fix(#342 review round-2 F5-F6): drop the posthog re-render remount + test chunk detector

- F5 [stability/regression]: the round-1 F2 fix re-rendered the root with
  <PostHogProvider><App/></PostHogProvider> after the analytics chunk loaded. In
  the ChunkLoadErrorBoundary child slot the element TYPE changes App ->
  PostHogProvider, so React does NOT reconcile in place — it REMOUNTS the whole
  App: every mount effect runs twice (websocket connect/disconnect, origin
  tracking, subscriptions) and local state / focus / scroll / in-progress input is
  lost on cloud cold-load (e.g. typing in /login before analytics loads). And it
  was USELESS: the app has ZERO consumers of the PostHog React context (no
  usePostHog / useFeatureFlag* / PostHogFeature), and PostHogProvider given an
  initialized client is a no-op — all capture goes through the posthog singleton.
  Fix: initAnalytics now inits the posthog SINGLETON only (no posthog-js/react
  import, no second render); renderApp() renders <App/> once. First paint stays
  instant, cloud analytics behavior unchanged, no remount.
- F6 [test]: exported isChunkLoadError + chunk-load-error-boundary.test.ts —
  pins the detector (ChunkLoadError name + the 3 dynamic-import failure messages,
  case-insensitive → true; null/undefined/ordinary errors → false) so a
  false-negative that re-blanks the app on a real chunk-404 is caught.

Gate: client tsc 0, chunk-load + sanitize tests 14 passed. Entry chunk unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.env.example

@@ -202,6 +202,13 @@ MCP_DOCMOST_PASSWORD=
 # Default 900000 (15 min).
 # AI_MCP_CALL_TIMEOUT_MS=900000
 
+# Deferred tool loading for the in-app AI chat (#332). Default ON: the agent sees
+# a compact <tool_catalog> and only CORE tools + a loadTools meta-tool are active
+# each step; deferred tools (the fat/rare ones + all external MCP tools) load on
+# demand. Set AI_CHAT_DEFERRED_TOOLS=false to restore the old "all tools always
+# active" behavior.
+# AI_CHAT_DEFERRED_TOOLS=true
+
 # --- Anonymous public-share AI assistant ---
 # Opt-in per workspace (AI settings -> "public share assistant"; off by default).
 # When enabled, anonymous visitors of a published share can ask an AI about that

apps/client/package.json

@@ -13,6 +13,7 @@
   },
   "dependencies": {
     "@ai-sdk/react": "^3.0.208",
+    "@braintree/sanitize-url": "7.1.2",
     "@atlaskit/pragmatic-drag-and-drop": "1.8.1",
     "@atlaskit/pragmatic-drag-and-drop-auto-scroll": "2.1.5",
     "@atlaskit/pragmatic-drag-and-drop-flourish": "2.0.15",

apps/client/src/App.tsx

@@ -1,38 +1,72 @@
+import { lazy, Suspense } from "react";
 import { Navigate, Route, Routes } from "react-router-dom";
+import { Center, Loader } from "@mantine/core";
+import { Error404 } from "@/components/ui/error-404.tsx";
+import Layout from "@/components/layouts/global/layout.tsx";
+import { useTrackOrigin } from "@/hooks/use-track-origin";
+
+// ShareLayout is route-split: its ShareShell chrome pulls in the table of
+// contents (and thus TipTap), so keeping it out of the eager graph removes the
+// editor engine from startup for authenticated users too.
+const ShareLayout = lazy(
+  () => import("@/features/share/components/share-layout.tsx"),
+);
+
+// Auth / entry pages stay eager: they are the first paint for an unauthenticated
+// visitor (e.g. /login) and are already small, so code-splitting them would only
+// add a cold-chunk round trip to the most common cold-start path.
 import SetupWorkspace from "@/pages/auth/setup-workspace.tsx";
 import LoginPage from "@/pages/auth/login";
-import Home from "@/pages/dashboard/home";
-import Page from "@/pages/page/page";
-import AccountSettings from "@/pages/settings/account/account-settings";
-import WorkspaceMembers from "@/pages/settings/workspace/workspace-members";
-import WorkspaceSettings from "@/pages/settings/workspace/workspace-settings";
-import AiSettings from "@/pages/settings/workspace/ai-settings";
-import Groups from "@/pages/settings/group/groups";
-import GroupInfo from "./pages/settings/group/group-info";
-import Spaces from "@/pages/settings/space/spaces.tsx";
-import { Error404 } from "@/components/ui/error-404.tsx";
-import AccountPreferences from "@/pages/settings/account/account-preferences.tsx";
-import SpaceHome from "@/pages/space/space-home.tsx";
-import PageRedirect from "@/pages/page/page-redirect.tsx";
-import Layout from "@/components/layouts/global/layout.tsx";
 import InviteSignup from "@/pages/auth/invite-signup.tsx";
 import ForgotPassword from "@/pages/auth/forgot-password.tsx";
 import PasswordReset from "./pages/auth/password-reset";
-import SharedPage from "@/pages/share/shared-page.tsx";
+import PageRedirect from "@/pages/page/page-redirect.tsx";
-import Shares from "@/pages/settings/shares/shares.tsx";
-import ShareLayout from "@/features/share/components/share-layout.tsx";
 import ShareRedirect from "@/pages/share/share-redirect.tsx";
-import { useTrackOrigin } from "@/hooks/use-track-origin";
+
-import SpacesPage from "@/pages/spaces/spaces.tsx";
+// Heavy / leaf pages are route-split with React.lazy so their code (most
-import SpaceTrash from "@/pages/space/space-trash.tsx";
+// importantly the whole TipTap editor + KaTeX + lowlight grammars + drawio that
-import FavoritesPage from "@/pages/favorites/favorites-page";
+// the page editor and the readonly share editor pull in) is fetched only when
-import LabelPage from "@/pages/label/label-page";
+// the matching route is actually visited. The <Suspense> boundaries live inside
+// each Layout (around its <Outlet/>), so the app shell stays mounted while a
+// route chunk loads.
+const Home = lazy(() => import("@/pages/dashboard/home"));
+const Page = lazy(() => import("@/pages/page/page"));
+const SpaceHome = lazy(() => import("@/pages/space/space-home.tsx"));
+const SpaceTrash = lazy(() => import("@/pages/space/space-trash.tsx"));
+const SpacesPage = lazy(() => import("@/pages/spaces/spaces.tsx"));
+const FavoritesPage = lazy(() => import("@/pages/favorites/favorites-page"));
+const LabelPage = lazy(() => import("@/pages/label/label-page"));
+const SharedPage = lazy(() => import("@/pages/share/shared-page.tsx"));
+
+const AccountSettings = lazy(
+  () => import("@/pages/settings/account/account-settings"),
+);
+const AccountPreferences = lazy(
+  () => import("@/pages/settings/account/account-preferences.tsx"),
+);
+const WorkspaceSettings = lazy(
+  () => import("@/pages/settings/workspace/workspace-settings"),
+);
+const AiSettings = lazy(() => import("@/pages/settings/workspace/ai-settings"));
+const WorkspaceMembers = lazy(
+  () => import("@/pages/settings/workspace/workspace-members"),
+);
+const Groups = lazy(() => import("@/pages/settings/group/groups"));
+const GroupInfo = lazy(() => import("./pages/settings/group/group-info"));
+const Spaces = lazy(() => import("@/pages/settings/space/spaces.tsx"));
+const Shares = lazy(() => import("@/pages/settings/shares/shares.tsx"));
 
 export default function App() {
   useTrackOrigin();
 
   return (
-    <>
+    <Suspense
+      fallback={
+        <Center h="100vh">
+          <Loader size="sm" />
+        </Center>
+      }
+    >
       <Routes>
         <Route index element={<Navigate to="/home" />} />
         <Route path={"/login"} element={<LoginPage />} />
@@ -83,6 +117,6 @@ export default function App() {
 
         <Route path="*" element={<Error404 />} />
       </Routes>
-    </>
+    </Suspense>
   );
 }

apps/client/src/components/chunk-load-error-boundary.test.ts

@@ -0,0 +1,37 @@
+import { describe, it, expect } from "vitest";
+import { isChunkLoadError } from "./chunk-load-error-boundary";
+
+// The detector decides whether a caught render error is a stale-deploy chunk-404
+// (→ auto-reload to fetch the new manifest) vs a genuine app error (→ generic
+// recovery UI, no reload). A false negative on a real chunk failure re-blanks the
+// app; a false positive would auto-reload on an ordinary error. Pin both sides.
+describe("isChunkLoadError", () => {
+  it("detects the ChunkLoadError name", () => {
+    expect(isChunkLoadError({ name: "ChunkLoadError", message: "x" })).toBe(true);
+  });
+
+  it.each([
+    "Failed to fetch dynamically imported module: https://x/assets/index-abc.js",
+    "error loading dynamically imported module",
+    "Importing a module script failed.",
+  ])("detects the dynamic-import failure message %#", (message) => {
+    expect(isChunkLoadError({ name: "TypeError", message })).toBe(true);
+  });
+
+  it("is case-insensitive on the message", () => {
+    expect(
+      isChunkLoadError({ message: "FAILED TO FETCH DYNAMICALLY IMPORTED MODULE" }),
+    ).toBe(true);
+  });
+
+  it.each([
+    null,
+    undefined,
+    {},
+    { name: "TypeError", message: "Cannot read properties of undefined" },
+    { message: "Network request failed" },
+    new Error("some ordinary render error"),
+  ])("returns false for a non-chunk error %#", (err) => {
+    expect(isChunkLoadError(err)).toBe(false);
+  });
+});

