by what stood in the way, `served`, the cooks answered from a held answer, and `notKeyed`, the kept plans whose watches are cooked again on every read, by why. - `engine` is the engine namespace, a constant. A member read or an item read out of constants, out of a record or a list a call wrote, is the ref written there and costs no node; `engine.require` of a module named by a constant is the module, required while the plan is made and asked for again each time the plan is read. - The plan is made in slices of `budgetMs` (5 ms) when the caller can yield; a cook over the plan hands a frame back after `cookBudgetMs` (100 ms). - A cook that shares no node cache with its caller keeps none: nothing is keyed or filed, what the plan cooked while it was made (or asked again) and what the cook makes stand in one table by node, and all of it is released as the cook ends but what the answer reaches. A Field that is the answer keeps the runs its leaves read; they stand as long as it does. - A node of a plan that stands is held to the contexts its producers operate in once, takes the label of its refusals from one string, and takes its params prepared once when every one is a constant. ### A cook after a change is run from a list (`cookrun`) A kept plan whose cone stands on numbers (graph inputs holding a number or a Bool) is cooked node by node while those numbers move. The recording wraps the plan's cone re-check and the cook together: the first such cook after the plan was made is recorded as a compute list and the next one is recorded beside it; the words that differ between the two recordings (a dispatch param, a word of a write) must each be the value of one leaf in both, and become the slots of the second list. A cook after that runs the list with the leaves' values in its slots, asks the recorded plan-time read-backs again, runs the watches that are not nodes (modules, animated values, the registry stamp) and answers a device copy of the list's columns (`cookrun.counts()`: `recorded`, `diffed`, `replayed`, `verified`, `mismatched`, `readbackChecks`, `dispatches`, `refused` by why): no node is cooked and no buffer is allocated but the copy. What makes a recording stand, and what ends it: - Two recordings are one cook when they make the same dispatches over the same programs and workgroups, bind buffers one for one, and write the same number of words; a run placed at another offset (`zero_base_*`, a write offset) is a placement. A first cook that uploads a mirror of a CPU column makes the pair differ once, so the comparison is tried again on the next cook (twice at most). A moved leaf that is no word of the list is refused by name, since its value reaches the device by a write the recording did not hear. - A cook that read the clock, a play's input or a live run, and an answer that cannot be copied are not recorded. - A `cpu.call` is first handed the geometries it is given with their columns on the device standing in as columns that answer their type, width, length and run (`substrate.structureOf`): a call that asks for no numbers (the names of the columns a class holds, the types a join decides) moves nothing off the device and records no read-back. A call that asks a stand-in for its numbers, fails, answers a stand-in, declares a return that is no plain value, or is handed a geometry counted on the device or a part is made again over the drained inputs, and its read-back is recorded. - Every node whose cook read back from the device (`transfer.readback`, a drained `cpu.call`) is recorded as a read-back row: its inputs, whether they were drained, and a digest of its plain outputs. After the list has run each row is asked again; one that answers another value than it did ends the recording and the cook is made node by node, answering correctly. A read-back whose outputs are not plain data is not recorded. - The first replay of a list is held to the cook of the same values made node by node, column for column; a difference ends the recording (`mismatched`). - A changed key of any other leaf (a Geometry input, a constant device value, an animated value), a plan made again and a registry change let go of the recordings; at most six plans hold recordings. - The recordings of two threads do not meet: a dispatch and a write belong to the recording of the thread that made it. - An answer a recording makes is copied word for word onto runs of their own (`cookrun.deviceCopy`) for the caller and for the held answer, so the buffers the list writes in place are never shared. `cookrun.setEnabled(false)` cooks every cook node by node. A `g:fold` of three passes or more over device state is planned as one `construct.fold_run` node (`foldrun`): the body is built once as a flat graph with the inputs `state` and `index`, the first pass is cooked while a compute list records it, and every later pass queues that list over a copy of the state, so a pass after the first cooks no node. `foldrun.counts()` answers how many passes were recorded, replayed and cooked, and how often a body was refused a list, by why. A body cooks each pass on its own when it reads `index`, when the state holds something the list cannot follow (a record, a text, a readback), when a pass makes a state of another shape, or when the fold sits inside another recording. `foldrun.setEnabled(false)` writes the folds of the plans made after out once per pass. ### How a step runs A step solver plans its body once (`proc.eval.stepplan`), when the state takes its first step. - A node of the body that depends on neither the state, the `step` record, an animated value nor the scene is STATIC: it is cooked once, with the first step, and read as it stands by every later step (`opts.fixed` of the evaluator). The plan reads only the nodes the next state stands on. - What is left is the device work of a step. The first step of each distinct set of `step` values the body reads (its `dt`, `first`, `progress`, ..., and the animated values it reads, as the device sees them: the f32 they round to) is cooked with every dispatch and every write it makes into a device buffer kept as one compute list over buffers the solver holds, and the next state written back into the state's own columns by the last dispatches of the list. Every later step with those values queues the list: the dispatches read the state's columns where they stand and write the next state into them. A step that runs from a list cooks no node and allocates no buffer. - The columns of the state are written in place, so what a step presents is a copy of them made on the device. - Up to six sets of `step` values are kept. A body the plan cannot stand on is cooked a step at a time with its static nodes still read as they stand, and says why in `report.steps[<step id>].reason`: a node of the body reads the clock, a play's input, a live run or the scene; the body reads values back from the device to the CPU; the next state is not a record of Columns the state's own shape; or a `step` value the body reads takes a new value in more than six steps (a body that adds `step.time` to a column, say). - `play:report().steps[<step id>]` holds `steps`, `replays`, `recorded`, `interpreted` (steps cooked node by node), `plannedDispatches` (what one list queues), `dispatches` (queued in all), `allocations` and `uploads` (buffers made and uploads done by steps), `ops` (nodes cooked in steps, by op), `staticNodes`, `variants`, `mode` (`replay` or `cook`), `reason` and `keyValues` (the `step` values of the last step). - `stepplan.setPlanning(false)` makes the solvers begun after cook every step node by node, which is how a planned body is held to the numbers a cooked one gives. A state column the next state names as a Field, a number or a vector is evaluated into a column of the state's own length and type. ### Constructs A construct is a node the evaluator itself knows: `construct.member`, `construct.item`, `construct.member_set`, `construct.column`, `construct.column_set`, `construct.column_remove` and `construct.engine` and `construct.select` are cooked; `construct.pick`, `construct.compare` and `construct.where` are written out in the plan (a compare of two constants is their answer, a `compare` otherwise calls the `math.compare` node, and a `where` over a condition the plan does not know is a `construct.select` node). `construct.select` chooses `a` where `cond` holds and `b` where it does not: numbers, vectors and Fields element by element (`math.where`), and a Geometry, a record or a text by a condition that is one value, which a condition over the elements of a column is not (it is refused naming `g.pick`). A Column or a Geometry a construct reads from a record or a list is a clone of it. `construct.member` reads `count` and `keys` of a plain list or record (the number of members, the sorted names of a record or the places of a list; a value that is nothing has none), `count`, `type` and `lanes` of a Field (the length of the longest column its chain reads, its element type, its components; a Field that reads no column has no length, and the read refuses naming `geometry.point.count`, `column.count` and `g.field.index({ domain })`), and `count`, `type` and `lanes` of a Column. `construct.member` also reads `x`, `y`, `z` and `w` of a vector (a Field of that component for a Field, a Column of that component for a Column of Vec2, Vec3, Vec4 or Quat), the classes of a Geometry, its `material` and its `names` (the attribute names each class holds). On a class of a Geometry a read of a `detail` column answers the value the detail holds. `count` of a Column or of a class of a Geometry is the live count when the elements are an appended or compacted output (a Column of one word, which a kernel with a `domain_count` binding reads the live count through, and a kernel without one runs over the capacity); `capacity` of either is the number of elements the run is allocated for. A record a graph holds under a name a class goes by (`point`, `vertex`, ...) is read and written as a record: `x.point.name` is the member of the member. `construct.column`, `construct.column_set` and `construct.column_remove` take `edge` beside `point`, `vertex`, `primitive` and `detail`: an edge is a unique point pair of the primitives, so the class counts them and a column written on it is one value per edge. `count_set` takes the four stored classes. `construct.column_set` writes a Column, a Field evaluated over the class it is written into, a vector or a number, a record or a text (one value for every element, a constant Record or String column) or a list of records or texts (one to an element, held in a Record or String column on the CPU); a record whose fields are numbers is a record, and a list of numbers is a vector. A column that is the structure of a geometry (`point.id`, `vertex.pt`, `primitive.vstart`, `vcount`, `curve` and `closed`) holds whole numbers: a number, a Field or a Column written to it is stored as an Int, and so is a value written to a column that already holds Int. A detail value is held as itself, a record among them, and `construct.column_remove` removes it. The inputs of every construct take nothing as a value: an optional input a call left out reaches them as nothing. `construct.scene { what }` answers where the graph stands in the scene it is cooked in (`entity`, `placement`, `inverse`, `rotation`); it is read as `g.clock.placement` and the other names, and as `g.transform`. ### What the plan answers A construct that answers while the plan is made cooks the nodes it reads (a wait in `report.waits`, a readback named as one) and records the answer, which is what the play compares to decide whether the plan still stands. - `check { cond } message`: a condition that is false refuses the play with the message. - `has { geometry, class, name }` is the Bool of whether the class holds the column; `names { geometry, class }` is the list of attribute names the class holds. A `for` over the names has one copy per name. - `fold` with `stop`: after each pass but the last the stop is answered, and the fold ends at the first pass whose stop holds. `"changed"` holds when the pass left its state as it found it (numbers by value, a Column or a Geometry by what its runs hold: a device run read for the comparison is a readback wait). A stop function's scope reads `fold_before` and `fold_after` and its result is the Bool. The report's fold scope states `copies`, `stop` (`"changed"` or `"function"`) and `stopped`; each answer is a wait named "the stop of fold '<id>' after pass <n>". - A node asset called with plain inputs (an input with options, or of type Bool or String) is the graph its source builds for those values (`builder.forInputs`); a play built for other plain inputs plans again. Constructs cooked as nodes: - `count_set { geometry, class, count }`: the geometry with the class's element count set. A number sets it; a live count Column (an appended output's) becomes the live count over the capacity it carries; arithmetic over a live count (`kept.indices_count * 6`, a Field over live counts) is evaluated on the device as one word, and the class is as large as that arithmetic reaches when every count it reads is full (`substrate.atCapacity`). - `material_set { geometry, material }`: the geometry carrying the material (a Material value, an asset reference or a name; nothing for none). - `step_between { step, index }`: how far from 0 to 1 the step has come through the span of graph time the present covers; `step_sample { step, value }` is the value, which a step body samples at the step's time. The step record carries `dt`, `time`, `graphTime`, `index`, `first`, `start` (the time the step began at) and `progress`. An `element` ref is the Field the graph's `g.element` stands for: `id` and `P` are fields the field compiler fuses into the chain they sit in; a column of the evaluation domain is `fieldc.element(field, class)`. ### Cooking Every play cooks into one store of the world, filed by content key (the op, its version, the seed, the params and the keys of the inputs): a node of any graph that cooks the same op over the same inputs is cooked once. A value that holds an asset reference is keyed by what the asset's files hold, so editing a module a node calls cooks the node again. The nodes a play pulls from and every step are the play's own. An output node answers a Handle (a table with a `release` function). `cpu.call` owns its list of results: a Handle in it is released when the cook is superseded and when the play ends.

