@particlr/runtime

The MIT-licensed particle runtime behind particlr — a framework-agnostic simulation core plus a PixiJS v8 adapter. It plays back .prt effect documents deterministically: the same document, seed, and sequence of dt values always produce the same frames. The particlr editor previews through this exact package, so what you tune is what you ship.

• Zero runtime dependencies in the core (pixi.js is an optional peer, used only by the /pixi adapter).
• Node-safe core — no DOM, no Math.random, no wall-clock reads.
• Ships the JSON Schema for .prt (@particlr/runtime/particle.schema.json).

Install

npm install @particlr/runtime pixi.js

pixi.js@^8 is a peer dependency (only needed if you use the /pixi adapter).

TypeScript note: if you compile with skipLibCheck: false, importing @particlr/runtime/pixi may surface duplicate-identifier errors (TS2300/TS2403) from pixi.js v8's bundled @webgpu/types colliding with TypeScript's own DOM WebGPU declarations. That's an upstream pixi.js × recent-TypeScript issue, not this package; the standard fix is "skipLibCheck": true (the common default).

Quick start (Pixi v8)

This is the whole integration — load a .prt, step it each frame, render it:

import { Application } from "pixi.js";
import { parseParticle, Effect } from "@particlr/runtime";
import { PixiParticleRenderer } from "@particlr/runtime/pixi";

const app = new Application();
await app.init({ background: "#000", width: 800, height: 600 });
document.body.appendChild(app.canvas);

const doc = parseParticle(await (await fetch("boom.prt")).text()).doc!;
const fx = new Effect(doc, { seed: 1337 });
const view = new PixiParticleRenderer(fx);
view.container.position.set(400, 300); // where the effect plays
app.stage.addChild(view.container);

app.ticker.add((ticker) => {
  fx.step(ticker.deltaMS / 1000); // advance the simulation
  view.sync();                    // copy state onto the Pixi particles
});

A complete runnable version is live at particlr.brac.dev/sample — click the canvas to spawn effects through this exact package.

Core API (@particlr/runtime)

Reading & writing documents

parseParticle(input: string | object): ParseResult

Parses (JSON string or object), migrates to the current schema version, and validates. Unknown fields are preserved for byte-stable round-tripping. Throws only if a string isn't valid JSON.

interface ParseResult {
  ok: boolean;
  doc: ParticleDoc | null;      // null when !ok
  errors: ValidationIssue[]; // { path, message, code? }
  warnings: ValidationIssue[];
}

serializeParticle(doc: ParticleDoc): string

Emits canonical JSON (declared key order, unknown keys preserved, 2-space indent, \n endings, trailing newline). serializeParticle(parseParticle(text).doc) equals text byte-for-byte for canonical input.

validateParticle(input: object): ValidationResult   // { ok, doc, warnings } | { ok:false, errors, warnings }

BUILTIN_TEXTURE_IDS — the built-in texture ids: circle-soft, circle-hard, square, spark, smoke.

Effect

class Effect {
  constructor(doc: ParticleDoc, opts?: { seed?: number; x?: number; y?: number }); // seed overrides doc.seed; x/y = initial emitter position
  step(dt: number): void;      // advance by dt seconds
  reset(seed?: number): void;  // rewind to t=0 (re-prewarms if configured); keeps the emitter position
  setEmitterPosition(x: number, y: number): void; // emitter position at the END of the next step (drives world-space trails)
  teleportEmitter(x: number, y: number): void;    // jump with no velocity/interpolation (respawns, screen wraps)
  setAttractor(x: number, y: number, strength: number, radius: number): void; // host-driven attractor (schemaVersion 4); radius <= 0 clears
  clearAttractor(): void;                          // remove the host attractor
  setParam(name: string, value: number): void; // drive an exposed SCALAR parameter (schemaVersion 6); clamps to [min,max]; unknown name / non-finite value = no-op
  getParam(name: string): number;              // current scalar param value (authored default until first set); unknown / color-kind name ⇒ NaN
  setColorParam(name: string, r: number, g: number, b: number, a: number): void; // drive an exposed COLOR parameter (schemaVersion 8); clamps each channel to [0,1]; unknown / scalar-kind name or any non-finite channel = no-op
  getColorParam(name: string): RGBAColor | null;  // current color param RGBA (a copy; authored default until first set); unknown / scalar-kind name ⇒ null
  timeScale: number;           // host playback rate; 1 = real time, 0 = frozen (hit-stop), <1 slow-mo; non-finite/≤0 ⇒ 0
  onDone: (() => void) | null; // fired once when a non-looping effect finishes (isDone); re-armed by reset()
  readonly emitterX: number;
  readonly emitterY: number;
  readonly time: number;       // effect-local seconds
  readonly seed: number;
  readonly isDone: boolean;    // non-looping effect finished and no particles remain
  readonly particleCount: number;
  readonly layers: readonly LayerSim[]; // per-layer: count, capped, typed-array state
}

