Files
editor/packages/mcp/CROSS_CUTTING.md
T
Adrian PerezandClaude Opus 4.7 e8d0b13ff5 feat(mcp,editor): Option A+B storage + 10 agent deliverables (Phase 7)
Ships the combined filesystem/Supabase storage adapter + MCP scene
lifecycle tools + Next.js API routes + editor /scene/[id] route, so
an MCP save is directly openable at /scene/<id> without any
injection hack. End-to-end verified: 10/10 e2e steps pass.

Storage (A1/A2/A3):
- SceneStore interface + error classes + slug helpers
- FilesystemSceneStore at $PASCAL_DATA_DIR (defaults XDG/~/.pascal)
  with atomic writes, .index sidecar, optimistic locking
- SupabaseSceneStore with scenes + scene_revisions tables, RLS
  migration SQL, mock-backed unit tests
- createSceneStore(env) auto-selects based on SUPABASE_URL +
  SUPABASE_SERVICE_ROLE_KEY

MCP tools (A4, A8, A9, A10):
- save_scene / load_scene / list_scenes / delete_scene / rename_scene
- list_templates / create_from_template (3 seed templates:
  empty-studio, two-bedroom, garden-house)
- generate_variants (7 mutation kinds, seeded RNG, save=true|false)
- photo_to_scene (vision sampling → scene graph → save)

Editor (A5, A6):
- /api/scenes + /api/scenes/[id] with RFC 7232 If-Match locking
- /scene/[id] and /scenes route pages with save button, SceneLoader
- Removed the window.__pascalScene dev injection hack

Security + UX edges (A7, A8):
- AssetUrl Zod validator: asset:// blob: data:image/ /path https:
  (http://localhost for dev) + PASCAL_ALLOWED_ASSET_ORIGINS env
  allowlist. Hardens scan.url, guide.url, item.asset.src,
  material.texture.url, MaterialMaps.*Map
- Auto-frame camera on empty→non-empty scene transition
  (camera-controls:fit-scene emitter event)

Shared utilities:
- rehydrateSiteChildren() extracted to packages/mcp/src/lib/ and
  used by both create-from-template and generate-variants to work
  around the SiteNode.children-as-objects vs. ids inconsistency
  (CROSS_CUTTING §2)
- Storage + MCP subpath exports added to packages/mcp/package.json
  (CROSS_CUTTING §4)

Tests: 293 pass / 0 fail across 40 files (was 142 pre-Phase-7).
Biome: clean.

Phase-7 e2e script at packages/mcp/test-reports/phase7-e2e.ts:
MCP HTTP + editor Next.js both point at $PASCAL_DATA_DIR =
/tmp/pascal-e2e, save_scene from MCP, GET /api/scenes/<id> from
editor server, /scenes list page renders all saved scenes, scene
page renders SceneLoader, delete_scene works.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-18 19:29:28 +02:00

10 KiB

Cross-cutting changes touching packages outside @pascal-app/mcp

Integrator review required. Each entry documents:

  • What was changed
  • Why (what blocked MCP without it)
  • Impact on existing consumers
  • Reversibility

1. packages/core/package.json — added subpath exports

What

Added these subpath entries to the "exports" map of @pascal-app/core:

  • ./schema./dist/schema/index.js
  • ./store./dist/store/use-scene.js
  • ./material-library./dist/material-library.js
  • ./spatial-grid./dist/hooks/spatial-grid/spatial-grid-manager.js
  • ./wall./dist/systems/wall/wall-footprint.js

The existing "." and "./clone-scene-graph" entries are unchanged.

Why

The main entry (.) re-exports every System* (WallSystem, SlabSystem, CeilingSystem, RoofSystem, ItemSystem, StairSystem, DoorSystem, WindowSystem, FenceSystem) which side-effect-imports three, three-mesh-bvh, and three-bvh-csg. In Node (no browser), three-mesh-bvh's CJS UMD build fails to resolve its three.* globals at module-load time, so merely import { WallNode } from '@pascal-app/core' crashes before any user code runs.

