Four issues in one pass:
1. Move snaps to grid
MoveOverlay's cursor is now snapped via snapPointToGrid([m.x, m.y],
GRID_STEP=0.5). Matches the 3D shelf tool's placement step so 2D
and 3D placement feel identical.
2. Move translates the actual rendered SVG, not a ghost
MoveOverlay no longer portals a 50%-opacity ghost. Instead it finds
the rendered [data-node-id] <g> inside the floor-plan scene and sets
its `transform` attribute imperatively each pointermove. The inner
group's translate(px pz) rotate(deg) stays untouched — the outer
transform composes as a pure delta. Same "smooth move" pattern as
the 3D MoveRegistryNodeTool: no React re-renders, no zundo bloat,
the actual shape follows the cursor with full fidelity.
3. Click commits the position (previously did nothing)
Switched from `window click` (with capture + composedPath check)
to `window pointerup`. Pointerup fires reliably regardless of
click-vs-drag semantics in the floor-plan panel's pointer-down
handlers (which can preventDefault on certain modes and suppress
the synthesized click). Target check uses
`target.closest('[data-floorplan-scene]')` instead of composedPath
for cross-browser SVG reliability.
4. Clicking in floor plan with shelf tool active creates a shelf
Root cause: `isFloorplanGridInteractionActive` is a hardcoded OR of
build/move modes that doesn't include registry kinds, so the panel
never emits `grid:click` / `grid:move` for them. Shelf tool listens
on those events; without them, clicks were silently dropped.
Fix: new `isRegistryToolBuildActive` derived from
`mode === 'build' && tool != null && nodeRegistry.has(tool)` — added
to the OR chain. Future Phase 5 kinds (fence, item, etc.) inherit
floor-plan placement automatically the moment they register a tool.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two new files in packages/editor/src/components/editor-2d/ keep the
work out of the 18k-line floorplan-panel.tsx monolith. The panel
itself gets only four tiny additions (two imports, two component
mounts, one data attribute).
<FloorplanRegistryActionMenu>
- Reads useViewer.selection — when a registered kind is selected and
we're not in a move state, queries the rendered [data-node-id] <g>
for its bounding rect (polled via rAF for pan/zoom/move reactivity).
- Portals an HTML overlay above the bounding box with the existing
<NodeActionMenu>. Buttons gated by def.capabilities:
* Move → setMovingNode(node)
* Duplicate → structuredClone + schema.parse + createNode + set
movingNode (placement cursor) — matches 3D duplicate UX.
* Delete → deleteNode(id) + clear selection.
- Same visual styling as the legacy <FloorplanActionMenuLayer> per
kind, but driven by registry data.
<FloorplanRegistryMoveOverlay>
- Activates when useEditor.movingNode is a kind with def.floorplan.
- Listens on window for pointermove (to track cursor in floor plan
meters via the scene <g>'s getScreenCTM — matches the legacy
getSvgPointFromClientPoint coordinate path so cursor → meters
accounts for pan/zoom/building rotation).
- Renders a 50%-opacity ghost via portal into the floor-plan scene
<g>. Builder reused from def.floorplan — no per-kind ghost code.
- Click commits via updateNode({ position: [cx, oldY, cz] }) and
clears movingNode. Clears `isNew` metadata on duplicates so they
don't loop. Esc cancels.
floorplan-panel.tsx touches:
- Two imports (action menu + move overlay).
- data-floorplan-scene="" attribute on the floorplanSceneRef <g>
so the overlay can find the scene without sharing a ref.
- <FloorplanRegistryActionMenu /> mounted alongside the legacy
action menu layer.
- <FloorplanRegistryMoveOverlay /> mounted inside the SVG tree
alongside the registry render layer.
FloorplanRegistryLayer: also stopPropagation on click events so the
outer SVG's onClick={handleBackgroundClick} doesn't deselect right
after our pointerDown sets selection. Fixes "click-in-2D doesn't
select" bug.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The drag-with-grab-cursor model was wrong design — 3D doesn't drag,
it uses select → inspector "Move" button → click-to-place. Floor
plan should match. Drag also had a coord-conversion bug (used outer
<svg> CTM instead of the floorplanSceneRef <g>, so screen→meter
conversion didn't account for the floor plan's pan/zoom/building-
rotation transforms).
Strips the drag pointerdown/move/up handlers, the imperative
transform override, the temporal pause bracketing, and the global
window listeners. Cursor goes back to 'pointer'. Only click-to-
select remains.
The right pattern (move via inspector / action menu + cursor-driven
placement) needs:
- Registry-aware FloorplanActionMenuLayer path
- Generic movingNode handler in floor-plan-panel for any registered
kind with capabilities.movable
- Shared floorplanSceneRef for accurate coord conversion
Both flagged in the plan as Phase 4 follow-on gaps with their
acceptance criteria. 3D-realtime-sync-while-moving is documented as
deferred (legacy doesn't do it for any kind either; design + ship
in a dedicated PR later).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three additions on top of the floor-plan registry contract:
1. def.toolHints + RegisteredToolHelper (registry contract for the
shortcut hint panel)
- New ToolHint type in core: { key, label } static array.
- Added `def.toolHints?: ToolHint[]` to NodeDefinition.
- New <RegisteredToolHelper hints={...}> in editor — same visual
styling as WallHelper / ItemHelper but data-driven.
- HelperManager: registry-first check before falling through to the
hand-written per-tool switch. Per-tool helper files get deleted
as their kind migrates `toolHints` in.
- Shelf + spawn definitions ship toolHints today; wall ports in
Phase 3 Milestone C alongside its tool/affordance port.
2. Floor-plan interaction layer (selection + drag-to-move)
- <FloorplanRegistryLayer> now wraps each entry in an interactive
<g>:
* Click → useViewer.setSelection({ selectedIds: [id] }).
Selection visual is a thicker accent-colored stroke applied
via withSelectionStyle() recursion through the FloorplanGeometry
tree — kinds don't author selection decoration.
* Drag → imperative SVG transform during the gesture, single
updateNode commit on pointerup. Same "smooth move" pattern as
MoveRegistryNodeTool for 3D drag: no per-tick store update,
no React re-render storm, no zundo bloat. Coordinate
conversion via svg.getScreenCTM().inverse().
* useScene.temporal.pause/resume brackets the gesture so one
drag = one undo step.