Simulation space & moving emitters (schemaVersion 2)

Each layer has a space: local particles ride the emitter (weld to the source — muzzle flash, aura); world particles spawn at the emitter's current position and then simulate independently, so a moving emitter leaves them behind — a trail (flamethrower, rocket smoke, comet, a flaming projectile).

Host contract: place view.container at the effect's world origin once, then drive motion through the emitter — never by moving the container (that drags world-space particles too). A projectile with a trail:

const fx = new Effect(fireballDoc, { seed });
const view = new PixiParticleRenderer(fx);
view.container.position.set(0, 0);      // fixed; the emitter moves, not this
fx.teleportEmitter(startX, startY);     // launch point, no start smear
app.stage.addChild(view.container);

app.ticker.add((t) => {
  x += speed * (t.deltaMS / 1000);
  fx.setEmitterPosition(x, y);          // advance the head; world layers shed a trail
  fx.step(t.deltaMS / 1000);
  view.sync();
});

inheritVelocity (per layer, [-2,2]) adds a fraction of the emitter's velocity to each spawned particle; emission.rateOverDistance spawns particles per pixel traveled (uniform trail density at any speed). Both are world-space only. Determinism extends to identical (document, seed, sequence of (dt, emitter position)).

Each LayerSim exposes read-only per-particle typed arrays (x, y, velX, velY, age, lifetime, rotation, …) plus count and capped. Use computeRenderState(layerSim, buffers) to evaluate size / color / flipbook-frame into your own buffers if you're writing a custom renderer — keeping all renderers identical to the preview.

Schema v3 (Tier-1 feature surface)

schemaVersion 3 adds per-layer feature modules, each null = off: noise (curl turbulence), bySpeed (speed-driven size/color/rotation remaps), startColor (per-particle spawn tint — two-gradient lerp or a ≤16-color palette), randomFlip, render (velocity alignment + speed stretch), collision (floor/rect planes), subEmitters (depth-1 refs to sibling layers), and trail (per-particle ribbons; v9 adds a connect mode). overLifetime.velocity gains additive x/y/orbital/radial tracks; circles gain innerRadius/arc/ arcMode/arcSpeed (donut + arc sweeps; cones get the arc modes too); bursts gain cycles/interval/probability. The layer cap is 8 (was 4). Exported types: NoiseConfig, BySpeedConfig, StartColor, RandomFlip, RenderConfig, CollisionConfig, SubEmitterRef, TrailConfig, ArcMode, SubTrigger, RGBAColor. v1/v2 documents migrate forward losslessly and behave bit-identically. Feature behaviors land milestone-by-milestone (TIER1_PLAN); a document using a not-yet-implemented module validates with an "unimplemented" warning and the field stays inert. The machine-readable reference for every field is the JSON Schema shipped with this package (@particlr/runtime/particle.schema.json).

Schema v4 — point attractor / vortex + host hook

schemaVersion 4 adds a per-layer attractor (point attractor / vortex; null = off) and a per-layer attractorInfluence ([-2, 2], 0 = off). The document attractor: { x, y, strength, tangential, radius, falloff, killRadius } applies a radial (strength) and tangential (tangential, orbiting) acceleration in px/s² over particle ageNorm to particles within radius, in the layer's sim frame (local ⇒ the point rides the emitter). Positive tangential orbits clockwise on screen. killRadius (0 = off) consumes particles that reach it — firing death-trigger sub-emitters. The force is applied to stored velocity (like gravity, so it composes with drag/collision/bySpeed), after the gravity add and before drag. strength/tangential are constant/curve only (zero PRNG draws). Exported types: AttractorConfig, AttractorFalloff, ATTRACTOR_FALLOFFS, DissolveConfig.