# fieldc A Field is a per-element expression. `fieldc` holds what a Field is made of and how it runs: a chain of templates over a position leaf, an element number, column reads and numbers is written as one WGSL function and runs as ONE dispatch of `kernels/map`, wherever it is consumed. Nothing dispatches per operator. ## Where a Field is evaluated ```luau local column = fieldc.evaluate(field, domain, opts) ``` The one entry point a consumer of a Field calls (a kernel's input, an attribute write, an output): `device.kernel` for a Field bound to a column input, `attribute.write` for a Field value, and every other node that needs a Column. - `field`: a Field, a Column, a `BufferRef` (a device column handed through `cpu.call`), a number or a vector record. - `domain`: a number of elements, or a Geometry (its `opts.class` elements, points by default), whose attributes the leaves reading the evaluation domain bind. - `opts`: `{ class, count, type, clock, where, label }`. `type = "Int"` writes i32 words (a Field tagged Int does so by default). `clock` is `{ time, frame, fps, duration }`: animated values are first sampled there. `where` names the caller in a refusal. - Answers a Column of the domain's count. Raises, naming what refused, when the chain cannot be written as a function. A chain holding an animated value or a live run answers a live column. `fieldc.liveRuns` finds it and `fieldc.tick(runs, clock)` rewrites it each frame: the animated values are sampled at the clock and written into the numbers the dispatch reads, and the dispatch runs again into the same run. An animated input therefore changes the result every frame without a cook. ## Fields the nodes make | Construct | Module | Answers | |---|---|---| | `fieldc.field(v)` | `fieldc` | any socket value as a Field: Field, Column, BufferRef, number, vector, animated value | | `operators.apply(name, ...)` | `fieldc.operators` | `math.*`: the plain result over numbers and vectors, a Field over a run or a Field | | `fieldc.element(name, class?, width?)` | `fieldc` | what `g.element` stands for: `"id"`, `"P"`, or a column of the domain (`g.element.point.N` is `element("N", "point")`) | | `leaves.elementId()` | `fieldc.leaves` | the element number, an Int Field (`g.element.id`) | | `fieldc.position` | `fieldc` | the position leaf `p` (`g.element.P`) | | `leaves.context(class, name, width?)` | `fieldc.leaves` | an attribute of the evaluation domain | | `leaves.rand(seed, scope, id)` | `fieldc.leaves` | the unit draw, equal to `hash.unitFloat` to the bit | | `leaves.call(fn, args)` | `fieldc.leaves` | a function of a prelude shader module of `kernels/` | | `leaves.gather(column, index)` | `fieldc.leaves` | a run read at computed elements | | `leaves.substitute(field, position)` | `fieldc.leaves` | the chain read at another point (`fieldc.withPosition`) | | `leaves.index(domain)` | `fieldc.leaves` | an Int column of 0 to n - 1 | A Field made by these answers `f.x`, `f.y`, `f.z`, `f.w` as the Field of that component. Components are Fields, so the Field of a vector is built from them with `operators.vec2`, `operators.vec3` or `operators.vec4`; Luau's `vector.create` takes numbers. ## How the leaves reach the kernel A chain's leaves are columns. `kernels/map` states every leaf's shape in its `cols` table and reads up to as many leaves as one dispatch has storage buffers for (`compute_abi.directLimit`: 8 per stage on every target, less the table, the numbers and the output, so 5 for `map`) straight from the buffer of the leaf's own run, at no cost beyond the binding. A chain of up to 5 leaf columns is the one dispatch of `map` and nothing else. Leaves past those are copied into the table, one copy dispatch each. `fieldc.bindSpec(spec, { direct })` makes the table; `direct` is the list of leaf slots read from their own runs (`compute_abi.directPlan(columns, limit)`), and the variant is made with the same list: `program:variant({ specialize = ..., direct = { cols = direct } })`. Without `direct` every leaf is copied into the table. `fieldc.operators` folds numbers: every operand a number or vector gives a number or vector. A narrower operand widens to the widest (a number applies to every component, a shorter vector reads 0 where it has none). Division and remainder by 0, the square root and the logarithm of a value at or below 0 answer 0. ## The walk step contract (for `kernels/walk`) A walk kernel takes its per-walker function as a Field spec, the way `kernels/map` takes `TEMPLATE`. ```luau local spec = fieldc.walkSpec(step, { width = 2, count = walkers, done = stop }) -- a state of 8 floats: lanes 0 to 3 in `step`, lanes 4 to 7 in `high` local wide = fieldc.walkSpec(low, { width = 8, count = walkers, high = high }) local bound = fieldc.bindSpec(spec) -- specialize = spec.specialize; bind cols = bound.cols:buffer(), u = bound.u ``` - `width`: floats per state, 1 to 8. `step` is the next state's lanes 0 to 3, a Field of `min(width, 4)` components; above 4, `opts.high` is the Field of the `width - 4` components of lanes 4 to 7. - The state is read through `fieldc.walkState(count, first?)`: `count` (1 to 4) lanes from lane `first` (default 0), all inside lanes 0 to 3 or all inside lanes 4 to 7. A state of 8 floats is read as two leaves, `walkState(4, 0)` and `walkState(4, 4)`. The nodes `field.walk_state({ count, first })` and `field.walk_index()` are these leaves in a graph. - `step` and `high` are written over the state leaves, `fieldc.walkIndex()` (the step number k, from 0), `leaves.elementId()` (the walker's number), Column reads (one value per walker, or a gather at a computed index) and numbers. - `opts.done`: an optional Field<Float> over the same leaves; the walker stops before step k when it is not 0. - `spec.specialize` is a table of the specialization values the walk kernel takes, ready to hand to `device.kernel` / `program:variant` as its `specialize` value: `WALK_STEP` (`zero_walk_chain_step(s, s_hi, k, id)`), `WALK_STEP_HI` (`zero_walk_chain_step_hi(s, s_hi, k, id)`, or `s_hi` when `width` is 4 or less), `WALK_DONE` (`zero_walk_chain_done(s, s_hi, k, id)`, or `false` when there is no `done`) and `WALK_PRELUDE`, the module-scope WGSL that defines those functions, preceded by the preludes the chain calls. Lanes 0 to `width - 1` of `s` (lanes 0 to 3) and `s_hi` (lanes 4 to 7) are the state. - The functions read its leaves through a `cols` columns binding (`cols_f32(slot, i, 0u)`, `cols_vec3(slot, i)`, `cols_i32` for Int runs) and its numbers through `u_get(k)`: the kernel declares those two bindings exactly as `kernels/map` does. `fieldc.bindSpec(spec)` answers the table (`cols`) and the numbers buffer (`u`) to bind; `release()` lets both go. - A read past a run's end reads its last element. ## Other entry points `fieldc.mapSpec(chain, count, domain, { position })` writes a chain as one function for `kernels/map` and `fieldc.runMap(spec)` runs it. `fieldc.compile` / `fieldc.column` write the same chain as a kernel of its own; they stay for the consumers that bind buffers directly.

# proc.fields Value wrappers for the procedural-generation type system: `Field<T>`, `PointCloud` and `Population`. Part of `@builtin::systems.procgen.proc`. Typed values backed by plain strided Luau arrays. - **`Field<T>`** — a lazy per-element value. `Field.constant(t, values)`, `Field.fn(t, stride, f)`, `Field.buffer(t, stride, buf)`; `:sample(id)` → the `stride` components, `:materialize(count)` → a flat `count*stride` array. - **`PointCloud`** — the value a `.procNode` input declared `PointCloud` receives (`proc.substrate.toPointCloud`); every other input, and every op, takes points as a `Geometry` (`proc.geometry.new`). `PointCloud.new(positions)` (flat stride-3 array); `:count()`, `:position(id)`, `:setAttribute(name, field)` / `:getAttribute(name)` / `:attributeNames()`. - **`Field.transform(offset, orient, scale)`** — a per-element transform whose three parts are fields of their own, so each compiles to the device; `:transformParts()` hands them back. Every value carries a `__proc_type` tag that `proc.types.typeOf` reads (`"Field<Vec3>"`, `"PointCloud"`, `"Population"`).