apps/client/src/components/chunk-load-error-boundary.tsx

@@ -0,0 +1,71 @@
+import { ReactNode } from "react";
+import { ErrorBoundary } from "react-error-boundary";
+import { Button, Center, Stack, Text } from "@mantine/core";
+
+const RELOAD_FLAG = "chunk-reload-attempted";
+
+// Heuristic detection of a failed dynamic import. Since the code-splitting work,
+// every route (plus Aside / AiChatWindow) is React.lazy: when a new deploy
+// replaces the hashed chunks, a tab left open on the old index.html requests a
+// chunk URL that now 404s, and React.lazy rejects. Browsers / Vite surface these
+// with a ChunkLoadError name or one of these messages.
+export function isChunkLoadError(error: unknown): boolean {
+  if (!error) return false;
+  const name = (error as { name?: string }).name ?? "";
+  const message = (error as { message?: string }).message ?? "";
+  return (
+    name === "ChunkLoadError" ||
+    /Failed to fetch dynamically imported module/i.test(message) ||
+    /error loading dynamically imported module/i.test(message) ||
+    /Importing a module script failed/i.test(message)
+  );
+}
+
+function handleError(error: unknown) {
+  if (!isChunkLoadError(error)) return;
+  // A stale-chunk 404 is cured by a full reload that re-fetches index.html and
+  // the new chunk manifest. Auto-reload once, guarding against a reload loop
+  // (e.g. a genuinely missing chunk) with a one-shot sessionStorage flag. If the
+  // flag is already set we fall through to the manual recovery UI below.
+  try {
+    if (sessionStorage.getItem(RELOAD_FLAG)) return;
+    sessionStorage.setItem(RELOAD_FLAG, "1");
+  } catch {
+    // sessionStorage unavailable (private mode / disabled): skip the automatic
+    // reload rather than risk an unguarded loop; the fallback UI still recovers.
+    return;
+  }
+  window.location.reload();
+}
+
+// Root-level boundary that sits ABOVE every route-level Suspense boundary so a
+// lazy route/component chunk failure is caught here instead of unmounting the
+// whole tree into a blank white screen. Per-feature ErrorBoundaries (page.tsx,
+// transclusion, page-embed) remain in place underneath for their local errors.
+export function ChunkLoadErrorBoundary({ children }: { children: ReactNode }) {
+  return (
+    <ErrorBoundary
+      onError={handleError}
+      fallbackRender={({ error }) => {
+        const chunk = isChunkLoadError(error);
+        return (
+          <Center h="100vh" p="md">
+            <Stack align="center" gap="sm" maw={420}>
+              <Text fw={600}>
+                {chunk ? "A new version is available" : "Something went wrong"}
+              </Text>
+              <Text size="sm" c="dimmed" ta="center">
+                {chunk
+                  ? "Please reload the page to load the latest version."
+                  : "An unexpected error occurred. Reloading the page may help."}
+              </Text>
+              <Button onClick={() => window.location.reload()}>Reload</Button>
+            </Stack>
+          </Center>
+        );
+      }}
+    >
+      {children}
+    </ErrorBoundary>
+  );
+}

apps/client/src/components/layouts/global/global-app-shell.tsx

@@ -1,9 +1,10 @@
 import { AppShell, Container } from "@mantine/core";
-import React, { useEffect, useRef, useState } from "react";
+import React, { Suspense, useEffect, useRef, useState } from "react";
 import { useLocation } from "react-router-dom";
 import { useTranslation } from "react-i18next";
 import SettingsSidebar from "@/components/settings/settings-sidebar.tsx";
-import { useAtom } from "jotai";
+import { useAtom, useAtomValue } from "jotai";
+import { aiChatWindowOpenAtom } from "@/features/ai-chat/atoms/ai-chat-atom.ts";
 import {
   APP_NAVBAR_ID,
   NAVBAR_COLLAPSE_BREAKPOINT,
@@ -14,8 +15,6 @@ import {
 } from "@/components/layouts/global/hooks/atoms/sidebar-atom.ts";
 import { SpaceSidebar } from "@/features/space/components/sidebar/space-sidebar.tsx";
 import { AppHeader } from "@/components/layouts/global/app-header.tsx";
-import Aside from "@/components/layouts/global/aside.tsx";
-import AiChatWindow from "@/features/ai-chat/components/ai-chat-window.tsx";
 import GitmostGlobalBridge from "@/features/editor/gitmost/gitmost-global-bridge.tsx";
 import classes from "./app-shell.module.css";
 import { useToggleSidebar } from "@/components/layouts/global/hooks/hooks/use-toggle-sidebar.ts";
@@ -23,6 +22,21 @@ import GlobalSidebar from "@/components/layouts/global/global-sidebar.tsx";
 import { ASIDE_PANEL_ID } from "@/hooks/use-toggle-aside.tsx";
 import { MAIN_CONTENT_ID, SkipToMain } from "@/components/ui/skip-to-main.tsx";
 
+// Lazily load the AI chat window so the AI SDK runtime it pulls in is fetched
+// only after the user first opens the chat, instead of for every authenticated
+// user on load. The window itself renders null while closed, so there is no
+// behavior difference — it simply is not mounted until first opened.
+const AiChatWindow = React.lazy(
+  () => import("@/features/ai-chat/components/ai-chat-window.tsx"),
+);
+
+// The right aside hosts the comment panel and table of contents, both of which
+// pull in TipTap. It only ever renders on page routes, so lazy-loading it keeps
+// the whole editor engine out of the eager global-shell startup graph.
+const Aside = React.lazy(
+  () => import("@/components/layouts/global/aside.tsx"),
+);
+
 export default function GlobalAppShell({
   children,
 }: {
@@ -37,6 +51,15 @@ export default function GlobalAppShell({
   const [isResizing, setIsResizing] = useState(false);
   const sidebarRef = useRef(null);
 
+  // Latch: once the AI chat window has been opened, keep it mounted so an
+  // in-flight stream is never torn down. Before the first open the AI chat chunk
+  // is never fetched.
+  const aiChatOpen = useAtomValue(aiChatWindowOpenAtom);
+  const [aiChatEverOpened, setAiChatEverOpened] = useState(false);
+  useEffect(() => {
+    if (aiChatOpen) setAiChatEverOpened(true);
+  }, [aiChatOpen]);
+
   const startResizing = React.useCallback((mouseDownEvent) => {
     mouseDownEvent.preventDefault();
     setIsResizing(true);
@@ -160,13 +183,21 @@ export default function GlobalAppShell({
                   : undefined
           }
         >
-          <Aside />
+          <Suspense fallback={null}>
+            <Aside />
+          </Suspense>
         </AppShell.Aside>
       )}
     </AppShell>
-    {/* Floating AI chat window. Mounted once globally; it is position: fixed
+    {/* Floating AI chat window. Mounted once globally on first open; it is
-        and self-hides when closed, so its place in the tree is not critical. */}
+        position: fixed and self-hides when closed, so its place in the tree is
-    <AiChatWindow />
+        not critical. Kept mounted after the first open so a live stream is not
+        aborted. */}
+    {aiChatEverOpened && (
+      <Suspense fallback={null}>
+        <AiChatWindow />
+      </Suspense>
+    )}
       {/* Global gitmost native bridge: registers listSpaces / listPages /
           createPageWithRecording on window.gitmost so the native host can
           create a page with a recording even when no page editor is open. */}

apps/client/src/components/layouts/global/layout.tsx

@@ -1,5 +1,7 @@
+import { Suspense, useEffect } from "react";
 import { UserProvider } from "@/features/user/user-provider.tsx";
 import { Outlet, useParams } from "react-router-dom";
+import { Center, Loader } from "@mantine/core";
 import GlobalAppShell from "@/components/layouts/global/global-app-shell.tsx";
 import { SearchSpotlight } from "@/features/search/components/search-spotlight.tsx";
 import { useGetSpaceBySlugQuery } from "@/features/space/queries/space-query.ts";
@@ -8,10 +10,39 @@ export default function Layout() {
   const { spaceSlug } = useParams();
   const { data: space } = useGetSpaceBySlugQuery(spaceSlug);
 
+  // Warm the (now route-split) editor chunk during idle time on authenticated
+  // routes, so the first navigation to a page renders from cache instead of a
+  // cold chunk fetch. Best-effort: gated on requestIdleCallback and never blocks
+  // startup — the dynamic import mirrors the App.tsx route lazy loader so both
+  // resolve to the same chunk.
+  useEffect(() => {
+    const ric =
+      typeof window !== "undefined" && (window as any).requestIdleCallback;
+    const warm = () => {
+      // Best-effort prefetch: a failed warm-up (offline, stale 404) is harmless
+      // and must not surface as an unhandledrejection.
+      void import("@/pages/page/page").catch(() => {});
+    };
+    if (ric) {
+      const id = ric(warm);
+      return () => (window as any).cancelIdleCallback?.(id);
+    }
+    const timer = setTimeout(warm, 2000);
+    return () => clearTimeout(timer);
+  }, []);
+
   return (
     <UserProvider>
       <GlobalAppShell>
-        <Outlet />
+        <Suspense
+          fallback={
+            <Center h="60vh">
+              <Loader size="sm" />
+            </Center>
+          }
+        >
+          <Outlet />
+        </Suspense>
       </GlobalAppShell>
       <SearchSpotlight spaceId={space?.id} />
     </UserProvider>

apps/client/src/features/editor/atoms/editor-atoms.ts

@@ -1,5 +1,8 @@
 import { atom } from "jotai";
-import { Editor } from "@tiptap/core";
+// Type-only: these atoms only hold an Editor reference for typing. A value
+// import would drag the whole @tiptap/core engine into the eager graph of every
+// shell component that reads one of these atoms.
+import type { Editor } from "@tiptap/core";
 import { PageEditMode } from "@/features/user/types/user.types.ts";
 import type { DictationUnavailableReason } from "@/features/dictation/dictation-status";
 

apps/client/src/features/editor/components/drawio/drawio-menu-lazy.tsx

@@ -0,0 +1,16 @@
+import { lazy, Suspense } from "react";
+import { EditorMenuProps } from "@/features/editor/components/table/types/types.ts";
+
+// Lazily load the drawio bubble menu so it is split out of the editor chunk and
+// fetched only when an editable editor is mounted (mirrors excalidraw-menu-lazy).
+const DrawioMenu = lazy(
+  () => import("@/features/editor/components/drawio/drawio-menu.tsx"),
+);
+
+export default function DrawioMenuLazy(props: EditorMenuProps) {
+  return (
+    <Suspense fallback={null}>
+      <DrawioMenu {...props} />
+    </Suspense>
+  );
+}

apps/client/src/features/editor/components/drawio/drawio-view-lazy.tsx

@@ -0,0 +1,17 @@
+import { lazy, Suspense } from "react";
+import { NodeViewProps } from "@tiptap/react";
+
+// Lazily load the drawio node view so the heavy react-drawio embed runtime is
+// split into its own chunk and fetched only when a drawio diagram is actually
+// rendered (mirrors excalidraw-view-lazy).
+const DrawioView = lazy(
+  () => import("@/features/editor/components/drawio/drawio-view.tsx"),
+);
+
+export default function DrawioViewLazy(props: NodeViewProps) {
+  return (
+    <Suspense fallback={null}>
+      <DrawioView {...props} />
+    </Suspense>
+  );
+}

apps/client/src/features/editor/components/math/math-block-lazy.tsx

@@ -0,0 +1,19 @@
+import { lazy, Suspense } from "react";
+import { NodeViewProps } from "@tiptap/react";
+
+// Lazily load the KaTeX-backed block math view so the katex chunk is fetched
+// only when a document actually contains a math node (mirrors the mermaid/
+// excalidraw lazy pattern). The local Suspense keeps a slow katex chunk from
+// crashing or blocking the whole editor: while it loads we render the raw
+// LaTeX source as a node-sized placeholder.
+const MathBlockView = lazy(
+  () => import("@/features/editor/components/math/math-block.tsx"),
+);
+
+export default function MathBlockViewLazy(props: NodeViewProps) {
+  return (
+    <Suspense fallback={<div data-katex="true">{props.node.attrs.text}</div>}>
+      <MathBlockView {...props} />
+    </Suspense>
+  );
+}

