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>
This commit is contained in:
Adrian Perez
2026-04-18 19:29:28 +02:00
co-authored by Claude Opus 4.7
parent 42bd05db9c
commit e8d0b13ff5
81 changed files with 8933 additions and 1213 deletions
@@ -0,0 +1,24 @@
import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
import type { SceneBridge } from '../../bridge/scene-bridge'
import type { SceneStore } from '../../storage/types'
import { registerPhotoToScene } from './photo-to-scene'
/**
* Register the `photo_to_scene` orchestrator tool. Chains the vision
* (`analyze_floorplan_image`-equivalent sampling call) → SceneGraph
* synthesis → optional `SceneStore.save` → `bridge.setScene` so callers get a
* navigable Pascal scene from a single photo upload.
*/
export function registerPhotoToSceneTool(
server: McpServer,
bridge: SceneBridge,
store: SceneStore,
): void {
registerPhotoToScene(server, bridge, store)
}
export {
photoToSceneInput,
photoToSceneOutput,
registerPhotoToScene,
} from './photo-to-scene'
@@ -0,0 +1,194 @@
import { describe, expect, test } from 'bun:test'
import { Client } from '@modelcontextprotocol/sdk/client/index.js'
import { InMemoryTransport } from '@modelcontextprotocol/sdk/inMemory.js'
import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
import { CreateMessageRequestSchema } from '@modelcontextprotocol/sdk/types.js'
import { SceneBridge } from '../../bridge/scene-bridge'
import { InMemorySceneStore } from '../scene-lifecycle/test-utils'
import { registerPhotoToScene } from './photo-to-scene'
type Handler = (req: unknown) => unknown | Promise<unknown>
/**
* Build a connected client/server pair for the `photo_to_scene` orchestrator.
* Optionally advertises the `sampling` capability on the client and installs
* a mock sampling handler that returns a caller-provided reply.
*/
async function makeWiredPair(opts: { withSampling: boolean; samplingHandler?: Handler }): Promise<{
client: Client
bridge: SceneBridge
store: InMemorySceneStore
}> {
const bridge = new SceneBridge()
bridge.setScene({}, [])
const store = new InMemorySceneStore()
const server = new McpServer({ name: 'test', version: '0.0.0' })
registerPhotoToScene(server, bridge, store)
const [srvT, cliT] = InMemoryTransport.createLinkedPair()
const client = new Client(
{ name: 'test-client', version: '0.0.0' },
{
capabilities: opts.withSampling ? { sampling: {} } : {},
},
)
if (opts.withSampling && opts.samplingHandler) {
const handler = opts.samplingHandler
client.setRequestHandler(
CreateMessageRequestSchema,
async (request) =>
// Cast to unknown — tests return arbitrary shapes to exercise
// parse/validation paths in the tool handler.
(await handler(request)) as never,
)
}
await Promise.all([server.connect(srvT), client.connect(cliT)])
return { client, bridge, store }
}
const VALID_VISION_JSON = {
walls: [
{ start: [0, 0], end: [5, 0], thickness: 0.2 },
{ start: [5, 0], end: [5, 4] },
{ start: [5, 4], end: [0, 4] },
{ start: [0, 4], end: [0, 0] },
],
rooms: [
{
name: 'Living Room',
polygon: [
[0, 0],
[5, 0],
[5, 4],
[0, 4],
],
approximateAreaSqM: 20,
},
],
approximateDimensions: { widthM: 5, depthM: 4 },
confidence: 0.82,
}
const VALID_REPLY = {
model: 'mock-model',
role: 'assistant',
content: {
type: 'text',
text: JSON.stringify(VALID_VISION_JSON),
},
}
describe('photo_to_scene', () => {
test('happy path: vision reply → walls + rooms + scene in bridge + saved', async () => {
const { client, bridge, store } = await makeWiredPair({
withSampling: true,
samplingHandler: () => VALID_REPLY,
})
const result = await client.callTool({
name: 'photo_to_scene',
arguments: {
image: 'aGVsbG8=',
scaleHint: '1 cm = 1 m',
name: 'Test Scene',
},
})
expect(result.isError).toBeFalsy()
const structured = result.structuredContent as {
sceneId?: string
url?: string
walls: number
rooms: number
confidence: number
}
expect(structured.walls).toBe(4)
expect(structured.rooms).toBe(1)
expect(structured.confidence).toBe(0.82)
expect(typeof structured.sceneId).toBe('string')
expect(structured.url).toBe(`/scene/${structured.sceneId}`)
// Bridge was swapped.