# proc.geometry CPU geometry container (positions/indices/normals/uvs) + bounds, normal recompute, and renderer.mesh.create args. Part of `@builtin::systems.procgen.proc`.

# proc.graph The graph data model and its definition entry: the nodes a graph holds, the normalized `InputRef` edges between them, its inputs and outputs, and its JSON (de)serialization. `proc.graph.define(function(g) ... end)` builds a graph in the call form, whose authoring surface is `proc.builder`: nodes called by asset identity (`g.points.disc({ ... })`), inputs typed by their defaults (`g:input`), sockets that take Luau's operators, members, columns and `for`, and the declarations a graph states about itself: `g:describe` (what it makes), `g:output` (what it gives), `g:family`, `g:timeline`, `g:still` and `g:cost`. The result is a graph with `surface = "call"` that records each node by the guid of the asset it calls. `proc.graph.refusedInput(graph, name, value)` names why a value is not one an input holds. `proc.graph.fromJSON` reads a graph document back and `toJSON` writes it. Part of `@builtin::systems.procgen.proc`.

# proc.hash Deterministic, portable scoped hashing for the procgen evaluator. `scoped(seed, graphId, compositePath, nodeId, channel, elementId) -> number` folds its arguments with 32-bit FNV-1a (via an exact 16-bit-split `mul32`) and a final avalanche, producing a bit-identical `u32` on native, WASM, and Android. `unitFloat(...)` maps the same scope to a value in `[0, 1)` for deterministic random draws. No `math.random` / `os.time` / float transcendentals — determinism is the contract. The per-element scope (`elementId`) keeps a given element's draw stable as unrelated element counts grow.

# proc.noise Deterministic, portable value/gradient noise. Part of `@builtin::systems.procgen.proc`. Built on the integer-lattice hash in `proc.hash` (`latticeCell2/3` / `unitCell2/3` — the corner hashes of one cell, each bit-identical to the matching `hashInts` call): every lattice corner's random value/gradient is integer-derived, and the only continuous math is the quintic fade + linear interpolation. No float transcendentals and no `math.random`, so the result is bit-identical on native, WASM, and Android — which is what makes a procedural texture reproducible everywhere. Exposes `value2`/`value3`, `gradient2`/`gradient3`, and `fbm`.

# proc.ops Op-family loader: registers every procedural op against the registry. Part of `@builtin::systems.procgen.proc`.

# proc.parity The machinery behind the procgen rewrite's parity harness (`procgenParity`), its sheet tests (`procgenSheet`) and the `procgen.coverage` tool. - `parity.json`: `entries` holds one entry per current op. Each names its new composition (`new`, node ids as `nodebox.name`, or `nodebox.*` for any node of a nodebox), its class (`exact`, `meaning` with an `invariant`, `gone` for a named decision with a `note`, `unmapped` where the sheet names no counterpart), and `tolerance` the comparison reads. An entry that states `decided` (the text of the ruling that accepts the difference between the op and its node) keeps running: the value differences its comparison finds are reported as accepted, under the verdict `decided`, and are held to nothing, while a node that raises and the cost findings stay findings (`accepting`, `decidedLines`). `budgets` states, per op, the read-backs its node may make (`{ "<op>": { "readbacks": n } }`); an op without one may make as many as the current op does. `types` maps a socket's type to the fixture a case builds for it. `cases` holds a case per op that needs more than the defaults and names only content the library ships. A parity suite states the cases whose inputs are its test fixtures in the `cases.json` of its folder (`{ "cases": { "<op>": case } }`), since the library ships its test content nowhere; `cases()` is both together, one case per op, and an engine whose library leaves the suite out finds no case for those ops. - A case states `params`, `fixtures` (a fixture by input socket), `node` (the new node a multi-node entry compares against), `rename` (`inputs`, `params`, `outputs`: the new node's names where they differ, `false` leaves a name out), `compare` (`bounds`: a Geometry is held to its bounds alone, for an output whose element layout differs by design; `ramp`: two ramps are held to the values they read at the same positions), `sink` (`asset` or `handles`: the op writes to the world, each side under its own name and what it wrote is compared and removed), `world` (entities spawned for the run; `scene` takes the run's root scene, `settle` waits until the op runs on them), `tolerance` (bounds that replace the global ones for this case, `{ attribute = 0.05 }`: for an op whose values come from the device's sampler, whose fixed-point weights and decode answer within a fraction of a percent of the exact blend a node computes), `ignore` (outputs left out), `unreachable` (the sentence that says why the harness has no fixture for the case; the entry then counts as a harness error) and, for an op that is a solver over a timeline, `suite` and `describe` (the group of an existing suite that holds its comparison; `deferred` leaves that group to the suite's own run and states the entry pending, for a group that costs the suite's whole time a second time inside this one). - A fixture is a name (`mesh`, `points`, `curve`, `positionField`, or a type named in `types`), a current op run on its own fixtures (`{ op, params, inputs, output }`; the new side builds it through the node that replaces the op when the entry for that op names exactly one, so a ramp, a group, a prefab, a raster or an Sdf a node reads is one that library made; `shared: true` makes both sides take what the current op makes, for inputs the entry under test is held to and the op that made them is not), a node run on its own fixtures (`{ node, params, inputs, output }`, a new node by id), a list (`{ list }`), a record (`{ record }`) or a literal (`{ value }`). A case's `newInputs` are fixtures only the new side takes, for an input the current op states as a param (a texture the node reads where the op names the asset). - `sheet.json`: the composition sheet's expectations: the four ops, the 23 kernel families with their tallies (a kernel asset `<kernel>_<pass>` belongs to the family `<kernel>`), the layers, the L2 ids as the core package names them, `universalUses` (the L2 nodes every row may call), one row per node family (the kernel families, L2 nodes and ops a node of the row may name; `authorKernel` for rows whose nodes run a kernel the author names; which nodes may stand on one kernel family alone), and the reported numbers. `scan()` finds the new parts in the library: the nodes of every built-in nodebox wherever its package keeps it, the kernel families with their pass assets, and the modules that sit beside nodes. `analyze(src, ...)` reads a node's source: the nodes it calls (`g.<name>(...)`), the ops it names (`op.<name>`), the kernel families it names, a structural hash per call, and WGSL text in its strings. `rowViolations(scan, node)`, `dispatchViolations(scan, node)` and `directCallers(scan)` state what a node breaks of the sheet. `valuesHold(result)` reads a comparison's values alone, which is what lets an existing suite run on the nodes of an op whose cost is over. `assess(scan, entry)` states where an entry stands (`decided`, `unmapped`, `pending`, `ready`); `compare(scan, entry)` runs a ready entry's current op beside its node; `checkCase(entry)` runs the current op on the entry's case. `redirect(map, fn)` runs `fn` with `registry.get` answering a new node for each op id in `map`; `runSuite(name, redirect)` runs an existing suite's cases through the test framework's own runner, and `runSuiteOnNodes(scan, name, ops)` runs it as it stands and with its ops answered by their new nodes. `report(opts)` is the coverage report. A comparison states cost as separate numbers (`result.cost`): `oldMs` (the current op's CPU time), `planMs` (the evaluation outside the nodes' own cooks: expanding the call form, lowering refs and constructs, scheduling, with the plan's share of the waiting taken out), `cookMs` (the CPU time the nodes' cooks spend, with their share of the waiting taken out), `waitMs` (the time in `task.wait`, a read-back's frames, shared between the cooks and the plan by the read-backs each made), `dispatches` (device dispatches queued), `readbacks` (read-backs the substrate counted) and `frames` (frames that passed during the evaluation). A Field in an output is read over the Geometry its side was given, a Column by its values, and a Bool answers 1 or 0. The cost ratio of `tolerance` is applied to `cookMs` against `oldMs`; the read-back count is asserted against the op's budget on its own, so a read-back removed from a node shows as `frames` falling to zero. Times are the median of three runs; counts are the last run's. A comparison states its values and its cost apart. `result.valueFindings` are the differences in what the node answers and `result.costFindings` the cost and read-back findings (`result.findings` holds both). `verdict(assessed, result)` reads them as `value` (`equal`, `differ`, `missing` for an attribute, layer or column the new output lacks, `raised` for a node that failed on the case's inputs, `harness` for a case whose inputs could not be made, `skipped`, `decided`) and `cost` (`within`, `over`, `none`); `summary(scan)` counts every entry under exactly one of `pass`, `costOver`, `differ`, `missing`, `raised`, `harness`, `skipped` and `decided` and reports the same per family (the nodebox of the entry's first new node) with the median and worst cost ratio. What is compared where the two libraries carry a value differently: a Geometry by its live counts (a domain allocated at a capacity compares its live front), a Raster layer by layer at the grid's cell centres, a Texture by size, channel count and texels (bytes scaled to 0 to 1 and held to half a byte step), a count the new library holds on the device as a one-word column by its value, a Population against the instanced points by instance count, and a list of numbers against a list of records of numbers as one layout finding that says whether the numbers agree.