- Global pointermove / pointerup listeners so the gesture survives
the cursor leaving the entry's bounding box (matches the legacy
elevator-resize-drag and item-drag patterns in floorplan-panel).
3. Spawn floor-plan builder (deferred wiring)
- buildSpawnFloorplan written but NOT wired on the definition —
spawn already renders in the legacy floorplan-panel.tsx via
`floorplanSpawnEntries`, and wiring def.floorplan now would
double-render. The pure builder lives in nodes/src/spawn/
floorplan.ts ready to wire when the legacy inline branch is
removed (Phase 5 spawn-floorplan migration PR — same shape as
wall's feature flag, but per kind inside the legacy panel).
Plan updated: floor-plan interaction section locks the click/drag
contract in, wall-floor-plan-as-legacy note flags everything advanced
the user sees today as legacy that ports alongside Milestone C.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the floor-plan side of the three-checkbox composition model
documented in wiki/architecture/node-definitions.md. Mirrors the 3D
side (def.geometry → <GeometrySystem> → <ParametricNodeRenderer>) but
emits SVG primitives instead of three.js Object3Ds, and runs inside
the floor-plan panel.
Type-side (packages/core/src/registry/types.ts):
- New FloorplanGeometry tagged union covering path / polygon /
polyline / rect / circle / line / group. FloorplanStyle props map
straight to SVG attributes. Coordinates are level-local meters;
rotations are radians (three.js convention).
- New `def.floorplan?: (node, ctx) => FloorplanGeometry | null` field
on NodeDefinition, independent of `geometry` and `renderer`. Re-
exported via packages/core/src/registry/index.ts.
Runtime (packages/editor):
- <FloorplanGeometryRenderer> walks the FloorplanGeometry tree and
emits the matching React-SVG elements. Pure data → DOM; no per-kind
logic.
- <FloorplanRegistryLayer> reads the active levelId, walks the level
subtree, looks up each node's def.floorplan, builds a
GeometryContext, calls the builder, and renders the output via
FloorplanGeometryRenderer.
- Mounted in floorplan-panel.tsx just before <FloorplanMarqueeLayer>
so registry-driven kinds layer above legacy inline content.
Shelf migration (proof port):
- New nodes/src/shelf/floorplan.ts — buildShelfFloorplan(node) emits
a group with the rotation/translation transform and a width × depth
rectangle in the shelf's color. Brackets omitted (hidden under top
board from above).
- Wired to shelfDefinition.floorplan. Shelf now appears in the floor
plan view for the first time (was missing from the legacy panel's
inline switch).
Pattern proven; every future kind migrating in Phase 5 follows the
same shape: a pure (node, ctx) => FloorplanGeometry function. As kinds
register their floor-plan builders, the corresponding inline branches
in floorplan-panel.tsx become dead code and can be deleted in the
same PR.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lands the three-checkbox composition runtime documented in
wiki/architecture/node-definitions.md. A kind with only a pure
geometry function now needs zero per-kind React or system code.
Type-side additions (packages/core/src/registry/types.ts):
- New `GeometryContext` (resolve / children / siblings / parent) — read-
only scene access for builders that reference other nodes by ID
(wall miters, door cutouts). Most kinds ignore it.
- New `geometry?: (node, ctx) => Object3D` field on NodeDefinition,
independent of renderer/system. Three orthogonal opt-ins replace the
v0 RendererSource union.
- Re-exported via packages/core/src/registry/index.ts (consumed by
nodes packages through `export * from './registry'`).
Runtime (packages/viewer):
- New <GeometrySystem> (systems/geometry/geometry-system.tsx) walks
dirtyNodes, builds a GeometryContext per dirty node, calls
def.geometry, disposes old children, attaches new ones, clearDirty.
Frame priority 2 (matches the priority shelf's per-kind system had).
Mounted in viewer/index.tsx alongside <RegisteredSystems>.
- New <ParametricNodeRenderer> (components/renderers/parametric-node-
renderer.tsx) — empty <group> + useRegistry + useNodeEvents +
markDirty-on-mount + useLiveTransforms. Mounts hosted children via
<NodeRenderer> recursively. The default renderer for any registered
kind without a custom def.renderer.
- <NodeRenderer> dispatch updated: custom renderer wins, else
geometry-only kinds fall through to ParametricNodeRenderer, else
null (legacy switch fallback). Documented inline.
Shelf migration (proof of the boilerplate collapse):
- Deleted nodes/src/shelf/renderer.tsx (was 45 lines of registry +
handler boilerplate).
- Deleted nodes/src/shelf/system.tsx (was 60 lines of dirty-loop +
dispose plumbing).
- shelfDefinition now: `geometry: buildShelfGeometry`. One line.
buildShelfGeometry is the pure function from geometry.ts that already
existed.
End-to-end effect: registry-driven shelf now mounts via the framework's
generic renderer + system. Parametric edits flow through the same
dirty-driven rebuild path, but the kind ships ~100 fewer lines of
boilerplate. Every future kind that fits the same shape (item, fence
segment, column, etc. as they migrate in Phase 5) follows the same
"one line, one pure function" pattern.
Wall stays on its dedicated def.renderer + def.system — its mitering
needs level-batch context (`ctx.levelData?.miters`, future extension)
that the generic system doesn't yet provide. Decided at Phase 3+, not
blocking Phase 4 acceptance.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Endpoint drags fire markDirty(wallId) on every pointermove tick. The
old behavior rebuilt the dragged wall AND every wall sharing a junction
on every tick — in a 4-corner room with doors, that's 4× the CSG
+miter pass per tick. Visible as drag lag.
New behavior: the dragged wall rebuilds every tick (so the drag tracks
the cursor with full fidelity, cutouts and all). Adjacent walls are
queued in pendingAdjacentByLevel and rebuilt on the trailing edge —
80ms after the dirty stream stops. The corners snap into their correct
miter joins ~80ms after release, which is the standard CAD-app
"rubber-band the dragged element, fix neighbors on commit" pattern.
Module-level singleton state for the queue + timestamp — WallSystem is
mounted exactly once globally, so module state is the right scope.
