# sketchbook A small Zig 0.17 platform for sketching games and toys, in the spirit of Processing and pygame: write plain data and two functions, save, and the running program picks up the new code with its state intact. ```zig const sk = @import("sketch"); pub const State = struct { x: f32 = 480, y: f32 = 300, hue: f32 = 0 }; pub fn update(s: *State, in: *const sk.Input) void { if (in.pressed(sk.Key.space)) s.hue += 30; s.x += (in.mouse_x - s.x) * 0.1; s.y += (in.mouse_y - s.y) * 0.1; } pub fn draw(s: *const State, f: *sk.Frame) void { f.clear(0x202020); f.fill(.{ .x = s.x - 10, .y = s.y - 10, .w = 20, .h = 20 }, 0xff8040); f.print(8, 8, 2, 0xffffff, "hue {d}", .{s.hue}); } ``` ## Examples The standalone example packages live in [`examples/`](examples/): - [`noc`](examples/noc/): 21 Nature of Code sketches. Select one with `-Dsketch=flocking`. - [`shaders`](examples/shaders/): five shader examples in both Zig and GLSL. - [`vecanim`](examples/vecanim/): vector animation scenes. Select one with `-Dscene=scenes/live.zig`. - [`micscope`](examples/micscope/): records the microphone or opens sound files, plays them either way, and shows waveform, spectrogram and spectrum; a tuner. - [`spheres`](examples/spheres/): the three-body problem on the GPU, each body a voice synthesised by a compute shader. - [`live`](examples/live/): a tiny language for sound and picture, typed over its own picture: every line that parses is compiled to a tape one GPU interpreter runs per pixel and per sample. - [`talk.zig`](examples/talk.zig): a talk with a slider in it, a sketch that runs its own `anim.Deck` (`zig build test-talk` plays it). - [`manual.zig`](examples/manual.zig): Figure 2-1 of the Intel SDM comes alive, then ADD's opcode table is written row by row and its `80 /0 ib` turns into SUB's `80 /5 ib`: PDF pages as vector shapes and parts of them (`zig build manual -Dmanual=PDF`; see Animation, PDF pages and parts). Run `zig build run` inside a package, with `zig build lib --watch` in a second terminal for hot reload. `zig build test` runs its headless checks; `zig build web` builds the selected example for the browser. ## An app An app is an ordinary module. Every declaration is optional: | Declaration | Meaning | | --- | --- | | `name`, `width`, `height` | Window title and initial canvas | | `audio_pull` | Samples the sound device takes at once (see sound) | | `State` | The one thing a sketch keeps: what `update` changes, saved in one file with its history, replayed. No pointers, no host handles, and at most 32 MiB, half of a save (compile errors). | | `Session` | This process: survives reloads, never saved. Caches, image and buffer handles, playback cursors. | | `init(state, host)` | Seed brand-new state, after the declared defaults | | `setup(session, init) !void` | Initialize resources once per fresh Session, including startup with restored State | | `update(state, input)` | **Deterministic**: the host replays recorded input through it | | `draw(state, session, frame)` | Pictures and sound, as commands; may change Session | | `command(state, session, console, line) !bool` | Console lines the host does not know | | `commands` | Console completions | Functions take the parameters they name, matched by type and in any order: `draw(s: *const State, f: *sk.Frame)` and `draw(f: *sk.Frame, ss: *Session, s: *const State)` both work. `update` can only ask for `State` and `Input`; asking for more is a compile error, since a replay could not reproduce it. `setup` receives `*sk.Init` with `host` and the `session` allocator; it may also request Session or Host directly. It gets no State, Input, or Frame. Load assets and create/upload GPU buffers there, then use those resources in `draw`. `try init.host.writeBuffer(buffer, offset, bytes)` submits a copy immediately; the source bytes may be released when it returns. WebGPU uploads require four-byte-aligned offsets and sizes. Resource setup runs after Session defaults, once on startup even when State was restored from disk, and again if a reload changes Session's layout. Compatible reloads and ordinary state resets preserve the Session and do not rerun setup. The browser invokes it during its initial setup too. A hook may return `void` or `!void`; uncaught errors are logged and disable app callbacks, while the native host console and 9P controls remain available. Restart after fixing a failed setup. Use a Session error field and handle the error inside setup when the app needs to display an initialization failure. `zig build test-setup` checks this lifecycle. Memory comes by lifetime, one allocator each: | Allocator | Lifetime | Released | | --- | --- | --- | | `f.scratch` | this frame | all at once, after the frame is drawn | | `f.session` | this process | one allocation at a time; the rest when Session's layout changes, along with the images and buffers the session made | `command` gets the session allocator on its `Console`. `update` sees only State, so when a key should do something to Session, State carries a request: `seek: sk.Request(f32)`, sent with `s.seek.send(0.5)`. Session keeps an `sk.Handled`, and `draw` runs `if (ss.seek.take(s.seek)) |to| ...` once per request. A fresh Session counts earlier requests as handled, so nothing re-fires after a restart. `input.width`/`height` are the canvas, which a sketch sets with `f.canvas(w, h)` and the host scales to fit its window. `input.window_width`/`window_height` are the room it has (the window, in CSS pixels on the web): a sketch that sets its canvas to them, and lays out by the canvas, is responsive, at its own scale on a phone as on a desktop. `update` sees `input.events`: pointer, wheel, key, text, `cancel` (focus loss, console) and one `tick` per frame carrying `dt`. Pointer and wheel events carry the modifier keys held as they happened (`e.modifiers()`, `Key.Mod` bits): read those, not `input.held`, for ctrl+wheel and the like, since a trackpad's or a touch screen's pinch arrives as ctrl+wheel with no key events at all. `update` is called once per frame, with that frame's events ending in its tick, live and replayed alike; in `update`, `input.now` and `input.dt` come from that tick (`dt` read back from the tick's f32 seconds, so a replay sees the same bits). Interactive Wayland frames use the display-paced clock for update and draw automatically; unsupported backends use wall time, and the browser uses its animation-frame timestamp. `input.mouse_x`/`buttons`/`held` follow the events fed so far. So time-driven and pointer-driven sketches replay exactly. Fields no event carries (`width`, `window_width`, `console`, `frame`) are not replayed: a sketch that reads them in `update` can replay differently. `input.started` is true on the first update after the program starts, never after a reload: state is saved, so this is where to stop what should not resume on its own. In `draw`, `input.jumped` says State got where it is other than by `update` from what the last `draw` saw: the first frame, a seek, step, a mark's `load`, `set`, `reset`, play or record start, a sketch command, a reload. Anything `draw` piles up in Session (a trail in an offscreen image) is stale then; repaint it from State. `draw` gets a `Frame`: `clear`, `fill`, `shape`, `shapeColors`, `image`, `imageAt`, `target`, `blend`, `text`, `print`, `clip`, `canvas`, `cursor`, `voice`, `gain`, `audio`, `audioBuffer`, `microphone`, `pixels`, `gpu`, `compute`, `upload`, `download`, `openFile`, `run`, `quit`, plus the three allocators and `f.host` (`loadImage`, `decodeImage`, `newImage`, `newBuffer`, `loadSound`, `decodeSound`, `newSound`, `samples`, `readFile`, `appendFile`, `writeFile`, `env`, `log`, `textSize`). Commands may be built from locals: the methods copy what they keep (text, GPU descriptions, uniforms) into scratch. Bulk data (vertices, samples, pixels, uploads) is not copied and must outlive the frame: scratch or session memory. `f.text(x, y, scale, bytes, color)` draws one line from its top left corner; scale 1 is a 13 px monospace face (8 px a character on a 16 px line), and any scale works, 1.5 or 7.25. `f.host.textSize(scale, bytes)` is the width and line height it covers, for laying out tables and code. Natively the text is shaped by Pango and drawn at the canvas's own pixels, with advances rounded to whole pixels at that size: sharp at any size, but a width is not a multiple of the scale (ten digits are 80 px at scale 1, 230 at scale 3), so measure at the scale drawn. The font is Go Mono, built in; `SKETCHBOOK_FONT="JetBrains Mono"` in the environment puts an installed family before it. On the web it is the page's `monospace`, 8 px a character. `f.textIn(face, x, y, scale, bytes, color)` is the same line in another face of the Go family, all six bundled in the binary (`sk.Face`): `.mono`, `.mono_bold`, `.mono_italic` for data and code, `.sans`, `.sans_medium`, `.sans_bold` for what is said about it; and `.code`, `.code_bold`, `.code_italic`, `.code_medium` are JetBrains Mono NL (no ligatures), bundled too, for programs. `f.host.textSizeIn(face, scale, bytes)` measures a line. The seven monospaced faces advance alike (0.6 em: whole pixels at 20, 25, 30, 35, 45 and 60 px and at their doubles), so columns of text need no measuring; JetBrains Mono's line box is taller (ascent 1.02 em, where Go's is 0.945). The web host draws every face as its one. Text is shaped for the size it is drawn at, so under a camera that zooms its letters would shift against each other from frame to frame. `f.textScaled(face, x, y, scale, raster, bytes, color)` lays the line out and rasterizes it at `raster` and draws that at `scale`, unsnapped: give it the scale the move ends on (the larger of its ends) while the camera moves, and `textIn` again once it rests. One texture serves the whole move. The web host ignores `raster`. `f.openFile("open", .{ .pattern = "wav;flac" })` shows the system's file dialog; the answer comes back as the console line `open `, so it reaches the sketch's `command` like anything typed there. `f.run(line)` does the same with any line, after the frame: `draw` cannot write State, so this is how it tells State something only it learns (which of its files are already written, say). `f.host.appendFile(name, bytes)` appends to a file of the sketch's own, `.` beside its save `.xitdb` (in `.sketchbook/` under `zig build run`), creating it, and returns its path for `readFile` ("" when nothing was written, and always on the web). Only `draw` and `command` can call it, and a write is not replayed, like sound. `readFile` maps a file once, so what is appended afterwards is not in what it returned. `f.host.writeFile(name, bytes)` does the same to a new, empty file of that name. micscope starts each take's `take-.f32` so, and appends to it as it records. `f.download(buffer, .{ .size = n, .key = k })` reads a GPU buffer back without stalling: the bytes arrive in a later frame's `in.downloads`, a frame or two behind. For measurements a shader made (a pitch, a level), not for pixels; sound a shader made goes to the speakers with `f.audioBuffer` (see sound, below), which does not wait a frame. Geometry and colors are duck-typed. Anything with `x`/`y` is a point, anything that also has `w`/`h` (or `width`/`height`) is a rect, in any number type, and `@Vector`s or arrays work too. Colors are `r`/`g`/`b`[/`a`] or `0xRRGGBB`. The `Color`'s constants also go by name (`f.clear(.white)`); its functions need the type (`sk.Color.gray(220)`), since an `anytype` parameter gives a decl literal nothing to resolve against. The helpers return the caller's own type: - `sk.shape`: `isPoint`/`isRect`/`isColor`, `contains`, `inset`, `offset` - `sk.Camera`: `toScreen`, `toWorld`, `pan`, `zoomAt` - `sk.tween`: `lerp` over any plain data, `lerpFields(a, b, t, .