const rootIds = bridge.getRootNodeIds()
expect(rootIds.length).toBe(1)
const rootId = rootIds[0]!
const root = bridge.getNode(rootId)
expect(root?.type).toBe('site')
// Walls and zones exist in the flat dict.
const allNodes = Object.values(bridge.getNodes())
const walls = allNodes.filter((n) => n.type === 'wall')
const zones = allNodes.filter((n) => n.type === 'zone')
expect(walls.length).toBe(4)
expect(zones.length).toBe(1)
// Scene was persisted in the store.
const saved = await store.load(structured.sceneId!)
expect(saved).not.toBeNull()
expect(saved?.name).toBe('Test Scene')
})
test('sampling unavailable → sampling_unavailable error', async () => {
const { client } = await makeWiredPair({ withSampling: false })
const result = await client.callTool({
name: 'photo_to_scene',
arguments: { image: 'aGVsbG8=' },
})
expect(result.isError).toBe(true)
const text = (result.content as Array<{ type: string; text: string }>)[0]!.text
expect(text).toContain('sampling_unavailable')
})
test('invalid JSON reply → sampling_response_unparseable', async () => {
const { client } = await makeWiredPair({
withSampling: true,
samplingHandler: () => ({
model: 'mock-model',
role: 'assistant',
content: { type: 'text', text: 'not json at all' },
}),
})
const result = await client.callTool({
name: 'photo_to_scene',
arguments: { image: 'aGVsbG8=' },
})
expect(result.isError).toBe(true)
const text = (result.content as Array<{ type: string; text: string }>)[0]!.text
expect(text).toContain('sampling_response_unparseable')
})
test('save=false → returns graph inline, no sceneId', async () => {
const { client, store } = await makeWiredPair({
withSampling: true,
samplingHandler: () => VALID_REPLY,
})
const result = await client.callTool({
name: 'photo_to_scene',
arguments: {
image: 'aGVsbG8=',
save: false,
},
})
expect(result.isError).toBeFalsy()
const structured = result.structuredContent as {
sceneId?: string
url?: string
walls: number
rooms: number
confidence: number
graph?: { nodes: Record<string, unknown>; rootNodeIds: string[] }
}
expect(structured.sceneId).toBeUndefined()
expect(structured.url).toBeUndefined()
expect(structured.graph).toBeDefined()
expect(Array.isArray(structured.graph?.rootNodeIds)).toBe(true)
expect(structured.graph?.rootNodeIds.length).toBe(1)
expect(structured.walls).toBe(4)
expect(structured.rooms).toBe(1)
// Nothing persisted.
const list = await store.list()
expect(list.length).toBe(0)
})
})
@@ -0,0 +1,427 @@
import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
import { ErrorCode, McpError } from '@modelcontextprotocol/sdk/types.js'
import type { SceneGraph } from '@pascal-app/core/clone-scene-graph'
import type { AnyNodeId, AnyNode as AnyNodeT } from '@pascal-app/core/schema'
import {
AnyNode,
BuildingNode,
LevelNode,
SiteNode,
WallNode,
ZoneNode,
} from '@pascal-app/core/schema'
import { z } from 'zod'
import type { SceneBridge } from '../../bridge/scene-bridge'
import type { SceneStore } from '../../storage/types'
/**
* Input shape for the `photo_to_scene` orchestrator. `image` matches the
* contract documented on `analyze_floorplan_image` — base64 or http(s) URL.