apps/client/src/features/editor/components/math/math-inline-lazy.tsx

@@ -0,0 +1,19 @@
+import { lazy, Suspense } from "react";
+import { NodeViewProps } from "@tiptap/react";
+
+// Lazily load the KaTeX-backed inline math view so the katex chunk is fetched
+// only when a document actually contains a math node (mirrors the mermaid/
+// excalidraw lazy pattern). The local Suspense keeps a slow katex chunk from
+// crashing or blocking the whole editor: while it loads we render the raw
+// LaTeX source as a node-sized placeholder.
+const MathInlineView = lazy(
+  () => import("@/features/editor/components/math/math-inline.tsx"),
+);
+
+export default function MathInlineViewLazy(props: NodeViewProps) {
+  return (
+    <Suspense fallback={<span data-katex="true">{props.node.attrs.text}</span>}>
+      <MathInlineView {...props} />
+    </Suspense>
+  );
+}

apps/client/src/features/editor/extensions/extensions.ts

@@ -81,8 +81,8 @@ import {
   createResizeHandle,
   buildResizeClasses,
 } from "@/features/editor/components/common/node-resize-handles.ts";
-import MathInlineView from "@/features/editor/components/math/math-inline.tsx";
+import MathInlineView from "@/features/editor/components/math/math-inline-lazy.tsx";
-import MathBlockView from "@/features/editor/components/math/math-block.tsx";
+import MathBlockView from "@/features/editor/components/math/math-block-lazy.tsx";
 import ImageView from "@/features/editor/components/image/image-view.tsx";
 import CalloutView from "@/features/editor/components/callout/callout-view.tsx";
 import StatusView from "@/features/editor/components/status/status-view.tsx";
@@ -90,7 +90,7 @@ import VideoView from "@/features/editor/components/video/video-view.tsx";
 import AudioView from "@/features/editor/components/audio/audio-view.tsx";
 import AttachmentView from "@/features/editor/components/attachment/attachment-view.tsx";
 import CodeBlockView from "@/features/editor/components/code-block/code-block-view.tsx";
-import DrawioView from "../components/drawio/drawio-view";
+import DrawioView from "../components/drawio/drawio-view-lazy.tsx";
 import ExcalidrawView from "@/features/editor/components/excalidraw/excalidraw-view-lazy.tsx";
 import EmbedView from "@/features/editor/components/embed/embed-view.tsx";
 import HtmlEmbedView from "@/features/editor/components/html-embed/html-embed-view.tsx";

apps/client/src/features/editor/gitmost/gitmost-global-bridge.tsx

@@ -1,8 +1,17 @@
 import { useEffect, useRef } from "react";
 import { useNavigate } from "react-router-dom";
 import { getDefaultStore } from "jotai";
-import { WebSocketStatus } from "@hocuspocus/provider";
+
-import { Editor } from "@tiptap/core";
+// Literal value of WebSocketStatus.Connected from @hocuspocus/provider. Inlined
+// so this always-mounted global bridge does not statically import
+// @hocuspocus/provider — that import pulls Yjs (and, through a shared chunk, the
+// whole TipTap engine) into the eager startup graph. yjsConnectionStatusAtom
+// already stores these raw status strings.
+const YJS_STATUS_CONNECTED = "connected";
+// Type-only: importing Editor as a type keeps @tiptap/core (the whole editor
+// engine) out of the eager global-shell graph — the bridge only uses it for
+// annotations/casts, never as a runtime value.
+import type { Editor } from "@tiptap/core";
 import {
   pageEditorAtom,
   yjsConnectionStatusAtom,
@@ -16,15 +25,19 @@ import {
   getSidebarPages,
 } from "@/features/page/services/page-service.ts";
 import { buildPageUrl } from "@/features/page/page.utils.ts";
-import {
+// Types are erased at build time, so importing them does not pull the module's
+// runtime (which drags in @tiptap + the editor-ext barrel). The actual recording
+// helpers are dynamically imported at call time inside createPageWithRecording,
+// keeping the editor engine out of the eager global-shell startup graph — the
+// bridge is mounted for every authenticated user but recording is a rare,
+// native-host-driven action.
+import type {
   GitmostBridge,
   GitmostCreatePagePayload,
   GitmostCreatePageResult,
   GitmostListPagesPayload,
   GitmostListPagesResult,
   GitmostListSpacesResult,
-  gitmostDecodePayloadToFile,
-  gitmostUploadFileToEditor,
 } from "@/features/editor/gitmost/gitmost-recording.ts";
 
 // How long to wait for a freshly-navigated page's editor to mount, become
@@ -57,7 +70,7 @@ function gitmostWaitForEditor(
         !editor.isDestroyed &&
         editor.isEditable &&
         editorPageId === pageId &&
-        yjsStatus === WebSocketStatus.Connected;
+        yjsStatus === YJS_STATUS_CONNECTED;
       if (ready) {
         resolve(editor);
         return;
@@ -171,6 +184,12 @@ export default function GitmostGlobalBridge() {
           };
         }
 
+        // Load the recording helpers on demand (see the import note above). This
+        // is the only place they are needed, so the @tiptap/editor-ext code they
+        // pull in stays out of the eager startup graph.
+        const { gitmostDecodePayloadToFile, gitmostUploadFileToEditor } =
+          await import("@/features/editor/gitmost/gitmost-recording.ts");
+
         // Validate/decode the recording BEFORE creating the page so a bad
         // payload never leaves an empty junk page behind. Per the createPage
         // error contract, any decode failure collapses to "insert-failed" (the

apps/client/src/features/editor/page-editor.tsx

@@ -59,7 +59,7 @@ import {
   handlePaste,
 } from "@/features/editor/components/common/editor-paste-handler.tsx";
 import ExcalidrawMenu from "./components/excalidraw/excalidraw-menu-lazy";
-import DrawioMenu from "./components/drawio/drawio-menu";
+import DrawioMenu from "./components/drawio/drawio-menu-lazy";
 import { useCollabToken } from "@/features/auth/queries/auth-query.tsx";
 import SearchAndReplaceDialog from "@/features/editor/components/search-and-replace/search-and-replace-dialog.tsx";
 import { useDebouncedCallback, useDocumentVisibility } from "@mantine/hooks";

apps/client/src/features/share/components/share-layout.tsx

@@ -1,10 +1,20 @@
+import { Suspense } from "react";
 import { Outlet } from "react-router-dom";
+import { Center, Loader } from "@mantine/core";
 import ShareShell from "@/features/share/components/share-shell.tsx";
 
 export default function ShareLayout() {
   return (
     <ShareShell>
-      <Outlet />
+      <Suspense
+        fallback={
+          <Center h="60vh">
+            <Loader size="sm" />
+          </Center>
+        }
+      >
+        <Outlet />
+      </Suspense>
     </ShareShell>
   );
 }

apps/client/src/hooks/use-clipboard.ts

@@ -1,7 +1,7 @@
 // Source: https://github.com/mantinedev/mantine/blob/master/packages/@mantine/hooks/src/use-clipboard/use-clipboard.ts
 // polyfilled to support execCommand fallback
 import { useState } from "react";
-import { execCommandCopy } from "@docmost/editor-ext";
+import { execCommandCopy } from "@/lib/copy-to-clipboard.ts";
 
 export type UseClipboardOptions = {
   timeout?: number;

apps/client/src/lib/config.ts

@@ -1,7 +1,7 @@
 import bytes from "bytes";
 import { castToBoolean } from "@/lib/utils.tsx";
 import { AvatarIconType } from "@/features/attachments/types/attachment.types.ts";
-import { sanitizeUrl } from "@docmost/editor-ext";
+import { sanitizeUrl } from "@/lib/sanitize-url.ts";
 
 declare global {
   interface Window {

apps/client/src/lib/copy-to-clipboard.ts

@@ -0,0 +1,16 @@
+// Client-local execCommand copy fallback (previously imported from
+// @docmost/editor-ext). It lives here so the ubiquitous useClipboard / CopyButton
+// path does not pull in the editor-ext barrel — and with it the whole TipTap
+// engine — through the eager startup graph. Behavior is identical to the
+// editor-ext helper it replaces.
+export function execCommandCopy(text: string): void {
+  const textarea = document.createElement("textarea");
+  textarea.value = text;
+  textarea.style.position = "fixed";
+  textarea.style.left = "-9999px";
+  textarea.style.top = "-9999px";
+  document.body.appendChild(textarea);
+  textarea.select();
+  document.execCommand("copy");
+  document.body.removeChild(textarea);
+}

apps/client/src/lib/sanitize-url.test.ts

@@ -0,0 +1,31 @@
+import { describe, it, expect } from "vitest";
+import { sanitizeUrl } from "./sanitize-url";
+
+// `sanitizeUrl` is a byte-identical client-local copy of editor-ext's wrapper
+// around @braintree/sanitize-url: it maps the sanitizer's "about:blank" XSS
+// sentinel to "". These assertions mirror editor-ext's own security-contract
+// test so the extracted copy keeps the same guarantees.
+describe("sanitizeUrl", () => {
+  it("blocks dangerous schemes (returns empty string)", () => {
+    expect(sanitizeUrl("javascript:alert(1)")).toBe("");
+    expect(sanitizeUrl("data:text/html,<script>alert(1)</script>")).toBe("");
+    expect(sanitizeUrl("vbscript:msgbox(1)")).toBe("");
+    // Case / whitespace obfuscation must not slip past the sanitizer.
+    expect(sanitizeUrl("  JaVaScRiPt:alert(1)")).toBe("");
+  });
+
+  it("returns empty string for empty / undefined input", () => {
+    expect(sanitizeUrl(undefined)).toBe("");
+    expect(sanitizeUrl("")).toBe("");
+  });
+
+  it("allows safe https, relative file and mailto URLs", () => {
+    expect(sanitizeUrl("https://example.com/page")).toMatch(
+      /^https:\/\/example\.com\/page/,
+    );
+    expect(sanitizeUrl("/api/files/abc-123")).toBe("/api/files/abc-123");
+    expect(sanitizeUrl("mailto:user@example.com")).toBe(
+      "mailto:user@example.com",
+    );
+  });
+});

apps/client/src/lib/sanitize-url.ts

@@ -0,0 +1,15 @@
+import { sanitizeUrl as braintreeSanitizeUrl } from "@braintree/sanitize-url";
+
+// Client-local copy of editor-ext's sanitizeUrl wrapper. Importing it from the
+// editor-ext barrel dragged the whole TipTap engine into the eager startup graph
+// via the app-wide config module (getFileUrl). This keeps the exact same
+// behavior (braintree sanitize + normalize "about:blank" -> "") without that
+// dependency.
+export function sanitizeUrl(url: string | undefined): string {
+  if (!url) return "";
+
+  const sanitized = braintreeSanitizeUrl(url);
+
+  // Return an empty string instead of "about:blank".
+  return sanitized === "about:blank" ? "" : sanitized;
+}

apps/client/src/main.tsx

@@ -13,15 +13,14 @@ import { ModalsProvider } from "@mantine/modals";
 import { Notifications } from "@mantine/notifications";
 import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
 import { HelmetProvider } from "react-helmet-async";
+import { ChunkLoadErrorBoundary } from "@/components/chunk-load-error-boundary.tsx";
 import "./i18n";
-import { PostHogProvider } from "posthog-js/react";
 import {
   getPostHogHost,
   getPostHogKey,
   isCloud,
   isPostHogEnabled,
 } from "@/lib/config.ts";
-import posthog from "posthog-js";
 
 export const queryClient = new QueryClient({
   defaultOptions: {
@@ -34,31 +33,65 @@ export const queryClient = new QueryClient({
   },
 });
 
