Agent skill

Pascal Architecture PR Review

by pascalorg in pascalorg/editor

Reviews a pull request against the Pascal editor's architectural rules: package boundaries, registry-driven node composition, hook hygiene and selector performance.

MITAuto-check passedDevelopment

Install Pascal Architecture PR Review

skills CLI
$ npx skills add pascalorg/editor --skill review-architecture -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install pascalorg/editor review-architecture --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/pascalorg/editor.git skills-src && mkdir -p .claude/skills && cp -r skills-src/.agents/skills/review-architecture .claude/skills/review-architecture && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
review-architecture
GitHub stars
25k
Token cost
~7.5k tokens
SKILL.md length
3,510 words
Files
1
Skills in repo
4
Repo updated
First seen
Licence
MIT

At a glance

Reviews a pull request against the Pascal editor's architectural rules: package boundaries, registry-driven node composition, hook hygiene and selector performance.

  • Works in 6 steps: Load the rules (required — do not skip) → Fetch the diff → Layer classification — do this BEFORE… → …
  • Reviewing a Pascal pull request for architectural violations
  • SKILL.md covers 1. Load the rules (required —…, 2. Fetch the diff, 3. Layer classification — do… and 4. Review checklist, plus 2 more sections
  • Calls gh and git

What it does

The agent first loads the architecture wiki pages as the source of truth, always reading the core set on layers, systems, renderers, tools, viewer isolation, node definitions, plugin authoring and agent surfaces, and further pages only when the diff touches their area. It then fetches the diff with the gh CLI or git, lists the changed files to map each to a rule, and classifies each by layer before running the checklist.

Checks cover package boundaries between core, viewer, editor and nodes, the three-part registry composition of geometry, renderer and system, regressions to legacy dispatch, the slots and world-scale-UV convention for new nodes and geometry, hook hygiene around useEditor, useScene and useViewer, selector performance, and limits on inspector field bounds. It takes a PR URL, a branch name or the current branch.

When your agent uses it

  • Reviewing a Pascal pull request for architectural violations
  • Auditing a branch before merge against the package boundary rules
  • Checking that a new node follows the registry composition model
  • Looking for selector or hook performance problems in a diff

Example prompts

  • “Review this Pascal pull request against our architecture rules.”
  • “Audit my current branch for viewer code that depends on the editor.”
  • “Check that my new node kind uses geometry, renderer and system definitions correctly.”

Requirements

  • The gh CLI and git
  • The Pascal repository's architecture wiki pages
  • Pre-approved tools (allowed-tools): Bash(git *), Bash(gh *), Read, Grep, Glob

Workflow steps

6 steps, taken from the step headings in SKILL.md.

  1. Load the rules (required — do not skip)
  2. Fetch the diff
  3. Layer classification — do this BEFORE the checklist
  4. Review checklist
  5. Output format
  6. Final summary

What it can do on your machine

Read from SKILL.md and the folder at commit a860e19. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves these tools, so the agent can use them without asking each time:

    • Bash(git *)
    • Bash(gh *)
    • Read
    • Grep
    • Glob

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    Shell commands in SKILL.md call:

    • gh
    • git

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md. Its commands use gh and git, which can reach the network depending on how they are called.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Pascal Architecture PR Review loads about 7.5k tokens when it runs. Until then it costs about 119 tokens; SKILL.md has 3,510 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~119
When it runs · the whole SKILL.md, loaded when a task matches
~7.5k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from pascalorg/editor at commit a860e19, republished under its MIT licence (© pascalorg). 3,510 words, ~7,539 tokens.

Download SKILL.mdSave it as .claude/skills/review-architecture/SKILL.md (or your agent's skills folder).
name
review-architecture
description
Review a PR against the Pascal architectural rules — package boundaries (core/viewer/editor/nodes), the registry-driven composition model (def.geometry / def.renderer / def.system), legacy-dispatch regressions, the slots + world-scale-UV convention for new nodes/geometry, hook hygiene (useEditor/useScene/useViewer), and selector performance. Use when the user asks to review a PR, audit a branch, or check that changes respect the codebase's architecture.
allowed-tools
Bash(git *), Bash(gh *), Read, Grep, Glob
metadata.internal
true

Architectural review for Pascal PRs. The user will provide a PR URL, branch name, or ask to review the current branch.

1. Load the rules (required — do not skip)

Read these before reviewing any diff. They are the source of truth, not your training data:

  • wiki/architecture/layers.md
  • wiki/architecture/systems.md — core systems vs viewer systems, what each may do
  • wiki/architecture/renderers.md — renderer responsibilities and prohibitions
  • wiki/architecture/tools.md — editor tools live only in apps/editor/components/tools/ or packages/nodes/src/<kind>/
  • wiki/architecture/viewer-isolation.md — viewer must stay editor-agnostic
  • wiki/architecture/node-definitions.md — the three-checkbox composition model (geometry / renderer / system)
  • wiki/architecture/plugin-authoring.md — public contract for external node packs
  • wiki/architecture/agent-surfaces.md — MCP ↔ hosted chat ↔ skill parity. Read whenever the diff touches packages/mcp/** or skills/**.