*/
export const photoToSceneInput = {
image: z.string().describe('Base64 or https URL of the floor-plan photo'),
scaleHint: z.string().optional().describe('e.g. "1 cm = 1 m" or "approx 80 m²"'),
name: z.string().default('Scene from photo'),
save: z.boolean().default(true),
defaultWallThickness: z.number().default(0.2),
defaultWallHeight: z.number().default(2.6),
}
export const photoToSceneOutput = {
sceneId: z.string().optional(),
url: z.string().optional(),
walls: z.number(),
rooms: z.number(),
confidence: z.number(),
notes: z.string().optional(),
graph: z.any().optional(),
}
/**
* Shape of the vision JSON we consume. Kept in-sync with
* `analyze_floorplan_image`'s output schema (walls / rooms /
* approximateDimensions / confidence).
*/
const VisionResponseSchema = z.object({
walls: z.array(
z.object({
start: z.tuple([z.number(), z.number()]),
end: z.tuple([z.number(), z.number()]),
thickness: z.number().optional(),
}),
),
rooms: z.array(
z.object({
name: z.string(),
polygon: z.array(z.tuple([z.number(), z.number()])),
approximateAreaSqM: z.number().optional(),
}),
),
approximateDimensions: z.object({
widthM: z.number(),
depthM: z.number(),
}),
confidence: z.number().min(0).max(1),
})
type VisionResponse = z.infer<typeof VisionResponseSchema>
/**
* System prompt mirrors `analyze_floorplan_image` — the contract between
* orchestrator and host is identical, so we keep the prompt verbatim to
* guarantee wire-compatible responses.
*/
const SYSTEM_PROMPT = `You are a vision assistant that extracts structured floor-plan data from an image.
Your ONLY job: return a JSON object that exactly matches this schema — no prose, no markdown fences.
{
"walls": [{ "start": [x, z], "end": [x, z], "thickness": number? }, ...],
"rooms": [{ "name": string, "polygon": [[x,z], ...], "approximateAreaSqM": number? }, ...],
"approximateDimensions": { "widthM": number, "depthM": number },
"confidence": number 0..1
}
Coordinates are in metres. Origin can be the floor plan's centre or bottom-left — be consistent.
If the image is unclear, lower the confidence score but still produce your best attempt.
DO NOT wrap the JSON in markdown. DO NOT explain. Just output the raw JSON.`
const DATA_URI_RE = /^data:(image\/[a-z0-9.+-]+);base64,(.+)$/i
type ImageBlock = {
type: 'image'
data: string
mimeType: string
}
/**
* Resolve the `image` input into a sampling-ready image block. Follows the
* same fetch/data-uri/raw-base64 rules as the vision tool so the user gets
* consistent behaviour whether they call `photo_to_scene` or
* `analyze_floorplan_image` directly.
*/
async function resolveImageBlock(image: string): Promise<ImageBlock> {
if (/^https?:\/\//i.test(image)) {
const res = await fetch(image)
if (!res.ok) {
throw new McpError(
ErrorCode.InvalidParams,
`failed to fetch image: ${res.status} ${res.statusText}`,
{ url: image, status: res.status },
)
}
const buf = Buffer.from(await res.arrayBuffer())
const data = buf.toString('base64')
const mimeType = res.headers.get('content-type') ?? 'image/jpeg'
return { type: 'image', data, mimeType }
}
const dataUriMatch = image.match(DATA_URI_RE)
if (dataUriMatch) {
return {
type: 'image',
mimeType: dataUriMatch[1]!,
data: dataUriMatch[2]!,
}
}
return { type: 'image', mimeType: 'image/jpeg', data: image }
}
/** Collect all text content blocks returned by the sampling host into one string. */
function extractText(
content:
| { type: 'text'; text: string }
| { type: 'image' | 'audio'; data: string; mimeType: string }
| Array<
| { type: 'text'; text: string }
| { type: 'image' | 'audio'; data: string; mimeType: string }
| { type: string; [k: string]: unknown }
>,
): string {
const blocks = Array.isArray(content) ? content : [content]
const texts: string[] = []
for (const block of blocks) {
if (block && typeof block === 'object' && (block as { type?: string }).type === 'text') {
const t = (block as { text?: unknown }).text
if (typeof t === 'string') texts.push(t)
}
}
return texts.join('\n').trim()
}
/**
* Call the host's sampling capability to analyse a floor-plan photo. Throws
* `sampling_unavailable` when the host has not advertised the capability and
* `sampling_response_unparseable` / `sampling_response_invalid` when the
* reply cannot be mapped onto `VisionResponseSchema`.
