@volter/editor-live 0.5.57

package/src/editor.ts

@@ -0,0 +1,663 @@
+/** Editor and document automation over the shared session command wire.
+ * Modeling executes in the editor tab; this client owns no model state. */
+
+import type {
+  ActiveDocumentCapture,
+  AssetKind,
+  AssetPreviewCapture,
+  AssetPreviewOptions,
+  AssetPreviewShotSetDefinition,
+  AssetPreviewSource,
+  CaptureDimensions,
+  DocumentLookOutcome,
+  EditorChromeCapture,
+  EditorChromeCaptureOptions,
+  EditorClient,
+  EditorState,
+  EditorView,
+  EditorWorkspaceName,
+  HistoryStep,
+  InspectedFieldWrite,
+  InspectedHierarchy,
+  InspectedInspection,
+  LabeledShotSetCapture,
+  OpenedDocument,
+  PresentedEditorView,
+  ShadingMode,
+  StructureOp,
+  StructureOpOptions,
+  StructureOpResult,
+  ViewPreset,
+  ViewportCapture,
+} from '@volter/editor-sdk';
+import { LiveEditorDocument } from './editor-document.js';
+
+/** A panel `showPanel` can focus: the viewport tabs, the console, the build
+ *  surface, or any key the editor's own static-panel registry holds
+ *  (`hierarchy`, `assets`, `asset-library`, `inspector`, `history`, …) — which
+ *  is why this is open: the registry, not this union, is the vocabulary. */
+export type PanelName =
+  | 'viewport-edit'
+  | 'console'
+  // Keeps the literals above in autocomplete while admitting every key the
+  // editor's registry holds — the registry answers, this union only hints.
+  | (string & {});
+
+const EXTENSION_KIND: Record<string, AssetKind> = {
+  '.glb': 'model',
+  '.gltf': 'model',
+  '.png': 'image',
+  '.jpg': 'image',
+  '.jpeg': 'image',
+  '.webp': 'image',
+  '.gif': 'image',
+  '.svg': 'image',
+  '.hdr': 'image',
+  '.exr': 'image',
+  '.mp4': 'video',
+  '.webm': 'video',
+  '.mp3': 'audio',
+  '.ogg': 'audio',
+  '.wav': 'audio',
+  '.flac': 'audio',
+  '.glsl': 'source',
+  '.vert': 'source',
+  '.frag': 'source',
+  // PROJECT SCRIPTS ARE SOURCE. Without these the guess below falls through to
+  // `'json'`, the asset-document router sends the file to the generic JSON
+  // viewer (`asset-documents.tsx#assetDocumentViewerRoute`: `spec.kind ===
+  // 'json'` is decided before any content routing), and the LIVE MODELING
+  // DOCUMENT never mounts — `editor.openAsset('src/lib/fox/fox.model.ts')`
+  // silently shows a text pane instead of the model. Only `kind: 'source'`
+  // reaches `SourceAssetViewer`, which is what content-routes a project script
+  // to `LiveModuleDocument`. The set matches that viewer's own
+  // `isProjectScriptPath` regex, `/\.(?:[cm]?[jt]sx?)$/`.
+  '.ts': 'source',
+  '.tsx': 'source',
+  '.mts': 'source',
+  '.cts': 'source',
+  '.js': 'source',
+  '.jsx': 'source',
+  '.mjs': 'source',
+  '.cjs': 'source',
+};
+
+/**
+ * Lightweight extension-based `AssetKind` guess for `openAsset`'s optional
+ * `kind` argument. Deliberately independent of (not shared with) the
+ * editor's own `AssetBrowser.tsx#getAssetKind` — that function is a private,
+ * React-component-local helper of a package `@volter/editor-live` has no dependency
+ * on. Callers with an unusual extension can always pass `kind` explicitly.
+ */
+export function inferAssetKind(path: string): AssetKind {
+  // `.prefab.json` is retired (PRs #578/#581/#589) but the `'prefab'`
+  // AssetKind itself remains in `@volter/editor-sdk` for view-link compatibility.