Host attractor hook. effect.setAttractor(x, y, strength, radius) drives a transient attractor from game code; effect.clearAttractor() removes it (a non-positive radius clears too). Coordinates are parent-frame — the same frame as setEmitterPosition — and are converted per layer (world layers use them as-is; local layers subtract the step-end emitter position). The host force uses a fixed smooth falloff, is radial only (no tangential, no kill), and is scaled per layer by attractorInfluence — so it is inert on any layer with influence 0 (the migration default, hence a no-op on every v1–v3 document). The last call before a step() wins, and the value persists across step() and reset() until cleared.

const fx = new Effect(doc, { seed });
fx.setAttractor(pointerX, pointerY, 800, 240); // suck particles toward the cursor
// …later…
fx.clearAttractor();                            // release them

Playback control & completion (host API)

Two host-only handles on Effect — no document surface, no schema change.

effect.timeScale multiplies the dt you pass to step(). It is applied as the very first thing step() does, so it composes with the existing guards:

fx.timeScale = 0.25;  // slow-motion
fx.timeScale = 3;     // fast-forward (each step still clamps to MAX_DT, so a
                      // fast-forward can never explode emitters)
fx.timeScale = 0;     // hit-stop: freeze the effect in place. `timeScale = 0` is
                      // exactly `step(0)` every frame — the clock and every
                      // particle hold, and a pending setEmitterPosition target
                      // stays queued for the next un-paused step.

The setter normalizes: any non-finite, negative, or zero value stores 0 (paused) — it never throws. timeScale is sugar over the dt you already control: timeScale = s; step(dt) is bitwise-identical to timeScale = 1; step(dt·s). One consequence worth knowing: under slow-motion the emitter still traverses its full pending segment within the (smaller) scaled step, so its implied velocity (target − start) / scaledDt rises — a host that moves the emitter in real time correctly reads as fast in particle-time (inheritVelocity, world trails). That is authentic game slow-mo, not a bug. timeScale persists across step() and reset() until reassigned.

effect.onDone is a single completion callback (assign a function or null; a host wanting fan-out wraps it). It fires synchronously at the end of the step() on which the effect first reads isDone (a non-looping effect past its duration with zero live particles) — after all state is committed, so inside the callback isDone === true and particleCount === 0. It fires at most once until reset(), which re-arms it (the onDone property itself survives reset(), like timeScale). Attaching a callback after the effect has already finished still fires it on the next step() (latch-on-fire, not on transition — friendlier to hosts that attach late). A paused step (dt ≤ 0, including timeScale = 0) runs no completion check. Looping effects never fire (isDone is always false), and prewarm never fires. Calling ordinary API from inside the callback — including reset() — is legal and safe; the callback runs after the step body, so it can never affect the state that step produced.

fx.onDone = () => {
  view.destroy();          // tear down the renderer
  removeFromActiveList(fx);
};
app.ticker.add((t) => { fx.step(t.deltaMS / 1000); fx.view.sync(); });

Exposed parameters (schemaVersion 6, A9)

One boom.prt, N weapons. A document may declare named parameters, and well-known knobs may bind to one. The host drives them per instance — no document mutation — so a pistol muzzle flash and a rocket blast are the same .prt at intensity = 0.3 and intensity = 1.0. Parameters come in two kinds, each with its own typed accessor pair (mirroring VFX Graph's SetFloat/SetVector4 split):

• scalar ({ kind: "scalar", name, default, min, max }) — driven by setParam(name, value) / getParam(name); multiplies one of seven scalar knobs.
• color ({ kind: "color", name, default: RGBAColor }, schemaVersion 8) — driven by setColorParam(name, r, g, b, a) / getColorParam(name); drives the layer-level tint knob (layer.tintParam). Channels are inherently [0,1]-clamped, so a color param has no min/max. One spell document, N element colors: a single color param bound to tintParam on several layers recolors the whole effect (fire → frost → poison) from one setColorParam.