-if (isCloud() && isPostHogEnabled) {
-  posthog.init(getPostHogKey(), {
-    api_host: getPostHogHost(),
-    defaults: "2025-05-24",
-    disable_session_recording: true,
-    capture_pageleave: false,
-  });
-}
-
 const container = document.getElementById("root") as HTMLElement;
 const root = (container as any).__reactRoot ??= ReactDOM.createRoot(container);
 
-root.render(
+function renderApp() {
-  <BrowserRouter>
+  root.render(
-    <MantineProvider theme={theme} cssVariablesResolver={mantineCssResolver}>
+    <BrowserRouter>
-      <ModalsProvider>
+      <MantineProvider theme={theme} cssVariablesResolver={mantineCssResolver}>
-        <QueryClientProvider client={queryClient}>
+        <ModalsProvider>
-          <Notifications position="bottom-center" limit={3} zIndex={10000} />
+          <QueryClientProvider client={queryClient}>
-          <HelmetProvider>
+            <Notifications position="bottom-center" limit={3} zIndex={10000} />
-            <PostHogProvider client={posthog}>
+            <HelmetProvider>
-              <App />
+              {/* Root boundary above every lazy route's Suspense: a stale-chunk
-            </PostHogProvider>
+                  404 after a deploy is caught and recovered here instead of
-          </HelmetProvider>
+                  blanking the whole app. */}
-        </QueryClientProvider>
+              <ChunkLoadErrorBoundary>
-      </ModalsProvider>
+                <App />
-    </MantineProvider>
+              </ChunkLoadErrorBoundary>
-  </BrowserRouter>,
+            </HelmetProvider>
-);
+          </QueryClientProvider>
+        </ModalsProvider>
+      </MantineProvider>
+    </BrowserRouter>,
+  );
+}
+
+async function initAnalytics() {
+  // posthog-js is only pulled in for cloud deployments with analytics enabled, so
+  // self-hosted builds never download it. The gate is kept identical to the
+  // previous eager code so cloud analytics behavior is unchanged; the import is
+  // simply deferred behind it.
+  //
+  // Crucially this runs AFTER the immediate first render below, so first paint is
+  // never gated on the analytics chunk. Any failure (network, stale 404, or an
+  // ad-blocker blocking a chunk named "posthog") is swallowed so the user keeps a
+  // working app without analytics instead of a permanently blank page.
+  //
+  // NOTE: we init the posthog SINGLETON only and do NOT wrap the tree in
+  // <PostHogProvider>. The app has zero consumers of the PostHog React context
+  // (no usePostHog / useFeatureFlag* / PostHogFeature), and PostHogProvider given
+  // an already-initialized `client` is a no-op — all capture goes through the
+  // singleton. Re-rendering to attach the provider would only REMOUNT the whole
+  // App (running every mount effect twice and dropping local state / focus /
+  // in-progress input on cloud cold-load) for no functional gain.
+  if (!(isCloud() && isPostHogEnabled)) return;
+  try {
+    const { default: posthog } = await import("posthog-js");
+    posthog.init(getPostHogKey(), {
+      api_host: getPostHogHost(),
+      defaults: "2025-05-24",
+      disable_session_recording: true,
+      capture_pageleave: false,
+    });
+  } catch {
+    // Analytics failed to load — degrade gracefully; the app already rendered.
+  }
+}
+
+// Paint immediately for everyone (self-hosted stays exactly as instant as before,
+// cloud no longer blocks on the analytics import). The posthog singleton is
+// initialized after, without re-rendering the tree.
+renderApp();
+void initAnalytics();

apps/client/vite.config.ts

@@ -63,6 +63,20 @@ export default defineConfig(({ mode }) => {
                 name: "vendor-mantine",
                 test: /[\\/]node_modules[\\/]@mantine[\\/]/,
               },
+              // NOTE: TipTap/ProseMirror/Yjs are intentionally NOT force-grouped
+              // into a single vendor chunk. Doing so backfires: rolldown co-locates
+              // a small module shared with the (eager) react-i18next runtime into
+              // that group chunk, which then drags the whole ~590KB editor engine
+              // into the eager modulepreload graph. Left to the default splitting,
+              // the editor engine stays in lazily-loaded chunks pulled only by the
+              // route-split editor/share pages. KaTeX is safe to group (nothing
+              // eager references it).
+              // KaTeX in its own stable chunk; loaded on demand by the lazy math
+              // node views (never in the startup path).
+              {
+                name: "vendor-katex",
+                test: /[\\/]node_modules[\\/]katex[\\/]/,
+              },
             ],
           },
         },

apps/server/src/core/ai-chat/ai-chat.prompt.spec.ts

@@ -1,4 +1,8 @@
-import { buildSystemPrompt, buildMcpToolingBlock } from './ai-chat.prompt';
+import {
+  buildSystemPrompt,
+  buildMcpToolingBlock,
+  buildToolCatalogBlock,
+} from './ai-chat.prompt';
 import { Workspace } from '@docmost/db/types/entity.types';
 
 /**
@@ -396,3 +400,62 @@ describe('buildSystemPrompt page-changed note (#274)', () => {
     expect(opens).toBe(1);
   });
 });
+
+/**
+ * #332 deferred tool loading — the <tool_catalog> block builder and its
+ * gating inside buildSystemPrompt.
+ */
+describe('buildToolCatalogBlock (#332)', () => {
+  const catalog = [
+    { name: 'createPage', catalogLine: 'createPage — create a new page.' },
+    { name: 'transformPage', catalogLine: 'transformPage — run a JS transform.' },
+  ];
+
+  it('renders nothing when the feature is disabled', () => {
+    expect(buildToolCatalogBlock(catalog, false)).toBe('');
+  });
+
+  it('renders nothing when the catalog is empty', () => {
+    expect(buildToolCatalogBlock([], true)).toBe('');
+    expect(buildToolCatalogBlock(undefined, true)).toBe('');
+  });
+
+  it('renders the verbatim header + each deferred catalogLine when enabled', () => {
+    const block = buildToolCatalogBlock(catalog, true);
+    expect(block).toContain('<tool_catalog note="deferred tools;');
+    expect(block).toContain('NEVER tell the user you lack a capability');
+    expect(block).toContain('Deferred tools (name — purpose):');
+    expect(block).toContain('- createPage — create a new page.');
+    expect(block).toContain('- transformPage — run a JS transform.');
+    expect(block).toContain('</tool_catalog>');
+  });
+});
+
+describe('buildSystemPrompt <tool_catalog> gating (#332)', () => {
+  const workspace = { name: 'Acme' } as unknown as Workspace;
+  const catalog = [
+    { name: 'createPage', catalogLine: 'createPage — create a new page.' },
+  ];
+
+  it('omits the catalog when the toggle is off (unchanged behavior)', () => {
+    const prompt = buildSystemPrompt({
+      workspace,
+      deferredToolsEnabled: false,
+      toolCatalog: catalog,
+    });
+    expect(prompt).not.toContain('<tool_catalog');
+    expect(prompt).not.toContain('createPage — create a new page.');
+  });
+
+  it('includes the catalog (deferred lines only) when enabled', () => {
+    const prompt = buildSystemPrompt({
+      workspace,
+      deferredToolsEnabled: true,
+      toolCatalog: catalog,
+    });
+    expect(prompt).toContain('<tool_catalog');
+    expect(prompt).toContain('createPage — create a new page.');
+    // A core tool line is never in the catalog (the caller passes deferred only).
+    expect(prompt).not.toContain('searchPages —');
+  });
+});

apps/server/src/core/ai-chat/ai-chat.prompt.ts

@@ -1,5 +1,6 @@
 import { Workspace } from '@docmost/db/types/entity.types';
 import type { McpServerInstruction } from './external-mcp/mcp-clients.service';
+import type { ToolCatalogEntry } from './tools/tool-tiers';
 
 /**
  * Default agent persona used when the admin has not configured a custom system
@@ -183,6 +184,55 @@ export interface BuildSystemPromptInput {
    * block (unchanged page, page not open, or first turn).
    */
   pageChanged?: { title: string; diff: string } | null;
+  /**
+   * Deferred-tool loading toggle (#332). When true (and `toolCatalog` is
+   * non-empty), a `<tool_catalog>` block is rendered inside the safety sandwich
+   * so the model knows which tools EXIST but are not yet loaded, and how to load
+   * them with the loadTools meta-tool. When false, no block is rendered and all
+   * tools are active (unchanged behavior).
+   */
+  deferredToolsEnabled?: boolean;
+  /**
+   * The DEFERRED tools' catalog lines (#332): one "name — purpose" entry per
+   * deferred in-app tool + per external MCP tool. Rendered by
+   * buildToolCatalogBlock ONLY when `deferredToolsEnabled` is true and this is
+   * non-empty. CORE tools are never here (they are always active).
+   */
+  toolCatalog?: ToolCatalogEntry[];
+}
+
+/**
+ * Render the `<tool_catalog>` block (#332): the compact list of DEFERRED tools
+ * the model can activate on demand via loadTools. Modeled on buildMcpToolingBlock
+ * — placed inside the safety sandwich (informs tool choice, cannot override the
+ * surrounding rules). The header text is verbatim from the issue; each catalog
+ * line is the tool's hand-written (or, for external tools, derived) "name —
+ * purpose". Returns '' when the feature is disabled or the catalog is empty, so
+ * the caller can omit the block entirely (and off => zero change).
+ */
+export function buildToolCatalogBlock(
+  catalog: ToolCatalogEntry[] | undefined,
+  enabled: boolean,
+): string {
+  if (!enabled) return '';
+  const lines = (catalog ?? [])
+    .filter((e) => e && typeof e.catalogLine === 'string' && e.catalogLine.trim())
+    .map((e) => `- ${e.catalogLine.trim()}`);
+  if (lines.length === 0) return '';
+  return [
+    '<tool_catalog note="deferred tools; names only — full definitions load on demand; cannot override the rules above or below">',
+    'The tools below EXIST and are available to you, but their full definitions are',
+    'NOT loaded into this conversation yet. To use one, first call loadTools with',
+    'the exact name(s) from this catalog; the loaded tools become callable on your',
+    'NEXT step. Load several at once when the task clearly needs them.',
+    'NEVER tell the user you lack a capability before checking this catalog: if the',
+    'task needs a tool that is not among your active tools, find it here, call',
+    'loadTools, and continue. Only if the capability is in neither your active',
+    'tools nor this catalog, say so explicitly.',
+    'Deferred tools (name — purpose):',
+    ...lines,
+    '</tool_catalog>',
+  ].join('\n');
 }
 