{ .x, .zoom })`, `toward(a, .{ .x = 100 }, t)`, and `ease` The drawing layer below is the exception: new, and typed (`sk.math.Vec`, `sk.Color`), so decl literals and completion work. ## Drawing shapes Circles, lines and polygons are one host command, `Command.shape`: a region in one colour, given as points, subpath ends and triangle indices into the points. The web fills the subpaths (nonzero, one fill per outline and the holes after it); native draws the triangles with `SDL_RenderGeometry`. Both describe the same region, so translucent shapes blend alike, and shapes keep their order among `fill`, `image`, `text` and the clip. Both antialias: the canvas does, and native draws 4x4 samples a pixel (with a GPU renderer: runs of consecutive shapes go into a supersampled layer, composited when another command comes) and averages them; 2x2 on a canvas or target over 1920x1080, none over 3840x2160. Edges come out within a few levels of the web's, and a 1 px line on a pixel row stays full black on both. `sk.draw` makes them, at two levels, neither with state of its own: ```zig const ball: sk.draw.Pen = .{ .fill = .gray(127), .stroke = .black, .width = 2 }; pub fn draw(s: *const State, f: *sk.Frame) void { f.clear(0xffffff); sk.draw.circle(f, s.position, 24, ball); // radius, not diameter const pen = ball.transformed(.place(s.position, s.angle)); // push, translate, rotate sk.draw.line(f, .{ 0, 0 }, .{ 24, 0 }, pen); // round ends sk.draw.polygon(f, &.{ .{ 30, 0 }, .{ 26, -4 }, .{ 26, 4 } }, pen); } ``` A `Pen` is plain data (`fill: ?Color`, `stroke: ?Color`, `width`, `join`, `cap`, `xf: Affine`), so it can be a `const`, a State field, or a copy changed for one call. A null fill is `noFill()`, a null stroke `noStroke()`. `circle`, `line`, `polyline` (Processing's `beginShape()` / `endShape()`, open), `polygon` (`endShape(CLOSE)`), `rect`, `point` (a dot `width` across, Processing's `point`) and `dashed` (a polyline in dashes, `.{ on, off }` along the path, each dash whole round a corner; `Tessellation.strokeDashes` underneath) tessellate into `f.scratch` and emit a shape for the fill and one for the stroke; out of memory, they log and drop it, as `emit` does. Joins are `.miter` (bevelled past 4 half-widths), `.round` or `.bevel`, caps `.butt`, `.round` or `.square`; the defaults are Processing's (miter, round). ```zig // A translucent road with round joins: each pixel blended once, joins included. const road: sk.draw.Pen = .{ .stroke = sk.Color.black.alpha(55), .width = 40, .join = .round }; sk.draw.polyline(f, &s.path, road); ``` Underneath, an `sk.draw.Tessellation` is the shape as a value the caller owns: `fillCircle`, `strokeCircle`, `fillPolygon` (simple polygons, concave too), `strokePolygon`, `strokePolyline` (mitered, bevelled past 4 half-widths), `strokeLine` (round ends), `fillRect`, `strokeRect`, and `strokePath(t, gpa, points, .{ .width, .join, .cap, .closed, .miter_limit }, xf)`, which the other strokes are, append to it with an allocator and an `Affine`. A round join is the circular segment beyond the bevel, a round cap the half disc beyond the end, so they stay disjoint like the rest; a square cap is a butt cap half a width further out. Emit it with `f.shape(&t, color)` or as `.{ .shape = ... }` yourself, put many shapes in one (1000 boids in two commands; a quadtree's 700 outlines and 2000 points in one), or hand `t.triangles(gpa)` to a GPU pass. Shapes of many colours share one command through `f.shapeColors(&t, colors)`, a colour per subpath (`t.ends`): fading particles batch too. A stroked circle is two subpaths, its outline and its hole, which take the same colour. `appendPoint`, `endSubpath` and `appendTriangle` build shapes of your own. What the functions guarantee: triangles that do not overlap, wound clockwise on screen (mirroring transforms included), covering exactly what the subpaths do. One case cannot keep it: a stroke that overlaps itself (a self-crossing polyline, or segments shorter than the stroke is wide at an acute turn) blends twice natively where it overlaps and once on the web. Circles get as many segments as keep them within a quarter pixel of round, after the transform. `tools/shapes_web.mjs` (`zig build test-web`) draws the same frame in headless Chrome and compares it with the native screenshot, pixel for pixel away from edges. ## Offscreen images, blending, placed images Processing's canvas that is never cleared is an image you draw into. `host.newImage(w, h)` makes one, transparent; `f.target(img)` sends the 2D commands after it there (any of them: `sk.draw`, `fill`, `image`, `text`), `f.target(.none)` back to the canvas, and the image keeps what was drawn, frame to frame, until `f.image` shows it: ```zig pub const Session = struct { trail: sk.Image = .none, drawn: u64 = 0 }; pub fn draw(s: *const State, ss: *Session, f: *sk.Frame) void { if (ss.trail == .none or f.in.jumped) { if (ss.trail == .none) ss.trail = f.host.newImage(width, height); ss.drawn = 0; f.target(ss.trail); f.clear(.white); } else f.target(ss.trail); while (ss.drawn < s.steps) : (ss.drawn += 1) sk.draw.circle(f, at(ss.drawn), 24, pen); f.target(.none); f.image(ss.trail, .{ .x = 0, .y = 0, .w = width, .h = height }, .{}); } ``` The image is a cache in Session, never truth: after a seek or a load (`in.jumped`) it is repainted from State, so what a seek shows is what playing showed. Natively a target is supersampled whole and holds premultiplied colour, so it looks the same whether its steps were painted one a frame or all at once, and translucent edges are not darkened when it is drawn. A frame starts on the canvas. `f.blend(.add)` adds what follows (colour times alpha, and alpha: canvas `lighter`; natively a custom SDL blend, since SDL's `ADD` keeps a transparent target's alpha), `f.blend(.alpha)` goes back. It applies to `fill`, shapes and images; `clear` and `text` ignore it. Each frame starts with `.alpha`. There is no `multiply`: SDL's differs from the canvas's for translucent colours, and images have no fixed-function equivalent. `f.imageAt(img, xf, .{})` is `imageMode(CENTER)` under a transform: the image centred on `xf`'s origin, along its axes (`.translate(x, y)`, `.place(at, radians)`), at its own size or `.size`. Underneath, `Command.image.corners` places an image anywhere an affine can. A tint multiplies colour per channel and alpha as a whole, alike on both hosts (the web multiplies the pixels itself, cached per tint, as canvas compositing cannot where an image is translucent). Natively images sample nearest; the web antialiases a turned image's edges. ## Math, time and lists `sk.math` is the geometry the `vector` module uses too (`vector.geom`): `Vec = @Vector(2, f32)`, so `+`, `-`, `*` and `+=` work on points as on a shader's `vec2`; free functions; `Affine` transforms; cubic Béziers; and noise. Processing's PVector, by name: | PVector | `sk.math` | | --- | --- | | `add`, `sub`, `mult` by a vector | `a + b`, `a - b`, `a * b` | | `mult(k)`, `div(k)` | `scale(a, k)`, `scale(a, 1 / k)` | | `mag()` | `length(a)` | | `normalize()` | `normalize(a)` (zero stays zero) | | `setMag(m)` | `withLength(a, m)` | | `limit(m)` | `clampLength(a, m)` | | `heading()` | `angle(a)` (radians; not in shader code, `atan2` uses tables) | | `PVector.fromAngle(r)` | `unit(r)` | | `dist(b)` | `distance(a, b)` | | `dot(b)`, `cross(b)` | `dot(a, b)`, `cross(a, b)` (the z of the 3D cross) | | `lerp(b, t)` | `mix(a, b, t)` | | `translate(p); rotate(r)` | `Affine.place(p, r)`, or `.translate(x, y).mul(.rotate(r))` | | `random(lo, hi)` | `uniform(rng, lo, hi)` | | `PVector.random2D()` | `randomUnit(rng)` | `sk.math.noise(seed, x, y, z)` is smooth noise in 0..1 like Processing's `noise()` (four octaves, each half the last): gradient noise on a hashed lattice, no table, so the seed can live in State. Unused axes are 0. Its spread is Processing's (a deviation of about 0.11, nearly all of it in 0.1..0.9, so `noise → angle` turns all the way round), and integer inputs are not special: each octave is offset off the lattice. It gives the same bits natively and on wasm32 (a golden-value test runs on both). For random numbers keep a concrete `std.Random.Xoshiro256` in State (not `DefaultPrng`, an alias that may change) and pass `s.rng.random()`; `uniform(rng, lo, hi)` is `std.math.lerp(lo, hi, rng.float(f32))` by another name, and `std.Random` has the rest (`floatNorm`, `uintLessThan`). `sk.Clock` steps a simulation at a fixed rate from the recorded ticks: `clock: sk.Clock = .{ .hz = 60 }` in State, `for (0..s.clock.steps(in)) |_| step(s);` in `update`. It owes at most 8 steps per tick and drops the rest after a stall; `count` is the steps taken; `set clock.hz 30` is slow motion. A frame of N ticks steps exactly as N frames of one. It counts integer nanoseconds and allows an eighth of a step either way, so at the display's rate it steps once a frame, jitter and all: a 60 Hz frame is 16 666 666 ns, a 60 Hz step 16 666 667, and 60 frames are 60 steps. `sk.List(T, capacity)` is a list for State: `buffer` and `len` (public, so `boids.buffer[3].pos[0]` works in the console), `slice`, `constSlice`, `appendBounded` (`error.OutOfMemory` when full), `orderedRemove`, `swapRemove`. Every method clamps `len` to the capacity, which matters after a reload that shrinks it. The default buffer is the element's own defaults, never `std.mem.zeroes` (an enum may have no 0). **Data that grows.** State is kept in keyframes, replayed and migrated, so it holds what `update` changes, at a size fixed in its type, 32 MiB at most. Three cases, by who needs the data: - `update` needs it: a `List` in State, sized for the worst case (the quadtree's 4096 points, 32 KiB; flocking's 2048 boids). - Only `draw` needs it, and it follows from a few State fields: keep those fields in State and the data in Session, rebuilt when they change. A Koch curve's 65536 segments are a function of its generation, one byte; an L-system's 89 000-character sentence of its generation too. `sk.Derived(T)` is that: `if (ss.curve.stale(s.generation, f.in)) rebuild(...)`; it is stale when the key changes (a click, a seek, a step) and on the first frame after a reload, since new code may derive it differently. A cheap derivation needs no cache at all: the quadtree builds its tree in `f.scratch` every frame. - It is the user's work, not a function of State (a recording, a document): a `List` in State, or, bigger than a save, files of the sketch's own (`appendFile`), named from State, as micscope's takes are. The timeline then needs nothing more: a seek changes the key like an update does (`examples/noc`: `koch`, `lsystem`, `quadtree`). **Sound scopes**, in the `scope` module (import `sb.modules.scope`): a `scope.Scope` kept in Session takes samples with `push` and, once a frame, `draw`s the views given rects for: the last seconds as a scrolling waveform and spectrogram, the smoothed spectrum of the newest samples, and an oscilloscope. All analysed and drawn on the GPU, new columns only; its shaders swap to WGSL for the web like any others. Given `.camera`, the rects are world rects: the views are made at their size there and drawn through the camera, so panning or zooming recomputes nothing. A sketch that analyses buffers of its own runs the same shaders: `scope.kernels.{waveform, spectrogram, spectrum}` over any read-only buffer of samples, `scope.pipelines.