# proc.placement Where a graph stands in the world, as a 4x4 matrix held as 16 numbers column after column: the layout `entity:worldMatrix()` and `camera.viewData().viewProj` answer in. A graph is evaluated in its own space, whose origin is the entity it stands on. A Generator draws what it makes under its entity, so the graph's origin stands where the entity does, and an effect built at the origin works wherever it is played. A node that pulls from the scene (`scene.fetch`, `terrain.ground`, `sim.scene_colliders`) reads that origin from the scene the evaluation hands it, `ctx:scene().origin()`, and takes what it reads from the world into the graph's space. A graph evaluated standing on no entity has the world's origin for its own (`proc.placement.IDENTITY`). `mul`, `inverse`, `point`, `scale`, `rotation`, `turn` and `conjugate` are the matrix and quaternion steps such a node takes.

# proc.records Record kinds: the values a graph moves that become assets when saved (a material, a mesh, a texture, a user's own). `proc.record({ name, asset, applies?, fields?, live?, release?, save?, load? })` defines a kind; a `.procRecord` asset's `init.luau` returns one. A record carries its kind's guid as `kind`; the node that writes it states the kind on the geometry's detail attribute `records` (`{ [guid] = reference }`), and output, export and import find the definition there. `save` and `load` default to the asset type's own `graph.write` and `graph.read`. A kind whose `live` value holds something (a scratch asset) states `release`, which output calls once nothing stands on the value. Part of `@builtin::systems.procgen.proc`.

# proc.registry The operation registry: register/lookup/schema/search/help for procedural ops. Part of `@builtin::systems.procgen.proc`.

# proc_field_noise (compute shader) 2D value-noise field: `dst[i] = value2(seed, coords[i] * freq)`. One GPU thread per element. Re-implements `proc.noise.value2` — the integer FNV-1a lattice hash (`proc.hash.hashInts`) is bit-exact-portable, so the `u32` hash here matches the Luau `bit32` hash exactly; the quintic-faded bilinear interpolation runs in `f32` (vs the CPU's `f64`), bounded to `f32` rounding.

# proc_field_scale (compute shader) Element-wise scale of an `f32` field: `dst[i] = src[i] * scale`. One GPU thread per element. Reads a storage buffer (`src`), writes a storage buffer (`dst`), and multiplies by the `scale` params uniform.

# proc_raster_apply_accumulator Folds one tile's fixed-point accumulator into a layer, clearing the accumulator for the next batch. `proc.raster`'s `Raster:resolveAccumulator(layerName, scratchName, work?)` runs this after `proc_raster_edge_add` has folded in what a neighbour pushed across the seam, then refreshes `layerName`'s apron so the next step reads a consistent neighbourhood.

# proc_raster_apron Refreshes one tile's apron ring of one layer from the neighbour tile that owns those cells, one dispatch per (destination tile, neighbour direction). `proc.raster`'s `Raster:exchangeAprons(name, work?)` drives it, over every neighbour pair a work set touches. Any iterative raster kernel op whose per-cell result depends on a previous step's neighbourhood needs its written layer's apron refreshed this way between steps, or a later step reads a neighbour's stale interior instead of what that neighbour actually computed.

# proc_raster_copy Copies one raster tile into another, cell for cell, apron included. `proc.raster`'s channel set uses it for checkpoints, clears, and layer copies. A checkpoint is a copy of every live channel tile taken at a stack index, so an edit below that index restores the copy instead of re-running the ops above it. A clear is the same copy with a zeroed tile as the source. A layer copy duplicates a raster's cells so a mutating op leaves the original untouched. `dst`'s length decides how many cells are copied, so the two tiles must be the same shape.

# proc_raster_edge_add Adds a neighbour tile's apron-ring fixed-point accumulator into this tile's interior, one dispatch per (destination tile, neighbour direction). `proc.raster`'s `Raster:resolveAccumulator(layerName, scratchName, work?)` drives it, before folding the accumulator into `layerName` and refreshing `layerName`'s apron. Droplet and thermal erosion both accumulate this way — any tiled kernel op whose per-cell result can spill across a tile seam needs the same fold.

# proc.types The procgen type system: core type tags + socket assignability for graph validation. Part of `@builtin::systems.procgen.proc`.