By adding subpath exports that point at modules which don't transitively pull graphics code, the MCP server package (and any future Node consumer) can import just the Zod schemas and the Zustand store without dragging in three and its GPU-bound dependencies.

Impact

Zero on existing consumers. This is purely additive. apps/editor and @pascal-app/viewer continue to import from the main entry and get the full surface — they currently don't use these subpaths and don't need to. No types, runtime behavior, or bundle composition is affected.

Reversibility

Remove the 5 new entries from exports and the change is undone. @pascal-app/mcp would then have to ship its own shim or the core team would need to split @pascal-app/core into a "core-data" package and a "core-systems" package — a larger refactor.

Suggested follow-up (upstream)

Long-term, consider moving systems/ into a separate package @pascal-app/systems so that @pascal-app/core stays data-only. That's a breaking change and out of scope for this PR; the subpath exports are the non-breaking interim fix.


2. SiteNode.children inconsistency (observed, not fixed)

What

packages/core/src/schema/nodes/site.ts:36-38 declares:

children: z.array(z.discriminatedUnion('type', [BuildingNode, ItemNode]))
  .default([BuildingNode.parse({})])

SiteNode.children therefore holds full node objects. Every other container node (building, level, wall, ceiling, roof, stair) stores string[] (IDs) in children.

Why this is a problem

  • Data duplication: the building exists both in nodes[building.id] and embedded inside site.children[0]. Updates to the building in the dict don't propagate to the embedded copy.
  • Traversal asymmetry: "get children of a container" needs site-specific branching.
  • duplicate_level, find_nodes({ parentId }), and scene-serialisation round-trips all need a special case for site.

Why we didn't fix it

Changing the schema is a breaking change to serialised scene data and would require a migration pass inside setScene. Out of scope for a non-breaking MCP addition.

Workaround (inside MCP)

MCP tools resolve node children through the flat nodes dict by scanning for nodes whose parentId matches. This is correct regardless of which representation the schema chose.

Suggested follow-up (upstream)

Align SiteNode.children to z.array(z.string()) + migration in setScene.migrateNodes that extracts embedded building/item objects into the flat dict and replaces them with IDs.


3. .github/workflows/mcp-ci.yml — new CI workflow

What

Adds a CI workflow that runs on pushes to main and on pull requests touching packages/mcp/, packages/core/, or bun.lock. The job installs deps with Bun, builds @pascal-app/core then @pascal-app/mcp, runs bun test in the mcp package, and runs bunx biome check packages/mcp.

Why

The existing .github/workflows/release.yml is workflow_dispatch-only (manual releases for core / viewer). There was no automated pre-merge check for MCP builds/tests. A new workflow is needed so that PRs touching mcp/core are verified before merge.

Impact

None on existing workflows; purely additive. The new workflow only triggers for paths under packages/mcp/, packages/core/, or bun.lock, so unrelated PRs remain unaffected. release.yml is untouched.

Reversibility

Delete .github/workflows/mcp-ci.yml.


4. packages/mcp/package.json — added ./storage subpath export

What

Added a ./storage entry to the "exports" map of @pascal-app/mcp, pointing at the built dist/storage/index.{js,d.ts}. The existing "." entry is unchanged.

Why

The Next.js editor (apps/editor) needs access to createSceneStore() + the SceneStore types/errors in server-only code (API route handlers + lib/scene-store-server.ts). The main entry . pulls in the full MCP server surface (tools, transports, MCP SDK), which is overkill for a consumer that only needs the storage adapter. The subpath export lets apps/editor do import { createSceneStore, SceneVersionConflictError } from '@pascal-app/mcp/storage' without dragging the rest of the package.

Impact

Zero on existing consumers. Purely additive. The . entry continues to export SceneBridge, createPascalMcpServer, etc., exactly as before.

Reversibility

Remove the ./storage entry from exports and update apps/editor to inline the types / use a different factory. No data or behavior changes — pure module-graph shaping.