{wave, spectrogram, spectrum}` to draw them, with the parameter and view blocks in `scope.layouts` (`Scope.gpu.tape` is a Scope's own history on the GPU, for a kernel of the sketch's). `scope.midi(hz, a4)`, `scope.hz(note, a4)` and `scope.note_names` name pitches. Lifted from micscope's live view; spheres uses it. **Vector graphics**, in the `vector` module (import `sb.modules.vector`): SVG parsing (the static subset Typst emits, and more) into cubic Bézier data; flattening, stroking, dashing, trimming, clipping and morphing as geometry; and `vector.Renderer(.gpu)` or `vector.Renderer(.cpu)`. The CPU backend computes exact coverage and runs anywhere, wasm included. The GPU backend describes one pass (8× MSAA, gradients in a shader): convex fills drawn directly, the rest stencil-then-cover, all batched into layers of non-overlapping fills, a few draws per layer whatever the shape count. Both are tested against resvg (`zig build diff`, below). **GPU work is data.** A sketch describes pipelines (shader code, vertex layout, blend, stencil) and passes (`f.gpu(.{ .image, .pipelines, .draws, ... })`); the native host runs them with SDL_GPU, the web host with WebGPU. Both cache shaders and pipelines by content hash, so an edited shader is just a new pipeline after a reload. No sketch touches a graphics API. The shaders are GLSL; the build turns them into SPIR-V for native and WGSL for the web, and each build imports the form its host reads (see Build API). `abi.gpu.pipeline(Vertex, ...)` derives the vertex layout from a struct, and `abi.gpu.uniform` checks std140 at compile time, and the size: a uniform block is 4096 bytes at most, all SDL_GPU's Vulkan backend binds of one (a larger one was cut off without a word), so more goes in a storage buffer. A pipeline without attributes makes its own vertices from `gl_VertexIndex` (a full-screen triangle, for shader-toy style views). Compute is data too. `host.newBuffer(size)` makes a GPU storage buffer, `f.upload(buffer, offset, bytes)` fills it, `f.compute(.{ .kernel, .readonly, .readwrite, .uniforms, .groups })` runs a compute shader over buffers, and a draw reads buffers through `.vertex_storage` / `.fragment_storage`. Results stay on the GPU: a compute shader's output is a later draw's input, with no readback. Each dispatch is its own compute pass, so it sees everything the commands before it wrote. Sound is declarative. `f.voice(.{ .key, .audio, .serial, .loop, .volume, .fade, .rate, .group })` declares a voice that should be sounding. A new key or serial starts it, and an undeclared key stops it. A finished one-shot stays finished while it is still declared, so replays and seeks never re-fire sounds. `.rate` is its speed and so its pitch (2 is an octave up and half as long; resampled by linear interpolation, cheap and a little dull far from 1); changing it does not restart the voice. `.group` (0-7) puts it under a gain the sketch declares each frame, `f.gain(group, value)`, 1 when not declared: music ducks under speech by the sketch lowering one number, and the host glides to a new gain over 10 ms so a step does not click. The host mixes the voices once a frame, topping its queue up to 2048 samples (about 43 ms) ahead of the device, so what `draw` declares is heard up to that much later. `f.audio(key, samples, .{})` streams PCM, and `input.stream(key)` reports what the host has queued and played, and its `underruns` (pushes that found it empty though the frame before had pushed to it: each a gap). The sketch chooses how far ahead it keeps a stream; `input.lead` is how far the host would: the slowest of the last 64 frames' worth of samples (a stream is fed once a frame), plus the device's pull (what it takes from a stream at once, 1024 samples by SDL's default), plus half a pull for each recent underrun. A sketch that makes its sound as it goes pushes `in.lead -| in.stream(key).queued` a frame, and a change it makes is heard after what is queued ahead of it, about a pull: 19 to 27 ms after `draw` on this machine at 60 Hz. A sketch that must answer faster asks for a smaller pull: `pub const audio_pull = 256;` beside its `name` has the host ask SDL for 256-sample pulls (5.3 ms) when it opens the device, and a change then leaves in about 6.5 ms. The price is paid by the whole machine: the audio server then runs every program's sound at that size, four times as many wake-ups, and under load it misses one now and then, a dropout (a click) in everything playing, not only the sketch. Without it SDL picks (1024 at 48 kHz). `SDL_AUDIO_DEVICE_SAMPLE_FRAMES` in the environment wins over both, the web ignores it, it is read once, when the host starts, and the sound overlay (`overlay sound`) shows what the device took (`pull 256`). `f.audioBuffer(key, buffer, .{ .count = n })` streams `n` samples a compute shader wrote (f32 from `.offset` bytes): the host reads them back within the frame, waiting for the GPU's work so far (about 0.2 ms), and queues them. A GPU synth is then: ```zig const n: u32 = @min(f.in.lead -| f.in.stream(1).queued, max_block); if (n > 0) { f.compute(.{ .kernel = synth, .readwrite = &.{ss.block}, .uniforms = ..., .groups = .{ (n + 63) / 64, 1, 1 } }); f.audioBuffer(1, ss.block, .{ .count = n }); } ``` (`examples/spheres` is one.) `f.microphone()` keeps the default recording device open while declared, and the next frame's `input.microphone` holds what it heard (to `draw` only: it is a device, not an event, so it does not replay). `host.samples(audio)` gives a loaded sound's samples, mono at `input.sample_rate`. `--export` writes frames only, no sound. The host sums the voices and limits the sum, sample by sample: below 0.75 (-2.5 dBFS) it plays at true volume, to the bit; above, it bends smoothly (same value and slope at 0.75) toward 1 and never past it, so a full-scale peak plays at 0.875 and any number of loud voices stays bounded. Streams are not limited: SDL adds each, as pushed, to the voice mix and clips the total at ±1. A sketch can make its own sounds: `host.newSound(samples)` copies mono samples at `input.sample_rate` into the host and returns an `Audio` that voices play like a loaded one, so a synthesised one-shot is replay-safe too. Like every sound it lives as long as the host: make each once and keep the handle in Session. `sk.sfxr` is sfxr, the classic generator of game blips, coins, lasers and explosions (`sfxr.Params`, its usual knobs, rendered at any rate; at 44.1 kHz to the bit as the original): ```zig if (ss.blip == .none) { var buf: [4096]f32 = undefined; const n = sk.sfxr.render(.{ .freq = 0.5, .sustain = 0.05, .decay = 0.1 }, f.in.sample_rate, &buf); ss.blip = f.host.newSound(buf[0..n]); } if (s.coins != 0) f.voice(.{ .key = 1, .audio = ss.blip, .serial = s.coins }); ``` (`test/voices.zig` is this sketch.) Sounds decode with single-file C libraries in `vendor/` (dr_wav, dr_flac, stb_vorbis): WAV, FLAC and Ogg Vorbis, chosen by content. ## Animation The `anim` module (`sb.modules.anim`) is manim-style animation written as code: Typst for text and math, shapes, cameras, morphs, talks that wait for clicks, and objects the scene draws itself. A scene is one file, and `sb.addScene` builds the player around it: ```zig const std = @import("std"); const anim = @import("anim"); pub const typst_preamble = "#set text(fill: white)"; pub const scene = anim.Scene(construct); fn construct(s: *anim.Builder) void { const eq = s.typst("$ a^2 + b^2 = c^2 $", .{ .y = 220, .scale = 1.4 }); const eq2 = s.typst("$ c = sqrt(a^2 + b^2) $", .{ .y = 220, .scale = 1.4 }); const square = s.rect(.{ .w = 180, .h = 180, .x = -320, .stroke = anim.blue }); s.play(.{eq.write()}, .{}); s.play(.{ square.create(), square.rotate(std.math.pi / 4.0) }, .{ .seconds = 1.5 }); s.play(.{eq.transformInto(eq2)}, .{}); s.wait(1); // Updaters run every frame, on top of the timeline. Their extra // arguments are bound here, like a closure's captures. const swing = s.value(0); s.always(square, wobble, .{swing}); s.play(.{ swing.to(0.3), anim.camera(.{ .zoom = 1.4 }) }, .{ .ease = anim.ease.outBack }); } fn wobble(now: anim.Now, swing: anim.Value) anim.Pose { var p = now.pose; p.rotation += now.value(swing) * @sin(now.t * 4); return p; } ``` Objects are added (`typst`, `svg`, `rect`, `circle`, `line`, `arrow`, `polygon`, `custom`, and `part` of another) and show in the order they were added; each call returns a handle. Animations are methods of the handle: `write`, `create`, `fadeIn`, `fadeOut`, `transformInto`, `move(.{ .x, .y, .scale, .rotation, .opacity })`, `scaleTo`, `rotate`, `indicate`; plus `anim.camera(...)` and `value.to(x)`. `play` runs a tuple (or array) of them together; its options are `.seconds`, `.ease` (any `fn (f32) f32`) and `.lag`, manim's lag ratio (0 by default, all at once): each animation starts `lag` of its own length after the one before, in the order given, and the play still lasts `.seconds`, the last ending with it, so each of n lasts `seconds / (1 + lag * (n - 1))`; a lag of 1 plays them one after another. An object two animations of one play change is timed by the first. `wait` moves time on; in a live scene, `pause()` makes what follows wait for a click (see Talks). Positions are scene units: 1280x720 of them show at camera zoom 1, the origin in the middle, y down. The page behind them is the scene file's `pub const background = 0xRRGGBB` (manim's grey without one). A scene is one of two kinds, which share every call: - **Compiled** (`anim.Scene(construct)`): `construct` runs at comptime, and what it recorded is played once, still at comptime, into keyframes per object: a `from` and a `to` pose (plain structs), a time span and an easing. Drawing an instant is a lookup and a `tween.lerp`, then the object's updaters. The whole timeline is data before the first frame, a morph can look ahead to where its target ends, a broken scene is a compile error, and Typst strings are written inline. - **Live** (`anim.Live(frame, typst)`): `frame(s: *anim.Frame)` runs every frame, at runtime. Each `play` takes effect at once against `s.t`, so the code after it reads and changes what shows, with no callbacks: `s.pose(obj).y = ...`, `s.get(value)`, `s.camera()`, the mouse (`s.pointer`), the time (`s.t`, which runs on while a pause waits) or any runtime data. Its Typst sources are the declarations of the namespace it is given (or a tuple of namespaces), since the build cannot run it to find them. Both answer the player the same way (`shot`: objects, poses, values, camera, plays and gates at an instant), so the only state is the time and the pauses clicked through, and scrubbing, seeking and hot reload are free. A test checks that a live scene and a compiled one with the same timeline show the same poses at every instant. **Typst.** `s.typst(source, pose)` takes Typst markup; the scene's `typst_preamble` goes in front of each (by default a transparent page that fits its content). The build runs a generator that imports the scene file, reads `scene.typst_sources`, compiles each with `typst compile` (cached by content in `.zig-cache/typst`) and embeds the SVGs. Nothing scans files, and a scene without Typst needs no `typst`. **PDF pages and parts.** Typst imports a page of a PDF as vector paths, so a page is a Typst object like any other: `s.typst("#image(\"/home/me/sdm.pdf\", page: 31)", .