*/
async function callVisionSampling(
server: McpServer,
image: string,
scaleHint: string | undefined,
): Promise<VisionResponse> {
const caps = server.server.getClientCapabilities()
if (!caps?.sampling) {
throw new McpError(ErrorCode.InvalidRequest, 'sampling_unavailable')
}
const imageBlock = await resolveImageBlock(image)
const instruction = scaleHint
? `Analyze this floor plan. Scale hint: ${scaleHint}. Return ONLY the JSON described by the system prompt.`
: 'Analyze this floor plan. Return ONLY the JSON described by the system prompt.'
const response = await server.server.createMessage({
systemPrompt: SYSTEM_PROMPT,
temperature: 0,
maxTokens: 2000,
messages: [
{
role: 'user',
content: [imageBlock, { type: 'text', text: instruction }],
},
],
})
const text = extractText(response.content as Parameters<typeof extractText>[0])
if (!text) {
throw new McpError(ErrorCode.InternalError, 'sampling_response_unparseable', {
reason: 'no text content returned by host',
})
}
let parsed: unknown
try {
parsed = JSON.parse(text)
} catch (err) {
throw new McpError(ErrorCode.InternalError, 'sampling_response_unparseable', {
raw: text,
reason: err instanceof Error ? err.message : String(err),
})
}
const validation = VisionResponseSchema.safeParse(parsed)
if (!validation.success) {
throw new McpError(ErrorCode.InternalError, 'sampling_response_invalid', {
raw: text,
errors: validation.error.issues,
})
}
return validation.data
}
type BuildResult = {
nodes: Record<AnyNodeId, AnyNodeT>
rootNodeIds: AnyNodeId[]
walls: number
rooms: number
warnings: string[]
levelId: AnyNodeId
}
/**
* Build a SceneGraph (flat `nodes` dict + `rootNodeIds`) from the vision
* response. Uses the schema factories for every node so IDs, defaults, and
* parent linkage match what the core store would produce. Each node is
* revalidated via `AnyNode.safeParse`; failures are dropped with a warning.
*/
function buildSceneGraphFromVision(
vision: VisionResponse,
defaultWallThickness: number,
defaultWallHeight: number,
): BuildResult {
const warnings: string[] = []
// Build the skeleton: site → building → level.
const building = BuildingNode.parse({})
const level = LevelNode.parse({ level: 0 })
const site = SiteNode.parse({ children: [building] })
// Link parent ids so downstream traversal works.
const siteId = site.id as AnyNodeId
const buildingId = building.id as AnyNodeId
const levelId = level.id as AnyNodeId
const linkedBuilding: AnyNodeT = {
...(building as AnyNodeT),
parentId: siteId,
}
const linkedLevel: AnyNodeT = {
...(level as AnyNodeT),
parentId: buildingId,
}
// BuildingNode children stores level ids (string[]).
;(linkedBuilding as BuildingNode).children = [levelId as BuildingNode['children'][number]]
// Collect level children (ids of walls/zones we create below).
const levelChildren: string[] = []
const nodes: Record<AnyNodeId, AnyNodeT> = {}
// Validate + add site, building, level in that order.
const siteValidated = AnyNode.safeParse(site)
if (!siteValidated.success) {
warnings.push(`site node failed schema validation: ${siteValidated.error.message}`)
}
nodes[siteId] = (siteValidated.success ? siteValidated.data : site) as AnyNodeT
const buildingValidated = AnyNode.safeParse(linkedBuilding)
if (!buildingValidated.success) {
warnings.push(`building node failed schema validation: ${buildingValidated.error.message}`)
}
nodes[buildingId] = (
buildingValidated.success ? buildingValidated.data : linkedBuilding
) as AnyNodeT
// Walls.
let wallsAdded = 0
for (let i = 0; i < vision.walls.length; i++) {
const w = vision.walls[i]!