+  if (path.endsWith('.prefab.json')) return 'prefab';
+  const dot = path.lastIndexOf('.');
+  const ext = dot >= 0 ? path.slice(dot).toLowerCase() : '';
+  return EXTENSION_KIND[ext] ?? 'json';
+}
+
+export class LiveEditor {
+  /** `#`-private, not `private`: `volter-editor eval --list` enumerates this object's
+   *  real runtime members, and TypeScript's erased `private` would leave the
+   *  raw `EditorClient` advertised beside them. */
+  readonly #client: EditorClient;
+
+  /**
+   * The ACTIVE center document's own DOM: read it, click it, key it, paste
+   * into it. The one door onto editor chrome that is not play-mode gated, and
+   * deliberately scoped to that document alone —
+   * `packages/editor/src/editor-document-probe.ts` carries the design and the
+   * refusal contract. Screenshotting the same subject is
+   * {@link LiveEditor.captureActiveDocument}, not a fifth verb here.
+   */
+  readonly document: LiveEditorDocument;
+
+  constructor(client: EditorClient) {
+    this.#client = client;
+    this.document = new LiveEditorDocument(client);
+  }
+
+  /**
+   * THE BLENDER LANE'S VERBS, from `volter-editor eval`.
+   *
+   * Blender runs headless in the editor tab's worker (ARCHITECTURE-CORE, "THE
+   * BLENDER IN THE TAB IS BLENDER") and answers `blender-start`,
+   * `blender-execute`, `blender-scene-info`, `blender-object-info`,
+   * `blender-screenshot-view`, `blender-read-file`, `blender-write-file`,
+   * `blender-list-files`, `blender-stop` and `blender-status`. They were
+   * reachable from `@volter/editor-sdk` and through `vgai blender-mcp` but from
+   * no GENERAL door, so driving a session meant writing an MCP client script
+   * per question — the same discovery failure `eval-surface.ts`'s header
+   * records, in a lane that had not noticed it yet.
+   *
+   *   volter-editor eval "await editor.blender('blender-execute', { code: 'import bpy; print(len(bpy.data.objects))' })"
+   *
+   * `blender-status` is the only verb that creates nothing: it answers whether
+   * this tab already has a session without starting one.
+   */
+  async blender<T extends object = Record<string, unknown>>(
+    type: `blender-${string}`,
+    fields: Record<string, unknown> = {},
+  ): Promise<T> {
+    return this.#client.blender<T>(type, fields);
+  }
+
+  /**
+   * The active authoring adapter's persistence destination — where a save would
+   * land (`status().savePath`). A read only: a three root has no scene document
+   * to open, and its root is activated instead.
+   */
+  async scene(): Promise<string | null> {
+    const state = await this.#client.getState();
+    return state.savePath;
+  }
+
+  /**
+   * Make the connected human editor show the same subject/view as the agent.
+   * The returned URL is a compact, shareable projection — not a serialized
+   * workspace or document payload.
+   */
+  async present(view: EditorView): Promise<PresentedEditorView> {
+    return this.#client.present(view);
+  }
+
+  /** The human editor's actual active document, selection, camera and utility. */
+  async currentView(): Promise<EditorView> {
+    return this.#client.currentView();
+  }
+
+  /**
+   * Capture the same center document the human is currently looking at.
+   *
+   * A number is a SQUARE of that size — the default, and the right shape for
+   * an unstaged look at a model. `{width, height}` asks for a shaped frame, so
+   * a video-aspect look needs no crop afterwards. Both are bounded by the
+   * relay budget (64-1024 per side, total no larger than a 1024 square); see
+   * `@volter/editor-sdk`'s `CaptureDimensions`.
+   * Supply a view to present and photograph it in one editor request.
+   */
+  async captureActiveDocument(
+    size?: CaptureDimensions,
+    view?: EditorView,
+  ): Promise<ActiveDocumentCapture> {
+    return this.#client.captureActiveDocument(size, view);
+  }
+
+  /**
+   * Photograph the editor PAGE — every panel, tab strip and viewport as the
+   * person sees it. `vgai screenshot editor` is this verb from the shell. The
+   * one door for judging chrome sighted: a skin, a workspace arrangement or a
+   * contributed panel is looked at through this, never guessed at from DOM
+   * probes. The page at its own layout, `scale` output pixels per CSS pixel
+   * (default `devicePixelRatio`) — a 1 px border or a glyph stroke is only
+   * judgeable at the scale the reference it is compared against was captured
+   * at, and the result reports its own `size` and `scale`.