# Compute shader (asset type) A `.computeShader` is a GPU compute (GPGPU) program with a **zero-scaffolding** authoring contract: you write only `@compute fn main(...)` plus a declarative `bindings.yaml`, and the engine GENERATES every `@group/@binding` declaration — storage buffers, textures, samplers, storage textures, and a `params:` uniform — expands `#include`s, naga-validates the result, and registers it. The single source of truth for the binding layout is `bindings.yaml`; there are no hand-written `@binding` lines to keep in sync. > New to how content lives across the CPU and GPU? Read `core/resource-model` > first: it covers the asset / CPU / GPU forms every resource takes, the upload > path, and how compute fits the performance ladder (`core/performance`). ## Folder shape A `.computeShader` is a folder asset: ``` <name>.computeShader/ shader.wgsl # the WGSL body — only `@compute fn main` + helpers (required) bindings.yaml # the binding + params schema (required) README.md # a one-paragraph description of THIS shader (required) .metadata # asset metadata ``` Each file carries a committed `.meta` sidecar pinning its stable guid. The identity is `<name>` (the `.computeShader` suffix strips). Resolve the shader once, then dispatch through the reference: ```lua local carve = asset.resolve("carve", "computeShader") carve:dispatch({ buffers = { heights }, workgroups = { 64 } }) ``` `asset.resolve` is what records your script's dependency on the shader, so the shader travels with the world your script lives in — packed, pulled, or published. It compiles on first dispatch; there is no setup step. ## bindings.yaml One ordered `bindings:` list — the binding index is the list position (0, 1, 2, …) — plus an optional `params:` scalar-uniform block. `kind` defaults to `buffer` when omitted. ```yaml bindings: - { name: clouds, kind: storage3d, format: rgba16f } # texture_storage_3d<rgba16float, write> - { name: occupancy, kind: texture3d } # texture_3d<f32> - { name: smp, kind: sampler } # sampler - { name: out_tex, kind: storage2d, format: rgba16f } # texture_storage_2d<rgba16float, write> - { name: U, kind: buffer, access: read, element: f32 } # var<storage, read> U: array<f32> - { name: Uniforms, kind: buffer, access: read, element: VolumeUniforms, array: false } # var<storage, read> Uniforms: VolumeUniforms params: - { name: scale, type: float, default: 2.0 } - { name: count, type: int, default: 0 } ``` | `kind` | extra keys | generated WGSL declaration | |---|---|---| | `buffer` (default) | `access: read \| read_write` (default `read_write`), `element: <wgsl type>` (default `f32`), `array: true \| false` (default `true`) | `var<storage, ACCESS> NAME: array<ELEMENT>;` (or `… NAME: ELEMENT;` when `array: false`) | | `texture3d` | — | `var NAME: texture_3d<f32>;` | | `texture2d` | `sample: float \| uint \| sint` (default `float`) | `var NAME: texture_2d<f32>;`, or `<u32>` / `<i32>` for the other classes | | `texture_depth` | — | `var NAME: texture_depth_2d;` — the one sampled shape with no `sample:` of its own; a depth-aspect channel (`@scene.depth`) binds here, not through `texture2d` | | `storage3d` | `format`, `access: write \| read \| read_write` (default `write`) | `var NAME: texture_storage_3d<FMT, ACCESS>;` | | `storage2d` | `format`, `access: write \| read \| read_write` (default `write`) | `var NAME: texture_storage_2d<FMT, ACCESS>;` | | `sampler` | — | `var NAME: sampler;` | | `textureCubeArray` | — | `var NAME: texture_cube_array<f32>;` — the environment cube array, bound to `@scene.environment` by a render feature's pass | | `vertices` | `access: read \| read_write` (default `read_write`) | `var<storage, ACCESS> NAME: array<u32>;` plus the mesh's layout uniform and the `zeroVertex*` accessors | | `acceleration_structure` | — | `var NAME: acceleration_structure;` (hardware ray tracing) | - **Eight storage buffers per shader** — the WebGPU guarantee, and what the engine asks every device for. `buffer` and `vertices` bindings spend that budget; textures, samplers, storage textures and `params:` do not. An `acceleration_structure` brings engine-owned buffers with it — **three** where the device has hardware ray query (the tables `zeroHitSurface` / `zeroMaterial` read) and **five** on the software backend (those, plus the scene triangles and the hierarchy over them) — so a traced pass has three slots of its own on the backend web users get. Over budget, the engine refuses the shader and logs which buffers it declared. - A `buffer` defaults to `array<ELEMENT>`. Set `array: false` for a single NON-array storage buffer — a struct binding (e.g. `element: VolumeUniforms`, `array: false` → `var<storage, read> NAME: VolumeUniforms;`). The struct definition itself lives in `shader.wgsl`, not `bindings.yaml`. - **A sampled texture binds under the class its FORMAT puts it in**, so `sample` states that class and the generated `texture_2d<..>` matches it. An integer channel — `@scene.material`, which is `Rgba8Uint` — is `uint`; everything else is the `float` default. Bind an integer channel to a `float` slot and the engine refuses the pass and names both classes, because the device would otherwise reject the bind group and invalidate the pipeline. - A storage texture's `access` decides both the generated declaration and the layout the engine registers, so the two always agree. It defaults to `write`, the access every target grants every storage-compatible format. `read_write` is granted on `r32f` alone; declaring it on any other format is refused at the compile, with the reason on `compute.programState`. A pass that needs to read a wider target reads it through a separate `texture3d` / `texture2d` binding, or through a second storage binding declared `read`. - `format` tags (storage textures): `r8`, `r16f`, `r32f`, `rgba8`, `rgba16f`, `rgba32f`. Storage textures only allow storage-compatible formats (`r32f` / `rgba8` / `rgba16f` / `rgba32f`). - An `acceleration_structure` binding resolves to the scene acceleration structure when the shader runs as a render-pass with ray tracing enabled (`renderer.setRaytrace(true)`). Declaring one makes the engine emit the helper functions the body can call: `zeroTraceClosest(origin, dir, t_min, t_max) -> ZeroHit` (struct: `hit, t, instance, primitive, bary, attr`, where `instance` is the hit object's render slot), `zeroTraceAny(origin, dir, t_min, t_max) -> bool` (terminate-on-first-hit, for shadow/occlusion rays), and `zeroHitSurface(hit) -> ZeroSurface` (`normal, uv, material`) + `zeroMaterial(index) -> ZeroMaterial` (`base_color, emissive, metallic, roughness`) to shade what a hit landed on. The same surface works on both backends — hardware ray query and the software/compute path — so the shader doesn't branch on `renderer.raytraceCapability()`. Bind it from a render-pass, which is where the scene acceleration structure exists. `params:` declares scalar uniforms (`type: float | int`). When non-empty the engine appends a `ZeroComputeParams` uniform as the trailing binding (read it in WGSL as `params.<name>`) and tunes its values at runtime with `shaderRef:setParam(prop, value)`. ## The kernel ABI: `columns`, counted and appended outputs, `specialize`, `determinism` One shader asset can serve many shapes of data. `bindings.yaml` states the shape with these extensions, and the engine's compile path lowers them to the plain bindings and WGSL above (`modules/compute_abi`, whose README has the whole surface): ```yaml workgroup: 64 # the size the body declares (default 64) bindings: - { name: cols, kind: columns, access: read } # a table of columns in one binding - { name: src, kind: buffer, access: read, element: T } - { name: dst, kind: buffer, access: read_write, element: T, count: "len(src)" } - { name: hits, kind: buffer, access: read_write, element: vec4<f32>, capacity: "n * 4", append: true } specialize: - { name: T, options: [f32, i32, vec2<f32>, vec3<f32>, vec4<f32>] } - { name: FOLD, options: [add, min, max] } - { name: BODY, kind: source, default: "0.0" } # text that replaces the identifier determinism: exact # or float_order ``` - A `buffer` of a scalar or vector element gets `NAME_get(i)`, `NAME_len()` and (unless `access: read`) `NAME_set(i, v)`. A `vec3` element is stored as a flat scalar array, so a column of vec3 keeps its 12-byte stride. `zero_n()` is the dispatch size. - `columns` binds a table of columns (`compute_abi.table`) as one storage buffer; `cols_f32(c, i, lane)`, `cols_vec3(c, i)` and the rest read it. - `count` / `capacity` give the size, in elements, of an output the run allocates. `append: true` adds a live counter binding, `NAME_count`, with `NAME_store(slot, v)`, `NAME_set_count(n)` and, under `determinism: float_order`, `NAME_append(v)`. - `specialize` parameters are compile-time: type options become `alias NAME`, named options become `const NAME` and `NAME_<option>`, `kind: source` text replaces the identifier. Each specialization compiles once. - `determinism: exact` refuses `NAME_append`: an exact output is written at ranks the kernel computed. Dispatch such a shader through its program: ```lua local variant = shaderRef:program():variant({ specialize = { T = "vec3<f32>" } }) local run = variant:run({ n = 64, inputs = { src = srcBuffer } }) local values = require("@builtin::modules.compute_abi").read(run.outputs.dst.buffer) ``` `shaderRef:program()` answers the same program, with the variants it has lowered and compiled, for as long as `shader.wgsl` and `bindings.yaml` hold the bytes they held when it was made. A write to either file, from any origin (`vfs.contentVersion` moves), makes the next `program()` a new one, and a dispatch through it runs the new text. The schema `getBindings()` answers is the one the program was made over. ## Writing a mesh's vertices — `kind: vertices` A mesh's vertices are raw bytes on the GPU, and the engine stores a mesh in one of two layouts: 92 bytes a vertex with every attribute a `f32`, or 32 bytes a vertex with normals and tangents quantized to bytes and UVs to shorts. Which one a mesh got depends on the mesh. Declare `kind: vertices` and the shader is handed both the buffer and the layout that mesh is in, plus accessors that address it: ```yaml bindings: - { name: verts, kind: vertices } # the mesh, writable - { name: heights, kind: buffer, access: read, element: f32 } params: - { name: lift, type: float, default: 1.0 } ``` ```wgsl @compute @workgroup_size(64) fn main(@builtin(global_invocation_id) id: vec3<u32>) { if (id.x >= zeroVertexCount()) { return; } var p = zeroVertexPosition(id.x); p.y = p.y + heights[id.x] * params.lift; zeroSetVertexPosition(id.x, p); zeroSetVertexNormal(id.x, vec3<f32>(0.0, 1.0, 0.0)); } ``` ```lua carve:dispatchOnVertices({ model = meshRef:handle().guid, buffers = { heights }, workgroups = { 64 } }) ``` The accessors, over a binding declared `read_write`: | read | write | |---|---| | `zeroVertexCount() -> u32` | — | | `zeroVertexPosition(i) -> vec3<f32>` | `zeroSetVertexPosition(i, vec3<f32>)` | | `zeroVertexNormal(i) -> vec3<f32>` | `zeroSetVertexNormal(i, vec3<f32>)` | | `zeroVertexTangent(i) -> vec4<f32>` | `zeroSetVertexTangent(i, vec4<f32>)` | | `zeroVertexUv(i) -> vec2<f32>` | `zeroSetVertexUv(i, vec2<f32>)` | | `zeroVertexColor(i) -> vec4<f32>` | `zeroSetVertexColor(i, vec4<f32>)` | | `zeroVertexNodeIndex(i) -> u32` | — | | `zeroVertexWord(i, offset) -> u32` | — | `access: read` gives you the readers only — a `var<storage, read>` binding cannot be assigned to, so the writers are left out rather than handed to you broken. Once a pass has written a mesh's vertices, the engine bounds that mesh by what the pass wrote: it reduces the vertices to an AABB every frame and culls against that, so geometry a pass moves across the scene needs no oversized placeholder to survive culling. `renderer.mesh.boundsSource(guid)` reports `"compute"` for such a mesh and `"geometry"` for one still bounded by what it was created with. A shader declares at most one `vertices` binding; the accessors are generated for that mesh. It can be declared anywhere in the list — the mesh is bound by the dispatch, and `buffers` fills the other storage slots in declaration order. The tangent's `w` carries handedness, and quantized normals/tangents come back in -1..1 while quantized UVs and colours come back in 0..1, so the values are the same ones a surface shader sees whichever layout the mesh is in. ## shader.wgsl Write only `@compute fn main` (and any helpers / `#include`s). The names declared in `bindings.yaml` are in scope — `data`, `out_tex`, `params.scale`, … — with no `@group`/`@binding` lines of your own. ```wgsl // @domain: compute // Bindings are declared in bindings.yaml — do NOT write @group/@binding here. @compute @workgroup_size(64) fn main(@builtin(global_invocation_id) id: vec3<u32>) { if (id.x < arrayLength(&data)) { data[id.x] = data[id.x] * 2.0; } } ``` ## Dispatch & runtime Everything a compute shader does, it does through its reference: | method | what it does | |---|---| | `ref:dispatch(opts)` | `{ buffers, workgroups, params? }` — bind one buffer per declared storage binding, in order | | `ref:dispatchOnVertices(opts)` | `{ model, buffers?, workgroups, params? }` — bind the mesh `model` names to the `vertices` binding, `buffers` filling the rest | | `ref:dispatchEx(opts)` | `{ resources, workgroups }` — for shaders mixing textures, samplers and storage textures: one `{ kind, name }` per declared binding, in order | | `ref:setParam(prop, value)` | write one `params:` scalar uniform | | `ref:createBuffer(name, opts)` | `{ type, len, usage? }` — a GPU buffer owned by this shader | | `ref:buffer(name)` | the buffer this shader last made under `name` | | `ref:getBindings()` | the parsed `bindings.yaml` schema, read and parsed again only after the file is written | | `ref:getSource()` / `ref:setSource(wgsl)` | read / overwrite `shader.wgsl` | | `ref:compile()` | compile now instead of on the first dispatch | | `ref:status()` | what the engine did with this shader's dispatches, one record per target | The `params:` uniform is engine-owned and takes no entry in `buffers` or `resources`, and neither does a `vertices` binding's layout. `setParam` writes the uniform every later dispatch reads. A dispatch that passes `params = { name = value }` reads those values instead, in a copy of the uniform only it binds, with every name it leaves out at the value the shader holds — so one shader dispatched twice in a frame over different data reads different values each time: ```lua lift:dispatch({ buffers = { a }, workgroups = { 4 }, params = { height = 1.0 } }) lift:dispatch({ buffers = { b }, workgroups = { 4 }, params = { height = 3.0 } }) ``` A name the shader does not declare fails that dispatch, and the failure names what it does declare. ### `dispatchEx` resources Each `resources` entry fills the binding at its position. Its `kind` says what the resource is, in the words `bindings.yaml` uses, and the binding decides how it is read. A slot takes the kind it is declared with; a `texture2d` slot also reads a storage texture, a side of a history pair or the scene depth, and a `storage2d` slot also writes this frame's side of a history pair: | binding's `bindings.yaml` kind | `resources` entry | `name` | |---|---|---| | `buffer` | `{ kind = "buffer", name = buf }` | a buffer handle from `ref:createBuffer` | | `texture3d` | `{ kind = "texture3d", name = vol }` | a 3D texture handle from `ref:createTexture3D` | | `storage3d` | `{ kind = "storage3d", name = vol }` | a 3D texture handle from `ref:createTexture3D` made with `storage = true` | | `storage2d` | `{ kind = "storage2d", name = target }` | a storage texture handle from `ref:createStorageTexture2D` | | `storage2d` | `{ kind = "history_curr", name = history }` | a history pair from `ref:createTextureHistory`: this frame's side | | `texture2d` | `{ kind = "texture2d", name = tex }` | a handle `ref:copyBufferToTexture` returns, or a texture asset's guid once `texRef:handle()` has made it GPU-resident | | `texture2d` | `{ kind = "storage2d", name = target }` | a storage texture handle, read as a texture | | `texture2d` | `{ kind = "history_prev", name = history }` / `{ kind = "history_curr", name = history }` | a history pair: the previous frame's side or this frame's | | `texture2d` | `{ kind = "scene_depth" }` | none: the frame's scene depth | | `sampler` | `{ kind = "sampler", name = smp }` | a sampler handle from `ref:createSampler` | `texture3d` and `storage3d` name the same resource, a 3D texture, so either one fills either slot. The engine's own spellings name the same resources and keep working: `texture_3d` for `texture3d` and `storage3d`, `texture_2d` for `texture2d`, and `storage_2d` for `storage2d`. Any other `kind` is refused at the call, naming every kind it accepts. A `vertices` binding is filled by `ref:dispatchOnVertices`, and `texture_depth`, `textureCubeArray` and `acceleration_structure` bindings by a render feature's pass. ```lua local march = asset.resolve("volume_march", "computeShader") -- bindings.yaml: density (storage3d), out_tex (storage2d), U (buffer) march:dispatchEx({ resources = { { kind = "storage3d", name = density }, { kind = "storage2d", name = target }, { kind = "buffer", name = uniforms }, }, workgroups = { 8, 8, 8 }, }) ``` `workgroups` is `{ x, y, z }`, `{ x }`, or a bare `x`, and every dimension it leaves out is 1. A zero in any dimension is refused: the dispatch is dropped and recorded as a failure, so it surfaces in `ref:status()` and `compute.failing()` with the reason. A count derived from how much data there is passes through `math.max` first, so an empty set dispatches one workgroup rather than none: ```lua carve:dispatch({ buffers = { heights }, workgroups = { math.max(1, math.ceil(#points / 64)) } }) ``` ### Buffers and readback (the CPU side) A buffer is created through the shader that owns it and comes back as a handle. Pass the handle to a dispatch — including another shader's — to bind it there: ```lua local double = asset.resolve("@builtin::shaders.compute_double", "computeShader") -- `type` is the element the shader declares, `len` how many of them. local data = double:createBuffer("data", { type = "f32", len = 4, usage = { "readback" } }) data:write({ 1.0, 2.0, 3.0, 4.0 }) -- :writeU32 for integer words, :writeBytes for packed bytes double:dispatch({ buffers = { data }, workgroups = { 1, 1, 1 } }) local pending = data:read() -- starts an async GPU->CPU readback -- a frame or two later, drain it (nil until it has arrived): local out = pending:result() -- :resultU32 / :resultBytes read the same bytes another way data:destroy() -- free it when done ``` `type` is one of `"f32"`, `"vec3"`, `"vec4"`, `"quat"`, `"mat4"`, and `usage` adds `"readback"`, `"vertex"`, `"index"` or `"indirect"` on top of the storage a buffer always has. An `"indirect"` buffer holds a draw's arguments where the GPU reads them, so a shader that compacts geometry writes the drawn count into the same dispatch that produced it and a render feature's `kind = "draw"` pass submits exactly that — the `renderFeature` asset type's README documents that pass and the `storage.indirect` name it reads. The handle's key is the shader's guid plus the name you gave it and a serial, so two shaders that both want a buffer called `data` get one each — and so do two callers asking this same shader for one. `ref:buffer(name)` returns the last one made under that name, which is how two callers share one on purpose. A buffer that no shader computes — a vertex or index run that is only ever drawn — is allocated straight from the primitive instead: `substrate.createBuffer({ name = ..., type = ..., len = ..., kind = "gpu", usage = { "vertex" } })`. The first dispatch of a shader compiles it, so the very first readback can still reflect the un-run shader; dispatch again once it is compiled, or `ref:compile()` to pre-warm. This create/write/dispatch/readback loop is the CPU side of the resource model in `core/resource-model`. ### Is it compiled? `ref:compile()` queues the compile; the pipeline is built on the render side frames later, so that call cannot say whether it produced a program. `compute.programState(ref)` is where that lands: ```lua if compute.programState(carve) == "ready" then carve:dispatch(opts) end local state, reason = compute.programState(carve) if state == "failed" then print(reason) end ``` - `"absent"` — the engine holds nothing under this key and nothing is in flight: never asked for, or released. - `"pending"` — a compile was asked for and has not reached the device. A recompile of a resident program reads pending too, because what it produces is a different program from the one bound now — which is what makes this the precondition to wait on after an edit. - `"ready"` — compiled and resident, so a dispatch naming it binds it. - `"failed"` — the most recent compile produced no program, and the second return says why. A dispatch issued while the program is `"pending"` is held behind the compile and dropped if it never lands, so a readback taken before `"ready"` reports the buffer as it was. Wait for `"ready"`, not for a count of frames. ### Did it run? A dispatch is recorded into a command encoder frames after the call that asked for it returned, so that call's `true` means "queued", not "ran". `ref:status()` is where the outcome lands — one record per target, cumulative: ```lua for _, d in ipairs(carve:status()) do print(d.target, d.ok, d.dispatches, d.failures, d.lastError) end ``` `target` is the mesh guid for a `dispatchOnVertices` and the buffers it bound for a dispatch that writes only those, so two systems driving one shader over their own buffers read as the two passes they are. `ok` is the most recent outcome; `lastError` says why the last failure failed and is kept after a recovery. An empty result means nothing has dispatched this shader at all. `compute.failing()` answers the same question across every shader at once. A failing dispatch also posts a notice, so a pass that stops running reports itself rather than waiting to be found. ### Running one every frame A `kind = "compute"` substrate job queues one dispatch per frame: ```lua jobs.register({ phase = "main", executor = { kind = "compute", shader = asset.resolve("carve", "computeShader"), buffers = { "heights" }, workgroups = { 64 }, }, }) ``` ## Registration & hot reload Writing `shader.wgsl` or `bindings.yaml` fires the `.computeShader` assetType's `onChange`, which calls `__compute.compile` — that generates the binding declarations, expands `#include`s, naga-validates, and registers. This happens on create and on every edit, so saving recompiles with no world reload. A shader with invalid WGSL (a type mismatch, an undefined name) **fails to compile** with the reason logged (`[compute] Compile of '<name>' failed: …`) and is **not** registered. You don't need to register a shader before dispatching it — the first dispatch auto-compiles it on use. `ref:compile()` remains as an optional explicit pre-warm, only to avoid the one-frame first-dispatch warm-up in a latency-critical spot. ## What compute is holding, and why a key names nothing Every buffer, volume, storage target, history pair and sampler a dispatch can bind is filed under a key, and `compute.observe()` lists all of them with what each costs: ```lua local r = compute.observe() print(r.totals.count, r.totals.bytes) for _, res in ipairs(r.resources) do print(res.key, res.kind, res.bytes, res.owner.name) end ``` A row carries the key verbatim — the same string a dispatch names it by and a dispatch-failure message prints — plus its `kind`, its `bytes`, its extent and `format` where it has them, the `usage` bits it was created with, and the frame it was created on. `owner` is read back out of the key (`<shader-guid>::<name>#<serial>`), so a resource nothing has ever dispatched over still names the shader that made it. `compute.resources(ref)` narrows to one shader's — including the `<shader-guid>::__params` uniform the compile created for a `params:` block, which the shader owns like any other resource. `r.totals.bytes` is the `compute` figure of `renderer.gpuMemory()` — the two are read off the same registries, so a listing accounts for the category exactly. The same reading is served at `/zero/runtime/compute`. When a key names nothing, `compute.diagnose(key)` says which of the states the inventory distinguishes it is in, from the closed set `compute.absentReasons()` enumerates: | reason | what the inventory read | |---|---| | `resident` | a resource is filed under exactly this key; the row is on `resource` | | `replaced` | the owner holds the same name at a different serial — `current` names the live key | | `nameUnused` | the owner holds resources, none under this name | | `ownerAbsent` | no live resource carries this key's owner | | `unowned` | the key carries no owner segment, and nothing is filed under it | | `notPublished` | no renderer has published an inventory, so no key can be answered for | So a key out of `[compute] dispatch '…' failed` resolves against what exists, and a typo reports which part of it the engine could not match rather than coming back empty. A reason is the inventory at the moment the call read it. The reading is republished on a frame where a registry gained or lost an entry, so a resource created and diagnosed inside one script sees the reading from before it was made; wait a frame and ask again. ## Discovery - `asset.list("computeShader")` — every registered compute shader. - `asset.inspect("<name>")` — status, source, parse errors, this README. - `computeRef:getBindings()` — the parsed `bindings.yaml` schema (`{ bindings, params }`). - `compute.programState(ref)` — whether the program is compiled, still compiling, or failed, and why. - `compute.observe()` — every GPU resource compute holds, and what it costs. - `compute.diagnose(key)` — whether a key names a live resource, and why not. - `profiler.gpuFrame()` — what each dispatch of this shader spends on the GPU. A dispatch is timed under `compute.<identity>` — the same identity `asset.resolve` takes, so a timing row names the shader that produced it. ## Common pitfalls - **Compute shaders don't render** — verify output via a buffer readback or a debug pass, not a screenshot. - **`ref:compile()` returns before the program exists** — the pipeline is built on the render side; `compute.programState(ref)` says when it is `"ready"`, and a dispatch issued before that is dropped. - **A dispatch call returning true means queued** — ask `ref:status()` whether it ran. - **A destroyed buffer is refused where it is bound** — a handle outlives its storage, and `buf:alive()` reports which. - **Binding order is the contract** — `buffers` and `resources` bind in `bindings:` list order; reordering the list reorders the bindings. - **Don't write `@group`/`@binding` in `shader.wgsl`** — the engine owns them; re-declaring is the fastest way to a layout mismatch. - **Don't infer a mesh's vertex stride** — declare `kind: vertices` and use the accessors. A wrong stride writes positions correctly and normals into the wrong word, which renders as flat lighting rather than as broken geometry. - **Storage-texture format must be storage-compatible** — `r32f` / `rgba8` / `rgba16f` / `rgba32f`; others are rejected at compile. - **A key that comes back empty is not the same as one that failed** — `compute.diagnose(key)` separates a name never used from a generation that was replaced, and names the live key in the second case. ## Related types - `.shader` — render-domain shaders (surface, sky, post-process, screen) and the legacy `@domain: compute` fallback. ## In the Inspector Selecting a compute shader opens its **Compute** section: whether its program is ready (and why not), how many dispatches it has run and failed in this session (and the last error), tables of the bindings and params its `bindings.yaml` declares, and openers for `shader.wgsl` and `bindings.yaml`.