Required on every review. Read the remaining pages on demand when the diff touches their subject area:

  • wiki/architecture/selection-managers.md
  • wiki/architecture/scene-registry.md
  • wiki/architecture/spatial-queries.md
  • wiki/architecture/node-schemas.md
  • wiki/architecture/inspector-field-limits.md — no arbitrary min/max on dimension fields. Read whenever the diff adds or edits parametrics.ts, a kind panel.tsx, or <SliderControl> bounds.
  • wiki/architecture/events.md
  • wiki/architecture/interaction-scope.md — the interaction state machine + the unified snapping/modifier convention. Read whenever the diff touches a tool, a move-tool / selection / endpoint / reshape file, lib/interaction/**, lib/snapping-mode.ts, or use-interaction-scope.

If anything in the diff looks like a new dispatch surface or registry concept, also read DECISIONS.md E-002 and wiki/architecture/node-definitions.md — they own the composition model and the capability vocabulary.

2. Fetch the diff

bash
# If the user gave a PR URL or number:
gh pr diff <pr-number-or-url>

# If reviewing the current branch:
git diff main...HEAD

Also list changed files so you can map each to the relevant rule:

bash
gh pr view <pr> --json files --jq '.files[].path'
# or
git diff --name-only main...HEAD

3. Layer classification — do this BEFORE the checklist

For every new file, new type, new store field, or new exported helper introduced by the diff, answer one question: which package does this belong to — core, viewer, editor, or nodes? If the answer is "editor" but the code lives in packages/core or packages/viewer (or vice versa), or if kind-specific code lands anywhere other than packages/nodes/src/<kind>/, flag it as a blocker. This is the most common and most damaging class of violation, and the checklist below won't reliably catch it on its own — do this pass explicitly.

The four packages and what they own

packages/core — domain data + pure logic. Owns: node schemas, the scene store (useScene), live transforms store, core systems (wall mitering, slab polygons, space detection), event bus, plain 2D/3D math helpers, sceneRegistry, the registry primitives (nodeRegistry, registerNode, loadPlugin, discoverPlugins/setPluginDiscovery, SceneApi, Plugin/NodeDefinition types). Consumed by every downstream package, including read-only embeds. Must not know about: Three.js/R3F, packages/viewer, apps/editor, packages/nodes, any rendering or UI concept, any tool/mode/phase concept, or any view-specific concept (floorplan, paint preview, cursor indicators, selection outline styling).

packages/viewer — the 3D canvas, shippable standalone. Owns: <Viewer>, the generic <NodeRenderer> / <ParametricNodeRenderer> / <GeometrySystem> / <RegisteredSystems> / <FloorplanRegistryLayer> plumbing, viewer systems (cutouts, zones, level positions, scans), the viewer store (useViewer) for genuine presentation state only (selection path, camera/level/wall/view modes, theme, display toggles, hover id), useNodeEvents. Consumed by both the editor and the read-only /viewer/[id] route. Must not know about: editor state (useEditor, tools, phases, modes), editor-only names baked into presentation modes ('delete', 'paint-ready'), editor-only state types (material preview, active paint target, floorplan anything), packages/nodes.

packages/editor (and apps/editor) — the editing experience. Owns: the tool framework (useDragAction, ParametricInspector, <MoveRegistryNodeTool>, the registry-aware dispatchers in tool-manager.tsx / MoveTool / panel-manager.tsx / helper-manager.tsx), useEditor, action menus, panels, the floorplan panel and its helpers, paint mode, selection-manager phase/mode logic, cursor badges, command palette, keyboard shortcuts — anything absent from the read-only viewer route. Injects itself into <Viewer> via children and props, never the reverse. Must not import from packages/nodes.

packages/nodes — the built-in plugin (pascal:core). Owns: one folder per node kind (packages/nodes/src/<kind>/) containing definition.ts, schema.ts, optionally geometry.ts / renderer.tsx / system.tsx / floorplan.ts / tool.tsx / move-tool.tsx / panel.tsx / parametrics.ts / preview.tsx. Exports builtinPlugin. Depends on editor, viewer, and core via their public surfaces — the same surfaces a third-party plugin uses (peer-dep style). Nothing in core/, viewer/, or editor/ may import from @pascal-app/nodes. The dependency arrow is one-way: framework code consults nodeRegistry, never reaches into a specific kind's folder.