+   */
+  async captureEditorChrome(options?: EditorChromeCaptureOptions): Promise<EditorChromeCapture> {
+    return this.#client.captureEditorChrome(options);
+  }
+
+  /** `'all'` -> `EditorClient.selectAll()` (mirrors `vgai select --all`); otherwise `EditorClient.select(id)` (mirrors `vgai select <entityId>`). */
+  async select(id: string | 'all'): Promise<void> {
+    if (id === 'all') {
+      await this.#client.selectAll();
+      return;
+    }
+    await this.#client.select(id);
+  }
+
+  /** Mirrors `vgai deselect`. */
+  async deselect(): Promise<void> {
+    await this.#client.select(null);
+  }
+
+  /** No `id` -> focus the current selection (mirrors bare `vgai focus`); `id` given -> focus that entity. */
+  async focus(id?: string): Promise<void> {
+    if (id !== undefined) {
+      await this.#client.focusEntity(id);
+      return;
+    }
+    await this.#client.focusSelection();
+  }
+
+  /**
+   * Frame the edit viewport camera on one entity. Same framing as
+   * `focus(id)`, but an entity id the scene does not know THROWS, naming the
+   * id — where `focus` quietly does nothing. Reach for this whenever the next
+   * step reads the viewport (`screenshot()`, an
+   * `assetPreview(..., { stage: 'scene' })`): a framing that silently missed
+   * would otherwise be indistinguishable from one that worked.
+   */
+  async frame(entityId: string): Promise<void>;
+  /**
+   * Bare `frame()` frames the OPEN Object3D document's subject instead — its
+   * selection if it has one, else the whole model: the toolbar's own Frame
+   * button, reachable from a script. `fit` scales the fitted distance (1 is
+   * that button's tight fit, 1.5 stands back a little for a shot).
+   */
+  async frame(options?: { readonly fit?: number }): Promise<void>;
+  async frame(target?: string | { readonly fit?: number }): Promise<void> {
+    if (typeof target === 'string') {
+      await this.#client.frameEntity(target);
+      return;
+    }
+    await this.#client.frameDocument(target?.fit);
+  }
+
+  /**
+   * WATCH THE AGENT LOOK AROUND THE MODEL.
+   *
+   * Swings the open Object3D document's camera — the camera the human's tab is
+   * showing — around the framed subject by `azimuth`/`elevation` RADIANS,
+   * animated over `duration` seconds (default 0.6), and resolves when the move
+   * ends. This is deliberately not a jump cut: the point of the verb is that a
+   * person watching sees the agent walk around the thing it is working on.
+   *
+   * `await editor.orbit({ azimuth: Math.PI / 2 })` — a quarter turn to the right.
+   *
+   * There is ONE camera, and the human owns it: a drag during the move cancels
+   * it exactly where it is, and the resolved outcome says `cancelledBy:
+   * 'human'` rather than throwing. A second look verb supersedes the first.
+   * The move is drawn by the document's own frame loop, so a document that
+   * isn't being drawn (background tab, inactive panel) doesn't orbit.
+   */
+  async orbit(options: {
+    /** RADIANS, relative to where the camera is now. `Math.PI / 2` is a
+     *  quarter turn; degrees are not accepted and `90` is fourteen turns. */
+    readonly azimuth?: number;
+    /** RADIANS, relative to where the camera is now. */
+    readonly elevation?: number;
+    /** SECONDS the move takes (default 0.6). */
+    readonly duration?: number;
+  }): Promise<DocumentLookOutcome> {
+    // AN UNKNOWN KEY IS REFUSED BY NAME. Every member here is optional, so a
+    // misspelling — `yaw` for `azimuth`, `pitch` for `elevation` — used to
+    // orbit by nothing at all while the outcome still reported a plausible
+    // ABSOLUTE azimuth, which reads exactly like a move that happened
+    // (measured 2026-09-21, and it cost a round). The keys and their unit are
+    // in the refusal because that is the moment the caller needs them.