  • apps/editor/package.json adds @pascal-app/mcp as a workspace dependency so the subpath resolves.
  • apps/editor/lib/scene-store-server.ts and apps/editor/app/api/scenes/** consume this subpath.

5. packages/core/src/schema/asset-url.ts — URL scheme allowlist on scene URL fields

What

Introduced a shared AssetUrl Zod validator and replaced the bare z.string() on every URL-bearing field in core's schemas:

  • scan.url (packages/core/src/schema/nodes/scan.ts)
  • guide.url (packages/core/src/schema/nodes/guide.ts)
  • item.asset.src (packages/core/src/schema/nodes/item.ts)
  • material.texture.url (packages/core/src/schema/material.ts)
  • material.maps.* (albedoMap, normalMap, roughnessMap, metalnessMap, aoMap, displacementMap, emissiveMap, bumpMap, alphaMap, lightMap)

The validator accepts asset://…, blob:…, data:image/…, /… app-relative paths, https://…, and http://localhost|127.0.0.1/…. Optional origin narrowing via process.env.PASCAL_ALLOWED_ASSET_ORIGINS (comma-separated). Rejects javascript:, file:, ftp:, ws:, data:text/html, data:application/*, link-local / private IPs over bare http, empty strings, and non-URL garbage.

Why

Phase 3 security audit (packages/mcp/test-reports/research/R9-production-readiness.md entry "URL validation in scenes"): an attacker-crafted scene containing javascript:alert(1) or http://169.254.169.254/latest/meta-data/ for a texture URL would beacon or exfiltrate when the editor renders it. AnyNode.safeParse, used by the MCP bridge, now rejects those payloads at the schema boundary.

Impact

  • Existing scenes: localStorage-resident scenes bypass strict validation on load (the store's setScene only runs safeParse on stair-type via migrateNodes), so this is not a breakage for returning users. Legacy URLs will keep loading; only explicit MCP-bridge safeParse calls reject.
  • MCP consumers: one existing test (packages/mcp/src/bridge/scene-bridge.test.ts, previously using src: 'data:model/gltf-binary;base64,') now fails because data:model/ is not in the allowlist. Replaced with asset://test/chair.glb — the only sanctioned scheme for an in-repo ItemNode fixture.
  • Other packages: @pascal-app/viewer, @pascal-app/editor, and material-library.ts all continue to work because every built-in URL is a /material/… app-relative path (allowlisted).

Known gaps / follow-ups

  1. item.asset.thumbnail stayed untyped — the field is still bare z.string() in item.ts. The Phase 3 audit called it out alongside src, but the Phase 7 task scope only required src. Follow-up: apply AssetUrl to thumbnail as well. Verify the place-item tool's default thumbnail: '' (currently empty string) gets a proper fallback first.
  2. dist/ pollutionpackages/core/tsconfig.json includes src and doesn't exclude **/*.test.ts, so the new asset-url.test.ts is emitted to dist/schema/. Harmless (nothing imports it), but should be excluded for a clean publish. Mirror the exclude: ["**/*.test.ts"] pattern used in packages/mcp/tsconfig.json. Out of scope for A7 because tsconfig is not in the ownership list.
  3. bun:test typing — the test file uses @ts-expect-error on its bun:test import because @pascal-app/core does not depend on @types/bun. Adding it as a dev dep (or, preferred, excluding tests from the core tsc build per gap 2) would remove the directive.
  4. data:image/svg+xml loophole — passes the validator because it starts with data:image/, but SVG can carry inline scripts. If the editor ever renders SVG via unsanitised HTML-injection APIs or <foreignObject>, this becomes an injection vector. Consider a stricter variant (data:image/(png|jpe?g|webp|gif)) for texture slots where SVG isn't needed.
  5. Same-origin HTTP sceneshttp://localhost is allowed for dev, but a scene persisted in dev and shared in prod will still validate. Consider gating on NODE_ENV once we have a stable env-flag story.

Reversibility

Delete packages/core/src/schema/asset-url.ts and revert the five imports in scan.ts, guide.ts, item.ts, and material.ts to z.string(). The scene-bridge test update is self-contained.