Expected speedup:
- t-junction drag: ~3× (was 3 walls/tick, now 1)
- 4-corner room with door per wall: ~4×
The trailing flush condition (!hasDirtyWalls && now - lastWallDirtyAtMs
>= DRAG_FLUSH_MS) means single edits (non-drag) pay an 80ms latency
before neighbors miter correctly. Acceptable for now; the real fix is
the affordance/tool port (Milestone C) which will explicitly signal
"drag in progress" so we can drop the heuristic. Until then this is a
substantial drag-perf win for zero risk.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three one-shot console.info calls so the Phase 3 milestone-B parity
check is unambiguous from the browser console alone:
- [wall:registry] system bundle mounted — fires when RegisteredSystems
lazy-loads nodes/src/wall/system.tsx (exactly once per viewer mount
when the flag is on).
- [wall:registry] first WallRenderer mounted — fires once when the
first registry-driven WallRenderer mounts.
- [wall:legacy] first legacy WallRenderer mounted — fires once if the
legacy path is active (flag off, or kind not registered).
Module-level booleans gate the renderer logs so they don't spam in
scenes with many walls. Drop all three alongside the feature flag at
Phase 3 sign-off.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Brings the wall kind onto the registry path when
NEXT_PUBLIC_USE_REGISTRY_FOR_WALL=true; default-off keeps wall on its
legacy path unchanged.
Files added:
- nodes/src/wall/renderer.tsx — thin placeholder-mesh mount point.
Identical pattern to the legacy WallRenderer: registers ref via
useRegistry, marks dirty on mount, renders hosted children
recursively via NodeRenderer. The legacy WallSystem fills geometry
on the next frame regardless of which mount path is active.
- nodes/src/wall/system.tsx — a bundle component that renders
<WallSystem /> + <WallCutout /> (both re-exported from viewer).
Registered via def.system with priority 4 to mirror the legacy
WallSystem's useFrame priority. Zero logic duplication — the
~970 lines of CSG/mitering/cutaway code stays in viewer.
Files changed:
- packages/viewer/src/index.ts — new exports for WallSystem, WallCutout,
and NodeRenderer. The first two so the registry-driven system bundle
can compose them; NodeRenderer so any parent kind (wall, slab,
ceiling, building) can recursively render hosted children without
reaching into viewer internals.
- nodes/src/wall/definition.ts — adds renderer + system fields. Tool
field stays absent (wall placement / endpoint drag remain bespoke
for now; the affordance port is a later milestone).
- nodes/src/index.ts — conditionally appends wallDefinition to
builtinPlugin.nodes based on isWallRegistryEnabled(). With the flag
off, the array is identical to before this commit; with it on,
Phase 0 dispatch shims switch wall to the registry path:
* <LegacySystem kind="wall"> around WallSystem returns null
* <LegacySystem kind="wall"> around WallCutout returns null
* <NodeRenderer> takes the registry-first branch and mounts the
new renderer instead of the legacy switch case for 'wall'
* RegisteredSystems mounts the new system bundle, which re-mounts
the same WallSystem + WallCutout components from viewer
No behavior change with the flag off. With the flag on, behavior should
be byte-identical (same components, same priority, same geometry path).
Manual verification next: place walls, t-junctions, walls-with-doors
with the flag toggled both ways; confirm visual + interactive parity.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lays down the wall folder under @pascal-app/nodes with everything needed
to register the kind, but intentionally without runtime wiring:
- schema.ts re-exports WallNode from core (door/window/item still type
their parentId against WallNode.shape.id, so the schema stays canonical
there for now).
- parametrics.ts declares thickness / height / curveOffset for the Phase 4
inspector. Endpoints and host children are edited via affordances, not
number inputs, so they're not in parametrics.
- definition.ts encodes capabilities (surfaces, selectable, duplicable,
deletable — no movable since wall's move is bespoke endpoint-drag),
relations (hosts doors/windows/items, affectsSpatial slabs/ceilings/
zones, linkedBy endpoint-match, cascadeDelete descendants), and the
presentation metadata for the palette. Renderer / system / tool fields
are deliberately absent — the existing wall-renderer.tsx and
wall-system.tsx keep serving wall until milestone B.
- feature-flag.ts gates the eventual registration via
NEXT_PUBLIC_USE_REGISTRY_FOR_WALL (same pattern Phase 2 used for spawn).
- wallDefinition is NOT yet appended to builtinPlugin.nodes — registration
is what flips the Phase 0 dispatch shims, and we don't want that until
the runtime port lands. Until then this file is metadata-only.
Two type-side changes pulled forward from Phase 4 to make a metadata-only
definition compile:
- NodeDefinition.renderer becomes optional (the three-checkbox model
documented in wiki/architecture/node-definitions.md already promises
this). RegistryRenderer in node-renderer.tsx gains a null-guard so an
undefined renderer cleanly falls through to the legacy switch.
No runtime behavior change. Walls render and behave exactly as before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New page covering the geometry/renderer/system trio that registry-driven
kinds opt into. Documents:
- The three optional fields on NodeDefinition and when each applies
- Generic <GeometrySystem> + <ParametricNodeRenderer> runtime
- GeometryContext shape (resolve / children / siblings / parent)
- Combination matrix for shelf / spawn / zone / door / window / GLB items
- Migration recipe from custom renderer+system files to def.geometry
- Rules around purity, dispose-on-rebuild, register-once
renderers.md and systems.md gain "prefer registry-driven" banners and
link out to the new page. Architecture README adds the page to the
index so review-architecture skill picks it up.
The pattern was validated by the shelf spike: inline-JSX geometry was
visibly laggy on parametric edits; moving to a per-kind system reading
dirtyNodes (mirroring door/wall/item) restored smoothness. The three-
checkbox model generalises that win so most future kinds need only a
pure geometry function.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Follows the renderer/system split documented in wiki/architecture/
renderers.md and systems.md: the renderer must not run geometry
generation. Mirrors the door-renderer/door-system pattern.
- New ShelfSystem reads dirtyNodes in useFrame, retrieves the shelf's
registered Group from sceneRegistry, swaps its children with the
output of buildShelfGeometry(node), then clears the dirty flag.
Geometry rebuild is fully imperative — no React work involved.
- ShelfRenderer is now a thin empty <group> that registers with
sceneRegistry, marks the node dirty on mount, and carries the
pointer-event handlers + live transform overrides at the root.
- Wired system into shelfDefinition so RegisteredSystems mounts it
alongside the renderer.
Net effect: dragging shelf parametric sliders no longer re-renders
the renderer per tick — the system rebuilds meshes at frame cadence
based on dirtyNodes, the inspector's per-field subscription only
re-renders the dragged field, and the rest of the React tree stays
quiet.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Subscribe to node.type at the top and to each field's value individually
inside FieldRenderer. Slider drags previously re-rendered the entire
inspector + every field every tick because the panel subscribed to the
whole node object (which gets a new reference on every updateNode).
