API reference

Part of the renderer SDK contract for extension renderers.

Called after this container's child list changed - an insert, a removal or a reorder - once the list and the caches derived from it are consistent again. Override in containers whose own state is derived from their children, such as a layout container that re-flows them.

Iterate the same frozen, cached document-order snapshot children returns - mutating the container after obtaining this iterator does not change what it yields (see children's doc comment for the full snapshot/invalidation contract).

Append one or more children to the end of the child list. Each child is detached from its previous parent (if any) before being added.

Insert child at index in the child list. The child is detached from any previous parent first. Throws if index is out of bounds, if child has already been destroy()ed, or if child is an ancestor of this container (would create a cycle). Self-as-child is a no-op. When child already belongs to this container the call is a pure reorder: the node keeps its parent, its stage and its keyboard focus.

Register a font by its raw bytes for use inside this node's SVG render. The family name must match the font-family value used in your CSS. Registering the same family twice replaces the previous entry. ts const bytes = await loader.load(Asset.type('binary', 'roboto.woff2')); label.addFont('Roboto', bytes, 'woff2');

Release keyboard focus from this node if it currently holds it.

Contribute this node to the render plan under construction: skip it when it is destroyed, invisible, or outside the builder's cull rect, otherwise emit it as one plan entry. Custom node types override this to add their own admission rules and must call super.collect(builder, seq) to keep the skip semantics. seq orders the entry within its parent; omit it to append. Part of the renderer SDK contract for extension renderers.

Compute a full CollisionResponse between this shape and target. Returns null in two cases: - the shapes do not overlap, **or** - the specific shape-pair combination does not support response generation (e.g. Line against any shape, Ellipse against Ellipse or Polygon). Use intersectsWith for a universal boolean overlap check that works across all supported shape pairs.

A container has no geometry of its own, so a point hits it when it hits any child. A RenderNode.hitArea replaces that union with the shape, which is how a bare layout container becomes clickable in its own right.

Destroy this container and every node beneath it. Ownership is by containment: detaching a subtree only unlinks it, so the descendants' GPU-backed resources (cached textures, filters, render textures) and signal listeners would outlive the tree that owned them and leak on every scene change. Idempotent: a second call is a no-op.

Request keyboard focus for this node through its owning focus service.

Axis-aligned bounding box of this node in its GLOBAL-transform space. That is world space for ordinary nodes, but GROUP-LOCAL space for nodes inside an engaged RetainedContainer transform group (the group matrix is applied on the GPU, not here) - this is deliberate and matches the rendering convention. For a true world-space extent of such a node, lift this rect by the group's getWorldTransform matrix. Pass out to receive a copy you own. Without it the return value is this node's **cached** rectangle, rebuilt in place whenever the transform or the local extent changes - retaining it across frames hands you a value that silently moves. The cache is why the no-arg form does not allocate: this runs per node per frame for culling and hit-testing.

This node's untransformed extent in its own local coordinate space. Returns the LIVE internal rectangle, not a copy: it is read on hot paths (updateBounds re-reads it on every transform-dirty recompute, i.e. potentially every frame for a moving node), so copying it here would add a per-frame allocation. It is therefore typed as a ReadonlyRectangle - reads are unchanged, writes are rejected at compile time. Writing to it directly would skip the bounds/content invalidation the engine needs, leaving culling, hit-testing and retained render fragments on a stale extent. A custom Drawable that owns its own size sets it through setLocalBounds, which writes and invalidates in one step.

Return the outward-facing edge normals used by the SAT solver. The array should be cached and reused across calls.

The node's TRUE world-space transform, composed through every transform-group boundary (RetainedContainer) in the ancestor chain. getGlobalTransform deliberately stops at the nearest engaged boundary (descendants resolve group-RELATIVE transforms; the renderer multiplies the group matrix back in on the GPU), so it is the right space for rendering but the wrong one for spatial queries. Use THIS accessor whenever a real world position/orientation is needed - picking, spatial audio, physics, world-space math against nodes outside the group. Without any engaged boundary ancestor it returns the exact getGlobalTransform matrix (same instance, no extra work). With one, it lazily caches groupLocal × groupWorld and revalidates on read via version/stamp compares - including runtime space flips such as RetainedContainer's deep-barrier sub-branch escape.