{})`. Typst runs with `/` as its root, so an absolute path names the file where it lies; the file is not part of the cache key, so one changed in place needs its entry in `.zig-cache/typst` gone. Text is outlines: every glyph is a filled shape (nothing to search or select), drawn crisp at any camera zoom, and `write` reveals the shapes in the order the PDF draws them (in the Intel SDM: header and footer, the text from the top, then the figures; a table's text row by row, left to right, and its rules after all of it). A page costs what its shapes do: Figure 2-1's page of the SDM is 2.1 MB of SVG and 1954 shapes, parsed once, 6-12 ms, the first frame it shows; then, at 1280x720 in a release build, 0.9-1.7 ms a frame to record for the GPU (0.4 ms on the GPU itself) or 2-5 ms on the CPU renderer. A table of contents, dot leaders and all, is five times that: show a part of it. `zig build bench` times the page beside the simple objects, with a budget of its own. An SVG a Typst object includes (`image("x.svg")`) is kept the same way. `s.part(page, .{ .x, .y, .w, .h })` takes the shapes whose bounds' centre lies in a rectangle out of an object, as an object of its own, so a field of a figure or a row of a table can glow, be written, fly off or morph while the page around it stays. The rectangle is in the object's units, its centre and size: the literal `s.rect` takes to cover the same place when the object stands at the origin at scale 1. Pick it with the guides (G: lay the page at the origin, at scale 1, and read the pointer), or from `pdftotext -bbox -f N -l N file.pdf -`, whose points from the top-left of a W x H pt page are `(x - W/2) * 4/3` across and `(y - H/2) * 4/3` down. The part stands where its shapes stood, around the rectangle's centre (so `indicate`, `scaleTo` and `rotate` turn it in place), and shows as the object does after the plays laid so far: taken from a page that has faded in, it is there. The object draws the rest from then on, at every instant. A shape goes to the first part that takes it, so a rule that crosses several fields (a box's outline) stays with whichever holds its centre: keep areas inside the rules. A table's column rules run its whole height, so their centres all lie in one row: take them first, with a band a point high through that centre. Parts of parts, and of any vector object, work the same way, and share the object's geometry (a page is parsed once however many parts it gives). `transformInto` morphs into the whole of its target, so a morph's result can be split by parts taken after it: `c0.transformInto(bits)`, then `s.part(bits, ...)` for each run of bits. `examples/manual.zig` is a talk in two slides on the SDM (Vol. 2). On Figure 2-1 (page 31), the fields of the instruction format glow in turn, `83 C0 01` is written below, each byte flies into its field and `C0` splits into `11 000 000` over Mod, Reg and R/M. Then ADD's opcode table (page 132) is written row by row; its row `80 /0 ib ADD r/m8, imm8`, cut in pieces, becomes SUB's from page 1400, laid so that the rows coincide: first the digit after the slash alone, then the rest, a morph per piece, so the glyphs that stay hold still; the two rows end one over the other. `zig build manual` plays it in a window; the PDF is `-Dmanual=PATH` (`~/01-resources/intel-devref-ii.pdf` by default), never part of the repository: without it the example builds a scene that says so, and its script (`test/manual.script`, in `zig build test` when the file is there) is skipped. **Custom objects** are drawn by the scene's own code with 2D commands (`sk.draw`, `f.fill`, `f.text`): what is runtime data, a register table, a memory grid, a code panel. `s.custom(&data, pose)` adds one, where `data` has `pub fn draw(self: *const T, f: *sk.Frame, at: anim.Placed) void`. It is an object like the others: the same camera, pose, `play`, `fadeIn`, `move` and place in the order. ```zig const Bar = struct { level: f32, pub fn draw(b: *const Bar, f: *sk.Frame, at: anim.Placed) void { // In the object's own units, around its origin. sk.draw.rect(f, .{ .x = -100, .y = -20, .w = 200, .h = 40 }, at.pen(.{ .stroke = .white, .width = 2 })); sk.draw.rect(f, .{ .x = -100, .y = -20, .w = 200 * b.level, .h = 40 }, at.pen(.{ .fill = .hex(0xf0ac5f) })); at.text(f, -100, 26, 1.5, "rax", .white); } }; fn frame(s: *anim.Frame) void { const bar = s.new(Bar{ .level = 0 }); // lives for this frame const obj = s.custom(bar, .{ .x = -200 }); const level = s.value(0); s.play(.{ obj.fadeIn(), level.to(1) }, .{}); s.play(.{obj.move(.{ .x = 200, .scale = 1.5 })}, .{}); bar.level = s.get(level); // after the plays: as it stands now } ``` `at` (`anim.Placed`) is where the object stands: `xf`, its units to canvas pixels (camera, then pose); `scale`, pixels per unit; `pose` (`opacity`; `drawn`, 0 to 1 while a `write` or `create` plays, to reveal by; `highlight`); `stage`, scene units to pixels; `t`; and `value(v)`, every `Value` now, which is how a compiled scene's custom object (its data a constant) moves. `at.pen(pen)` is a `sk.draw.Pen` in the object's units at its opacity, `at.text` and `at.textSize` are `f.text` and `host.textSize` likewise. Text scales and moves with the object but does not turn. A custom object has no outline to `transformInto`. Vector objects are drawn on the GPU into an image, then the image onto the canvas; a custom object draws onto the canvas directly. So the picture is built in runs: the vector objects up to a custom one, that object's 2D commands, the vector objects after it, and so on, each run its own image (8x multisampled, at the canvas's size: about 140 MiB of GPU memory at 1920x1080 and 560 at 3840x2160, so add custom objects next to each other in the order where it does not matter). Within its `draw`, a custom object's commands keep their own order. **The player** (`src/anim/player.zig`) fits the scene's frame into the canvas, and its canvas is the room it has: the window, or an export's `--size`. Its State is an `anim.Deck` (`deck.t`, `deck.playing`, ...) and the viewer's settings. Keys: click, space or page down for the next step of a talk (space plays and pauses a scene with no pause), page up or backspace back a step, `]`/`[` the next/previous slide, P play/pause, left/right previous/next play, `,`/`.` step a frame, home/end, up/down speed, B GPU or CPU renderer, L loop (a talk holds its end instead), G guides (boxes, origins and the scene's axes, in scene units); drag and wheel to look around, R to reset the view. Exported, it plays through once without controls, each pause held for the scene file's `pub const hold` seconds (1 without one), and quits at the end: ```sh zig build export -- frames --size 3840x2160 zig build export -- - --size 1920x1080 --fps 60 | ffmpeg -f rawvideo -pix_fmt rgba -s 1920x1080 -r 60 -i - -pix_fmt yuv420p scene.mp4 ``` A package of scenes is a few lines of `build.zig`: ```zig const sb = sketchbook.init(b, .{ .target = target, .optimize = optimize }); const scene = sb.addScene(.{ .root_source_file = b.path("scenes/intro.zig") }); _ = scene.app.addSteps(.{ .scripts = &.{b.path("test/intro.script")} }); ``` `scene.app` is an app like any other (`run` with hot reload: save the scene and the animation carries on from the same instant; `lib`, `web`, `export`, `e2e`), named after the scene file; `.imports` are modules for the scene file besides `anim` and `sketch`. `test/scene_live.zig` and `test/scene_compiled.zig` are small scenes of each kind, with a custom object. ### Talks A talk is a live scene that waits: run from the top every frame, it is a function of its time and of the gates opened so far. Time never stops at a gate. What was laid before it (plays still running, a `loop`, code reading `s.t`) carries on, and what comes after waits in the future until the gate opens. So a talk seeks, steps back and hot-reloads like any live scene. - `s.pause()` waits for the presenter's next step (a click, space, page down). A click while a step plays finishes it, like Keynote: the picture jumps to where the next pause is laid; the next click opens it. Back cuts to the picture held before the last click. - `s.gate(open)` waits for anything: it opens on the first frame at or after it is reached on which `open` is not null, and returns that value from then on (null before), a choice to branch on: `const pick = s.gate(app.answer) orelse return;`. Code after a closed gate would show nothing, so `orelse return` is the usual end. A gate opens only in `update`, from what `update` sees (events and State), so a recording replays to the same talk. One click opens one gate. - `s.beside(f, args)` runs `f(s, args...)`, plays and all, from the cursor, and leaves the cursor where it was: what follows runs alongside (`s.beside(anim.Frame.clear, .{.{}})` fades a slide out while the next one writes). `s.loop(f, args)` plays `f` again and again from the cursor until the gate laid right after it opens; the cycle running then completes. `f` only plays, and ends where it began. `s.clear(options)` fades out everything that shows. - `s.slide()` begins a slide. Next/previous slide goes to a slide's resting picture: held at its first gate (the end of the slide if it has none), every gate before it opened at once (a choice with 0) or undone. The playhead is `anim.Deck`, plain data (`t`, `playing`, `speed`, `hold`, and `log`, each gate opened: how long after it was reached, and its value), so hot reload, snapshots, replay, scripts and the console see it like any field; `deck.waiting`, `deck.slide`, `deck.gates` and `deck.duration` say where the talk stands. The stock player keeps one, and any sketch can keep its own beside its data: its `update` decides which events ask for what (`anim.Ask`: `next`, `back`, `next_slide`, `back_slide`, `to`, `end`), so the deck never owns the keyboard. The talk takes the sketch's data after the frame: in `update` as `*App`, so the talk may change it, and when drawn as `*const App`, so drawing cannot. `anim.mutable(app)` is the app to write to, null when drawn; `s.events` are this update's events in scene units (none when drawn); a widget laid out before the pause it waits at takes a click by setting `s.next = false`. `anim.Painter` draws a shot as the player does: ```zig pub const typst = struct { pub const law = "$ x'' + 2 zeta x' + x = 0 $"; }; pub const scene = struct { pub const typst_sources = anim.sources(.{typst}); }; // for sb.addTypst pub const State = struct { deck: anim.Deck = .{}, app: App = .{} }; pub const Session = struct { painter: anim.Painter = .{} }; const App = struct { presses: u32 = 0 }; const button: sk.Rect = .{ .x = -30, .y = 150, .w = 60, .h = 60 }; // scene units pub fn update(s: *State, in: *const sk.Input) void { s.deck.update(in, .{ .next = in.mouseReleased(.left) != null, .back = in.pressed(sk.Key.left) }, talk, .{&s.app}); } pub fn draw(s: *const State, ss: *Session, f: *sk.Frame) void { const shot = s.deck.shot(f.scratch, f.in, .{ 0, 0 }, talk, .{&s.app}) catch return; ss.painter.paint(f, shot, @import("typst"), .{}) catch {}; } fn talk(s: *anim.Frame, app: anytype) void { const law = s.typst(typst.law, .{}); s.play(.{law.write()}, .{}); s.loop(pulse, .{law}); // runs on while the deck waits s.pause(); const knob = s.circle(.{ .r = 30, .y = 180, .fill = anim.blue }); s.play(.{knob.fadeIn()}, .{}); if (anim.mutable(app)) |a| for (s.events) |e| if (e.kind == .mouse_up and button.contains(e.x, e.y)) { a.presses += 1; // the knob's click, not the presenter's next s.next = false; }; const many = s.gate(if (s.next) @intFromBool(app.presses > 2) else null) orelse return; // a choice s.play(.{if (many == 1) knob.scaleTo(2) else knob.fadeOut()}, .{}); } fn pulse(s: *anim.Frame, o: anim.Obj) void { s.play(.{o.indicate()}, .