Primitive field values stay === equal across unrelated mutations, so
now only the dragged field re-renders.
Handlers (move/delete/update) use useScene.getState() inside callbacks
instead of subscribing — they only need the current value, not a
reactive read.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First slice of Phase 4 work pulled in: an inspector that reads any
registered node's `parametrics` descriptor and renders the right-side
panel automatically. Verifies the parametric descriptor design end-to-end
on shelf and proves the loop (descriptor → UI → store update → render).
ParametricInspector component:
- Reads `nodeRegistry.get(node.type)?.parametrics`.
- Renders one `<PanelSection>` per group, one control per field.
- Field kinds supported in v1:
* `number` → SliderControl with min/max/step/unit from descriptor
* `enum` → dark-themed <select> with prettified labels
* `color` → native color picker + hex input pair
* `vec3` → 3× SliderControl (X / Y / Z)
- Honors `field.visibleIf(node)` to gate conditional fields.
- Generic Actions footer with Move / Delete, gated on
`capabilities.movable` / `capabilities.deletable !== false`.
- Title from `presentation.label`, defaults to `node.type`.
Wired as the `default:` arm of panel-manager.tsx's switch — registered
kinds without a hardcoded case (shelf, future kinds) get the auto-derived
panel. Spawn keeps its hand-written panel (the legacy switch catches
it first); we'll switch to registry-first dispatch when Phase 4 finishes
and the hand-written panels can be deleted.
What you can verify after this:
- Click a shelf → right panel shows Dimensions (width/depth/thickness/
height sliders with units + bounds from the schema) + Style (bracket
style select + color picker) + Actions (Move / Delete).
- Drag a slider → mesh updates live (store update → renderer re-renders).
- Try to set width > 3.0 — schema rejects, no update fires (the
parametric bounds are enforced by Zod, same source of truth as MCP
bound generation in Phase 4's MCP work).
Not in scope for this commit:
- `parametrics.customPanel?` escape hatch.
- `material` / `ref` field kinds.
- `invariants` validation feedback in the UI.
- Migration of legacy panels (spawn/column/etc.) to the auto-generated
path — those keep working unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Shim ordering bug: the `nodeRegistry.has(movingNode.type)` check sat at
the END of the dispatch chain, AFTER the per-kind `if (movingNode.type
=== 'spawn') return <MoveSpawnTool>` branches. Spawn (now registered
in builtinPlugin) was therefore still routing to the legacy
MoveSpawnTool — which uses the broken useLiveTransforms pattern and
makes the spawn mesh disappear during drag.
Moved the registry check to the TOP, matching the registry-first
dispatch model the Phase 0 shims use everywhere else (NodeRenderer,
ToolManager, system guards). Now any kind registered via
@pascal-app/nodes routes to MoveRegistryNodeTool — same smooth
imperative drag for shelf, spawn, and every future kind. Legacy
per-kind movers below run only for kinds not yet in the registry.
This is exactly how the Phase 5 progressive consolidation works: as
kinds migrate to the registry, their legacy movers stop being reached
and can be deleted.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User feedback: updating the store per tick caused tons of React
re-renders → laggy drag. Switched to pure imperative.
MoveRegistryNodeTool now:
- Mutates `sceneRegistry.nodes.get(id).position` directly per
grid:move tick. No useScene.updateNode during drag. No store
change → no renderer re-render → R3F doesn't reapply
`position={node.position}` → the imperative mutation sticks.
- On commit: single tracked `useScene.updateNode(id, { position })`.
Undo replays one step (original → final), no per-tick spam.
- On cancel / unmount: imperatively snap the mesh back to original.
Store was never touched so no data revert needed.
Trade-off vs the items pattern (which does update the store per tick
and re-renders per tick): our approach is faster but assumes the
renderer doesn't re-render mid-drag. Items get away with constant
re-renders because their renderer is heavily optimized; for parametric
shelves (and future kinds) the imperative path is simpler and faster.
Cleanup: removed the dev `[shelf] rendered` and `[shelf] placed`
console.info logs from the shelf renderer and tool. They were Phase 2
verification scaffolding — no longer needed now that everything works.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The useLiveTransforms + sceneRegistry.position.set approach used by
MoveColumnTool is broken — column ALSO disappears during move, per
user observation. The mesh doesn't visibly follow the cursor.
Items work because their move tool directly updates the scene store's
node.position on every grid:move tick (with history paused), and the
renderer reads node.position. Matching that pattern here.
MoveRegistryNodeTool now:
- Snapshots the original position at mount (for cancel / commit
revert path).
- Pauses scene history so per-tick updateNode calls don't fill undo.
- On grid:move: `useScene.updateNode(id, { position })`. The kind's
registered renderer reads node.position and re-renders, so the
actual mesh visibly follows the cursor.
- On commit: revert to original while still paused → resume → final
update (single tracked action) → re-pause. Undo replays one step,
not the per-tick spam.
- On cancel / unmount-without-commit: restore original position with
history still paused (won't enter undo), then resume.
The cursor sphere stays as the aim indicator alongside the moving
mesh. No translucent ghost — the actual mesh IS the preview now.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User feedback: the move tool shouldn't render a separate translucent
preview. The actual mesh (registered with sceneRegistry through the
kind's renderer) should follow the cursor — that's what's already
happening via useLiveTransforms + the imperative position.set on the
registered Object3D.
Removed from MoveRegistryNodeTool:
- The lazy `def.preview` load + Suspense-wrapped <Preview> render.
- Now only CursorSphere shows as the aim indicator. The shelf's
actual rendered mesh follows the cursor via:
- `useLiveTransforms.set(...)` triggers ShelfRenderer to re-render
with `position={liveTransform.position}`.
- `sceneRegistry.nodes.get(node.id).position.set(...)` is a
defensive imperative update so motion feels snappy.
Added: `sfx:grid-snap` emit on grid-cell cross, matching the placement
tools' behavior. Move now sounds like placement.
The `preview` slot on NodeDefinition stays — still used by ShelfTool
for the placement cursor (where no real mesh exists yet). Phase 4 may
consolidate placement preview with the renderer too.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User feedback: the move tool's CursorSphere is just a vertical line,
hard to tell what you're moving. Placement gets a translucent shape
that follows the cursor; move should too.