Triggers that mean "this is probably in the wrong package"
  1. Would the read-only /viewer/[id] route need this? If no, it belongs in apps/editor / packages/editor.
  2. Does the name contain an editor-specific word? (Floorplan, Paint…, Draft…, Marquee, CursorBadge, HoverMode, …Tool, Moving…, Curving….) Default to editor and justify loudly if it's anywhere else.
  3. Does the type or field reference a tool/mode/phase vocabulary? ('delete', 'paint-ready', 'material-paint', 'site'/'structure'/'furnish', 'build'/'edit'.) Belongs in useEditor, not useViewer or core.
  4. Does the helper compute something only a 2D editor view needs? (Floorplan transforms, measurement offsets, SVG path builders, marquee bounds scoped to floorplan.) Editor. Generic 2D geometry that any view could use (polygon math, rotation, clamping, line thickening) can live in core as long as its names are generic — no Floorplan prefix.
  5. Does a new store field have a setter that no part of the target layer ever calls? (e.g. setMaterialPreview in useViewer that only the editor would ever invoke.) That's a layering smell — the state belongs in the caller's layer.
  6. Does the new file mention a specific kind by name? (door-…, wall-…, item-…, etc.) Then it belongs in packages/nodes/src/<kind>/, not under packages/viewer/src/components/renderers/<kind>/, packages/viewer/src/systems/<kind>.ts, packages/editor/src/components/tools/<kind>/, or packages/editor/src/components/ui/panels/<kind>-panel.tsx. Those legacy locations were deleted at Phase 6 cleanup — reintroducing one is a regression to the dispatch model.
  7. Does an import line read from '@pascal-app/nodes' inside core/, viewer/, or editor/? Blocker. The Biome noRestrictedImports overrides in biome.jsonc already ban this, along with the rest of the arrow (core → three/R3F/viewer/editor, viewer → editor); if one slipped through, a framework package is reaching down into the plugin or up the arrow.

Write the classification down before writing findings. If core gains "Floorplan" types, the viewer gains paint-mode vocabulary, a renderer grows editor awareness, or a kind-specific file appears outside packages/nodes/src/<kind>/ — those are the blockers to lead with, not downstream symptoms.

4. Review checklist

A. Package boundaries
  • packages/viewer/** does not import from @pascal-app/editor, apps/editor, or @pascal-app/nodes, and does not reference useEditor, tool state, phase, or mode.
  • packages/core/** does not import Three.js, react-three-fiber, @pascal-app/viewer, @pascal-app/editor, or @pascal-app/nodes.
  • packages/editor/** does not import from @pascal-app/nodes.
  • packages/core/** does not introduce types or helpers named after an editor view (Floorplan*, Paint*, Draft*). Generic plan-geometry helpers are fine; view-specific vocabulary is not.
  • No new case '<kind>': clauses (or equivalent kind-specific branching keyed on node.type) inside packages/viewer/** or packages/editor/**. Phase 6 deleted these; the dispatch happens via nodeRegistry. The exceptions left in tree are treeNodeByType (a lookup map, not a switch) and unit-formatting switches (centimeters / feet / inches). Any new case 'door'|'wall'|'item'… in a framework package is a blocker — the behavior belongs on the kind's NodeDefinition.
  • Tools mutate useScene (committed state) and useLiveTransforms (ephemeral drag state); direct sceneRegistry mesh transforms are allowed only under the live-drag exception in wiki/architecture/tools.md. No business logic, no imports from packages/viewer.
B. Node registry & composition (packages/nodes)

If the PR adds or modifies a node kind, check against wiki/architecture/node-definitions.md and wiki/architecture/plugin-authoring.md:

  • Three independent fields: def.geometry?: (node, ctx) => Object3D, def.renderer?: () => Promise<{ default }>, def.system?: () => Promise<{ default }>. There is no discriminator — presence is participation. Setting all three is fine if the kind genuinely needs them; setting a def.system whose only job is to rebuild geometry on dirty is a smell — collapse to def.geometry and let <GeometrySystem> do the work.
  • Builders must be pure. A def.geometry function must not import useScene, must not mutate the store, and must not depend on React context. Read other nodes via GeometryContext (ctx.resolve / ctx.children / ctx.siblings / ctx.parent).
  • Builders emit local-space children. <ParametricNodeRenderer> binds <group position={liveTransform?.position ?? node.position}> in JSX. A builder that bakes world position into vertex coords, or a system that imperatively writes group.position / group.rotation, will desync R3F's prop binding — the node will snap to (0,0,0) after rebuild. Flag any imperative group.position.set(...) inside def.geometry or a registered system. (Tool-driven sceneRegistry.nodes.get(id).position.set(...) during a live drag is fine and is the documented pattern — see hook hygiene below.)
  • Tag geometry-built children. <GeometrySystem> only disposes children carrying userData.__fromGeometry = true. Custom systems that imperatively add children to a registered group must follow the same convention if the group can host React-mounted children (e.g. shelf surfaces hosting items).
  • One registered mesh per node ID. If a custom renderer mounts multiple objects, register the parent group (or whichever object the system needs to address via sceneRegistry.nodes.get(id)).
  • Previews must clone cached materials. If def.preview calls the geometry builder and then sets material.opacity = 0.5, but the builder caches materials at module scope (most do, keyed on material / materialPreset), the mutation leaks into every committed instance. Clone, mutate the clone, reassign mesh.material, dispose only the clone on unmount. Reference: nodes/src/shelf/preview.tsx.
  • Schema changes must keep old scenes loadable. Any diff that adds, renames, removes, or retypes a property on a node schema needs a load path for scenes saved before the change (parsed through AnyNode in SceneState.setScene). A new field needs a Zod .default() / .optional(). A rename / removal / retype needs a migrateNodes entry in packages/core/src/store/use-scene.ts that rewrites the legacy shape before parse — a .default() alone silently drops the old value. A schema diff that does neither is a blocker: it breaks every existing scene. See wiki/architecture/node-schemas.md § Schema Evolution.
  • Host kinds need children on the schema. If def.relations.hosts is set, the schema must declare children: z.array(z.string()).default([]) (and migrateNodes must patch existing scenes). Otherwise useScene.createNode(child, parentId) writes a parent.children entry into nothing and the host never sees the new child.
  • Movable opt-in. MoveTool dispatches to MoveRegistryNodeTool only when def.capabilities.movable is set. Kinds with bespoke move semantics (wall endpoint drag with linked-wall cascade, slab vertex edit, etc.) deliberately omit movable and supply def.affordanceTools.move instead. Force-routing a bespoke-move kind through generic dispatch (nodeRegistry.has(kind) instead of def.capabilities.movable) is a regression — call it out. See wiki/architecture/node-definitions.md § "Opting out of a generic path".
  • Paint dispatch lives on def.capabilities.paint. A paintable kind declares resolveRole / buildPatch / applyPreview (+ optional getEffectiveMaterial) on PaintCapability; the editor's selection-manager routes hover / click / preview through the generic dispatcher. A PR that adds an if (node.type === '<kind>') arm to paint-mode handling, paint-preview application, or material picker resolution is a regression — the behaviour belongs on the kind's paint capability. See packages/core/src/registry/types.ts (PaintCapability).
  • Slots + world-scale UVs for paintable surfaces. A new kind (or geometry change) that exposes paintable parts must follow the unified slot convention, not reinvent it:
    • Paintable parts are slots, carried on the node. Overrides live in a slots record (slotId → MaterialRef, scene:/library:) on the schema, resolved via def.capabilities.paint — not ad-hoc per-surface material / materialPreset fields, and not a parallel store. A new paintable kind whose schema lacks slots (or whose duplicate / preset / clone path drops it) is a blocker: it silently loses painted materials. (Slots are plain data — generic clone/parse preserves them; bespoke draft-rebuild placement paths must thread them through explicitly. Reference bug: item duplicate rebuilt the draft from asset and dropped slots.)
    • Texturable geometry emits UVs in metres (1 UV unit = 1 m). Any def.geometry producing a surface a finish can tile onto must generate UVs at the same world scale walls / slabs / roofs use, because catalog finishes set repeat as tiles-per-metre. Unitless, bounding-box-normalised, or hardcoded UVs that don't scale with the surface are a blocker — finishes won't tile consistently. Flat-colour-only surfaces need no UVs. GLB item authoring follows the same contract via slot_-prefixed materials (case-insensitive, slot_ → slot id). See wiki/architecture/materials-and-themes.md § "Texture world scale" and wiki/architecture/item-authoring.md.
  • Floor elevation lives on def.capabilities.floorPlaced. Kinds that rest on a level and lift over overlapping slabs declare a footprint (and optional applies predicate) on FloorPlacedConfig; the generic <FloorElevationSystem> writes slabElevation + node.position[1] onto the registered mesh on each dirty mark. A new per-kind useEffect / per-kind system that recomputes Y from slab overlap is a regression — the per-kind block was lifted out of ItemSystem in Phase 6.1. See packages/core/src/registry/types.ts (FloorPlacedConfig).
  • Render-mode behaviour goes through def.surfaceRole. Solid / Rendered / Clay viewer modes look up the kind's surfaceRole on NodeDefinition to pick the right material strategy (clay overrides, edge passes, theme overlays). New if (node.type === '<kind>') arms inside the render-mode pipeline or theme application are a regression — the behaviour belongs on def.surfaceRole. See packages/core/src/registry/types.ts:529.
  • Capability names must describe verbs, not host kinds. movable / paint / floorPlaced / cuts / selectable describe what the node does. A new capability named after a host kind (slabAccessory, wallAccessory, siteAccessory, anything Xaccessory / Xhosted shaped) couples the registry's type surface to one specific host and reads as precedent for the next reviewer. Blocker. Push back: generalise into a paired host-side capability ("I merge subtractive accessories from my children") + accessory-side capability ("I provide a cut geometry, cascade my dirty mark to my host's parent"); cuts go through capabilities.cuts (cut intents, packages/core/src/schema/cut.ts). The single existing case — capabilities.roofAccessory (packages/core/src/registry/types.ts:791, consumed by packages/viewer/src/systems/roof/roof-system.tsx) — is documented tech debt; do not extend the pattern.