# luau_introspect (module) Parses Luau/asset source TEXT into structured declarations. Pure string/pattern parsing (`string.find` / `string.match` / `string.gmatch`) over the literal source — no compiling, no VM spawn, no runtime state, no engine FFI. Given source text, it extracts what is written in the file. This is the shared parser every per-assetType `M.ref.inspect` hook uses to turn a component/module/tool's own source into the structured record `asset.inspect` returns. ## Exports - `docstrings(src) -> DocMap` — scans `--!desc` / `--!arg` / `--!return` / `--!example` comment runs and binds each run to the name of the declaration on the next non-comment line (the identifier after `function`, `public:`, `local function`, or `M.`). - `publicFields(src) -> { FieldEntry }` — parses `public = { name = Field.<kind>(default, mode), ... }` table-literal entries and module-scope `public.<name> = Field.<kind>(...)` assignments. - `events(src) -> { EventEntry }` — parses `events = { name = Event(payloadSchema?, syncMode?), ... }` table-literal entries. - `methods(src) -> { MethodEntry }` — parses `function public:<name>(...)` / `typed function public:<name>(...)` declarations (params + optional return type). - `lifecycleHooks(src, catalog) -> { string }` — top-level `function <name>(` declarations whose name is in the caller-supplied `catalog` array. - `moduleExports(src) -> { ExportEntry }` — parses `M.<name> = function` assignments, `function M.<name>(...)` declarations, and a trailing `return { foo = foo, ... }` export table. - `forSource(src, computeFn) -> any` — memoises `computeFn(src)` keyed on `src` itself in a module-local table. The same source text returns the cached result without re-running `computeFn`; different source text recomputes. Keyed on the source content (not an asset checksum) because composite/container assets (component folders, scene folders, etc.) all hash to the empty-string checksum — a checksum-keyed cache collapses every such asset's detail onto whichever one was inspected first. ## Usage ```luau local Introspect = require("modules.luau_introspect") local src = self:getInitScript() local detail = Introspect.forSource(src, function(s) return { fields = Introspect.publicFields(s), events = Introspect.events(s), methods = Introspect.methods(s), hooks = Introspect.lifecycleHooks(s, lifecycleCatalog), } end) ``` ## Notes - Every function is a pure string→table transform: same source text in, same structured table out, every time. No reads of live component state, no entity/world queries. - Structural scanning runs over a comment- and string-aware **mask** of the source (a same-length copy where comment bodies and string-literal contents are blanked to spaces, delimiters and newlines preserved). Pattern matches and brace/paren balancing run on the mask — so a brace, `Field.`, `Event(`, or a whole entry that appears inside a comment or a string literal never registers — while captured values are sliced from the original source at the same (aligned) indices. This makes the parser robust to apostrophes in comments, commented-out entries, string defaults containing `}` / `Field.` / `Event(`, and `[[ ]]` long strings. - `default` / `category` are raw trimmed source-text slices (with any trailing comment stripped), not evaluated values. - For a **ref constructor** (`Field.assetRef` / `dataRef` / `resource` / `componentRef`) the FIRST positional argument is the category / contract / componentType selector, captured under `category`; the SECOND argument is captured under `default`. - For every **other** Field kind the first argument is the `default` and `category` is nil. - A trailing `Sync` / `NoSync` identifier is captured as `sync`, tolerating a trailing comma or a trailing comment after the last argument. - Field/Event entries whose key is a computed or indirect expression (`[expr] = Field.number(...)`) are omitted — only literal identifier keys are recognised. - `methods` captures a `...` variadic parameter (as `name = "..."`), with its `: <type>` annotation when present. - A `--!desc` written above a field/event declaration (`<name> = Field.<kind>(...)` / `<name> = Event(...)`) or a `public.<name> =` assignment binds to that name in `docstrings`, and is folded into the corresponding `publicFields` / `events` entry's `desc`. - `lifecycleHooks` never hardcodes the lifecycle-callback catalog; the caller passes the list of names to match against. - `forSource`'s memo is bounded (oldest-key eviction) so a long session touching many distinct asset versions can't grow it without limit.