Test whether this shape overlaps target using a fast boolean algorithm (no penetration depth or normal computed). Prefer this over collidesWith when only the yes/no result is needed.

Mark this node's visual content dirty without going through a standard setter - e.g. a custom Drawable subclass backed by externally mutable data (the pattern TileChunkNode in @codexo/exojs-tilemap already uses via its own _chunk.revision compare). Call this after mutating such state so the Track-B retained-plan skip does not serve a stale frame for this node.

Project this shape onto axis and write the scalar min/max into interval. Used internally by the SAT solver.

Remove child from this container. No-op if not present.

Remove children in the half-open range [begin, end). Defaults to the entire child list. Throws if the range is invalid.

Remove a previously registered font and trigger a re-render.

Raw rendering entry point. Direct backend access - bypasses the RenderPlan pipeline machinery. Prefer the high-level RenderingContext.render path via the owning RenderingContext wherever possible.

Change the logical dimensions and rebuild the mesh. Prefer this over setting width and height individually when changing both at once - it triggers a single re-render instead of two.

Write this node's local extent and run the bounds invalidation the change implies - the node's own bounds flag, the ancestor bounds cascade, and the content-dirty stamp that keeps retained fragments from replaying the old extent. This is the only supported way to resize a node from outside getLocalBounds; the rectangle itself is handed out read-only so the invalidation cannot be forgotten. Built-in drawables (Sprite, Text, BitmapText, Mesh, ...) and custom ones alike go through here. Part of the renderer SDK contract for extension renderers.

When true and interactive is also true, this node will be automatically repositioned to follow the pointer during a drag gesture. The framework captures the pointer offset at drag-start so the node doesn't snap to the cursor position. Both interactive and draggable must be set for dragging to work - a draggable but non-interactive node will never receive pointerdown and therefore cannot start a drag.

When true, this node can receive keyboard focus - via focus, Tab traversal, or app.interaction.focus(node) - and is delivered key events through onKeyDown / onKeyUp while focused, or while any of its descendants holds focus (key events bubble up the parent chain like pointer InteractionEvents do). A Widget additionally has to be enabled: disabling one takes it out of the Tab order and rejects programmatic focus, without touching this flag.

Optional pick shape in this node's LOCAL space, replacing the bounds test contains would otherwise perform. Defaults to null. The world-space point is mapped through the inverse of the node's global transform before the shape is tested, so the region follows the node's position, rotation, scale and skew like the rendered output does - the shape itself is never re-transformed and never needs updating when the node moves. Affects picking only. Bounds, culling and rendering ignore it entirely, and because the interaction system finds candidates by their bounds, a hit area reaching outside the node's bounds is only reliably picked where the two overlap. Use it to shrink or reshape a pick region, not to grow one. The shape is the caller's: it is read live on every hit test, so mutating it in place takes effect immediately, and the node never destroys it.

Optional human-readable identity for this node. Defaults to null. Purely a label the engine never interprets: useful for debugging, find-by-name lookups, prefab references, and as a stable key when merging serialized state back onto an existing tree. Not required to be unique.

Tab-traversal order among focusable nodes in the same focus scope. Lower values are visited first; equal values keep document (tree) order.

Bottommost edge of the subtree - see left.

Bake this node's subtree into a RenderTexture once and replay that texture until the subtree changes, instead of walking and drawing it every frame. Worth it for a subtree that is expensive to draw and rarely changes. The cache is invalidated by anything that moves the node's world bounds - the node's own transform included - so a node that animates re-bakes every frame and is strictly slower than not caching it at all. Setting it to false frees the texture immediately.

Resolution the cacheAsTexture texture is baked at, in device pixels per logical unit. 'inherit' (the default) matches the surface the cache is composited into, so enabling the cache does not soften the picture on a HiDPI display. Pin it to a number to trade sharpness for memory and bake cost - a cache is resolution² texels, so 1 on a DPR-3 phone is a ninth of the VRAM and a ninth of the fill per re-bake. Changing it invalidates the cache.

Snapshot of the current children in document order. Frozen and cached - repeated reads return the same array reference until the next structural change (addChild/removeChild/setChildIndex/ swapChildren/etc.), which invalidates it. A reference to a previous snapshot is unaffected by later changes - it keeps reflecting the child list as it was at the time of the read. Mutating methods (push, splice, ...) throw in normal (strict-mode) usage; go through addChild/removeChild instead so parent linkage, stage propagation, and bounds invalidation stay consistent.