  • Per-kind files in legacy locations are a regression. New viewer/src/components/renderers/<kind>/*, viewer/src/systems/<kind>-system.tsx, editor/src/components/tools/<kind>/*, editor/src/components/ui/panels/<kind>-panel.tsx, editor/src/components/ui/helpers/<kind>-helper.tsx, or inline useMemo floor-plan entry-builders inside editor/src/components/editor/floorplan-panel.tsx — all of these were systematically deleted at Phase 6. The behavior belongs on the kind's NodeDefinition (def.renderer / def.system / def.geometry / def.tool / def.affordanceTools / parametrics.customPanel / def.toolHints / def.floorplan).
  • Floor-plan output via def.floorplan. New per-kind floor-plan rendering must return FloorplanGeometry from def.floorplan(node, ctx) and be rendered by <FloorplanRegistryLayer>. New inline branches in floorplan-panel.tsx are a blocker.
  • Plugin contract surface. A PR that extends the v1 plugin surface — adding plugin.materials, plugin.systems, plugin.panels, or making plugins extend host stores (useScene / useEditor / useViewer) — is out of scope for the v1 contract documented in wiki/architecture/plugin-authoring.md. Either the change belongs as a new field on NodeDefinition (additive, doesn't bump apiVersion) or it needs its own plan.
Show full SKILL.md (1,355 more words)Show less
C. Hook hygiene (useEditor, useScene, useViewer)
  • Stores hold state + setters only. No business logic, side effects, async work, or derived computations inside the store definition.
  • Derived values belong in selectors or systems, not in the store body.
  • No cross-store coupling: a store's action should not call another store's actions inside itself.
  • New state added to useViewer must be presentation-only (selection, camera, level mode, display toggles). Editor-only state (active tool, phase, edit mode, paint preview, floorplan state) goes in useEditor.
  • Node code does not import useScene directly. A kind's geometry / system / tool should read and write through SceneApi (passed in by the framework) or GeometryContext. Direct useScene.getState() calls inside packages/nodes/src/<kind>/ are a smell — they bypass the registry's IoC point and make the code harder to test.
  • Live drag motion is imperative, not store-driven. Tools must not call useLiveTransforms.set(...) per grid:move tick to animate registered parametric kinds — the selector path doesn't reliably re-render and the mesh visibly disappears mid-drag. Use sceneRegistry.nodes.get(node.id)?.position.set(x, y, z) instead, and commit once at the end via useScene.temporal.getState().resume() → updateNode → pause(). The reference implementation is MoveRegistryNodeTool. This is the only sanctioned use of imperative mesh transforms by a tool; flag any other location that does the same.
  • Data-driven drags preview via useLiveNodeOverrides, never per-tick useScene. A kind whose geometry is recomputed from data fields (wall start/end, opening host-cut, endpoint reshape) previews by publishing field patches to useLiveNodeOverrides (merged by getEffectiveWall / getEffectiveNode), writing the scene store once on commit. A tool that calls useScene.updateNodes/updateNode on grid:move (or any per-pointer-move tick) is a blocker — it swaps the nodes map ref and re-renders every useScene(s => s.nodes) subscriber app-wide each frame. markDirty per tick is fine for bounded gestures (drag marks drain every frame); a useFrame/animation loop that marks dirty for as long as something animates is a blocker — the scene can then never settle to DIRTY 0. Animations signal rebuilds through their own records (useInteractive animations), marking dirty once on completion; see wiki/architecture/node-definitions.md § "geometry + system". Grep tell: updateNode(s)?( in an onGridMove/onMove/applyPreview path under packages/nodes/src/<kind>/. See wiki/architecture/tools.md § "Data-driven live drag".
D. Selector performance
  • Top-level components (pages, layouts, providers, <Viewer> siblings) must not subscribe to large or frequently-changing slices — e.g. useScene(s => s.nodes), useScene(s => s). Flag these: they re-render the whole subtree on every mutation.
  • Selectors that return new object or array references each call (e.g. s => ({ a: s.a, b: s.b }), s => s.items.filter(...)) without a custom equality function (shallow or custom) are re-render hazards.
  • Prefer subscribing by ID deep in the tree (one node per renderer) over subscribing to the full collection high up.
  • Inside a <XxxPanel> (legacy or parametrics.customPanel-mounted), avoid useScene(s => s.nodes[selectedId]) as a callback dep — it changes every tick and pushes useCallback into infinite-loop territory. The recipe is in wiki/architecture/node-definitions.md § "Custom panels: keep handlers stable during slider drags".
  • Per-node list renderers subscribe per-node, not to the whole live Map. A list that draws one child per node (FloorplanRegistryLayer → FloorplanRegistryEntry) must have each child subscribe to its own slice (useLiveTransforms(s => s.transforms.get(id)) / overrides.get(id)) and be memo'd with referentially stable props; the parent subscribes only to the stable id list. Subscribing the parent or a child to the whole transforms/overrides Map, dropping a memo, or passing unstable props re-renders all N children every drag tick — a flood that type-checks and passes tests. Sibling invalidation goes through a per-node epoch, not a whole-layer re-render. See wiki/architecture/tools.md § "Floorplan registry: per-node subscriptions".
E. Separation of concerns
  • Viewer and core stay unaware of editor-specific concepts (tools, phases, active modes, editor UI state, view-specific helpers).
  • Editor-only overlays and systems are injected as children of <Viewer>, not added inside the viewer package.
  • Editor overlay meshes must carry the editor layer. Any new <mesh> / <line*> / <points> / <sprite> an editor overlay or tool component adds to the 3D scene (gizmos, handles, guides, previews, cursor meshes, marquees) must set layers={EDITOR_LAYER} — or GRID_LAYER for the ground grid, ZONE_LAYER for zone fills. The thumbnail/snapshot camera renders only layer 0, so an untagged overlay leaks into exports (and gets inked / SSGI-darkened in the live view). Flag any overlay-component primitive that omits the layer assignment. See wiki/architecture/layers.md.
  • New node types are added by creating one folder under packages/nodes/src/<kind>/ and registering its definition in builtinPlugin.nodes. Adding to a hand-maintained list elsewhere is a sign the registry hasn't absorbed that surface yet. A few such lists remain (AnyNode, treeNodeByType in the outliner, the allTypes seeds in editor/src/components/editor/selection-manager.tsx, the built-in tool unions in use-editor.tsx): extending one of those is expected; a new one is a finding.
  • AnyNode is hand-maintained for now (full runtime derivation would lose static typing); packages/nodes/src/index.test.ts is the drift gate. If a PR adds a kind to AnyNode without adding it to builtinPlugin.nodes (or vice versa), the parity test catches it — but flag it in review too.
F. Interaction scope, snapping & modifiers