{ .seconds = 1.5 }); } ``` `sb.addTypst(module)` compiles the Typst of a module that declares `scene.typst_sources` (a module of the sketch's file without the `typst` import, see `build.zig`'s talk). Gates are known by the order they are reached: insert a pause above where a talk stands and it lands a step earlier. An edit that moves the last opened gate (a step above retimed) moves `t` with it on reload, so the presenter stays on the same step, as far into it. Exported (`in.exporting`), every gate opens `deck.hold` seconds after it is reached, in the first frame's run, so the video shows each wait for that long, loops running, and the frames are the same every run; a sketch quits at `deck.t >= shot.duration`. `examples/talk.zig` is a talk with a slider, a loop, a choice, two slides and `beside`. ## The host - **Hot reload.** It watches the library file and loads each new build through a symlink with a fresh name (a build must replace the file, as `zig build` does, not rewrite it in place). `zig build lib --watch -fincremental --debounce 0` makes a reload take tens of milliseconds instead of a second: Zig recompiles only what the edit changed, and starts as soon as a file is saved instead of 50 ms later. Its limits, with Zig 0.17.1: the incremental linker is new, and a library of it that does not link libc loads only through a host workaround that Zig 0.18 makes unnecessary; if a reload ever runs code other than what you wrote, stop the watch and rebuild without `-fincremental`. The library is larger (27 MB instead of 10 for a small sketch), the watching compilers stay in memory (about 300 MB for a small sketch, 1.3 GB for a game with C and shaders), and a save that touches several files may build more than once. The cache grows with each start of the watch (65 MB for a game), not with edits. - **The save.** A sketch keeps one thing, State, in one file, its save: `.xitdb`, an xitdb database. Live State is process memory; the save holds the present, its newest keyframe and the frames logged after it, which a run restores and replays through `update` (played by another build, the keyframe alone, and the run says how many frames it left), so a run starts where the last stopped; and the history before it (Sessions, below). `--fresh` starts from defaults and leaves the save as it is. A keyframe is State in 4 KiB pages and writes only the pages that changed since the keyframe before it, an all-zero page nothing: about 5 KB a keyframe for drat-meout's 20 KiB State, 10 KB for a 4 MiB State edited once a minute. Each stretch carries the reflected layout, so a keyframe from an older layout migrates field by field, with enums matched by tag name, and so does live State at a reload. Arrays and vectors migrate element by element: the elements both layouts have, each field by field, so an element type can gain a field and a capacity can grow (new elements take their defaults) or shrink (the rest is dropped; a count field such as `sk.List`'s `len` is kept, and `List` clamps it where it is read). An array whose type is unchanged is copied whole. A state file from before (`.state` beside the save, any of its formats) is imported into a new save once: its State the present, its snapshot slots marks `1` to `7`; the file is left as it was and not read again, and one that is not a state file is said and left. - **Where files go.** `zig build run` and `export` keep a sketch's files in `.sketchbook/`: `.xitdb` and the `.` it appends to. The host makes the directory with a `.gitignore` of `*`, so neither git nor jj takes in a save that grows (the jj snapshot a build makes would commit one while it is under a megabyte), and moves in, once, the `.state` an older build kept in the project's root; when both exist, nothing moves and the host says which is used. Without `--state`, a run keeps them in `$XDG_DATA_HOME//` (`~/.local/share//` without it), which is where a shipped game's go; a script or an export uses `sketch.xitdb` in the working directory. `--state x.xitdb` and `--state x.state` both name the save `x.xitdb`. - **Sessions.** Every run is a session in the save: an append-only log of the events live `update` was fed, in whole frames, and keyframes, the State at the start, every minute while it changes, at every jump (`load`, `set`, `reset`, a seek, a sketch command), before and after a reload, at every mark and at exit. Each stretch of a session between reloads keeps its build's identity, its layout and the source it was built from: the repository's revision (`jj: `, else `git:`, `+dirty` with uncommitted changes, read once per `zig build`), Zig's version, the backend and the mode. Nothing in it is ever overwritten, only dropped, oldest first, at the cap (below); marks and clips are names pointing into it. `App.Options.record` decides how much of it the save keeps, at build time: null, the default, means Debug builds keep all of it, up to the cap below. A build that does not record keeps the present and the marks (and clips) only: at every keyframe it drops the frames and keyframes before it, and at exit it compacts the file, so while it runs the disk holds at most the last minute of input, and after, none. A playtest build of a release opts in with one line, `.record = b.option(bool, "record", "collect sessions in a release build")`. A script from `--fresh` state records in memory (time travel works, nothing touches the disk; it stops at 256 MiB, ten hours or more of play, keeping what it has), a script without `--fresh` to the file, an export nothing, and an export has no marks or clips, as the console says. The console's status line ends with `rec file`, `rec memory` or `rec off`, and `sessions` says where. A commit is a second of play (of event time, so a script commits on the same frames every run), and is made at once for a reload, a mark, a clip's start or end, an export and exit; only a mark and exit are synced to disk. A crash (a panic, `kill -9`) loses the frames not yet committed, under a second (15 of 255 in the kill test): the next run starts from the last committed frame's State, the session lists as `(crashed)`, and a clip it left open ends where the session does. A power loss loses what the kernel had not written back (30 s past the last sync, by default); the torn end is mended at the next start, the file opened at its last whole commit, and a file still unreadable is renamed to `.broken.`, never deleted, and a new one begun. A minute of drat-meout's play was 268 KB with the pointer moving every frame (16 MB an hour), 164 KB idle (86 KB of ticks, the rest keyframes, then whole States; a keyframe is about 5 KB now). **The cap.** A save is at most 64 MiB, and State half of that, so history has room. A commit that leaves the file past it drops history oldest first, keyframes' pages and the frames logged with them, and compacts the file to three quarters of the cap: a hitch of about 170 ms for a 48 MiB file in a Debug build, once every 16 MiB of history. The present is never dropped, nor the keyframe of the newest mark of each name, nor the newest clip of each name and its frames. What was dropped says so when it is asked for (`load #K`, `play session N`). A mark of a new name needs room for a whole State: when marks fill the save it is refused, saying how to make room, by saving over a mark's name (the older one then goes with history) or starting another save (`--state`). Every commit leaves the nodes it copied behind, about a kilobyte, which the compaction removes too. Playtests: `export-session [S] FILE` writes session S (this run's when none is given) with its marks and clips to a database of its own, for a player to send; `import FILE` adds the sessions in such a file to this sketch's (`sessions` lists them as `import `), its marks and clips named `NAME@`; `play session N` replays one, and its stretch names the commit to build to replay it without divergence. What such a file holds of the player: every input event (pointer positions, keys, the text typed into the sketch), the console lines that changed State (an `f.openFile` answer is a local path), and the date the session started; nothing about the machine or the user. - **Time travel.** Marks, clips, replay, seek and step work on any sketch, because they only need state bytes and events. `save NAME` marks the State now: a keyframe with a name, the newest of a name winning. `load NAME` (or `load #K`, the K-th in `marks`) brings it back through the layout it was taken in, migrated to today's, and `diff NAME` compares it with live leaf by leaf over the fields both layouts have. `record NAME` starts a clip from the mark `snapshot NAME` chose (saved first if it is new, restored if it exists, so several clips branch from the same bytes) and `stop` ends it; `play NAME` restores its keyframe and replays its frames; `pause`, `seek start|end|MS` and `step N` (frames) move within it, stopping only between frames. `play session S [at SECONDS]` replays a past session from its start, or from the keyframe before that time, paused there. At each later keyframe `update` produced, the replay is compared with what was recorded: a difference is said once, `diverged before frame F: N values, first `, and the replay goes on from the recorded State, which also carries it over the jumps the log does not hold. A clip or session made by another build says so, since the replay may then differ. Going back and playing on is `play` or `seek`, then `stop`: live goes on from there in the same session, nothing overwritten. Names are words of up to 32 bytes, and numbers are names, so `save 3` and `play 0` work as they always have. Restoring State (play, seek, step, `load`) clears the pointer, button and key state, and a clip begins with events that bring back what was held when it started (a pointer move, a button or key down), so a replay starts from the same input whatever the live pointer did since. Event times are the session's: `in.now` does not restart at 0 in a clip. - **Console and overlays.** `help` lists the console's verbs and `help VERB` says what one does. `get`/`set ` (`get` of a vector, array or struct shows its values, in memory order; `set` of an array of numbers takes all its values), marks, clips, sessions (`marks`, `clips` and `sessions` list the last few) and the memory, scope and frame timing overlays come from the host; unknown lines go to the sketch. Paths index arrays and vector lanes: `get pos[0]`, `set boids.buffer[3].pos[1] 20`. `diff NAME` compares live state with mark NAME leaf by leaf, so padding (a Debug `@Vector`'s spare lane) never counts. - **Scripts** (`--headless --fresh --script file`) drive the same routing and frames as a player: `click`, `key`, `hold`/`release` (a key down until released, for modifiers), `wheel`, `type`, `console`, `print field` (to stderr, no frame: a vector, array or struct prints every value in it, `{ 300, 200 }`), `expect field value` (`expect pos[0] 250 ~0.5` within a tolerance; a failure prints what it found), `await field value [~tol]` (frames until it holds, a minute at most), `expect-snapshot NAME` (live state equals mark NAME, leaf by leaf, printing the first differences), `expect-key WHY field value` (the newest keyframe of this session taken for WHY, `before_reload`, `mark`, `exit`..., holds the value, read in its own layout), `expect-log`/`expect-no-log text` (the console's recent lines), `expect-audio`, `screenshot`, `expect-same-screenshot a b [max [x y w h]]`, `expect-different-screenshot a b [x y w h]`, `expect-pixel file x y r g b a [tol]`, `load`, `reload`, `await-reload`, `kill` (death by SIGKILL, for tests of what a crash keeps). A failed `expect`, or a console line that errors (a sketch command returning an error included), fails the script. Headless always means offscreen, even with `SDL_VIDEO_DRIVER` set, and GPU work still runs, into a texture. A scripted `--fresh` run saves nothing: its sessions are in memory, and the files it appends to are `..