// pistol / rifle / rocket from one document
const fx = new Effect(boomDoc, { seed });
fx.setParam("intensity", tier);   // 0.3 … 1.0
fx.getParam("intensity");         // read back the current (clamped) value

• Author surface. doc.params: { name, default, min, max }[]. Each bindable knob carries an optional sibling reference field (…Param: string | null); null/absent = unbound. The multiply is multiply-only and applies to the knob's evaluated value — never its stored track keys — so a bound knob at param value 1 is an IEEE-exact no-op (a params-carrying document renders byte-identically to one with params: []). The authoring identity is the value 1 ("default 1 = as authored"); the editor seeds a new param as { default: 1, min: 0, max: 2 }.
• Setter behavior. setParam clamps value into the param's authored [min, max]; a non-finite value is ignored (no-op) and an unknown name is a silent no-op — house tolerance, no throws (cf. timeScale normalizes, setAttractor clears). getParam returns the authored default until the first setParam, and NaN for an unknown name. setColorParam is the color sibling: it clamps each channel to [0,1], and any non-finite channel rejects the whole call (no partial write); an unknown name OR a scalar-kind name is a silent no-op. getColorParam returns a copy of the current RGBA (the authored default until the first setColorParam), or null for an unknown / scalar-kind name. The scalar and color namespaces are independent — the typed accessors never cross (a getParam on a color name is NaN, a getColorParam on a scalar name is null). For sim-consumed knobs (rate/speed/life/gravity) the last call before a step() wins; the render-path knobs (size/opacity/tint) are frame-live — a set shows in the very next render even while paused (timeScale = 0 hit-stop, a paused preview), no step needed. Param values (scalar and color) persist across reset() (like timeScale) — so "scrub with intensity = 0.7" replays exactly.

The bindable knobs, and whether a mid-flight change is live (already-alive particles respond next step/render) or future-spawn (baked at spawn):

Binding field	Knob	Kind	Mid-flight
emission.rateOverTimeParam	continuous emission rate	scalar	emission timing (future)
emission.rateOverDistanceParam	rate-over-distance (world trails)	scalar	emission timing (future)
initial.speedParam	launch speed	scalar	future spawns only
initial.lifeParam	particle lifetime	scalar	future spawns only
initial.sizeParam	particle size (render multiply)	scalar	live
overLifetime.velocity.gravityParam	gravity vector	scalar	live
opacityParam (layer-level)	particle alpha	scalar	live
tintParam (layer-level)	RGBA tint on the finished color chain	color	live

speed/life are baked into pool state at spawn, so changing them never rescales particles already on screen; size/gravity/opacity/tint are re-evaluated every step/frame, so they retune the whole live population for free. The tint multiplies each particle's finished RGBA (gradient × startColor × bySpeed × tint × opacity — tint before opacity; both commute) with an implicit base of white {1,1,1,1}, so an unbound tint is the untouched render path and a tint at white is byte-identical. Params join the determinism tuple (below): the vector of current scalar and color param values in force at each step() is a per-step host input, exactly like dt and timeScale.

Schema v4 — dissolve / alpha erosion

schemaVersion 4 adds a per-layer dissolve (null = off): { frequency, scroll, edgeWidth, edgeColor }. It is a renderer-only effect — the sim is untouched, no PRNG draws, no pool columns. The particle's final render alpha is the erosion progress: whatever modules drive per-particle alpha (the color gradient, bySpeed, startColor) drive the dissolve, so a fading puff erodes through an internal procedural noise tile instead of fading uniformly. frequency ∈ (0, 64] is the noise repeat across the sprite; scroll is UV/s over the effect clock (uTime = effect.time, so the burn is exact under scrub and golden replay); edgeWidth ∈ [0, 1] is the soft erosion band; edgeColor (null = off) is a hot-rim RGBA tint that glows at the erosion edge. It keeps one draw call per layer (no render targets, no extra passes) via a forked ParticleContainer shader. Exported type: DissolveConfig.

Renderer parity — WebGL only. The dissolve fork ships both a GLSL and a WGSL source, but only the WebGL (GLSL) path is verified: the golden-frame suite runs SwiftShader WebGL and there is no WebGPU golden lane yet, so L4 preview/runtime parity is attested on WebGL only; the WGSL path ships unverified. Pin preference: "webgl" in your host app.init if you rely on byte-exact parity (the editor preview, the golden harness, and the live sample host all do). Dissolve does not erode a trail ribbon (E25) — a trail on the same layer renders un-eroded through its separate mesh shader.