Apply when the diff touches a tool, a move-tool / selection / endpoint / reshape file, lib/interaction/**, lib/snapping-mode.ts, or use-interaction-scope. Source of truth: wiki/architecture/interaction-scope.md and wiki/architecture/tools.md.

  • No new useEditor interaction flag. "What the user is doing" is owned by useInteractionScope (begin / update / end / endIf). A new useEditor boolean for an in-flight interaction (moving…, curving…, dragging…, editing…, …InFlight) is a blocker — it goes through the scope. The legacy mirror flags are being retired, not extended.
  • Snapping is mode-driven; Shift is not a bypass. A tool / move-tool / selection file that reads event.shiftKey, event.nativeEvent?.shiftKey, or modifiers.shiftKey to bypass snapping (raw cursor, skip grid, skip angle) is a blocker — the convention is Shift = cycle the mode, Alt = force/free. Snap state must come from isGridSnapActive() / isMagneticSnapActive() / isAngleSnapActive(). Grep tell: shiftKey near a snap / step / projectToAngleLock / alignment expression in packages/nodes/src/<kind>/{tool,move-tool,selection}.tsx. (Shift for multi-select in select mode, or a documented topology opt-out, is fine — confirm which it is.)
  • No hardcoded, ungated grid step. A quantize that isn't gated on isGridSnapActive() — always useEditor.getState().gridSnapStep, or a constant WALL_GRID_STEP / 0.5 / getSegmentGridStep() applied unconditionally — ignores the active mode and is a blocker. The gated form is const step = isGridSnapActive() ? useEditor.getState().gridSnapStep : 0.
  • Snappable kinds declare snapProfile. A kind whose tool snaps but whose NodeDefinition omits snapProfile ('item' | 'structural') gets no contextual chip and the wrong default mode-set — flag it (suggestion, blocker if it ships a bespoke per-kind snapping switch instead).
  • Bespoke movers must not open a moving scope. useMovingNode() reads the scope, and tool-manager mounts the generic MoveRegistryNodeTool whenever it's non-null. A bespoke move-tool.tsx that calls begin(movingScope(...)) or setMovingNode(node) re-creates the dual-path double-handling (FPS collapse / teleport on move). Blocker. Mode-driven snapping inside a bespoke mover must resolve the mode without a global moving / reshaping scope (see interaction-scope.md § "Snapping mode & modifiers").
  • event.altKey is not an alignment bypass. A drafting/preview path that reads event.altKey to suppress Figma-alignment is a blocker in any new or touched tool — alignment follows the magnetic snap mode (bypass: !isMagneticSnapActive()). Alt is force/free for placement/move; it is not a snap/alignment modifier. The one sanctioned Alt use outside force is the wall/fence chain-mode toggle (clean Alt-tap → cycleWallChainMode / cycleFenceChainMode, via hooks/use-keyboard.ts isChainModeContext()), allowed only because wall/fence drafting has no force role. Grep tell: event.altKey near an align / bypass expression in a tool.tsx / floorplan preview path.
  • Known-legacy exceptions (migrate on touch). Tracked debt in wiki/architecture/interaction-scope.md (Known-legacy); a PR that touches one must migrate it, not extend it; a new tool on either legacy pattern is a blocker regardless. (1) shiftKey snap-bypass in the MEP move/endpoint tools (packages/nodes/src/{duct-segment,pipe-segment,liquid-line,lineset,duct-fitting}/{move-tool,selection}.tsx). (2) altKey alignment-bypass in the roof / polygon / slab pointer-move previews (components/editor/floorplan-panel.tsx) and the resolveSlabPlanPointSnap / resolveCeilingPlanPointSnap paths. Already migrated — do not regress: wall + fence drafting (both modifier patterns) and zone drafting (components/tools/zone/zone-tool.tsx — mode-driven grid/angle gates, no Shift bypass).