NodeDefinition.preview?: () => Promise<{ default: ComponentType<{ node }> }>
opt-in lazy component that renders a translucent ghost of the node.
Used by:
- The placement tool (ShelfTool) — renders the preview at the cursor
position so the user sees the shape they're placing.
- The move tool (MoveRegistryNodeTool) — renders the preview at the
drag target alongside the CursorSphere. Plus the original node is
also dragged via live transforms, so the user sees both: the actual
node moving + a translucent ghost at the same spot.
Implementation:
- New nodes/shelf/preview.tsx: ShelfPreview component. Renders the
same shape as ShelfRenderer but `transparent: true, opacity: 0.5`.
- shelfDefinition.preview = () => import('./preview').
- ShelfTool's placement preview now uses <ShelfPreview node={defaults} />
instead of an inline copy of the box geometry.
- MoveRegistryNodeTool lazy-loads `def.preview` (cached by loader,
Suspense-wrapped). If a kind doesn't define `preview`, only the
CursorSphere shows — matches today's behavior.
Phase 4 may merge `preview` with `renderer` behind an `opacity` prop
so kinds don't duplicate JSX between the solid and translucent
versions; until then defining `preview` is opt-in and one extra file.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Shelf had the floating action menu (move/delete) showing thanks to
the previous registry-driven selection commit, but clicking move
did nothing and duplicate silently failed. Two hardcoded chains:
1) FloatingActionMenu.handleMove guarded `setMovingNode` behind a
hardcoded `node.type === 'item' || ... || node.type === 'spawn'`
chain. Added `|| isRegistrySelectable(node.type)` so any
registry kind triggers the move flow.
2) MoveTool dispatched per-kind components (MoveItemContent,
MoveColumnTool, MoveWallTool, ...). The default fallback
mounted MoveItemContent, which assumes the node is an ItemNode
with asset/scale/metadata — crashes for shelf. Added a generic
MoveRegistryNodeTool (kind-agnostic clone of MoveColumnTool):
pure position+rotation drag with grid snap, re-parses orphan
re-creates via `nodeRegistry.get(kind).schema.parse(...)`.
MoveTool dispatches to it for any `nodeRegistry.has(movingNode.type)`
before the MoveItemContent fallback.
3) FloatingActionMenu.handleDuplicate had a hardcoded
`node.type === 'door' ? DoorNode.parse(...) : ...` chain. Added
a registry-driven fallback after it:
`const def = nodeRegistry.get(node.type); duplicate = def.schema.parse(duplicateInfo)`.
Then the createNode + setMovingNode branches also augment with
`nodeRegistry.has(duplicate.type)` so the new shelf gets
created in the scene and handed off to the move tool for
placement.
After this:
- Click shelf → move icon in floating menu → cursor follows mouse,
click to place at new position.
- Click shelf → duplicate icon → new shelf appears, offset by (1,0,1),
handed to move tool so the user can position it.
Phase 4 will collapse MoveRegistryNodeTool with the per-kind movers
once they all reduce to the same position+rotation shape, and read
`capabilities.movable` to gate handleMove instead of the OR chain.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two concerns from the spike:
1) Selection / floating-action-menu had hardcoded kind lists scattered
across 4 files. Adding 'shelf' to each one per migration was the
wrong abstraction — the user's question "did you make it generic
from the noderegistry?" was the right one. Done now.
Added to @pascal-app/core/registry:
- getSelectableKinds(): string[] — returns all registered kinds
whose definition declares `capabilities.selectable`.
- isRegistrySelectable(kind): boolean — predicate for OR-chains.
Refactored hardcoded sites to merge registry kinds at runtime,
keeping legacy hardcoded lists intact so existing kinds keep
working unchanged:
- editor SelectionManager: 4 subscription loops (enter/leave/click)
+ structure.isValid + getSelectionTarget — all augment with
registry kinds. Phase 6 deletes the hardcoded lists.
- viewer SelectionManager: subscription loop + SelectableNodeType
broadened with `(string & {})` to accept registry kinds.
- floating-action-menu: ALLOWED_TYPES OR'd with isRegistrySelectable.
- Removed the manually-added 'shelf' entries from previous commit
857ddd4; they were redundant once the registry-driven path landed.
Future built-in nodes that declare `capabilities.selectable` get
click-selection + hover + the floating action menu (move/delete
icons) for free, no editing of these 4 files.
2) Spawn parity is signed off. Drop the
NEXT_PUBLIC_USE_REGISTRY_FOR_SPAWN flag entirely; spawn registers
unconditionally in builtinPlugin.nodes. Restored SPAWN_COLOR to
the original #22c55e green (was #ef4444 red as a Phase 2
verification marker).
Pre-existing typecheck errors in editor (ceiling/fence/slab tree-node,
scene.ts buildingId) are unchanged.
630 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Shelf renderer was emitting `shelf:click` / `shelf:enter` / `shelf:leave`
via useNodeEvents from the previous commit, but no listener subscribed
— the SelectionManager components (one in editor, one in viewer) each
maintain hardcoded allTypes arrays that didn't include 'shelf'.
Adds 'shelf' to:
- editor/selection-manager: 5 allTypes arrays (one per selection strategy
— structure, structure-hover, furnish, site, deselect-also-listens-to).
- viewer/selection-manager: the SelectableNodeType union + allTypes
array.
Shelves can now be clicked / hovered in the 3D canvas and the
selection state updates correctly.
The hardcoded arrays are exactly the kind of cross-cutting friction
the registry is supposed to eliminate. Phase 4 should derive these
lists from `nodeRegistry.entries().filter(d => d.capabilities.selectable)`
so adding a new kind doesn't require editing two files.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Matches the wall / slab / curve-wall tool pattern: emit
sfx:grid-snap only when the snapped position changes (cursor crosses
a grid cell), not every frame of mouse movement within the same cell.
Tracked via a `previousSnapRef` per tool, reset when the tool
re-activates.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User-visible follow-ups after first running the Phase 2 spike.
Spawn tool now matches legacy UX:
- CursorSphere from @pascal-app/editor for the placement indicator
(ring + line + tool-icon tooltip) — was a plain sphere mesh.
- Emits sfx:structure-build on commit + setTool(null) + setMode
('select') to exit build mode, matching legacy spawn-tool.
Shelf tool placement:
- Emits sfx:structure-build on commit.