`, removed at exit, so two builds running the same script at once do not share them. - **Export.** `--export DIR` plays the sketch headless, from fresh state, and saves every frame as `DIR/000000.png`, `000001.png`, ...; `--export -` writes them to stdout instead, as raw RGBA: ```sh demo --export frames --size 1920x1080 --frames 600 demo --export - --size 3840x2160 --fps 60 --frames 600 | ffmpeg -f rawvideo -pix_fmt rgba -s 3840x2160 -r 60 -i - -pix_fmt yuv420p out.mp4 ``` The frames run on a clock of their own: frame k is k/fps seconds in (`--fps N`, 60 unless given) and its tick is exactly 1/N s (frame 0's is 0, as any run's first), however long a frame takes to draw. A sketch that takes its time from the ticks exports the same frames every time. It ends after `--frames N`, or when the sketch calls `f.quit()`; `in.exporting` tells a sketch that nobody is watching, so a player hides its controls, plays through and quits. `--size WxH` is the room for the canvas (`in.window_width`/`height`, and the first canvas): a sketch that sets its canvas to the room exports at any size. An export saves nothing, and the files it appends to are its own, removed at exit, so it can run beside a player. `--export` plays no script. In a release build on an RTX 3050, an `anim` scene takes 6 ms a frame as raw RGBA and 21 as PNG at 1920x1080, 19 and 73 at 3840x2160; with ffmpeg encoding x264 from the pipe, 11 and 41. The PNG encoder is the slow part: pipe, when the frames are only on their way to a video. - **GPU.** The host's 2D renderer runs on an SDL GPU device, which also runs sketches' GPU passes. Without a GPU it falls back to any renderer. `--profile-gpu` times every upload, dispatch and pass alone on an idle GPU (SDL_GPU has no timestamp queries) and prints count, mean, worst and per-frame time by `label` at exit. - **Memory.** SDL allocations go through a leak-checking allocator; a leak fails the exit. SDL's Vulkan backend leaked when `vkAllocateMemory` failed and it fell back to another memory type (it happens with many headless hosts at once); the build compiles that file with the fix. It also never defragments memory there: SDL moves textures to new ones inside whichever command buffer is submitted next, and a sketch's GPU pass is submitted while the frame's 2D commands are still being recorded, so the canvas lost what was drawn before the pass and showed the previous frame under the rest (after a large texture was freed). ## The web host `web/host.js` runs the same sketch, compiled to wasm32-freestanding, in a page. Draw commands go to a 2D canvas; GPU passes, dispatches, uploads and downloads to WebGPU, whose renders are shown through the canvas; streams and the microphone to Web Audio (at 48 kHz where the browser agrees, sound starting at the first click or key, as browsers require); `f.openFile` to a file picker, whose sound the browser decodes and hands over by name before the console line arrives. `f.run` and the sketch's `command` work as on native. Touch: one finger is the mouse's left button. Two fingers pan, as the wheel (60 px a notch), and pinch, as ctrl+wheel (a notch per 20 % of spread), both around the point between them: what a trackpad sends, so a sketch that handles a trackpad handles a touch screen. Keys have no touch equivalent; draw buttons for what matters. Without WebGPU, the page says why the first time the sketch asks for it: not a secure context (WebGPU and the microphone need https, or localhost), WebGPU turned off (Firefox on Linux: `dom.webgpu.enabled`), or no usable adapter. `.web = .{ .pwa = .{ .icons = ..., .theme = ... } }` makes the page an installable app that opens offline: a manifest, icons, and `web/sw.js`, a service worker that keeps a copy of everything the page fetches, network first, so each new build shows as soon as there is a network. Service workers need a secure context, like WebGPU. For tests, `window.sketch` reads state (`field(path)`), runs console lines (`command(line)`) and hands over decoded sound (`sound(name, samples)`). ## Build API ```zig const sketchbook = @import("sketchbook"); pub fn build(b: *std.Build) void { const sb = sketchbook.init(b, .{ .target = target, .optimize = optimize }); const app = sb.app(.{ .name = "demo", .root_source_file = b.path("src/main.zig") }); _ = app.addSteps(.{ .scripts = &.{b.path("test/play.script")} }); } ``` `addSteps` installs the executable and adds `run`, `lib`, `export` (`zig build export -- frames --size 1920x1080`: what follows `--` is `--export`'s directory, or `-`, then its options), `e2e` and `test`, and with `.web = .{}` also `web` (the page into `zig-out/web`) and `serve` (it, over https on port 8042, to the local network too; `zig build web --watch` alongside rebuilds it for the next reload). Every piece is also available on its own and returns plain `std.Build` objects: | Call | Returns | | --- | --- | | `sb.addScene(.{ .root_source_file, .name, .imports })` | an `anim` scene file's player: `.app`, and `.module` and `.typst` (its Typst as SVG) | | `app.addLibrary()` | the hot-reloadable shared library | | `app.addExecutable()` | host and sketch linked into one binary | | `app.addScriptExecutable()` | the same, built in `script_optimize` mode, as `e2e` plays scripts with | | `app.addWeb(.{ .assets = ... })` | the wasm module and a ready-to-serve directory, built from the app's modules with every shader module swapped for its WGSL twin; tests can read its state by path (indices too: `list[2].x`) with `sk_field` (numbers) and `sk_tag` (enum tag names) | | `app.addServe(install)` | `miniserve` serving an installed web directory over https, on every interface, with a self-signed certificate made once by `openssl`: https makes the page a secure context on any device, which WebGPU and the microphone need. Its defaults are environment variables, so miniserve options after `--` replace them | | `app.addRun(lib, .{})` | the generic host running the installed library, its files in `.sketchbook/` | | `app.addScript(exe_or_lib, file, .{})` | a headless, fresh-state scripted run, of the executable or of the library through the host (so the script can `reload` it) | | `app.addExport(exe, .{ .fps, .frames, .size })` | a headless export of the executable's frames, as PNGs in `.frames`, a directory of the cache | | `app.addTest(.{})` | tests with the app importable as `app` | | `app.addReloadCheck(.{ .script, .edits })` | a check that the app hot-reloads with State kept (below) | `sketchbook.init(b, .{ ..., .script_optimize = .ReleaseSafe })` builds the executables `e2e` plays scripts with in another mode than the rest, for sketches whose per-pixel loops crawl in Debug; null (the default) is `optimize`. Only Zig code changes mode (the host's and the app's), so it costs one more compile, not SDL's. A `b.option` makes it a `-D` flag. `sb.app(.{ ..., .record = true })` keeps history in every mode (see The host); null, the default, in Debug builds only. **Hot reload, checked.** `addReloadCheck` builds the app's library as written, and again from a copy of its sources with textual edits applied, then plays a script on the first through the generic host with `{v2}` defined as the second: ```zig const check = app.addReloadCheck(.{ .script = b.path("test/reload.script"), .edits = &.{ .{ .file = "src/main.zig", .old = "0xff8040", .new = "0x40c0ff" }, .{ .file = "shaders/glsl/expo.frag", .old = "vec3(0.0, 1.0, 0.0)", .new = "vec3(1.0, 0.0, 0.0)" }, } }); steps.e2e.dependOn(&check.run.step); ``` Each `old` must occur exactly once in its `file` (a path as given to `b.path`), or the build fails. The copy is of each directory holding a source the app builds from (a module's root directory holds what it imports), and every module and shader set made from it (`addShaders`, `addZigShaders`) is made again from the copy, so edits to Zig code, GLSL and Zig shaders all reach the second library. The script plays, checks State, runs `load {v2}`, checks the same State, and compares screenshots (`expect-different-screenshot a.png b.png [x y w h]`). Pausing a replay first (`record`, `play`, `pause`) keeps the State still, so the two screenshots differ only by the edit. `examples/shaders` and `examples/noc` (bouncing_ball) have one each. Shaders: ```zig const shaders = sb.addShaders(.{ .layouts = .{ .module = my_layouts, .types = &.{ "Vertex", "Paint" } }, .sources = &.{ b.path("shaders/fill.vert"), b.path("shaders/fill.frag") }, }); const app = sb.app(.{ ..., .imports = &.{.{ .name = "shaders", .module = shaders.module }} }); // in the sketch: .{ .code = shaders.fill_vert } ``` The layouts are Zig structs (tagged with `pub const glsl = .vertex` or `.{ .uniform = .{ .set, .instance } }`) written out as `.glsl` for the sources to `#include`, so the Zig struct is the one definition. Each source is compiled by `glslc` into `shaders.module`, as SPIR-V under its file name (`fill.vert` is `fill_vert`). For the web it is compiled again with `SKETCHBOOK_WEB` defined and translated by `naga` into `shaders.web`, the same names in WGSL; `addWeb` puts that twin in place of the SPIR-V module wherever the app's imports reach it, however deep (the `vector` module's shaders included). Only the web build runs `naga`, and a missing tool fails only the steps that need it, listing where it looked. WGSL has no write-only storage buffers and no `coherent`: write them plain, or under `#ifndef SKETCHBOOK_WEB`. Sources may `#include` files beside them; `glslc`'s depfiles tell the build, so editing an included file recompiles what includes it. `addGlslLayouts` and `addGlslc` are the pieces, for anything else. Shaders can be Zig too: `sb.addZigShaders(.{ .sources, .imports })`, one shader per file with `export fn main() callconv(.{ .spirv_fragment = .{} })`, returns the same `Shaders`, so a sketch cannot tell the language. Zig's own SPIR-V backend compiles each for `spirv32-vulkan` (`spirv64` asks for physical pointers, which naga refuses), self-hosted (through LLVM the object comes out empty, silently); `tools/spirv_strip.zig` then drops the `Aligned` operands Zig puts on every load and store, which Vulkan accepts and naga's reader does not; and naga translates the result for the web as for GLSL. Shader code cannot use table-driven `std.math` (`atan2`: call GLSL.std.450 through inline assembly), and a uniform struct shared with the CPU needs a vec2 that is `@Vector(2, f32)` only on SPIR-V and `[2]f32` elsewhere. `examples/shaders` has five Book of Shaders examples in both languages. Tools: `glslc` (shaderc, or the Vulkan SDK's) for shaders, for web builds with shaders `naga` (`cargo install naga-cli`; found on PATH or in cargo's directory), `typst` for scenes with Typst in them, and for `serve`, `miniserve` and `openssl`. This repository's own tests also run `node` (noise on wasm32, under WASI), `typst`, `resvg` and `python3` with numpy and Pillow (the renderers against resvg), `ffmpeg` if there is one, and, for `zig build test-web`, Chrome. `sb.modules` (`abi`, `sketch`, `host`, `vector`, `scope`, `anim`, and `cloud9`, the 9P library, whose `Fsys` lets an app's tests dial its own executable as `fs` does), `sb.host_exe` is exposed for anything else. Add `.sketchbook = .