+    const known = ['azimuth', 'elevation', 'duration'];
+    const unknown = Object.keys(options ?? {}).filter((key) => !known.includes(key));
+    if (unknown.length > 0)
+      throw new Error(
+        `editor.orbit: ${unknown.join(', ')} ${unknown.length === 1 ? 'is not a key' : 'are not keys'} ` +
+          'this verb takes. It takes azimuth and elevation in RADIANS (relative to where the ' +
+          'camera is now) and duration in SECONDS.',
+      );
+    return this.#client.orbitDocument(options);
+  }
+
+  /**
+   * A slow full revolution of the open document's subject — {@link orbit} with
+   * the turns spelled out and a constant angular rate. Resolves at the end of
+   * the last revolution.
+   */
+  async turntable(options?: {
+    readonly seconds?: number;
+    readonly revolutions?: number;
+  }): Promise<DocumentLookOutcome> {
+    return this.#client.turntableDocument(options);
+  }
+
+  async view(preset: ViewPreset): Promise<void> {
+    await this.#client.viewPreset(preset);
+  }
+
+  /**
+   * Switch the editor's NAMED WORKSPACE — `await editor.workspace('model')`.
+   *
+   * A workspace is a task-named LAYOUT MEMORY over the one dock
+   * (ARCHITECTURE-CORE §Editor chrome): `game` (the default, the editor's
+   * standing arrangement), `model`, `sculpt`, `texture`, `animate`, `look`.
+   * Switching is an EXPLICIT act — nothing in the editor moves chrome on its
+   * own, opening a document included — and this is the session door to it,
+   * beside `Window → Workspace` and the registered actions.
+   *
+   * Resolves once the dock has finished rebuilding, so a capture taken
+   * immediately after photographs the arrangement that was asked for. Each
+   * workspace remembers the user's own hand-tuning per project, so switching
+   * away and back is lossless.
+   */
+  async workspace(id: EditorWorkspaceName): Promise<void> {
+    await this.#client.setWorkspace(id);
+  }
+
+  /**
+   * Apply a STYLE BUNDLE by id — the chrome's palette, material, icon set and
+   * region defaults in one gesture, the session door beside
+   * `View → <Style> Style`. A bundle the open project does not offer refuses
+   * and names the vocabulary; `currentView().style` reports the one worn.
+   */
+  async style(id: string): Promise<void> {
+    await this.#client.setStyle(id);
+  }
+
+  /**
+   * Set the MATERIAL apart from the bundle that usually carries it.
+   * Appearance is palette × material, independent axes by ruling, so
+   * `style()` alone can never say whether a cost belongs to the blur or to
+   * the palette. This is the door that measures them apart; it answers with
+   * what the chrome wears afterwards (`style` is `null` when the mix matches
+   * no registered bundle).
+   */
+  async appearance(appearance: {
+    readonly material?: string;
+  }): Promise<{ material: string; style: string | null }> {
+    return this.#client.setAppearance(appearance);
+  }
+
+  /** Focus an editor panel: a viewport tab, the console, the build surface, or
+   *  any key the editor's static-panel registry holds — an unknown key refuses
+   *  naming the ones it does. */
+  async showPanel(name: PanelName): Promise<void> {
+    switch (name) {
+      case 'viewport-edit':
+        await this.#client.showViewport('edit');
+        return;
+      case 'console':
+        await this.#client.toggleConsole();
+        return;
+      default:
+        // Every other name is the EDITOR's to resolve, against its live panel
+        // registry. A list kept here could only ever be a copy going stale.
+        await this.#client.showPanel(name);
+        return;
+    }
+  }
+
+  /** `kind` inferred from `path`'s extension when omitted (`inferAssetKind`) — pass it explicitly to override. */
+  async openAsset(path: string, kind?: AssetKind): Promise<void> {
+    await this.#client.openAsset(path, kind ?? inferAssetKind(path));
+  }
+
+  /**
+   * SELECT a project asset — the browser's single click, which fills the
+   * Inspector without opening a document. `openAsset` is the double click.
+   *
+   * This is how a project's own `asset.inspector` section is reached: select
+   * the file it matches, then `inspect()` lists the verbs that section
+   * declares and `runAction(id)` runs one. Selecting a path nothing matches
+   * is not an error — the Inspector shows what it has, exactly as it does
+   * for a human.