- Cursor preview now shows top board + brackets (was just the top),
matching what gets placed.
Shelf selectable from the 3D canvas:
- ShelfEvent type added to @pascal-app/core/events/bus.
- 'shelf' added to NodeConfig in useNodeEvents.
- ShelfRenderer wires `useNodeEvents(node, 'shelf')` handlers onto
every mesh. Clicks/hovers now bubble through the editor's selection
manager and update useViewer.selection.
Shelf appears in the sidebar:
- ShelfTreeNode component (mirrors spawn-tree-node's shape +
selection/hover/rename wiring; lucide Layers icon).
- TreeNode dispatcher adds a `case 'shelf':` arm.
Framework changes:
- @pascal-app/editor exports CursorSphere alongside triggerSFX.
- @pascal-app/nodes now declares @pascal-app/editor as peer/dev dep.
Pre-existing typecheck errors in @pascal-app/editor (ceiling-tree-node,
fence-tree-node, slab-tree-node, scene.ts) are unchanged — present on
main and not introduced by this commit.
630 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three concrete bugs surfaced when first-running the spike in community:
1) NEXT_PUBLIC_USE_REGISTRY_FOR_SPAWN flag never detected:
The previous readEnvFlag used dynamic bracket access
(`env?.[name]`), which Next.js / Turbopack does NOT substitute at
build time. Only literal `process.env.NEXT_PUBLIC_FOO` references
get inlined into the client bundle. Switched to literal access
plus a `typeof process` guard. Spawn now toggles via the flag as
designed.
2) Shelf cursor appeared offset from the mouse:
The cursor mesh lives inside the ToolManager's building-local
group, but the tool was setting `cursorRef.current.position` to
level-local coordinates (computed via `worldToLocal(level)`).
Result: cursor shifted by (building-pos − level-pos) in worst
case. Switched cursor display to use `event.localPosition`
(already building-local) with grid snap — matches the legacy
spawn-tool pattern. The commit path keeps the worldToLocal(level)
conversion since the shelf node's `position` field is stored
relative to its level parent.
3) Shelf rendered invisibly after click (suspected):
The renderer used a useEffect-swap pattern where it mounted an
empty <group> and imperatively added Three.js children from a
buildShelfGeometry() Group. Plausibly fragile under StrictMode
double-invoke or fast HMR. Switched to inline R3F JSX — top
board + brackets as plain <mesh> primitives. The pure geometry
function still exists in geometry.ts for tests and AI-authored
consumers; renderer just doesn't go through it.
Diagnostics added (dev-only; removed once spawn parity ships):
- `[shelf] placed <id> level-local <pos> parent <levelId>` on click
- `[shelf] rendered <id> at <pos>` on mount
Also: types: ["node"] in nodes/tsconfig.json so the typeof process
guard typechecks cleanly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three small follow-ups after the first user-visible spike test.
shelf cursor stuck at origin:
The placeholder cursor is a translucent slab inside a <group ref>.
The tool was calling setCursor(state) on every grid:move, which
triggered a React re-render. R3F re-applies props on each render,
and since the <group> had no `position` prop, the implicit default
[0,0,0] clobbered the imperative `position.set` from the previous
tick. Result: cursor stuck at level origin instead of following
the mouse.
Fix: drop the unused useState entirely. Pure imperative position
updates via the ref. No re-renders, no clobbering. (The legacy
spawn tool gets away with the same pattern because CursorSphere
buffers its position prop differently — but for the spike, the
simpler model is fine.)
shelf snap step:
Was 0.1 (10cm) — much finer than the editor's default 0.5 grid.
Bumped to 0.5 (matches the toolbar grid setting and the legacy
half-meter snap pattern used by spawn/column).
registry-driven spawn renderer paints red:
Temporary verification marker. With NEXT_PUBLIC_USE_REGISTRY_FOR_
SPAWN=1, spawns rendered via the new path now appear in #ef4444
red. Legacy renderer stays in #22c55e green. Easy visual check
for "which dispatch path is this spawn on?" Reverted in the PR
that signs off spawn parity (alongside legacy file deletion).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Verification anchor for "which path is running this kind?":
[pascal:registry] loaded pascal:core v1 (1 kinds: shelf)
prints in the browser dev console on app boot. Empty array means every
kind is on the legacy dispatch path. A kind in the array means the
registry-first NodeRenderer / ToolManager shims own it (legacy path is
short-circuited).
Also exposes `globalThis.__pascalNodeRegistry` in dev so you can run
ad-hoc inspections like `__pascalNodeRegistry.has('spawn')` from the
console. Production builds skip both.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The first time registry-driven nodes actually run in the editor.
Spawn migration (under NEXT_PUBLIC_USE_REGISTRY_FOR_SPAWN flag):
- New packages/nodes/src/spawn/ folder with renderer, tool, schema
(re-exported from core), parametrics, definition, index.
- Spawn definition appended to builtinPlugin.nodes only when the flag
is set. With the flag off, the Phase 0 shims fall through and the
legacy SpawnRenderer / SpawnTool keep ownership.
- New no-props SpawnTool reads activeLevelId from useViewer directly,
matches legacy placement behavior (half-meter snap, singleton-per-
level, replace-on-reclick).
- Structural parity test (9 cases) validates definition shape +
schema identity. Pixel-diff defers to Phase 4 alongside more nodes.
New shelf node (no legacy — registered unconditionally):
- ShelfNode schema in core/schema/nodes/shelf.ts (hand-maintained
AnyNode union for now; Phase 6 derives the union from the registry
and moves the schema fully into nodes/shelf/).
- packages/nodes/src/shelf/ folder: pure geometry builder
(buildShelfGeometry returns a Three.js Group of top board +
brackets), R3F renderer that mounts the built group, no-props
placement tool, parametrics descriptor (width/depth/thickness/
height/bracketStyle/color), definition with surfaces.top stackable
surface for future stacking, and presentation metadata for the
palette.
- 13 unit tests across schema bounds and geometry behavior.
- Palette wiring: 'shelf' added to StructureTool union + an entry in
the structure-tools array (placeholder icon, replaced in Phase 4
when palette is registry-driven).
Framework changes:
- @pascal-app/viewer now exports useNodeEvents from its public barrel
so node bundles in @pascal-app/nodes can subscribe to node-specific
pointer events. (Used by spawn renderer; shelf renderer skips it
for now since useNodeEvents has a hardcoded kind list — Phase 4
generalizes it via the registry.)