5. Output format

Group findings by severity:

  • Blocker — violates a rule in wiki/architecture/ or breaks a layer/package boundary. Must be fixed before merge.
  • Suggestion — likely problem, worth discussing. Not a hard block.
  • Nit — minor, optional.

For each finding, include:

  1. File and line: path/to/file.ts:42
  2. The offending snippet (short — 1–5 lines)
  3. The rule it violates, linked to the wiki page (e.g. wiki/architecture/viewer-isolation.md, wiki/architecture/node-definitions.md)
  4. A concrete proposed fix

Skip formatting, import ordering, and anything CI already covers.

If the PR fully complies, say so explicitly — do not invent nits to appear thorough.

6. Final summary

End with:

  • Blocker count, suggestion count, nit count
  • One-sentence verdict: ready to merge / needs changes / needs discussion
  • If blockers exist, list the files the author should open first

© pascalorg, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

Just SKILL.md in .agents/skills/review-architecture of pascalorg/editor.

Open the folder on GitHubat commit a860e19

Compare with similar skills

Pascal Architecture PR Review next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Pascal Architecture PR Review compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Pascal Architecture PR Review this skillpascalorg/editor25k—~7.5kAutomated safety check: PassMIT
Code Review Skillawesome-skills/code-review-skill2.1k—~2.8kAutomated safety check: NotesMIT
Code Review SkillRain-kl/OpenFlare288—~2.3kAutomated safety check: NotesMIT
Code Review Excellenceandrew-yangy/gru-ai155—~1.7kAutomated safety check: NotesMIT
Code Reviewerrevfactory/harness-1001.3k—~1.8kAutomated safety check: PassApache-2.0
Conduct Code Review RoundFerroxLabs/wayland608—~3.9kAutomated safety check: PassApache-2.0

Similar skills

  • Code Review Skill

    awesome-skills/code-review-skill

    Provides comprehensive code review guidance for React 19, Vue 3, Angular 17+, Svelte 5, Rust, TypeScript, Java, Java 8, PHP, Ruby, Rails, Python, Django, FastAPI, Go, C/.NET, Kotlin, Swift, Dart…

    2.1k GitHub stars~2.8k tokensUpdated 29 days ago
    DevelopmentAuto-check: notes
  • Code Review Skill

    Rain-kl/OpenFlare

    Provides comprehensive code review guidance for React 19, Vue 3, Angular 17+, Svelte 5, Rust, TypeScript, Java, PHP, Python, Django, Go, C/.NET, Kotlin, Swift, NestJS, C/C++, and more.

    288 GitHub stars~2.3k tokensUpdated today
    DevelopmentAuto-check: notes
  • Code Review Excellence

    andrew-yangy/gru-ai

    Provides comprehensive code review guidance for React 19, Vue 3, Rust, TypeScript, Java, Python, and C/C++.