+   */
+  async selectAsset(path: string): Promise<void> {
+    await this.#client.selectAsset(path);
+  }
+
+  /**
+   * Captures the editor's native four-view preview. A bare string is a
+   * project-relative asset path (the common case); an explicit source object
+   * targets a path, a LIVE SCENE ENTITY (`assetPreview({ entityId }, …)`) —
+   * which is what makes `options.stage: 'scene'`, the entity photographed
+   * where it stands under the scene's own lighting, reachable from here — or
+   * RAW GLB BYTES (`assetPreview({ glbBase64 }, …)`), for a model that exists
+   * only in the calling Node process's memory and has never been written to
+   * disk. `stage` defaults to `'lab'`, the neutral Asset Lab staging this has
+   * always produced, and the bytes form is lab-only.
+   */
+  async assetPreview(
+    source: string | AssetPreviewSource,
+    options?: AssetPreviewOptions,
+  ): Promise<AssetPreviewCapture> {
+    return this.#client.captureAssetPreview(
+      typeof source === 'string' ? { assetPath: source } : source,
+      options,
+    );
+  }
+
+  /**
+   * The same subject photographed as a LABELED SHOT SET instead of the four
+   * views — a caller-supplied definition of turntable yaws and bone-anchored
+   * crops, rendered against the asset's own skeleton, with a contact sheet.
+   * Every source {@link assetPreview} takes works here, GLB bytes included:
+   * a shot set stages its own subject, so it needs no place to stand.
+   *
+   * Sole in-repo caller today: `project.bake.preview`'s `--orbit` lane.
+   */
+  async assetPreviewShots(
+    source: string | AssetPreviewSource,
+    definition: AssetPreviewShotSetDefinition,
+    options?: AssetPreviewOptions,
+  ): Promise<LabeledShotSetCapture> {
+    return this.#client.captureShotSetPreview(
+      typeof source === 'string' ? { assetPath: source } : source,
+      definition,
+      options,
+    );
+  }
+
+  async grid(on: boolean): Promise<void> {
+    await this.#client.setGrid(on);
+  }
+
+  async helpers(on: boolean): Promise<void> {
+    await this.#client.setHelpers(on);
+  }
+
+  async stats(on: boolean): Promise<void> {
+    await this.#client.setStats(on);
+  }
+
+  async shading(mode: ShadingMode): Promise<void> {
+    await this.#client.setShadingMode(mode);
+  }
+
+  /**
+   * READ the inspector, as data — the serialized inspection subject
+   * (`editor.inspect()`; design: `docs/ARCHITECTURE-CORE.md` §Editor chrome,
+   * "The Inspection Model"). This is the Figma-Inspect analog: whatever a
+   * human would see in the inspector right now — the subject's identity, its
+   * verbs, and every identified section in display order, with a `fields`
+   * section's CURRENT VALUES at their scriptable `path`s.
+   *
+   * Reach for it whenever the next step depends on what an object actually
+   * IS: `await editor.select(id)` then `await editor.inspect()` answers "what
+   * properties does this thing have, and what are they set to" in one call,
+   * against the same model the panel renders — no scene-graph reads, no
+   * guessing at property names.
+   *
+   * When the inspector is showing NOTHING — nothing selected on a surface
+   * with no empty-state subject of its own, which is most of them — the answer
+   * is `{none: true}`, so "the human sees no inspector" and "the read failed"
+   * are never the same value. A surface whose empty space IS a real thing (an
+   * open Asset Lab document) still answers with that subject, and never with
+   * another surface's.
+   *
+   * A `custom` section body is a named opaque: the editor renders it with
+   * React, so the wire reports its identity rather than pretending to describe
+   * its rendering — plus, when the section can say what it DISPLAYS, a `data`
+   * payload in its own vocabulary (`transform` carries
+   * `{position, rotation, scale}`, rotation in Euler XYZ degrees).
+   */
+  async inspect(): Promise<InspectedInspection> {
+    return this.#client.inspect();
+  }
+
+  /** Run one verb listed by `inspect().quickActions`, through the same action
+   * the human Inspector button invokes. */
+  async runAction(actionId: string): Promise<InspectedInspection> {
+    return this.#client.runInspectionAction(actionId);
+  }
+
+  /**
+   * Run ONE command by id — the door to everything the command palette lists.