- @pascal-app/nodes gains @pascal-app/viewer as a peer + dev dep so
node bundles can import from it.
630 tests pass across 76 files (22 new this phase). Editor app
continues to ship green with both legacy spawn and the new shelf
node co-existing through the Phase 0 dispatch shims.
To validate end-to-end in dev:
- bun dev:community → open editor → click 'Shelf' in structure
toolbar → click to place. Confirms full registry path
(NodeRenderer dispatch + ToolManager dispatch + sceneRegistry
byType Proxy).
- Set NEXT_PUBLIC_USE_REGISTRY_FOR_SPAWN=1, restart dev, place spawn
→ visually identical to legacy. Confirms parity.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Builds a 15200-node fixture (50×100 wall grid + 8000 hosted doors +
200 sparsely-indexed slabs) and runs `cascadeDirty` 1000 times with
warm-up. Reports p50/p95/p99/mean/max in ms.
Runs via `bun run bench:registry` from the core package.
Today: p95 measured at ~0.002ms — three orders of magnitude under the
Phase 1 gate of 2ms. Headroom is substantial; we'll only revisit this
if Phase 3 wall introduces `linkedBy: 'endpoint-match'` and pushes the
inner cascade past the gate.
Not wired into CI for v1 — regressions reviewed manually before phase
gates. Output is JSON so a future CI step can diff against a baseline.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pure orchestrator in core, thin React wrapper in editor:
- `core/services/drag-session.ts` — `createDragSession(action, scene,
options)` returns an imperative session with `start / move / commit
/ cancel / dispose / isActive / getDraft`. Pauses history on start,
resumes on terminate. Per-move runs preview → snap → apply, then
cascades dirty marks via the relations resolver (deduped across
ticks). Re-entry guard, idempotent dispose, fires onCommit/onCancel
callbacks. All tested in bun:test — no React needed.
- `editor/src/hooks/use-drag-action.ts` — wraps the session with the
editor's grid-event emitter and an Esc-to-cancel keyboard listener.
Builds a `SceneApi` once via `createSceneApi(useScene)` at module
init. The hook itself is small enough to read top-to-bottom; all
behavior lives in the session.
Tests (13 cases) cover the hard parts: history pause/resume bracket,
explicit cancel restoring all touched nodes, dispose mid-drag, commit
returning false short-circuiting to cancel, snap callback wired in,
re-entry rejected, deduped dirty-mark across multiple move ticks,
hosts cascade from the registry firing in apply.
No callers yet — Phase 2 column and shelf tools are the first
consumers.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pure constraint math built on the registry's `MovableConfig`:
- `resolveMovable(node)` — reads `def.capabilities.movable`, runs the
optional `override(ctx)` callback (returning null falls back to the
base config). Returns null when the kind isn't movable.
- `applyAxisLock(current, target, axes)` — projects 3D motion onto
the allowed axes; locked components fall back to current.
- `moveToward(node, current, target, options?)` — top-level helper
combining axis lock + (optional) grid snap. Returns null when the
node is not movable.
- `movePlanToward(node, currentY, current, target, options?)` —
X/Z-plane convenience for floor/plan-view placement.
- `isMovable(node)` — predicate for tools/UI gating.
Tests cover override callback, null-override fallback, axis lock
permutations, grid-snap on/off, and the 2D plan convenience.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pure math, no React, no scene access. Three primitives plus a facade:
- `snapScalar(value, step)` / `snapPointToGrid(point, step)` /
`snapVec3ToGrid(point, step)` — regular grid snapping. Default
step 0.25m matches the editor's wall tool.
- `snapPointToAngle(from, cursor, angleStep, gridStep?)` — locks a
cursor to the nearest angle multiple from a fixed point, preserves
distance, optionally re-grids the projected point. Default angle
step π/12 (15°).
- `snapAngleToList(angle, list, tolerance)` — snaps a free angle to
the nearest entry in a fixed list (e.g. 0/45/90/135) within a
tolerance; returns the original angle otherwise. Handles wrap.
- `snapServices` facade — `grid.*` + `angle.*` namespaces. Stable
contract that `DragAction.snap` callbacks receive. Phase 3 ports
the existing `snapWallDraftPoint` family from
`editor/.../wall-drafting.ts` under a `wall.*` namespace.
17 unit tests cover the math + the facade pass-through. No existing
callers re-wired yet — Phase 2 column/shelf tools are the first
consumers.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First service in `core/src/services/`: pure, no React or R3F, takes
SceneApi + node data, returns results.
Exports:
- `canAttach(childId, hostId, scene)` — validates host attachment.
Rejects self-host, cycles (host's ancestor chain contains child),
chains past MAX_HOST_DEPTH (6), and host kinds outside the child
def's `capabilities.hostable.parents` allowlist. Returns a typed
AttachError discriminated union so callers can render specific
messages.
- `getSurface(host)` / `getTopSurfaceHeight(host)` — reads
`def.capabilities.surfaces` from the registry; resolves
function-valued heights with the node.
- `clampYToHostTop(host, y)` — convenience for placement code.
- `pickHost({ point, candidates, placedKind, hitTest? })` — given
spatially pre-filtered candidates, returns the first hostable.
The runtime is responsible for spatial filtering; this function
stays pure.
MAX_HOST_DEPTH = 6: the explore earlier found today's editor has no
cap on item-on-item nesting. Cap is bounded by hostable depth, not
total tree depth (sites/buildings/levels don't count).
17 tests cover all rejection paths + happy paths + function-valued
surface heights.
Re-exported from `@pascal-app/core` via a new `services/` barrel.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pure traversal that walks a node's declared `relations` and returns the
full set of IDs that should be marked dirty alongside it. Two
implementations:
- `cascadeDirty(id, ctx)` — follows `hosts` (matching children) and
`affectsSpatial` (via injected spatialQuery). Phase 3 will add
`linkedBy: 'endpoint-match'`.
- `collectDescendants(id, ctx)` — pure subtree traversal for
`cascadeDelete: 'descendants'` and subtree deletion tools.
Both bounded by maxDepth (default 16) and visited-set so cycles in
bad data can't loop forever.
Context-based design: spatialQuery and childQuery are injected, so the
resolver itself stays pure — the DragAction runtime can plug in
spatialGridManager-backed queries; tests pass stubs.