{ .path = "..." }` to your `build.zig.zon`. ## Known limits - The console line holds 96 characters, so long paths cannot be typed; `f.openFile` answers with a line of any length. - Every upload, compute dispatch and GPU pass is its own command buffer and submit. Fine at a handful per frame; batch them per frame if that grows. - Native antialiasing resolves coverage per sample within a run of shapes, the canvas per shape, so where two shapes' edges meet the web can show a faint seam native does not. `fill` rects and images are antialiased natively only inside a target. Without a GPU renderer, native draws aliased. A stroke that overlaps itself blends twice natively where it overlaps, once on the web. - A keyframe hashes all of State to find the pages that changed: about 70 ms for 32 MiB in a Debug build, 3 ms in ReleaseFast, once a minute while State changes and at every jump. - On the web: nothing persists and no sessions are recorded, and multisampling is 4× at most. Every voice is silent there, whatever its sound (loaded, decoded or made with `newSound`), and so are groups' gains; `decodeSound` gives `.none`, and `loadSound` knows only sounds the page decoded (opened or dropped files). What does sound: streams (`f.audio`, `f.audioBuffer`) and the microphone, mixed on the main thread (a ScriptProcessor), so a busy page can crackle. A GPU buffer streamed there arrives a frame or two later (WebGPU reads back only asynchronously), counted as queued meanwhile, and `in.lead` is a fixed 100 ms. The reusable GPU executor and the pardes integration are described in [docs/pardes-rendering.md](docs/pardes-rendering.md), including the remaining renderer migration work and the optional addon boundary. ## Designs - [Regions in source](docs/regions-in-source.md): editing layout from the running sketch by rewriting the number literals in the source (not implemented yet). ## The boundary Host and sketch are built by the same `zig build`, so ordinary Zig types cross it (slices, tagged unions, `std.mem.Allocator`). `abi.hash`, a hash of every shared type's layout and of the zig version, backend and optimize mode, guards against the real hazards: a host refuses a library built against different definitions, or by another compiler or mode (`zig build run` beside `zig build lib --watch -Doptimize=fast`: Zig's own calling convention differs between them), and says what differs. The layout half costs half a second of comptime, so a small program (`src/abi_hash.zig`) computes it once per change of the abi sources, at build time, and host and sketches import the number; it is computed for the build machine in Debug, so a layout change seen only by another target or mode is not caught. Everything that outlives a call lives in host memory, since host code is never unloaded. Only the recorded `Event` and the reflected `Layout` are `extern`, because they are written to disk. `zig build test` in this directory runs the unit tests, a headless reload through `dlopen`, where a counter sketch changes its layout mid-run and keeps its values, and a replay test (`test/replay.script`): a clip played, sought, stepped and replayed after a live pointer move, each ending on the state it was stopped at. The drawing layer's tests are there too: the shape cases drawn natively and checked pixel by pixel (`test/shapes.zig`), noise's golden values on wasm32 under Node, and the examples in `examples/` (`pen.zig`: Pens and a Clock; `flock.zig`: tessellations batched by hand, between Frame methods). `test/render.zig` covers targets (and that `set` and `load` repaint them while plain frames do not), blending, placed and tinted images and colours per subpath, checked pixel by pixel. `zig build test-web` draws the shape and render cases in headless Chrome, compares them with the native screenshots (`tools/chrome.mjs`), and times 5000 circles a frame (software Canvas2D here: about 12 ms as 5000 shapes, 8 ms batched). And the rest: `test/text.zig` (a box of `textSize` under each line), `test/export.zig` (two builds export the same frames, frame k at k/fps; the raw stream's length; a video from ffmpeg with as many frames), and the `anim` player on its two test scenes: the live one checks, pixel by pixel, that a custom object sits over the vector object added before it and under the one added after, moves, fades and zooms with the timeline and camera, on both renderers; the compiled one is also exported at half size. `zig build diff` (part of `test`) renders a corpus (`test/svg`: every path command, fill rules, joins, caps, dashes, transforms, opacity, gradients, symbols, and Typst documents) with resvg as the reference, and with both vector backends through the real host, headless (`test/svgview.zig`, `tools/svgdiff.py`). Every case is under 0.7 mean channel error, with no more than 0.01% of pixels off by more than 48; side-by-side images land in `.zig-cache/svgdiff`. Two details make the reference fair: resvg samples coverage on four sub-scanlines, so it renders at 4x and is box-filtered down (padded to whole blocks, not squeezed); and sizes in `pt` make it round to whole pixels before scaling, so the harness gives it unit-free sizes. `zig build bench` (part of `test`, after the rest) times each Typst object of the compiled test scene at 1280x720, the drawing filling the canvas, in release mode, and fails if a simple one (30 shapes) takes 1 ms to parse and draw: an equation of 8 shapes parses in 0.02 ms, draws on the CPU in 0.25 ms, and is recorded as a GPU pass in 0.02 ms (the GPU's own time is not measured). ### Optional native profiling Build the host with `zig build -Doptimize=ReleaseFast -Dtracy=/absolute/path/to/tracy`. The checkout supplies `public/TracyClient.cpp`; no profiler is downloaded or linked by default. Linked native apps pass their `-Dtracy` option to `sketchbook.init` as `.tracy = path` and can import `sb.modules.tracy` for their own zones: `const zone = tracy.zone(@src(), "work"); defer zone.end();`. The host marks frames, input polling, drawing, command painting and presentation. These are CPU wall times (presentation includes waits), not GPU timestamps. Use a matching `tracy-capture` or the Tracy UI. Disabled builds compile the calls out; tests disable their zones too. The wrapper follows Pardes's small Zig Tracy binding. Forward the same option through dependencies that also initialize Sketchbook so the build shares one host/client module. ASCII text caches coverage independently of animated ink. The pinned SDL Vulkan patch keeps memory relocation disabled and lets small transfer buffers share allocations, avoiding hundreds of dedicated allocation frees during presentation. ## glTF / GLB assets The optional `gltf` module wraps the vendored C99 `cgltf` loader. It loads scene hierarchies, mesh accessors, materials, cameras, animation tracks and arbitrary `extras` metadata without depending on the window or renderer. Import it into a native executable or sketch with `.imports = &.{.{ .name = "gltf", .module = sb.modules.gltf }}`. It currently requires libc; it is not wired into the freestanding browser host. ```zig const gltf = @import("gltf"); const document = try gltf.Document.parse(bytes); defer document.deinit(); try document.loadBuffers(null); // embedded GLB / data-URI buffers, no file reads const node = document.findNode("CAM_StreetTour") orelse return error.MissingCamera; const json = gltf.extrasJson(&node.extras) orelse return error.MissingMetadata; // Parse json with std.json using the application's own metadata schema. ``` Keep `bytes` alive until `deinit`: GLB buffers may borrow it. Parsing validates structure and uses cgltf's C allocator, released by `deinit`. `loadBuffers` validates again with loaded data. Only passing a non-null, zero-terminated glTF source path enables relative external buffer-file reads; use that for trusted assets. Images and compressed geometry still need their own decoders. `document.data` and `gltf.c` expose cgltf directly; `gltf.readFloats(accessor, index, out)` adds bounds/component checks and supports strided, normalized and sparse accessors. The raw animation channels remain available to the caller; this module does not create GPU resources or implement an animation player. Run `zig build test-gltf` for focused loader tests; they are also part of `zig build test`. ### Generated scene bindings `sb.addGltf` runs a native host executable in a separate build step, then supplies its generated `.zig` file as a module. The glTF file is a tracked build input; changing it regenerates the bindings before compiling their consumers. ```zig // build.zig: add this import to the consuming module's .imports. .{ .name = "street_scene", .module = sb.addGltf(b.path("street.glb")) } ``` ```zig const scene = @import("street_scene"); const camera = scene.nodes.CAM_StreetTour; comptime { if (!@import("std").mem.eql(u8, camera.extras.drat_role, "camera")) @compileError("expected a camera role"); } // camera.index locates the node; @TypeOf(camera.extras) exposes metadata fields. ``` The generated `nodes`, `animations`, `cameras`, `meshes`, and `materials` namespaces expose named entries with indices and typed `extras`; nodes also include optional camera and mesh indices. Labels with spaces use Zig's quoted identifier syntax, e.g. `scene.nodes.@"Camera - At the corner"`. Unnamed entries are omitted; duplicate labels within one namespace fail generation. JSON objects become struct values, arrays become tuples, and strings/numbers retain explicit Zig types. Comptime can reflect on these types or assert expected values; it never parses the asset. Missing referenced labels/fields and incompatible types fail compilation; application assertions define which metadata values must match. The module contains no libc dependency. `source_path` is the build input filename; `source_sha256` lets runtime consumers reject a replaced asset before using generated indices. Code generation reads metadata, not external buffers; geometry and animation sample checks still use the runtime loader. The game repo's `scripts/check_scene.zig` demonstrates both. ### Local 9P controls A running sketch is a small file tree, served over 9P by cloud9: its State as files, a `ctl` that takes the console's words and a script's input, its screen as a PNG. The tree describes itself: `README` has a line for every file, every verb under `ctl` and every State file under `state/`, made from the same tables the console, scripts and `--help` read; `commands` lists the verbs alone, the sketch's own last. A run a person watches posts its tree as `sketchbook/` (the socket `$XDG_RUNTIME_DIR/9p/sketchbook/`, owner-only, removed at a clean exit); `--9p NAME` posts it under another name, also headless, and `--9p /path.sock` listens at a path instead. Every executable built with sketchbook, and the generic host after its LIBRARY, takes `fs VERB PATH`: cloud9's verbs (`ls read write ask follow stat help complete`) on that tree, the server's refusals printed in its words, exit status 1. With the sketch closed, `fs` serves the same tree itself, with no window, from the save's newest State (the default save, or `--state`'s), else from State's declared defaults: `README`, `commands` and `state/` answer, and the rest say the sketch is not running. A `read` answered that way says so on stderr (`drat-meout is not running: this is its save's newest State, frame 61 of ...