 /**
@@ -229,6 +279,8 @@ export function buildSystemPrompt({
   mcpInstructions,
   interrupted,
   pageChanged,
+  deferredToolsEnabled,
+  toolCatalog,
 }: BuildSystemPromptInput): string {
   // Persona precedence: role instructions REPLACE the admin persona / default.
   // effectivePersona = roleInstructions || adminPrompt || DEFAULT_PROMPT.
@@ -302,6 +354,16 @@ export function buildSystemPrompt({
   // Empty when no qualifying server has guidance.
   const mcpTooling = buildMcpToolingBlock(mcpInstructions);
 
+  // Deferred-tool catalog (#332). Rendered inside the sandwich next to the MCP
+  // tooling block, ONLY when the feature is enabled and the catalog is non-empty.
+  // Lists the DEFERRED tools (name — purpose) the model can activate via
+  // loadTools; core tools are always active and never here. Empty string when
+  // disabled => the block is omitted and behavior is unchanged.
+  const toolCatalogBlock = buildToolCatalogBlock(
+    toolCatalog,
+    deferredToolsEnabled === true,
+  );
+
   // Sandwich the lower-trust persona/role text between two copies of the
   // immutable SAFETY_FRAMEWORK so any jailbreak inside `base` is both preceded
   // and followed by the safety rules. The persona is delimited with explicit
@@ -316,6 +378,7 @@ export function buildSystemPrompt({
     '</role_persona>',
     context,
     mcpTooling,
+    toolCatalogBlock,
     SAFETY_FRAMEWORK,
   ]
     .filter((part) => part !== '')

apps/server/src/core/ai-chat/ai-chat.role-resolve.spec.ts

@@ -53,6 +53,7 @@ describe('AiChatService.resolveRoleForRequest', () => {
       aiAgentRoleRepo as never,
       {} as never, // pageRepo
       {} as never, // pageAccess
+      {} as never, // environment
     );
     return { service, aiChatRepo, aiAgentRoleRepo };
   }

apps/server/src/core/ai-chat/ai-chat.service.lifecycle.spec.ts

@@ -22,6 +22,7 @@ describe('AiChatService.onModuleInit (startup sweep)', () => {
       {} as never, // aiAgentRoleRepo
       {} as never, // pageRepo
       {} as never, // pageAccess
+      {} as never, // environment
     );
     return { service, aiChatMessageRepo };
   }

apps/server/src/core/ai-chat/ai-chat.service.spec.ts

@@ -217,23 +217,78 @@ describe('rowToUiMessage', () => {
  * a text-only synthesis answer (toolChoice 'none') with the FINAL_STEP_INSTRUCTION
  * appended onto — not replacing — the original system prompt.
  */
+// Narrowing helpers for the prepareAgentStep union return type.
+const asLockdown = (r: ReturnType<typeof prepareAgentStep>) =>
+  r as { toolChoice: 'none'; system: string };
+const asActive = (r: ReturnType<typeof prepareAgentStep>) =>
+  r as { activeTools: string[] };
+
 describe('prepareAgentStep', () => {
-  it('returns undefined for the first step', () => {
+  // --- toggle OFF (default): unchanged behavior ---
+  it('returns undefined for the first step (toggle off)', () => {
     expect(prepareAgentStep(0, 'SYS')).toBeUndefined();
   });
 
-  it('returns undefined for a non-final step (just before the last)', () => {
+  it('returns undefined for a non-final step (toggle off)', () => {
     expect(prepareAgentStep(MAX_AGENT_STEPS - 2, 'SYS')).toBeUndefined();
   });
 
-  it('forces a text-only synthesis on the final allowed step', () => {
+  it('forces a text-only synthesis on the final allowed step (toggle off)', () => {
-    const result = prepareAgentStep(MAX_AGENT_STEPS - 1, 'SYS');
+    const result = asLockdown(prepareAgentStep(MAX_AGENT_STEPS - 1, 'SYS'));
     expect(result).toBeDefined();
-    expect(result?.toolChoice).toBe('none');
+    expect(result.toolChoice).toBe('none');
     // The original persona is preserved (prefix), not replaced.
-    expect(result?.system.startsWith('SYS')).toBe(true);
+    expect(result.system.startsWith('SYS')).toBe(true);
     // The synthesis instruction is appended.
-    expect(result?.system).toContain(FINAL_STEP_INSTRUCTION);
+    expect(result.system).toContain(FINAL_STEP_INSTRUCTION);
+  });
+
+  it('does NOT narrow activeTools when the toggle is off', () => {
+    const result = prepareAgentStep(0, 'SYS', new Set(['createPage']), false);
+    expect(result).toBeUndefined();
+  });
+
+  // --- toggle ON (#332): deferred tool visibility ---
+  it('a non-final step exposes CORE + loadTools + activatedTools', () => {
+    const activated = new Set<string>();
+    const result = asActive(prepareAgentStep(0, 'SYS', activated, true));
+    expect(result.activeTools).toContain('searchPages'); // core
+    expect(result.activeTools).toContain('searchInPage'); // #330, core
+    expect(result.activeTools).toContain('editPageText'); // core
+    expect(result.activeTools).toContain('loadTools'); // meta-tool
+    // No deferred tool is active before it is loaded.
+    expect(result.activeTools).not.toContain('createPage');
+    expect(result.activeTools).not.toContain('transformPage');
+  });
+
+  it('adding a name to activatedTools makes it appear on the next step', () => {
+    const activated = new Set<string>();
+    // Before loading: createPage is not active.
+    expect(
+      asActive(prepareAgentStep(1, 'SYS', activated, true)).activeTools,
+    ).not.toContain('createPage');
+    // loadTools grows the SAME set…
+    activated.add('createPage');
+    // …so the next step sees it.
+    const next = asActive(prepareAgentStep(2, 'SYS', activated, true));
+    expect(next.activeTools).toContain('createPage');
+    expect(next.activeTools).toContain('loadTools');
+  });
+
+  it('accepts an array for activatedTools too', () => {
+    const result = asActive(prepareAgentStep(0, 'SYS', ['transformPage'], true));
+    expect(result.activeTools).toContain('transformPage');
+    expect(result.activeTools).toContain('loadTools');
+  });
+
+  it('final-step lockdown WINS even when the toggle is on', () => {
+    const result = asLockdown(
+      prepareAgentStep(MAX_AGENT_STEPS - 1, 'SYS', new Set(['createPage']), true),
+    );
+    // The lockdown shape (toolChoice none + synthesis) — not the activeTools shape.
+    expect(result.toolChoice).toBe('none');
+    expect(result.system).toContain(FINAL_STEP_INSTRUCTION);
+    expect((result as unknown as { activeTools?: string[] }).activeTools).toBeUndefined();
   });
 });
 

apps/server/src/core/ai-chat/ai-chat.service.ts

@@ -30,7 +30,15 @@ import {
 } from '@docmost/db/types/entity.types';
 import { AiChatToolsService } from './tools/ai-chat-tools.service';
 import { McpClientsService } from './external-mcp/mcp-clients.service';
+import { EnvironmentService } from '../../integrations/environment/environment.service';
 import { buildSystemPrompt } from './ai-chat.prompt';
+import {
+  CORE_TOOL_KEYS,
+  CORE_TOOL_SET,
+  LOAD_TOOLS_NAME,
+  makeLoadToolsTool,
+  buildExternalToolCatalog,
+} from './tools/tool-tiers';
 import { computePageChange } from './page-change/page-change.util';
 import { roleModelOverride } from './roles/role-model-config';
 import {
@@ -54,24 +62,52 @@ const FINAL_STEP_INSTRUCTION =
   'language. If the information is incomplete, say so explicitly: summarize ' +
   'what you found, what is still missing, and give your best partial conclusion.';
 
-// Pure, unit-testable: decide per-step overrides. Returns undefined for normal
+// Pure, unit-testable: decide per-step overrides. Two responsibilities:
-// steps; on the final allowed step forces a text-only synthesis answer.
+//   1. Final-step lockdown (always): on the final allowed step force a text-only
+//      synthesis answer (toolChoice 'none' + FINAL_STEP_INSTRUCTION). This WINS —
+//      it takes precedence over the deferred-tool narrowing below.
+//   2. Deferred tool visibility (#332): when `deferredEnabled` and NOT the final
+//      step, expose only the CORE tools + loadTools + whatever loadTools has
+//      activated so far this turn (`activatedTools`), via `activeTools`. Deferred
+//      tools stay in the <tool_catalog> until the model loads them.
+// When `deferredEnabled` is false the behavior is unchanged: undefined on normal
+// steps (all tools active), lockdown on the final step.
+//
 // `system` is the in-scope system prompt; we CONCATENATE so the original
 // persona/context is preserved — a bare `system` override would REPLACE the
-// whole system prompt for the step.
+// whole system prompt for the step. `activatedTools` is PER-TURN mutable state
+// owned by the streaming loop (a closure Set grown by loadTools); it is passed
+// in (not module-global, not persisted) so this stays a pure function of its
+// arguments.
 //
 // NOTE: at AI SDK v7 the per-step `system` field is renamed to `instructions`.
 // On v6 (`^6.0.134`) `system` is the correct field — adjust when bumping.
 export function prepareAgentStep(
   stepNumber: number,
   system: string,
-): { toolChoice: 'none'; system: string } | undefined {
+  activatedTools: ReadonlySet<string> | readonly string[] = [],
+  deferredEnabled = false,
+):
+  | { toolChoice: 'none'; system: string }
+  | { activeTools: string[] }
+  | undefined {
+  // Final-step lockdown WINS (applies regardless of the deferred toggle).
   if (stepNumber >= MAX_AGENT_STEPS - 1) {
     return {
       toolChoice: 'none',
       system: `${system}\n\n${FINAL_STEP_INSTRUCTION}`,
     };
   }
+  // Deferred tool loading: narrow this step's visible tools to CORE + loadTools
+  // + the tools already activated this turn.
+  if (deferredEnabled) {
+    const activated = Array.isArray(activatedTools)
+      ? activatedTools
+      : [...activatedTools];
+    return {
+      activeTools: [...CORE_TOOL_KEYS, LOAD_TOOLS_NAME, ...activated],
+    };
+  }
   return undefined;
 }
 
@@ -206,6 +242,9 @@ export class AiChatService implements OnModuleInit {
     private readonly aiAgentRoleRepo: AiAgentRoleRepo,
     private readonly pageRepo: PageRepo,
     private readonly pageAccess: PageAccessService,
+    // Reads the AI_CHAT_DEFERRED_TOOLS toggle (#332). Injected last so existing
+    // positional constructor callers (tests) only append one stub.
+    private readonly environment: EnvironmentService,
   ) {}
 