Today, registry has no kinds → cascadeDirty(id) always returns just
{id}. No behavior change. Phase 3 wall is the first real consumer.
11 unit tests cover empty/no-relations baseline, hosts cascade, depth
limit, spatial query, missing spatialQuery branch, cycle protection,
childQuery override, descendant collection.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds optional `presentation` field for tool palette metadata: sentence-
case label, optional description, icon (iconify reference, inline SVG,
or lazy React component), palette section override, sort order, and a
`hidden` flag for container kinds that exist but should not appear in
the palette.
Consumer arrives in Phase 4 (auto-derived palette buttons) — defining
the type now means Phase 2's `column` and `shelf` definitions ship with
the field already populated, no later round-trip.
Iconify is the encouraged form for built-ins and AI-authored nodes:
matches the @iconify-react setup the editor app already uses, and AI
emits a name string from a curated list (no asset upload step needed).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Without a "test" script the package was invisible to `turbo test` —
private-editor's CI runs `bun run test` (= `turbo test`), which only
walks workspace packages that declare a test runner. Mirrors mcp and
nodes which already do this.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Before mounting any legacy build tool, ToolManager checks whether the
active tool's kind has a registered NodeDefinition with a tool
contribution. If yes, the registry tool wins; the legacy tool map and
special-cased spawn/column/elevator branches are skipped for that kind.
Lazy-loaded via React.lazy (cached by loader) and wrapped in Suspense.
Today the registry is empty, so useRegistryTool is always false and
every code path renders unchanged. The moment a kind registers (Phase
2+), its registry tool takes over without further edits here.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two additions plus a viewer JSX rewire:
- legacy-system.tsx: <LegacySystem kind="..."> wrapper that renders its
children only when nodeRegistry.has(kind) is false. Lets one wrapper
cover all legacy systems for a kind (door has DoorSystem and
DoorAnimationSystem — both belong to 'door' so they yield together).
- registered-systems.tsx: <RegisteredSystems /> iterates the registry,
filters entries that contribute a system, sorts by system.priority
(default 5; e.g. wall mitering at 8 runs after door cuts at 3),
mounts each via React.lazy. Today empty registry = renders nothing.
- viewer/index.tsx: every existing per-kind system is wrapped in
LegacySystem. RegisteredSystems is mounted alongside.
With the registry empty (Phase 0), every LegacySystem passes through
unchanged and RegisteredSystems is a no-op — zero behavior change.
Once a kind registers in Phase 2+, its legacy systems yield and its
registry-contributed system runs in their place.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
NodeRenderer now checks the registry first. Registered kinds load
their renderer module via React.lazy (cached by RendererSource so the
Suspense boundary is stable across re-renders). Unregistered kinds
fall through to the legacy chain below.
Today the registry is empty (Phase 0 builtinPlugin.nodes is []), so
every node still hits the legacy chain — no behavior change. The
moment a kind registers in @pascal-app/nodes (Phase 2 onward), the
registry path takes over without further edits here.
GLB / instanced-GLB RendererSource kinds are typed but not yet
honored — they get their built-in renderers in Phase 5.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
byType was a hardcoded object keyed by the built-in node kinds. With
the registry, kinds can come from @pascal-app/nodes (or future
plugins) — so byType now wraps a Map via a Proxy that auto-creates an
empty Set the first time any kind is touched.
Built-in kinds are still pre-seeded at module init so the fast path
(no Proxy trap) is preserved. clear() iterates the backing Map.
useRegistry's `type` parameter widens from `keyof typeof byType` to
`KnownNodeKind | (string & {})` — preserves autocomplete for
built-ins while accepting plugin-supplied kinds.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds apps/editor/lib/bootstrap.ts that calls loadPlugin(builtinPlugin)
as a module-side-effect on first import. Imported from scene-loader.tsx
so it runs on the client side where the editor mounts.
Idempotency guard handles HMR re-execution (would otherwise throw on
duplicate registerNode). For the empty plugin this commit, the entire
call is a no-op — included now so future commits that add real node
kinds only need to push them onto builtinPlugin.nodes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Companion changes for the new nodes package: bun.lock entry from
`bun install`, and a Biome-auto-sort of the registry barrel.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New workspace package that owns built-in node bundles, one folder per
kind. Today the plugin is empty — no behavior change. Future commits
will add column/, shelf/, wall/, etc. and append each definition to
builtinPlugin.nodes.
The package depends on core (registry types) at v1; viewer, editor,
react, three are declared as peer deps so future node bundles can use
them without bumping their own version on every monorepo bump.
External plugins land as separate packages with the exact same shape
— this package is the dogfooded reference.
Tests: builtinPlugin shape + loadPlugin succeeds with zero kinds.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a Biome noRestrictedImports rule scoped to core/, viewer/, and
editor/ packages. Framework code must reach node-specific behavior via
nodeRegistry.get(kind), never via direct import. The @pascal-app/nodes
package doesn't exist yet (lands in the next commit) so the rule is a
no-op today; codifies the boundary ahead of node bundles.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Introduces the @pascal-app/core/registry surface that future node-bundle
packages (and external plugins) will use to register node kinds with the
host. No runtime behavior changes — registry is empty until subsequent
PRs populate it.
- types.ts: NodeDefinition, Capabilities, Relations, DragAction, Plugin,
ParametricDescriptor, Affordance, SceneApi, NodeRegistry. Capability
configs accept an override escape hatch; additive-only after v1.
- registry.ts: nodeRegistry singleton, registerNode, async loadPlugin.
Validates kind, schemaVersion, apiVersion; rejects duplicate kinds.
- scene-api.ts: createSceneApi factory wrapping the scene store with
copy-on-write snapshot semantics for pauseHistory/restore/resumeHistory.
- index.ts: barrel re-exporting the public surface.
- core/index.ts + package.json: export * from registry and add the
./registry subpath so consumers can import either way.
Tests (27 cases, all bun:test): registry registration / validation /
plugin loading; SceneApi read/write/dirty/history; lazy snapshot capture
with update/upsert/delete reversal via restore and restoreAll.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The file imports useEffect/useRef from React, which Next.js RSC builds
flag as client-only. Other core systems (e.g. elevator-runtime-system)
use useFrame from @react-three/fiber and slip through, but this one
needs the directive explicitly.
Fixes Turbopack build failure in private-editor community app:
"You're importing a module that depends on useEffect into a React
Server Component module."
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