    155 GitHub stars~1.7k tokensUpdated 7 mo ago
    DevelopmentAuto-check: notes
  • Code Reviewer

    revfactory/harness-100

    Full pipeline for automated code review. An agent skill from revfactory/harness-100.

    1.3k GitHub stars~1.8k tokensUpdated 6 mo ago
    DevelopmentAuto-check passed
  • Conduct Code Review Round

    FerroxLabs/wayland

    Orchestrates a thorough code review process by chaining four engineering skills into a structured review pipeline.

    608 GitHub stars~3.9k tokensUpdated yesterday
    DevelopmentAuto-check passed
  • PR Babysitter

    openinterpreter/openinterpreter

    Watches an open GitHub pull request until it merges, handling review comments, diagnosing CI failures and retrying flaky checks along the way.

    69k GitHub starsUsed in 3 repos~4.2k tokens
    DevelopmentAuto-check passed

More from pascalorg/editor

  • Furniture Fit Check

    pascalorg/editor

    Checks whether a sofa, table, bed or appliance fits in a measured Pascal room and reports only what the evidence supports, or asks for the missing measurements.

    25k GitHub stars~5.1k tokensUpdated yesterday
    Auto-check passed
  • Pascal Editor PR Opener

    pascalorg/editor

    Opens or refreshes a pull request on pascalorg/editor from the current branch, describing only what the branch's commits and diff actually contain.

    25k GitHub stars~619 tokensUpdated yesterday
    Auto-check passed
  • Pascal 3D Scene Editing

    pascalorg/editor

    Connects an agent to Pascal's MCP tools to create, inspect, edit, validate and save editable 3D building scenes, then return a verified editor link.

    25k GitHub stars~2.2k tokensUpdated yesterday
    Auto-check passed

Categories

Questions about Pascal Architecture PR Review

What does Pascal Architecture PR Review do?

Reviews a pull request against the Pascal editor's architectural rules: package boundaries, registry-driven node composition, hook hygiene and selector performance. The agent first loads the architecture wiki pages as the source of truth, always reading the core set on layers, systems, renderers, tools, viewer isolation, node definitions, plugin authoring and agent surfaces, and further pages only when the diff touches their area. It then fetches the diff with the gh CLI or git, lists the changed files to map each to a rule, and classifies each by layer before running the checklist.

When should I use Pascal Architecture PR Review?

Pascal Architecture PR Review fits situations like: reviewing a Pascal pull request for architectural violations; auditing a branch before merge against the package boundary rules; checking that a new node follows the registry composition model; looking for selector or hook performance problems in a diff.

How do I install Pascal Architecture PR Review in Claude Code?

Run `npx skills add pascalorg/editor --skill review-architecture -a claude-code`. Or copy the skill folder (.agents/skills/review-architecture in pascalorg/editor) into .claude/skills/review-architecture in your project. Claude Code loads it when a task matches its description.

How do I install Pascal Architecture PR Review in Codex?

Run `npx skills add pascalorg/editor --skill review-architecture -a codex`. Or copy the skill folder (.agents/skills/review-architecture in pascalorg/editor) into .agents/skills/review-architecture in your project. Codex loads it when a task matches its description.

Can I use Pascal Architecture PR Review in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add pascalorg/editor --skill review-architecture -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/review-architecture, .gemini/skills/review-architecture, .github/skills/review-architecture and .opencode/skills/review-architecture in your project.

What does Pascal Architecture PR Review need to run?

Going by SKILL.md and its folder, Pascal Architecture PR Review needs the command-line tools its instructions call (gh and git). Our summary lists: The gh CLI and git; The Pascal repository's architecture wiki pages. Its frontmatter pre-approves these tools: Bash(git *), Bash(gh *), Read, Grep, Glob.

Does Pascal Architecture PR Review access the network?

SKILL.md contains no URLs. Its commands use gh and git, which can reach the network depending on how they are called. This is read from the text; nothing was executed.

Is Pascal Architecture PR Review safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Pascal Architecture PR Review use?

Pascal Architecture PR Review is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Pascal Architecture PR Review use?

About 7.5k tokens (SKILL.md is roughly 30k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.

What are the alternatives to Pascal Architecture PR Review?

Skills that share tags, products or a category with Pascal Architecture PR Review: Code Review Skill (awesome-skills/code-review-skill, 2.1k stars), Code Review Skill (Rain-kl/OpenFlare, 288 stars), Code Review Excellence (andrew-yangy/gru-ai, 155 stars) and Code Reviewer (revfactory/harness-100, 1.3k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Pascal Architecture PR Review?

pascalorg (a GitHub organization) maintains it in pascalorg/editor, which has 24,695 GitHub stars. The repository holds 4 skills in this directory. The repository was last updated on October 6, 2026.

Source: pascalorg/editor on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.