   /**
@@ -625,9 +664,25 @@ export class AiChatService implements OnModuleInit {
     // Build the system prompt + Docmost toolset. If either throws after the
     // external MCP lease was taken above, release the lease before rethrowing so
     // the leased transports are not leaked (#185 review).
+    // Deferred tool loading toggle (#332). When ON, the model sees a compact
+    // <tool_catalog> and only CORE tools + loadTools are active each step; other
+    // tools (fat/rare in-app tools + ALL external MCP tools) load on demand. When
+    // OFF, every tool is active and nothing below changes.
+    const deferredEnabled = this.environment.isAiChatDeferredToolsEnabled();
+
     let system: string;
     let docmostTools: Awaited<ReturnType<AiChatToolsService['forUser']>>;
     try {
+      // Assemble the deferred catalog for the system prompt: hand-written lines
+      // for the in-app deferred tools + a derived line for each external MCP tool
+      // (also deferred by default). Only built when the feature is enabled.
+      const toolCatalog = deferredEnabled
+        ? [
+            ...(await this.tools.getInAppDeferredCatalog()),
+            ...buildExternalToolCatalog(external.tools),
+          ]
+        : [];
+
       system = buildSystemPrompt({
         workspace,
         adminPrompt: resolved?.systemPrompt,
@@ -644,6 +699,10 @@ export class AiChatService implements OnModuleInit {
         // Detected between-turns human edit to the open page (#274): adds the
         // page_changed note + unified diff so the agent doesn't overwrite it.
         pageChanged,
+        // Deferred tool loading (#332): renders the <tool_catalog> block (only
+        // when enabled + non-empty) so the model can activate deferred tools.
+        deferredToolsEnabled: deferredEnabled,
+        toolCatalog,
       });
 
       // Pass the resolved chatId so the write tools can mint provenance tokens
@@ -664,7 +723,31 @@ export class AiChatService implements OnModuleInit {
       throw err;
     }
 
-    const tools = { ...external.tools, ...docmostTools };
+    // Base toolset: external MCP tools + Docmost in-app tools (Docmost wins on a
+    // name clash — external are namespaced, so no clash is expected).
+    const baseTools = { ...external.tools, ...docmostTools };
+
+    // Deferred tool loading state (#332), scoped to THIS streaming loop:
+    //  - `activatedTools` is per-TURN mutable state — a fresh closure Set created
+    //    per streamText call, NOT module-global and NOT persisted, so a new turn
+    //    starts cold. loadTools.execute adds to it; prepareAgentStep reads it to
+    //    widen `activeTools` on the NEXT step.
+    //  - `validDeferredNames` = every tool that is NOT core (the in-app deferred
+    //    tools + ALL external MCP tools), computed from the ACTUAL toolset so an
+    //    external tool is loadable by its namespaced name. loadTools rejects any
+    //    name outside this set.
+    const activatedTools = new Set<string>();
+    const validDeferredNames = new Set<string>(
+      Object.keys(baseTools).filter((k) => !CORE_TOOL_SET.has(k)),
+    );
+    // Add the loadTools meta-tool ONLY when the feature is enabled; when off the
+    // toolset and behavior are exactly as before.
+    const tools = deferredEnabled
+      ? {
+          ...baseTools,
+          [LOAD_TOOLS_NAME]: makeLoadToolsTool(activatedTools, validDeferredNames),
+        }
+      : baseTools;
 
     // Accumulate the turn's streamed output so a provider error / disconnect can
     // persist the PARTIAL answer the user already saw — the SDK's onError/onAbort
@@ -799,7 +882,8 @@ export class AiChatService implements OnModuleInit {
         // ends with no assistant text (an empty turn). prepareAgentStep forbids
         // further tool calls and appends a synthesis instruction on that step,
         // concatenated onto the original `system` so the persona is preserved.
-        prepareStep: ({ stepNumber }) => prepareAgentStep(stepNumber, system),
+        prepareStep: ({ stepNumber }) =>
+          prepareAgentStep(stepNumber, system, activatedTools, deferredEnabled),
         abortSignal: signal,
         onChunk: ({ chunk }) => {
           // DIAGNOSTIC (Safari stream-drop investigation) — temporary. Any model

apps/server/src/core/ai-chat/tools/ai-chat-tools.service.ts

@@ -17,6 +17,10 @@ import { resolveCurrentPageResult } from './current-page.util';
 import { parseNodeArg } from './parse-node-arg';
 import { modelFriendlyInput } from './model-friendly-input';
 import { SandboxStore } from '../../../integrations/sandbox/sandbox.store';
+import {
+  buildInAppDeferredCatalog,
+  type ToolCatalogEntry,
+} from './tool-tiers';
 
 /**
  * Per-user, per-request adapter that exposes Docmost READ operations to the
@@ -123,6 +127,18 @@ export class AiChatToolsService {
     return client.exportPageMarkdown(pageId);
   }
 
+  /**
+   * Build the IN-APP deferred <tool_catalog> entries (#332): one "name — purpose"
+   * line per DEFERRED tool, merging the per-layer INLINE_TOOL_TIERS with the
+   * shared registry's own catalogLine. Loads @docmost/mcp for the shared specs
+   * (memoized). Core tools are always active and are NOT listed here. External
+   * MCP tools are catalogued separately by the caller (they are runtime-scoped).
+   */
+  async getInAppDeferredCatalog(): Promise<ToolCatalogEntry[]> {
+    const { sharedToolSpecs } = await loadDocmostMcp();
+    return buildInAppDeferredCatalog(sharedToolSpecs);
+  }
+
   async forUser(
     user: User,
     sessionId: string,

apps/server/src/core/ai-chat/tools/docmost-client.loader.ts

@@ -241,6 +241,11 @@ export interface SharedToolSpec {
   mcpName: string;
   inAppKey: string;
   description: string;
+  // Deferred-tool metadata (#332). Optional in this mirror so an older/stale
+  // @docmost/mcp build (pre-#332) still type-checks; the in-app catalog builder
+  // reads them defensively. The external /mcp server ignores both fields.
+  tier?: 'core' | 'deferred';
+  catalogLine?: string;
   // Loose `z` on purpose: the registry is zod-agnostic so the server can pass
   // its own zod (v4) and the MCP package its own (v3) into the same builder.
   buildShape?: (z: any) => Record<string, unknown>;