Schema v5 — cheap wins (limit-velocity · random-between-curves · hue jitter · flipbook upgrades)

schemaVersion 5 adds four companion features. Every one reuses an already-drawn per-particle uniform, so the batch adds zero new PRNG draws and zero pool columns — a v4 document migrates forward bit-identically (limitVelocity defaults null; a flipbook gains randomStartFrame: false, frameOverLife: null; the new track/startColor modes are opt-in enum values).

• limitVelocity: ScalarTrack | null (per layer; constant/curve only — no range/randomBetweenCurves, zero draws). A speed cap over particle ageNorm, applied to stored velocity right after drag and before the position write, so it persists and composes like drag: speed = √(velX²+velY²); when speed > cap the velocity is scaled down to cap. cap = 0 freezes particles in place (E27).
• randomBetweenCurves ScalarTrack mode — { mode, a: CurveKey[], b: CurveKey[] }. Each particle draws a stable blend factor r for life and evaluates lerp(evalCurve(a, t), evalCurve(b, t), r), so the population reads as a mix of two shapes. Valid only on the eight per-particle over-lifetime tracks that own a reserved spawn uniform — overLifetime.size, overLifetime.rotation, and overLifetime.velocity.{drag, speedMultiplier, x, y, orbital, radial} — consuming the SAME uniform range does (hence no new draw). The validator rejects it on emission.rateOverTime and every constant/curve-only track.
• hueJitter startColor mode — { mode, degrees } (degrees ∈ [0, 180]), mutually exclusive with the two-gradient/palette startColor modes. At spawn it reuses the existing startColor uniform (draw 19) to store a per-particle hue offset ∈ [−degrees, +degrees]; at render it hue-rotates the over-lifetime gradient colour per particle (alpha untouched) instead of multiplying a tint — so a warm base reads as a natural spread of hues.
• Flipbook randomStartFrame: boolean + frameOverLife: ScalarTrack | null (both render-only, deterministic, NOT folded into the statehash). Frame-index precedence (E30): frameOverLife — a deterministic 0..1 position across the sheet evaluated over ageNorm — overrides the mode entirely; else mode: "random" keeps its stable per-particle frame (randomStartFrame ignored); else loop/once with randomStartFrame adds a per-particle start offset (reusing the draw-13 frame uniform), then wraps (loop) or clamps (once). With randomStartFrame: false and frameOverLife: null the frame is exactly ⌊age·fps⌋, byte-identical to v4.

Schema v9 — connect-ribbon trail mode + sub-emitter inheritance

schemaVersion 9 adds trail.mode: "perParticle" | "connect" (migration injects "perParticle", so every v8 document behaves bit-identically).

• "perParticle" — the pre-v9 behavior: each particle carries its own polyline of recent positions.
• "connect" — the layer emits ONE ribbon threaded through all of its currently-live particles' current positions (energy beams, lightning, chains — a Shuriken connected ribbon / Effekseer track). No position history is kept, so the ring buffer is not allocated and maxPoints/minVertexDistance are ignored (E36). Vertices are ordered oldest → newest by a stable per-particle spawn ordinal, so a particle dying mid-ribbon (swap-removed from the pool) never reorders the survivors — the order is deterministic and independent of pool compaction. width/color sample over ribbon t with t = 0 at the newest particle (head), matching per-particle trails; color null ⇒ each vertex takes its own particle's current render RGBA (a non-null gradient scales its alpha by that vertex's particle alpha). Fewer than 2 live particles ⇒ no ribbon (empty geometry, never a degenerate quad). The stable ordinal is assigned from the per-layer spawn counter, so connect mode adds zero PRNG draws; it reuses the same mesh/blend render path as per-particle trails.