# assetType.shared.ref Builds the metatable every `AssetRef` (the `{ __ref, type, name, guid, identity, path }` envelope returned by `asset.resolve` / `asset.ref`) carries. Replaces the per-call C metatable that used to live in `crates/zero_scripting/src/ffi/bindings/asset.rs::attach_asset_ref_metatable`. ## Public API ```lua local assetRef = require("@builtin::assetTypes.assetType.shared.ref") -- Called by the Rust factory (push_asset_ref_handle) via -- _G.__build_asset_ref_proxy after the envelope fields are set. assetRef.build(envelope) -- attaches AssetRefMT, returns the table ``` The module exposes no invalidation surface — discovery is registry-driven and hot-reload is transparent (see "Discovery and hot-reload" below). ## Method dispatch order `ref.foo` is resolved in this order: 1. **Default methods** — `getSource` / `getBytes` / `getText` / `exists` / `inspect` and the lazy `meta` property. Per-type definitions cannot override these (engine contract from gh#1889). 2. **Per-type `ref` table** — discovered through `asset.resolve(<typename>, "assetType")` + `require(<identity>.behavior)`. The lookup resolves to whichever `<typename>.assetType/` folder is currently registered for the typename (built-in, user, or library-vendored) and pulls its `ref` methods. 3. **Nil** — unknown key, same shape as the original C metatable. ## Adding methods for a new asset type Drop a `behavior.luau` into the type folder: ``` src/lua/lib/assetTypes/<typename>.assetType/ ├── type.yaml -- structural spec (existing) ├── behavior.luau -- ref + global behaviour (new) └── template/ -- placeholder body (existing) ``` `behavior.luau` returns: ```lua return { ref = { --!desc Per-instance method on every `<typename>` AssetRef. myMethod = function(self, ...) ... end, }, global = { -- Reserved for the asset.<typename>.* surface — not consumed -- yet; included so the contract from gh#1889 lands cleanly. }, } ``` Within `ref` methods, `self` is the AssetRef envelope: `self.path`, `self.identity`, `self.guid`, `self.type`, `self.name` are available. Engine-required fields (`guid`, `path`, `identity`, `type`) must not be redefined. ## Discovery and hot-reload Type discovery uses the asset registry, so it doesn't matter where a `<typename>.assetType/` folder lives — engine built-ins under `@builtin::assetTypes.<X>`, user types dropped under `/zero/source/<X>.assetType/`, or types vendored inside an imported library all resolve identically. The dispatcher walks `asset.resolve(<typename>, "assetType")` → canonical identity → `require <identity>.behavior` on every miss; the asset registry is the cache for the resolve half and Luau's `_LOADED` is the cache for the require half. Hot-reload is transparent. When a `<typename>.assetType/behavior.luau` is rewritten through VFS, the engine's `invalidate_require_cache_for_state` re-runs the module and mutates the cached `_LOADED[<identity>.behavior]` table in place. Dispatch reads `mod.ref` on every access (rather than caching a sub-pointer), so existing AssetRefs see the new methods on the very next access — no Rust-side hook into this module is needed, and no invalidation API is exposed for callers to call. The same registry-driven path also handles type **registration** correctness: dropping a new `<typename>.assetType/behavior.luau` updates the asset registry, the next `asset.resolve(<typename>, "assetType")` returns the new ref, and AssetRefs of that type immediately gain the new methods. ## Why Luau, not Rust The asset surface needed per-category behaviour (`material:getProperties()`, `bundle:instantiate(entityId)`, `tool:run(args)`) and the old shape couldn't grow without an FFI change per method. Owning the proxy in Luau means: - Type-specific methods ship as content (the `<typename>.assetType/` folder), not engine code. - Adding a new asset type drops in a new folder; no Rust rebuild. - Methods can call any other Luau global (`vfs`, `asset`, `Material`, `entity`, etc.) directly, instead of going through narrow FFI primitives. ## Consumers - `crates/zero_scripting/src/ffi/bindings/asset.rs::attach_asset_ref_metatable` — the Rust factory that pushes envelopes invokes `_G.__build_asset_ref_proxy` once per push. - Every Luau call site of `asset.resolve` / `asset.ref` / `entity(id).component.X.material` / any binding declared `AssetRef<...>` — they all receive envelopes built through this module.