+   *
+   * ONE NAME (orchestrator ruling 2026-09-19). There were briefly TWO doors
+   * onto the one view-verb table — this one and `editor.viewVerb(view, verb)`,
+   * which addressed the same registry by its two halves. A second addressing
+   * of one table is a second name for one thing, and an agent reading
+   * `--list` had to choose between them with nothing to choose on. This door
+   * stays because it is strictly wider: it addresses a COMMAND ID, so under
+   * the frame it reaches everything the workbench knows — a `vgai.action.<id>`
+   * editor action, one of VS Code's own — and not only a view. A VIEW is
+   * reached by spelling its verb's command id:
+   *
+   *     await editor.command('vgai.blender-uv-view.state')
+   *     await editor.command('vgai.blender-uv-view.zoom', { to: 600 })
+   *
+   *     await editor.command('vgai.blender-node-view.view-all')
+   *     await editor.command('vgai.blender-node-view.look', { node: 'Principled BSDF' })
+   *
+   * Under the Code-OSS frame this is the workbench's own command service, so
+   * any command id works — ours and VS Code's alike. Standalone `vgai edit`
+   * has no command service and answers the `vgai.<view>.<verb>` shape off the
+   * SAME verb table the frame's commands call, refusing any other id by name.
+   * One table, two doors, exactly like the keymap's.
+   *
+   * Answers with whatever the command returned — a view verb's own state, or
+   * `null` for a command that returns nothing.
+   */
+  async command(commandId: string, args?: unknown): Promise<unknown> {
+    return this.#client.runCommand(commandId, args);
+  }
+
+  /**
+   * RESTRUCTURE the authored tree — the hierarchy context menu's own verbs.
+   *
+   * `create`, `delete`, `duplicate`, `reparent`, `reorder`, `wrap`, `unwrap`,
+   * `group`, `ungroup`, `copy`, `cut`, `paste`; `extractComponent` and
+   * `forkComponent` are the two that write whole new files and have their own
+   * doors below. All of them run the SAME `authoring/consumer-actions.ts`
+   * helpers the menu items call, so there is one implementation of each op and
+   * not a second that can disagree with what a human gets.
+   *
+   * It exists because the menu is a POINTER surface: every one of these ops was
+   * reachable only by right-clicking a hierarchy row, which is nothing an agent
+   * can do — so for an ingest root, whose only authoring surface IS the editor,
+   * structure was closed entirely.
+   *
+   * `id`/`ids` default to the current selection. The answer carries the same
+   * per-edit `write` ack `setField` does, so `write.persisted` tells a saved
+   * restructure from a live-only one. An op the active adapter does not provide
+   * REJECTS by name — never a silent no-op.
+   */
+  async structure(op: StructureOp, options?: StructureOpOptions): Promise<StructureOpResult> {
+    return this.#client.structureOp(op, options ?? {});
+  }
+
+  /**
+   * "Extract Component…" — lift the selected native subtree into its own
+   * component file (plus a story) and replace the callsite with it.
+   *
+   * Answers the action's own sentence, which NAMES both new files, because
+   * undo owns the callsite edit and will not remove them.
+   */
+  async extractComponent(options?: { id?: string; name?: string }): Promise<string> {
+    return (await this.#client.extractComponent(options ?? {})).hint;
+  }
+
+  /**
+   * "Fork Component…" — copy the selected instance's component definition to a
+   * new file and retarget THIS CALLSITE at it.
+   *
+   * One callsite is the unit of the edit; when that callsite sits inside a
+   * component rendered many times, every one of those renders now renders the
+   * fork.
+   */
+  async forkComponent(options?: { id?: string }): Promise<string> {
+    return (await this.#client.forkComponent(options ?? {})).hint;
+  }
+
+  /**
+   * READ the hierarchy panel, as data — the rows a human is looking at right
+   * now, nested exactly as the panel nests them.