Sub-emitter property inheritance. Each SubEmitterRef gains three booleans beside inheritVelocity — inheritColor, inheritSize, inheritRotation (migration injects false). At the trigger event the parent particle's state is captured and applied to each child at spawn; the child's own PRNG stream is never touched (inheritance modifies drawn RESULTS, zero new draws).

flag	captured from the parent (at the event moment)	applied to the child
inheritColor	the parent's sim-side RGBA — the over-life gradient at ageNorm × the startColor tint including any hueJitter hue rotation, excluding bySpeed and host-param tints (those are render-only)	a per-particle multiply on the child's finished color chain (a dedicated inherit-RGBA column, allocated only on a layer that is the target of ≥1 inheritColor ref) — after startColor, before bySpeed
inheritSize	the dimensionless over-life size FACTOR evalScalarTrack(overLifetime.size, ageNorm, rand0) (1 when the track is null) — not the px size (px × px is nonsense; the factor gives "a shrinking parent spawns smaller children")	the child's drawn size is multiplied by that factor (baked into sizeInit)
inheritRotation	the parent's current rotation in degrees	added to the child's drawn rotation

Capture is per-PARENT-layer gated: a parent whose refs all leave every flag false records the flat pre-v9 event and does no capture work, so a document that opts out is bit-identical to schemaVersion 8. The full color chain is gradient × startColor × inherit × bySpeed × tint × opacity.

Behavioral guarantees (edge cases)

#	Case	Behavior
E1	dt larger than 1/20 s (tab unhide)	clamped to 1/20 s; never sub-steps
E2	dt <= 0	no-op
E5	prewarm	simulates one full duration of continuous emission (bursts suppressed) before the first visible frame
E6	looping: false ends	emitters stop; live particles finish; isDone becomes true at 0 particles
E7	pool full (maxParticles)	new spawns dropped silently; layer.capped flags it
E8	seeking a time	reset(seed) then step(1/60) to the target — exact, thanks to determinism
E9	serializing while playing	serializeParticle uses the authored document; playback state never serializes
E15	teleportEmitter (respawn/wrap)	jump with no velocity and no spawn interpolation across the gap; resets the distance accumulator
E16	prewarm on a world-space layer	prewarm runs at the initial emitter position with zero velocity; particles pile there
E27	limitVelocity cap = 0	clamps stored velocity to zero — particles settle in place (valid, not an error)
E30	flipbook frame precedence	frameOverLife (deterministic) > mode:"random" > loop/once + randomStartFrame

Pixi adapter (@particlr/runtime/pixi)

class PixiParticleRenderer {
  constructor(effect: Effect, opts?: { renderer?: unknown });
  readonly container: Container; // one ParticleContainer per layer
  readonly warnings: string[];   // e.g. a missing user texture fell back to a built-in
  sync(): void;                  // call after effect.step(dt)
  destroy(): void;
}

Built-in textures are generated procedurally (pure math, no asset files, no canvas) so output is identical across browsers and GPUs. Blend modes map directly to Pixi's (normal/add/multiply/screen).

Embedded user textures ("user:<name>" refs backed by a data URL in doc.textures) decode asynchronously: the layer renders the soft-circle built-in until the image is ready, then swaps in. Decoded user textures are cached by data URL and shared across renderer instances for the lifetime of the page — the runtime does not evict them (per-texture refcounting is a v1.5 concern; a document embedding many large textures is out of scope for v1).

Determinism contract

Given identical (document, seed, sequence of (dt, emitter-position, host-attractor state, timeScale, param-values) tuples), output is bit-identical — the emitter position (setEmitterPosition), the host attractor (setAttractor parameters), timeScale, and the exposed-parameter values (scalar setParam, schemaVersion 6, and color setColorParam, schemaVersion 8) are per-step host inputs exactly like dt (schemaVersion 4 + host-API amendments). The runtime never reads wall-clock time, Math.random, or any global — the only randomness is a seeded mulberry32 PRNG, one stream per layer. This is what makes seek/scrub exact and enables golden-frame testing.

The .prt format

.prt is a small, versioned, declarative JSON document. Its machine-readable reference is the JSON Schema shipped with this package — particle.schema.json (import schema from "@particlr/runtime/particle.schema.json").

Because the format is small, documented, and agent-readable, a coding agent can author or tweak effects directly — "make the explosion 20% punchier and shift it orange" is a JSON edit against a published schema, no editor required.

License

MIT. See LICENSE.