try {
const wall = WallNode.parse({
start: w.start,
end: w.end,
thickness: w.thickness ?? defaultWallThickness,
height: defaultWallHeight,
})
const linkedWall: AnyNodeT = {
...(wall as AnyNodeT),
parentId: levelId,
}
const validated = AnyNode.safeParse(linkedWall)
if (!validated.success) {
warnings.push(`wall[${i}] dropped: ${validated.error.message}`)
continue
}
nodes[wall.id as AnyNodeId] = validated.data as AnyNodeT
levelChildren.push(wall.id)
wallsAdded++
} catch (err) {
const msg = err instanceof Error ? err.message : String(err)
warnings.push(`wall[${i}] dropped: ${msg}`)
}
}
// Rooms → zones.
let roomsAdded = 0
for (let i = 0; i < vision.rooms.length; i++) {
const r = vision.rooms[i]!
try {
const zone = ZoneNode.parse({
name: r.name,
polygon: r.polygon,
})
const linkedZone: AnyNodeT = {
...(zone as AnyNodeT),
parentId: levelId,
}
const validated = AnyNode.safeParse(linkedZone)
if (!validated.success) {
warnings.push(`room[${i}] dropped: ${validated.error.message}`)
continue
}
nodes[zone.id as AnyNodeId] = validated.data as AnyNodeT
levelChildren.push(zone.id)
roomsAdded++
} catch (err) {
const msg = err instanceof Error ? err.message : String(err)
warnings.push(`room[${i}] dropped: ${msg}`)
}
}
// Finalise the level's children array now that walls/zones are in the dict.
;(linkedLevel as LevelNode).children = levelChildren as LevelNode['children']
const levelValidated = AnyNode.safeParse(linkedLevel)
if (!levelValidated.success) {
warnings.push(`level node failed schema validation: ${levelValidated.error.message}`)
}
nodes[levelId] = (levelValidated.success ? levelValidated.data : linkedLevel) as AnyNodeT
return {
nodes,
rootNodeIds: [siteId],
walls: wallsAdded,
rooms: roomsAdded,
warnings,
levelId,
}
}
export function registerPhotoToScene(
server: McpServer,
bridge: SceneBridge,
store: SceneStore,
): void {
server.registerTool(
'photo_to_scene',
{
title: 'Photo to Pascal scene',
description:
'Orchestrator: analyse a floor-plan photo via MCP sampling, translate the structured vision result into a Pascal SceneGraph (site → building → level with walls and zones), optionally save it, and swap the bridge to the new scene. Requires host support for sampling.',
inputSchema: photoToSceneInput,
outputSchema: photoToSceneOutput,
},
async ({ image, scaleHint, name, save, defaultWallThickness, defaultWallHeight }) => {
// 1. Vision.
const vision = await callVisionSampling(server, image, scaleHint)
// 2. Build scene graph.
const built = buildSceneGraphFromVision(vision, defaultWallThickness, defaultWallHeight)
const graph: SceneGraph = {
nodes: built.nodes as SceneGraph['nodes'],
rootNodeIds: built.rootNodeIds as SceneGraph['rootNodeIds'],
collections: {} as SceneGraph['collections'],
}
// 5. Swap the bridge to the new scene so follow-up MCP calls operate on it.
bridge.setScene(graph.nodes, graph.rootNodeIds)
const notes = built.warnings.length > 0 ? built.warnings.join('; ') : undefined
// 4. Save or return inline.
if (save) {
const meta = await store.save({
name,
graph,
})
const payload: {
sceneId: string
url: string
walls: number
rooms: number
confidence: number
notes?: string
} = {
sceneId: meta.id,
url: `/scene/${meta.id}`,
walls: built.walls,
rooms: built.rooms,
confidence: vision.confidence,
}
if (notes) payload.notes = notes
return {
content: [{ type: 'text' as const, text: JSON.stringify(payload) }],
structuredContent: payload,
}
}
const payload: {
walls: number
rooms: number
confidence: number
notes?: string
graph: SceneGraph
} = {
walls: built.walls,
rooms: built.rooms,
confidence: vision.confidence,
graph,
}
if (notes) payload.notes = notes
return {
content: [{ type: 'text' as const, text: JSON.stringify(payload) }],
structuredContent: payload,
}
},
)
}