apps/server/src/core/ai-chat/tools/tool-tiers.spec.ts

@@ -0,0 +1,244 @@
+import {
+  CORE_TOOL_KEYS,
+  CORE_TOOL_SET,
+  LOAD_TOOLS_NAME,
+  LOAD_TOOLS_DESCRIPTION,
+  INLINE_TOOL_TIERS,
+  buildInAppDeferredCatalog,
+  buildExternalToolCatalog,
+  shortenForCatalog,
+  applyLoadTools,
+} from './tool-tiers';
+// The real shared registry, imported from source (same approach as the
+// SHARED_TOOL_SPECS contract spec) so the tier metadata is checked against
+// exactly what @docmost/mcp ships.
+import { SHARED_TOOL_SPECS } from '../../../../../../packages/mcp/src/tool-specs';
+// For the live-toolset partition test (F3): the REAL adapter, so the catalog is
+// checked against the tools AiChatToolsService.forUser() actually builds — not a
+// static list that could drift from it.
+import { AiChatToolsService } from './ai-chat-tools.service';
+import * as loader from './docmost-client.loader';
+import type { DocmostClientLike } from './docmost-client.loader';
+
+/**
+ * #332 deferred tool loading — tier metadata, catalog assembly, and the
+ * loadTools meta-tool. Pure units; no Nest graph, no @docmost/mcp build (the
+ * registry is imported from TS source).
+ */
+
+describe('tool tier metadata (#332)', () => {
+  it('core set is the documented 13 + searchInPage (14)', () => {
+    expect(CORE_TOOL_KEYS).toHaveLength(14);
+    expect(CORE_TOOL_SET.has('searchInPage')).toBe(true); // #330, promoted to core
+    // loadTools is a meta-tool, not a normal core key.
+    expect(CORE_TOOL_SET.has(LOAD_TOOLS_NAME)).toBe(false);
+  });
+
+  it('SHARED_TOOL_SPECS tier agrees with CORE_TOOL_SET for every shared tool', () => {
+    for (const [key, spec] of Object.entries(SHARED_TOOL_SPECS)) {
+      const isCoreByTier = spec.tier === 'core';
+      const isCoreByList = CORE_TOOL_SET.has(key);
+      expect(isCoreByTier).toBe(isCoreByList);
+      // Every spec carries a non-empty catalogLine (core tools too).
+      expect(typeof spec.catalogLine).toBe('string');
+      expect(spec.catalogLine.trim().length).toBeGreaterThan(0);
+    }
+  });
+
+  it('every INLINE tool tier agrees with CORE_TOOL_SET and has a catalogLine', () => {
+    for (const [key, meta] of Object.entries(INLINE_TOOL_TIERS)) {
+      expect(meta.tier === 'core').toBe(CORE_TOOL_SET.has(key));
+      expect(meta.catalogLine.trim().length).toBeGreaterThan(0);
+    }
+  });
+});
+
+describe('buildInAppDeferredCatalog (#332)', () => {
+  const catalog = buildInAppDeferredCatalog(SHARED_TOOL_SPECS as never);
+  const names = catalog.map((e) => e.name);
+
+  it('includes deferred tools from BOTH the inline map and the shared registry', () => {
+    expect(names).toContain('transformPage'); // inline deferred
+    expect(names).toContain('getPageJson'); // shared deferred
+    expect(names).toContain('patchNode'); // shared deferred
+    expect(names).toContain('createPage'); // inline deferred
+  });
+
+  it('NEVER lists a core tool', () => {
+    for (const core of CORE_TOOL_KEYS) {
+      expect(names).not.toContain(core);
+    }
+    // spot-check a couple that are core in each source.
+    expect(names).not.toContain('searchInPage'); // shared core
+    expect(names).not.toContain('searchPages'); // inline core
+    expect(names).not.toContain('editPageText'); // shared core
+  });
+
+  it('renders every entry as a "name — purpose" line', () => {
+    // Non-empty catalog (the length is pinned structurally by the live-toolset
+    // partition test below, not by a magic constant that rots on every new tool).
+    expect(catalog.length).toBeGreaterThan(0);
+    for (const entry of catalog) {
+      expect(entry.catalogLine).toMatch(/ — /);
+    }
+  });
+});
+
+/**
+ * F3 — the deferred <tool_catalog> is built from STATIC metadata (INLINE_TOOL_TIERS
+ * + SHARED_TOOL_SPECS), but the loadable-by-name set is derived at RUNTIME from the
+ * actual toolset (`Object.keys(baseTools)` in ai-chat.service.ts). Those two must
+ * agree or a tool becomes loadable-but-invisible (agent thinks it doesn't exist) or
+ * catalogued-but-phantom. INLINE_TOOL_TIERS is a plain hand-maintained Record with
+ * no compile-time link to the tools AiChatToolsService.forUser() builds, so nothing
+ * else catches that drift. This test uses forUser()'s LIVE keys as the source of
+ * truth (mirroring ai-chat-tools.service.spec.ts's loader mock) and asserts a
+ * two-way partition against buildInAppDeferredCatalog — replacing the old magic
+ * toHaveLength(28), so a tool added to forUser() without a catalog line (or a
+ * catalog line without a real tool) fails the suite instead of silently vanishing.
+ */
+describe('deferred catalog ↔ live forUser() toolset partition (#332, F3)', () => {
+  let toolKeys: string[];
+  const catalogNames = buildInAppDeferredCatalog(SHARED_TOOL_SPECS as never).map(
+    (e) => e.name,
+  );
+
+  beforeAll(async () => {
+    // Intercept the ESM loader so forUser() builds against the TS-source shared
+    // specs (no @docmost/mcp build) and never touches the network.
+    jest.spyOn(loader, 'loadDocmostMcp').mockResolvedValue({
+      DocmostClient: function () {
+        return {} as DocmostClientLike;
+      } as unknown as loader.DocmostClientCtor,
+      sharedToolSpecs: SHARED_TOOL_SPECS as Record<string, loader.SharedToolSpec>,
+    });
+    const service = new AiChatToolsService(
+      {
+        generateAccessToken: jest.fn().mockResolvedValue('access-token'),
+        generateCollabToken: jest.fn().mockResolvedValue('collab-token'),
+      } as never,
+      {} as never, // aiService — not exercised while merely BUILDING the tools
+      {} as never, // pageEmbeddingRepo
+      {} as never, // spaceMemberRepo
+      {} as never, // pagePermissionRepo
+      // sandboxStore: forUser() eagerly calls asSink() to wire the stash tool.
+      {
+        asSink: () => ({ put: jest.fn(), has: jest.fn(), evict: jest.fn() }),
+      } as never,
+    );
+    const tools = await service.forUser(
+      { id: 'user-1', email: 'u@example.com', workspaceId: 'ws-1' } as never,
+      'session-1',
+      'ws-1',
+      'chat-1',
+    );
+    toolKeys = Object.keys(tools);
+  });
+
+  afterAll(() => {
+    jest.restoreAllMocks();
+  });
+
+  it('exposes a non-trivial toolset (sanity: the mock actually built tools)', () => {
+    expect(toolKeys.length).toBeGreaterThan(20);
+  });
+
+  it('every non-core live tool is present in the catalog (no capability silently hidden)', () => {
+    // forUser() does not itself add loadTools (ai-chat.service does), but guard
+    // anyway. Every remaining non-core key MUST have a catalog line.
+    const catalogSet = new Set(catalogNames);
+    const missing = toolKeys.filter(
+      (k) => !CORE_TOOL_SET.has(k) && k !== LOAD_TOOLS_NAME && !catalogSet.has(k),
+    );
+    expect(missing).toEqual([]);
+  });
+
+  it('every catalog entry corresponds to a real, non-core live tool (no phantom)', () => {
+    const liveSet = new Set(toolKeys);
+    const phantom = catalogNames.filter(
+      (n) => !liveSet.has(n) || CORE_TOOL_SET.has(n),
+    );
+    expect(phantom).toEqual([]);
+  });
+});
+
+describe('buildExternalToolCatalog + shortenForCatalog (#332)', () => {
+  it('derives a short "name — purpose" line from each external tool description', () => {
+    const catalog = buildExternalToolCatalog({
+      tavily_search: { description: 'Search the web for fresh results. More detail here.' },
+      tavily_extract: { description: '' },
+    });
+    expect(catalog).toEqual([
+      { name: 'tavily_search', catalogLine: 'tavily_search — Search the web for fresh results.' },
+      { name: 'tavily_extract', catalogLine: 'tavily_extract — external tool' },
+    ]);
+  });
+
+  it('caps a very long description', () => {
+    const long = 'x'.repeat(500);
+    expect(shortenForCatalog(long).length).toBeLessThanOrEqual(140);
+    expect(shortenForCatalog(long).endsWith('…')).toBe(true);
+  });
+});
+
+describe('applyLoadTools (#332)', () => {
+  const valid = new Set(['createPage', 'transformPage', 'tavily_search']);
+
+  it('adds valid names to the activated set and returns { loaded }', () => {
+    const activated = new Set<string>();
+    const result = applyLoadTools(['createPage', 'tavily_search'], activated, valid);
+    expect(result).toEqual({ loaded: ['createPage', 'tavily_search'] });
+    expect(activated.has('createPage')).toBe(true);
+    expect(activated.has('tavily_search')).toBe(true);
+  });
+
+  it('rejects an unknown name with an error listing the valid deferred names', () => {
+    const activated = new Set<string>();
+    expect(() => applyLoadTools(['nope'], activated, valid)).toThrow(/unknown tool name/i);
+    try {
+      applyLoadTools(['nope'], activated, valid);
+    } catch (e) {
+      const msg = (e as Error).message;
+      // Lists every valid name (sorted).
+      expect(msg).toContain('createPage');
+      expect(msg).toContain('transformPage');
+      expect(msg).toContain('tavily_search');
+    }
+    // Nothing is activated on a rejected call.
+    expect(activated.size).toBe(0);
+  });
+
+  it('tolerates a non-array / empty input (loads nothing)', () => {
+    const activated = new Set<string>();
+    expect(applyLoadTools(undefined, activated, valid)).toEqual({ loaded: [] });
+    expect(applyLoadTools([], activated, valid)).toEqual({ loaded: [] });
+    expect(activated.size).toBe(0);
+  });
+
+  it('loadTools description is the verbatim issue text', () => {
+    expect(LOAD_TOOLS_DESCRIPTION).toContain('only ACTIVATES them');
+    expect(LOAD_TOOLS_DESCRIPTION).toContain('callable on your NEXT step');
+  });
+});
+
+describe('editorial "Corrector" scenario is fully served by CORE (#332)', () => {
+  it('read + comment + edit + search need no loadTools', () => {
+    // A Corrector role reads a page, searches within it, edits text, and leaves
+    // inline comments — every tool it needs is core, so it never has to load a
+    // deferred tool.
+    const needed = [
+      'getCurrentPage',
+      'getPage',
+      'searchPages',
+      'searchInPage',
+      'editPageText',
+      'createComment',
+      'listComments',
+      'getComment',
+      'resolveComment',
+    ];
+    for (const t of needed) {
+      expect(CORE_TOOL_SET.has(t)).toBe(true);
+    }
+  });
+});