+   *
+   * The companion to {@link inspect}: that one answers "what IS the selected
+   * thing", this one answers "what does the tree LOOK LIKE". It is the panel's
+   * own output, not a fresh walk of the scene — the adapter's tree after the
+   * component marks fold implementation subtrees (bones, particle renderers,
+   * instanced pools), after the internals reveal, the document promotion, the
+   * child cap, the collapse state, the search filter and the selection scope.
+   *
+   * Works in play mode and edit mode; the answer says which (`playState`,
+   * `activeViewportTab`), because the two are different adapters and a tree
+   * that looks wrong is very often the wrong adapter's tree.
+   *
+   * Prefer this over `status().entities`, which is deliberately a different
+   * question — the RAW adapter tree, unprojected. A panel that renders the
+   * wrong rows looks perfectly healthy in that facet.
+   *
+   * Each row carries `childCount` (what its caret opens), `internalChildCount`
+   * (what is folded behind "Reveal Internals") and `expandable` (whether the
+   * panel draws a caret at all), so "this subtree exists but nothing in the UI
+   * opens it" is a fact you can read rather than one you have to notice.
+   *
+   * Rejects, naming the panel, when no hierarchy panel is mounted — an empty
+   * tree would be a fabricated answer about a surface nobody is being shown.
+   */
+  async hierarchy(): Promise<InspectedHierarchy> {
+    return this.#client.hierarchy();
+  }
+
+  /** Expand every branch through the Hierarchy panel's own action. */
+  async expandHierarchyAll(): Promise<void> {
+    await this.#client.expandHierarchyAll();
+  }
+
+  /** Collapse every branch through the same panel action. Expanding is
+   *  persisted per project, so without this the tree's REST STATE — what a
+   *  person sees on opening the project — is unreachable once any reader has
+   *  expanded it. */
+  async collapseHierarchyAll(): Promise<void> {
+    await this.#client.collapseHierarchyAll();
+  }
+
+  /**
+   * Write one editable field from `inspect()` by its stable path, through the
+   * same Inspector IO and persistence boundary the human control uses.
+   *
+   * The answer is `{ subject, write }`, and `write` is the half worth reading
+   * first: a write with no persistence route open still succeeds — it lands on
+   * the live object and journals live-only — so `write.persisted` is how you
+   * tell a saved edit from one that will not survive the session, without
+   * diffing the tree. `write.destination` is the adapter's own words for where
+   * it went ("live-only (not saved)" is a destination, never silence).
+   */
+  async setField(path: string, value: unknown): Promise<InspectedFieldWrite> {
+    return this.#client.setInspectionField(path, value);
+  }
+
+  /**
+   * REMOVE one field's authored override — the revert arrow, as a command.
+   *
+   * Reach for this instead of `setField` whenever you are UNDOING an edit that
+   * added a property the source did not carry: `setField` can only write a
+   * value, so setting the default back leaves `position={[0, 0, 0]}` in the
+   * file where there was nothing before. Only this door restores the bytes.
+   *
+   * The answer is the same `{ subject, write }` shape, awaited past the bytes.
+   * It rejects with `code: 'REMOVAL_UNAVAILABLE'` when the field is not
+   * declared removable or the lane has no removal door — which is a missing
+   * seam to report, not a removal that failed.
+   */
+  async removeField(path: string): Promise<InspectedFieldWrite> {
+    return this.#client.removeInspectionField(path);
+  }
+
+  /** Open a document by its adapter-declared id, through its registered owner. */
+  async open(id: string): Promise<OpenedDocument> {
+    return this.#client.open(id);
+  }
+
+  /** Undo / redo one project transaction, through the session's own history
+   *  queue — the same one the keyboard shortcut drives. */
+  async undo(): Promise<HistoryStep> {
+    return this.#client.undo();
+  }
+
+  async redo(): Promise<HistoryStep> {
+    return this.#client.redo();
+  }
+
+  /** Mirrors `vgai status` — the full live editor state as JSON. */
+  async status(): Promise<EditorState> {
+    return this.#client.getState();
+  }
+
+  /** A live viewport PNG (`EditorClient.captureViewport`) — no direct CLI verb exists; this is the closest wire read. */
+  async screenshot(size?: number): Promise<ViewportCapture> {
+    return this.#client.captureViewport(size);
+  }
+}