When true, descendants are geometrically clipped to clipShape. Unlike mask (which is alpha/visibility masking), clip is a hard geometric boundary: - clipShape === null - clip to this node's world-space bounds (getBounds), using the GPU scissor fast path. - clipShape is a Rectangle - clip to that world-space rectangle via scissor. - clipShape is a Geometry - clip to the geometry's silhouette via the stencil buffer (WebGL2). Only fragments inside the shape survive. Clipping wraps the node's final (filtered/masked) output and acts as a render barrier: draw commands are never reordered or batched across the clip boundary.

Clip region used when clip is true. A Rectangle (or null for the node's bounds) maps to the scissor fast path; a Geometry maps to the stencil path. Has no effect while clip is false.

When false, this node is never culled by the viewport check and is always considered in-view. Defaults to true.

Custom rectangle used for viewport cull intersection test. When set, replaces the default node bounds in cull checks. Set to null to restore default bounds-based culling.

true once destroy has run on this node. A destroyed node has released its pooled resources (transform/bounds), has been unlinked from its parent, renders nothing, and must not be reused or re-attached. The render plan skips a destroyed node, so even one handed straight to a renderer as a detached root contributes nothing.

Rendered height of the whole subtree - see width.

Leftmost edge of the subtree in the node's global-transform space. These four edges are read straight off the aggregate bounds so they always span exactly width/height and agree with each other. They cannot be reconstructed from position and origin: origin is in local pixels and a container's own local bounds are empty, so the rendered extent comes entirely from the children.

The mask source that controls visibility of this node's render output. See MaskSource for accepted source types and their semantics. Setting to null removes any active mask. Setting a RenderNode that is this is rejected (a node cannot mask itself). Indirect cycles (a.mask = b; b.mask = a) are rejected as well: the candidate's mask chain is walked and any cycle - whether it closes on this or was already present in the chain - fails the assignment.

When true, material-aware overlap reordering is disabled for this node's draw-order scope. Draw commands are submitted in exact document order (after scope-local z-sorting), preserving the painter's guarantee irrespective of material compatibility or AABB safety analysis. Adjacency coalescing of consecutive same-material draws still applies; it does not change visual output order.

Resolves when the current render pass finishes writing to the canvas. Await before adding to a scene if you need the first frame to be visible immediately.

Rightmost edge of the subtree - see left.

Rotation angle in degrees. Wraps via trimRotation on assignment.

Horizontal skew angle in degrees. Shears the node along the X axis (positive values lean the top edge right). Combines correctly with rotation and scale.

Vertical skew angle in degrees. Shears the node along the Y axis (positive values lean the left edge downward). Combines correctly with rotation and scale.

Topmost edge of the subtree - see left.

Rendered width of the whole subtree, in the node's global-transform space. Unlike Sprite.width - which scales an unscaled texture frame - a container has no intrinsic local size, so this reads the aggregate bounds directly. Those are already scaled, so multiplying by scale again would count it twice. Writing rescales scale.x to make the subtree render at the requested width.

Fired when this node loses keyboard focus.

Fired when a pointer requests a context menu over this node - right-click, or a long-press/touch gesture that has an attributable pointer. Bubbles like the other pointer events, so a scene-wide fallback can listen on an ancestor. Carries no native event - whether the browser's own menu appears is decided by ApplicationOptions.input.allowNativeContextMenu, independently of this. Requires an attributable pointer: a pointerless keyboard-only request (the context-menu key, or Shift+F10, with no pointer ever having touched the surface - see ContextMenuRequest's doc comment) has nothing to hit-test or bubble with, so it never reaches this per-node event. It only ever reaches the engine-wide, scene-graph-independent app.input.onContextMenu.

Fired on every pointer-move while this node is being dragged. Does not bubble.

Fired when the drag gesture ends (pointer-up or cancel). Does not bubble.

Fired once when a drag gesture begins on this node. Does not bubble.

Fired when this node gains keyboard focus.

Fired for each key pressed while this node - or a descendant of it - holds focus. Bubbles; see KeyEvent.

Fired for each key released while this node - or a descendant of it - holds focus. Bubbles; see KeyEvent.