# yaml.module YAML decode + encode for authored engine content. Pure Luau, no engine dependencies — a general parsing library, not data-system plumbing. ```luau local Yaml = require("@builtin::modules.yaml") local doc = Yaml.decode(vfs.read(path)) vfs.write(path, Yaml.encode(doc)) ``` ## Supported subset - Block mappings (`key: value`) and block sequences (`- item`), nested to any depth. - Compact sequence-of-mapping entries (`- path: x` / ` kind: y`). - Flow collections — `{ a: 1, b: 2 }` and `[1, 2, 3]` — single line. - Plain scalars with type inference: `~` / `null` → nil, `true` / `false` → boolean, numeric literals → number, everything else → string. - Single- and double-quoted strings (`''` escapes a literal quote inside a single-quoted string; `\n` `\t` `\r` `\\` `\"` inside a double-quoted one). A quote opens a quoted scalar where a scalar begins — after `key:`, after a `- ` entry marker, or inside a flow collection — so a plain scalar carries an apostrophe as data (`note: the world's rules`). Quoted keys are read in flow mappings. - `#` comments — honored outside quotes, so a `#` inside a quoted string is kept as data. - Literal (`|`) and folded (`>`) block scalars — the bare indicators only. A block scalar's content is literal text: quotes (balanced or not), `#`, `key: value` and `---` inside it are data. - An optional leading `---` document marker. ## Loud-error contract Constructs outside the subset above raise instead of misparsing: anchors/aliases (`&`, `*`), tags (`!`), directives (`%`), multi-document streams (a second `---`), block-scalar chomping/indentation indicators (`|-`, `|+`, `>-`, `|2`, ...), tab indentation, and inconsistent indentation (both block structure and block-scalar bodies). Every error carries the offending line number (`yaml: line N: ...`), so a bad config points straight at the line to fix. Tab indentation is rejected everywhere — including block-scalar body lines whose leading whitespace contains a tab. A tab after the first non-space character of a line is content and passes through. ## API - `Yaml.decode(text: string): any` — decode a YAML document into a Luau value (table, scalar, or nil for an empty document). Raises on malformed input or an unsupported construct. - `Yaml.decodeWithComments(text: string): (any, boolean)` — decode as `decode` does, and report whether the text carries a comment by the decoder's own rule: a `#` at a line's start or after whitespace, outside a quoted scalar and outside a block scalar. Re-encoding the value drops such comments, so an editor that writes a document back reads this first. - `Yaml.encode(value: table): string` — encode a Luau table as YAML (block style, two-space indent, alphabetically sorted keys). Raises on values YAML can't represent (functions, userdata, non-string mapping keys). A top-level empty table encodes to an empty document, which decodes back to `nil` (the YAML empty-document ambiguity).

# json JSON encode/decode library for Luau. Encodes Lua values to JSON strings and decodes JSON strings back to Lua values. Used for communication with the Rust side of the engine, the VFS read/write bridge, and any wire-format that needs JSON. Pure Luau, no engine dependencies. Compact and pretty-printed encoders, plus a hand-rolled decoder that streams the input by position so it works under WASM as well as native. ## Exports - `Json.encode(value: any, indent?: string, currentIndent?: string) -> string` — compact encode. Functions / unknown types and NaN/Inf encode as `null`. - `Json.encodePretty(value: any, indentStr?: string) -> string` — pretty-printed encode with sorted object keys (diff-friendly). - `Json.encodeArgs(...: any) -> string` — encode varargs as a JSON array. - `Json.decode(str: string) -> any` — decode a JSON string. Returns the decoded value, or `nil` + error message on failure. ## Usage ```luau local Json = require("@builtin::modules.json") local widget = { type = "button", text = "Click Me" } local compact = Json.encode(widget) -- '{"text":"Click Me","type":"button"}' local pretty = Json.encodePretty(widget, " ") local decoded = Json.decode(compact) local v, err = Json.decode("oops") -- v = nil, err = error message ``` ## Notes - Object keys are sorted alphabetically in both encoders for consistent output across runs. - Numeric keys on objects are stringified at encode time (JSON has no numeric keys). Pure-integer key sets get detected as arrays via `isArray` and encoded with brackets. - NaN, +Inf, -Inf encode as `null` — JSON has no representation. Round trips through `decode` recover `null` (Lua `nil`), so they don't preserve. - Unicode `\uXXXX` escapes decode to UTF-8 by hand to stay WASM-safe. Only the BMP is covered; supplementary planes via surrogate pairs are not. - Functions encode as `null`. - Decode is character-streamed — no regex, no `string.match` patterns on the whole input — so the line-and-column information needs to be reconstructed from the position offset.

# assetType runtime The machinery that reads a `<typename>.assetType/behavior.luau` and turns its declarations into live behaviour. A `behavior.luau` is a table of declarations — `ref` methods, `modules` shared code, an `events` schema, a `namePattern`, `refShapes`, and the `onCreate` / `onRegister` / `onChange` / `onDelete` / `validate` / `settingsDefaults` / `settingsCorrections` hooks. None of it runs itself. This is the code that runs it, and it lives here because reading a type's declarations is the **assetType type's own behaviour** — the same way `material.assetType` owns what a `.material` does. ## Sub-modules | Module | Reads | Responsibility | |---|---|---| | `ref.module` | `ref`, `modules` | Builds every `AssetRef` envelope, attaches the shared metatable, and dispatches per-type methods and shared modules onto it. Also owns edit-mode write-behind persistence. | | `create.module` | `onCreate`, `namePattern`, `validate` | Backs `asset.create`: resolves the type, checks the name against its pattern, runs `onCreate` (or clones `template/`), and writes the instance. `validate.module` beside it checks a creation `opts` table against the type's `createSchema`. Before running `onCreate`, resolves the type's settings chain (below) and hands the result to the hook as `opts.settings`. | | `settings.module` | `settings`, `settingsDefaults`, `settingsCorrections` | A type's declared settings schema and where its values live (`type.yaml`'s `settings:` block). `Settings.resolveForCreate` layers the schema's own defaults, the type's own `settingsDefaults(name, opts, asked)` hook (a DEFAULT the schema can't express — a name-derived convention, a fact read off a payload the caller handed it, a value following a field `asked` shows a higher layer naming; any later layer overrides it), the Preset Manager, an importer's settings, the caller's `settings` table, then the type's own `settingsCorrections(values, name, opts)` hook (a CROSS-FIELD CONSTRAINT the schema can't express, applied last over the fully merged values). A correction over a default/preset/importer value lands silently; a correction that would override what the caller's own `settings` table explicitly asked for refuses the creation instead, naming the field and why. Neither hook is ever handed a table to mutate; each returns what it wants changed (`settingsCorrections` may also return a `{ [field]: reason }` table for that refusal message). `Settings.cloneValues` deep-copies a values table — `create.module` hands `onCreate` a copy, never the table it stamps. `Settings.valuesOf` / `writeValues` read and write an asset's own values file. | | `typeYaml.module` | `type.yaml` | The folder a type lives in and its parsed `type.yaml`, decoded once per content version of the file. `settings.module` and `inspector.module` read their blocks through it. `about` / `aboutAt` read a type's `about.yaml` (`intent`, `function`, `usage`), leaving out any key that is empty or still starts with `[[describe]]`; `aboutRows` and `aboutMarkdown` lay the texts out for the Inspector and the `types/<name>` guide. | | `inspector.module` | `type.yaml`'s `inspector:` block | Which inspector a type declares for its assets (`inspector: { type: ... }` or `{ instance: ... }`), and the declaration that file returns — the sections it adds, the generic ones it replaces, and their order — or the problem that kept it from loading. | | `onRegister.module` | `onRegister` | Fires an instance's once-only registration hook — on arrival, on world load, and for `@builtin` library instances. | | `changeDispatch.module` | `onChange`, `onDelete`, `refShapes` | Routes a VFS source write or a folder delete to the enclosing typed asset's type hooks. | | `refShapes.module` | `refShapes`, `ref` | Publishes what every `AssetRef<category>` answers to, so a member read on an asset-typed value is checked against that category's real surface. | | `events.module` | `events` | The per-asset event runtime behind `ref.events`. | | `instantiable.module` | `ref.instantiate` | The instantiation contract an asset type opts into: `root` stands the root entity, `place` applies the base placement opts to one the type adopted, and `result` shapes the `(root, idMap)` every `instantiate` returns. Also registers the `sceneInstantiable` field-constraint validator. | | `declaration.module` | `type.yaml` | The type's decoded declaration, read once and held until a write inside an `.assetType` folder; what every question about a type as a whole reads. | | `previewRecipe.module` | `type.yaml` `preview:` | How an instance's `preview.png` is produced: which of the four recipes (`instantiate`, `image`, `surface`, `custom`) the type declares, and what that recipe needs. | ## Reaching it ```luau local rt = require("@builtin::assetTypes.assetType.shared") rt.ref.loadTypeBehavior("material") ``` Fields resolve lazily. The prelude requires the sub-modules directly, in dependency order (`ref`, then `changeDispatch`, `onRegister.install()`, `refShapes.install()`) — a table that required all six on load would take that ordering away from it.