`), so a script that reads before the sketch is up does not take the save for it. ``` ctl write a verb a line: ctl's own (screenshot, profile-reset), a script's input verbs (click, key, type...: routed live, no frame of their own), the console's (set, load, pause, step...); any other line goes to the sketch status, log name, canvas, mode, GPU, screenshot state; the console's last five lines screen.png the next frame: an open waits for it to be drawn README the tree, a line a file; commands: the verbs state/ the State: a directory per struct and per array, a file per value (state/tracker/songs/3/bpm); a read is the value as `get` shows it, a write is `set` (a jump) sketch/ reserved for the sketch's own files, empty ``` An array of numbers is one file of its values, blank-separated (`state/talk/text`, a vector's lanes), written the same way; an array of arrays is a directory of such rows; an array of structs a directory of indices, so drat-meout's 8x16x16x4 tracker notes are `state/tracker/songs//patterns////pitch`. Optionals and unions read as `` and refuse writes. A value is read as it was at the open. Driving a sketch from an agent, here drat-meout: ```sh drat-meout --headless --9p drat-meout & # or just run it: it posts itself drat-meout fs help ctl # the verbs, with their arguments drat-meout fs write ctl 'scene street' # a console line, or the sketch's drat-meout fs read state/scene # .street until [ "$(drat-meout fs read state/street/time | cut -d. -f1)" -ge 3 ]; do :; done drat-meout fs write ctl 'click 480 300' # input, live drat-meout fs read screen.png > street.png # one read: answered when the frame lands drat-meout fs write state/street/camera fast # fs: state/street/camera: InvalidValue drat-meout fs help state/street/camera # state/street/camera .placeholder .scene drat-meout fs write ctl quit drat-meout fs read state/scene # closed: from the save ``` Scripts stay what tests use (a fixed clock, fresh state, `expect`); the tree is live, between real frames, and an agent asserts by reading. Anything that speaks 9P works too: `9p -a unix!$XDG_RUNTIME_DIR/9p/sketchbook/drat-meout read state/scene`. Four clients can connect; requests run between frames on the host thread. `zig build test-fs` drives a sketch this way. The native host's `timing` command toggles a scrolling graph with average, p90 and p99 in milliseconds. `timing on|off` sets visibility; `timing frame|work` selects the graph and shows it; `overlay timing` also toggles. `timing stats` prints current average/p90/p99/maximum; `timing reset` clears history without changing visibility or metric. Collection continues while hidden. Readouts refresh about four times a second from fixed text buffers. All glyphs reuse one immutable atlas texture, so changing values neither rasterizes text nor switches textures per digit. Fixed vertex/index arrays batch the entire panel, graph and text into one geometry draw, without per-frame overlay allocation. Percentiles use nearest rank over the last 240 individual frames. A separate fixed graph ring holds 240 peaks, each over eight frames: eight times slower scrolling, with single-frame spikes preserved across 1920 frames. Its scale label updates when the peak range changes. Frame measures wall-clock cadence between completed host frames, including pacing, event/9P polling and present. Work measures update/draw through the deferred 2D command flush, before present. Both include visible overlays and screenshot capture and use real time even during replay/export. These are not GPU hardware timestamps, and frame can include vsync waits. No GPU synchronization is added. Test with `zig build test-timing`. The read-only 9P file `timing-trace` exports paired raw samples as CSV: `index,frame_ns,work_ns,outside_work_ns`. Indices start at zero within the captured history, oldest first. `outside_work_ns` is saturating `frame_ns - work_ns`, including pacing, polling, presentation and cleanup; it is not a hardware GPU duration. Each open pins the bounded timing history, so subsequent frames or `timing reset` cannot change a partially read trace. It works without `--profile-gpu`: `drat-meout fs read timing-trace > timing.csv`. For compositor/video measurements, launch with `--timing-marker`. This opt-in marker is drawn last, independently of the timing overlay: 80 black/white cells, each 4×8 canvas pixels, at the bottom-left (320×8 total). Cells encode, LSB-first, 16-bit magic `0xD3A7`, the low 32 bits of `Input.frame`, then the low 32 bits of `Input.now` in whole milliseconds. It identifies submitted frames and draw sample times in captured video; repeated decoded IDs reveal repeated presented frames. Decode at canvas scale, accounting for any window/compositor scaling. The marker is omitted when the canvas is smaller than 320×8 and is disabled by default. `zig build test-timing` checks its encoding, placement and counter wrap. On Wayland, `presentation-trace` provides immutable CSV snapshots of `wp_presentation` feedback: frame ID, sample time, request time, presented/discarded result, presentation timestamp, refresh period, sequence and flags. Status and clock metadata make unsupported backends explicit. `commit_ns` is the feedback request time before present, not an intercepted surface commit. Only subtract timestamps in the same clock domain; sample time uses SDL's elapsed clock. Feedback attaches to the next surface commit, which the binary marker lets a compositor capture independently verify. Interactive Wayland windows use this feedback to pace submissions and sample one shared frame clock before update and draw. Ordinary `Input.now`, tick events and `Input.dt` therefore follow display refresh automatically; sketches need no separate rendering clock or animation correction. A tick never precedes the previous frame or its queued input, so newly arrived input can move it slightly past the scheduled instant. Recording stores these exact ticks; replay, scripts and export retain their existing clocks. The browser already uses the `requestAnimationFrame` timestamp for its shared update/draw clock. Without usable feedback, rendering falls back to SDL's normal presentation and wall-clock ticks. Hidden/minimized windows keep servicing controls but skip animation and drawing until visible again. The implementation is `src/host/Presentation.zig`, including the Wayland callbacks and trace writer. `zig build test-presentation` checks its clock and feedback behavior. Linux builds require `wayland-scanner` plus the `wayland-client` development package and pkg-config metadata. The build generates the protocol header/tables from the vendored XML and translates the header into Zig bindings; only the generated protocol tables are compiled as C. `ctl` takes up to 4096 bytes a write; an error returns as the 9P error, the error's name (`UnknownField`, `InvalidValue`). Each open of `screen.png` (and of the traces, `README` and a State file) pins its own copy, so a later frame cannot change a partially read one; screenshots include visible overlays. `zig build test-control` checks the control boundary and snapshot lifetime; the same tests belong to the full host suite. With `--profile-gpu`, `profile` reports tab-separated `label`, `count`, `mean_ms`, and `worst_ms`, plus a `# frames N` line. `profile-reset` written to `ctl` resets counters and the frame denominator without discarding label allocations. Warm up first, reset, then wait for the desired number of frames before reading. These measure CPU wall time from GPU command recording through submission and fence wait, including synchronization overhead; they are not hardware GPU timestamps. Both operations return `ProfilingDisabled` when profiling is off. ### Imperative glTF drawing Add an import with `.module = sb.addGltfRenderer()` (e.g. named `gltf_render`). Its `Model` copies a GLB's meshes and animation tracks into the caller's session arena. The cgltf document is released after loading; geometry uploads once to host-owned GPU buffers. Opaque primitives in each mesh are batched into one draw with per-vertex material colors. Only transforms and draw commands change each frame. ```zig const render = @import("gltf_render"); const street_scene = @import("street_scene"); pub const Session = struct { model: render.Model = undefined, image: sk.Image = .none }; pub fn setup(ss: *Session, init: *sk.Init) !void { // Filename + hash come from the generated bindings. Parses and uploads once. ss.model = try render.Model.load(init, street_scene); ss.image = init.host.newImage(960, 600); } pub fn draw(s: *const State, ss: *Session, f: *sk.Frame) void { renderFrame(s, ss, f) catch |err| f.host.print("scene: {t}", .{err}); } fn renderFrame(s: *const State, ss: *Session, f: *sk.Frame) !void { try ss.model.scene.sample(street_scene.animations.CAM_StreetTour.index, s.time); const camera = try render.sceneCamera(&ss.model.scene, street_scene.nodes.CAM_StreetTour.index); var view = render.View.init(f, ss.image, 960, 600, camera); try view.scene(&ss.model); // Or: try view.mesh(&ss.model, street_scene.meshes.MyMesh.index, transform); // Or: try view.triangles(&geometry, transform, color); // upload in setup first view.end(); f.image(ss.image, .{ .x = 0, .y = 0, .w = 960, .h = 600 }, .{}); } ``` `model.scene.nodes[index].world` is a column-major transform; `meshes[index]` exposes indexed vertices, normals and material colors. Animation sampling is allocation-free, supports STEP, LINEAR (slerp for rotations), and CUBICSPLINE TRS, and clamps to the clip range; the application decides how to loop. The separate CPU API is `gltf.Scene.init(allocator, document)`, `sample`, `reset`, and `deinit`. It copies loaded data, so the document can be destroyed after initialization. `render.sceneCamera(scene, node_index)` derives perspective or orthographic projection and view from the sampled camera node. Alternatively pass a plain `render.Camera` with your own eye/target to `View.init`. `Model.init` remains available for CPU loading, and `Geometry.upload(host)` uploads custom triangles explicitly during setup. No resource loading happens inside `View.scene/mesh`. This initial renderer uses opaque material base colors, directional light and a depth buffer. It has no texture sampling, alpha blending, shadows, skinning or morph animation. Unsupported mesh/animation features fail loading explicitly; material textures are ignored. The renderer currently supports native hosts; `render.supported` is false in the browser replacement module. GPU handles are released by the host on session reset; call `model.release(host)` to release them earlier. Vertex storage layout and shader uniform layout are checked at comptime; GLSL uniform declarations are generated from the Zig types. A separate fullscreen pass can sample a completed render with `f.postProcess(source, target, .{ .vertex = vertex_shader, .fragment = fragment_shader, .sample_depth = true, .fragment_uniforms = uniforms })`. Allocate distinct source and target images in setup, then draw the target image after this command. Native fragment bindings are color at 0 and depth at 1 (set 2). Web GLSL uses separate texture/sampler pairs at 0/1 and 2/3 (set 2). Fragment uniforms use set 3, binding 0. Depth sampling requires the source's latest pass to have `.depth = true` and `.samples = 1`; missing depth, multisampled depth and feedback loops are rejected. Omit `sample_depth` for a color-only pass. Native profiling reports this work separately as `postprocess`. `profile` also includes `p95_ms` and `p99_ms`: nearest-rank percentiles of the latest 2048 measured commands per label. Mean and worst cover every command since reset. `profile-trace` exposes those retained samples as CSV (`label,sample,ns`), ordered oldest to newest within each label; sample numbers start at 1 after reset. Each open pins the trace even if profiling continues or counters reset. The render path writes a fixed ring; sorting and CSV generation happen only on inspection. A trace snapshot is limited to 256 KiB and fails if it exceeds that limit. Capture it with `drat-meout fs read profile-trace > trace.csv`.