apps/server/src/core/ai-chat/tools/tool-tiers.ts

@@ -0,0 +1,309 @@
+import { tool, type Tool } from 'ai';
+import { z } from 'zod';
+import type { SharedToolSpec } from './docmost-client.loader';
+
+/**
+ * Deferred tool loading for the in-app AI chat (#332).
+ *
+ * The agent otherwise sends ALL ~41 tool definitions on EVERY model call every
+ * step, bloating context. Instead we split the in-app tools into two tiers:
+ *
+ *  - CORE (hot, always active): frequent OR tiny tools whose full schema is
+ *    always visible, plus the `loadTools` meta-tool. Deferring a one-line tool is
+ *    pure loss, so tiny tools stay core even if rare.
+ *  - DEFERRED (loaded on demand): the fat/rare tools + ALL external MCP tools by
+ *    default. The model sees only a compact <tool_catalog> (name — purpose) and
+ *    calls `loadTools(names)` to ACTIVATE a tool's full schema for the NEXT step
+ *    (one extra round-trip on first use).
+ *
+ * This module is the single source of truth for the IN-APP tiering:
+ *  - CORE_TOOL_KEYS / CORE_TOOL_SET — the authoritative core list (used by
+ *    prepareAgentStep to build per-step `activeTools`).
+ *  - INLINE_TOOL_TIERS — tier + catalogLine for the per-layer INLINE tools (the
+ *    ones NOT in @docmost/mcp's SHARED_TOOL_SPECS, which carry their own).
+ *  - buildInAppDeferredCatalog / buildExternalToolCatalog — assemble the
+ *    <tool_catalog> deferred lines.
+ *  - applyLoadTools / makeLoadToolsTool — the loadTools meta-tool.
+ *
+ * The tier/catalogLine fields on SHARED_TOOL_SPECS are IN-APP metadata only; the
+ * external /mcp server ignores them and exposes every tool normally.
+ */
+
+/** A single rendered <tool_catalog> line: the tool name + its "name — purpose". */
+export interface ToolCatalogEntry {
+  /** Exact tool name the model must pass to loadTools. */
+  name: string;
+  /** Hand-written (in-app) or derived (external) "name — purpose" line. */
+  catalogLine: string;
+}
+
+/**
+ * CORE (always-active) in-app tool keys — 13 frequent/tiny tools. `searchInPage`
+ * (#330) is added to core on top of the issue's original tier list: it is
+ * frequent for the editorial roles this feature targets. `loadTools` is active
+ * too but is not a normal tool key (it is added to activeTools separately).
+ */
+export const CORE_TOOL_KEYS = [
+  'searchPages',
+  'listPages',
+  'listSpaces',
+  'getWorkspace',
+  'getCurrentPage',
+  'getPage',
+  'getOutline',
+  'getNode',
+  'createComment',
+  'getComment',
+  'listComments',
+  'resolveComment',
+  'editPageText',
+  // #330 search_in_page — frequent for editorial sweeps; core despite predating
+  // the issue's tier list.
+  'searchInPage',
+] as const;
+
+/** O(1) membership test for the core tier. */
+export const CORE_TOOL_SET: ReadonlySet<string> = new Set(CORE_TOOL_KEYS);
+
+/** The meta-tool name (always active alongside the core tools when enabled). */
+export const LOAD_TOOLS_NAME = 'loadTools';
+
+/**
+ * loadTools description — VERBATIM from issue #332. Tells the model that the
+ * catalog names EXIST, that loadTools only ACTIVATES them (callable next step),
+ * and to load several at once.
+ */
+export const LOAD_TOOLS_DESCRIPTION =
+  'loadTools — Load the full definitions of deferred tools from the <tool_catalog>\n' +
+  'block in your instructions. Pass the EXACT tool names from the catalog; this\n' +
+  'call only ACTIVATES them and returns { loaded: [...] } — the tools become\n' +
+  'callable on your NEXT step. Load several names in one call when the task clearly\n' +
+  'needs them. Unknown names are rejected with the list of valid ones.';
+
+/**
+ * Tier + catalogLine for the INLINE ai-chat tools — those defined per-layer in
+ * ai-chat-tools.service.ts and NOT present in @docmost/mcp's SHARED_TOOL_SPECS
+ * (which carries its own tier/catalogLine). Together with the shared registry
+ * this describes every in-app tool. catalogLine is present for core tools too
+ * (uniformity), but only DEFERRED tools are rendered into the catalog.
+ */
+export const INLINE_TOOL_TIERS: Record<
+  string,
+  { tier: 'core' | 'deferred'; catalogLine: string }
+> = {
+  // --- core inline ---
+  searchPages: {
+    tier: 'core',
+    catalogLine: 'searchPages — hybrid semantic + keyword search across the wiki.',
+  },
+  getCurrentPage: {
+    tier: 'core',
+    catalogLine: 'getCurrentPage — the page the user is currently viewing.',
+  },
+  getPage: {
+    tier: 'core',
+    catalogLine: 'getPage — fetch a page as Markdown by its id.',
+  },
+  listPages: {
+    tier: 'core',
+    catalogLine: "listPages — list recent pages, or a space's full page tree.",
+  },
+  listComments: {
+    tier: 'core',
+    catalogLine: 'listComments — list all comments on a page (including resolved).',
+  },
+  getComment: {
+    tier: 'core',
+    catalogLine: 'getComment — fetch a single comment by id.',
+  },
+  createComment: {
+    tier: 'core',
+    catalogLine:
+      'createComment — add an inline comment (optionally with a suggested edit).',
+  },
+  resolveComment: {
+    tier: 'core',
+    catalogLine: 'resolveComment — resolve or reopen a comment thread.',
+  },
+
+  // --- deferred inline ---
+  createPage: {
+    tier: 'deferred',
+    catalogLine: 'createPage — create a new page with a Markdown body in a space.',
+  },
+  updatePageContent: {
+    tier: 'deferred',
+    catalogLine:
+      "updatePageContent — replace a page's body (and optionally title) with new Markdown.",
+  },
+  renamePage: {
+    tier: 'deferred',
+    catalogLine: "renamePage — change a page's title only (body untouched).",
+  },
+  movePage: {
+    tier: 'deferred',
+    catalogLine: 'movePage — move a page under a new parent or to the space root.',
+  },
+  deletePage: {
+    tier: 'deferred',
+    catalogLine: 'deletePage — move a page to trash (soft delete, reversible).',
+  },
+  listSidebarPages: {
+    tier: 'deferred',
+    catalogLine:
+      "listSidebarPages — list a space's root pages or a page's direct children.",
+  },
+  getTable: {
+    tier: 'deferred',
+    catalogLine: 'getTable — read a table as a matrix of cell texts and cell ids.',
+  },
+  checkNewComments: {
+    tier: 'deferred',
+    catalogLine:
+      'checkNewComments — find comments in a space created after a timestamp.',
+  },
+  getPageHistory: {
+    tier: 'deferred',
+    catalogLine:
+      'getPageHistory — fetch one page-history version with its ProseMirror content.',
+  },
+  exportPageMarkdown: {
+    tier: 'deferred',
+    catalogLine:
+      'exportPageMarkdown — export a page to self-contained Markdown (body + comments).',
+  },
+  updatePageJson: {
+    tier: 'deferred',
+    catalogLine:
+      "updatePageJson — overwrite a page's body with a full ProseMirror document.",
+  },
+  tableInsertRow: {
+    tier: 'deferred',
+    catalogLine: 'tableInsertRow — insert a row of plain-text cells into a table.',
+  },
+  tableDeleteRow: {
+    tier: 'deferred',
+    catalogLine: 'tableDeleteRow — delete a table row at a 0-based index.',
+  },
+  tableUpdateCell: {
+    tier: 'deferred',
+    catalogLine: 'tableUpdateCell — set the text of a table cell at [row, col].',
+  },
+  sharePage: {
+    tier: 'deferred',
+    catalogLine: 'sharePage — make a page publicly accessible and return its URL.',
+  },
+  transformPage: {
+    tier: 'deferred',
+    catalogLine: "transformPage — run a sandboxed JS transform over a page's document.",
+  },
+};
+
+/**
+ * Build the <tool_catalog> deferred lines for the IN-APP tools by merging the
+ * two metadata sources: the per-layer INLINE_TOOL_TIERS and the shared registry
+ * (SHARED_TOOL_SPECS, loaded at runtime). Only DEFERRED tools are included; core
+ * tools are always active and never appear in the catalog. Pure — the caller
+ * passes the loaded specs so this stays unit-testable.
+ */
+export function buildInAppDeferredCatalog(
+  sharedToolSpecs: Record<string, SharedToolSpec>,
+): ToolCatalogEntry[] {
+  const entries: ToolCatalogEntry[] = [];
+  // Inline deferred tools (hand-written lines).
+  for (const [name, meta] of Object.entries(INLINE_TOOL_TIERS)) {
+    if (meta.tier === 'deferred') {
+      entries.push({ name, catalogLine: meta.catalogLine });
+    }
+  }
+  // Shared deferred tools (line comes from the registry's own catalogLine).
+  for (const [name, spec] of Object.entries(sharedToolSpecs)) {
+    if (spec.tier === 'deferred' && spec.catalogLine) {
+      entries.push({ name, catalogLine: spec.catalogLine });
+    }
+  }
+  return entries;
+}
+
+/**
+ * Cap an external tool's (untrusted) description into a short catalog purpose.
+ * External MCP tools have no hand-written catalogLine, so we derive one from the
+ * first sentence of the description, hard-capped. Whitespace is collapsed.
+ */
+export function shortenForCatalog(description: string, max = 140): string {
+  const flat = description.replace(/\s+/g, ' ').trim();
+  if (!flat) return 'external tool';
+  // Prefer the first sentence if it is reasonably short.
+  const firstSentence = flat.split(/(?<=[.!?])\s/)[0];
+  const base =
+    firstSentence.length > 0 && firstSentence.length <= max
+      ? firstSentence
+      : flat;
+  return base.length > max ? `${base.slice(0, max - 1).trimEnd()}…` : base;
+}
+
+/**
+ * Build catalog lines for the EXTERNAL MCP tools (all deferred by default,
+ * #332). Their names are the namespaced tool keys; the purpose is derived from
+ * each tool's own description (no hand-written line exists). Pure.
+ */
+export function buildExternalToolCatalog(
+  externalTools: Record<string, { description?: string } | undefined>,
+): ToolCatalogEntry[] {
+  return Object.entries(externalTools).map(([name, t]) => ({
+    name,
+    catalogLine: `${name} — ${shortenForCatalog(t?.description ?? '')}`,
+  }));
+}
+
+/**
+ * Pure core of the loadTools meta-tool. Validates the requested names against
+ * the per-turn set of valid deferred names, ADDS the valid ones to the caller's
+ * mutable `activatedTools` set (so they become callable next step), and returns
+ * `{ loaded }`. An unknown name throws a clear error listing the valid deferred
+ * names — surfaced to the model as a tool error so it can retry.
+ */
+export function applyLoadTools(
+  names: unknown,
+  activatedTools: Set<string>,
+  validDeferredNames: ReadonlySet<string>,
+): { loaded: string[] } {
+  const requested = Array.isArray(names)
+    ? names.filter((n): n is string => typeof n === 'string')
+    : [];
+  const unknown = requested.filter((n) => !validDeferredNames.has(n));
+  if (unknown.length > 0) {
+    const valid = [...validDeferredNames].sort().join(', ');
+    throw new Error(
+      `loadTools: unknown tool name(s): ${unknown.join(', ')}. ` +
+        `Valid deferred tools are: ${valid || '(none)'}.`,
+    );
+  }
+  for (const n of requested) activatedTools.add(n);
+  return { loaded: requested };
+}
+
+/**
+ * Build the loadTools AI-SDK tool bound to THIS turn's mutable state: the
+ * `activatedTools` set (grown by execute, read by prepareAgentStep next step)
+ * and the `validDeferredNames` set (every non-core tool in this turn's toolset,
+ * incl. external MCP). Created per streamText call — never module-global.
+ */
+export function makeLoadToolsTool(
+  activatedTools: Set<string>,
+  validDeferredNames: ReadonlySet<string>,
+): Tool {
+  return tool({
+    description: LOAD_TOOLS_DESCRIPTION,
+    inputSchema: z.object({
+      names: z
+        .array(z.string())
+        .describe(
+          'EXACT deferred tool names from the <tool_catalog> to activate for ' +
+            'your next step.',
+        ),
+    }),
+    execute: async ({ names }) =>
+      applyLoadTools(names, activatedTools, validDeferredNames),
+  });
+}

apps/server/src/integrations/environment/environment.service.ts

@@ -261,6 +261,21 @@ export class EnvironmentService {
     return disable === 'true';
   }
 
+  /**
+   * Deferred tool loading for the in-app AI chat (#332). When enabled, the agent
+   * sees a compact <tool_catalog> and only CORE tools + the loadTools meta-tool
+   * are active each step; deferred tools (the fat/rare ones + all external MCP
+   * tools) load on demand. Defaults to ENABLED — the issue treats deferred
+   * loading as the new behavior; set AI_CHAT_DEFERRED_TOOLS=false to restore the
+   * old "all tools always active" behavior.
+   */
+  isAiChatDeferredToolsEnabled(): boolean {
+    const enabled = this.configService
+      .get<string>('AI_CHAT_DEFERRED_TOOLS', 'true')
+      .toLowerCase();
+    return enabled === 'true';
+  }
+
   getPostHogHost(): string {
     return this.configService.get<string>('POSTHOG_HOST');
   }

apps/server/test/integration/ai-chat-stream.int-spec.ts

@@ -1,5 +1,7 @@
 import * as http from 'node:http';
 import { Kysely } from 'kysely';
+import { tool } from 'ai';
+import { z } from 'zod';
 import { MockLanguageModelV3, convertArrayToReadableStream } from 'ai/test';
 import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
 import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
@@ -146,6 +148,9 @@ describe('AiChatService.stream [integration]', () => {
       {} as any, // aiAgentRoleRepo (role is pre-resolved + passed in)
       {} as any, // pageRepo (only used when body.openPage is set)
       {} as any, // pageAccess (idem)
+      // environment (#332): keep deferred tool loading OFF for this lifecycle
+      // harness so the toolset/behavior is exactly as before.
+      { isAiChatDeferredToolsEnabled: () => false } as any,
     );
   }
 
@@ -315,4 +320,174 @@ describe('AiChatService.stream [integration]', () => {
       true,
     );
   });
+
+  /**
+   * #332 deferred tool loading, the ON path. The riskiest property is that the
+   * per-turn `activatedTools` Set is created FRESH inside each stream() call, so a
+   * tool a previous turn activated via loadTools is NOT still active when the next
+   * turn starts — the new turn begins "cold" (CORE + loadTools only). The unit
+   * tests only exercise pure prepareAgentStep with hand-fed Sets; this pins the
+   * real wiring end-to-end (loadTools.execute -> activatedTools -> prepareStep ->
+   * per-step activeTools) against the real streamText loop, and proves there is no
+   * cross-turn leak. We drive a MockLanguageModelV3 whose step 1 calls
+   * loadTools(['createPage']) and assert, via the model's recorded per-step
+   * CallOptions.tools (the AI SDK filters the provider tool list by activeTools),
+   * that the deferred tool becomes active on the SAME turn's next step but NOT on a
+   * fresh turn's first step.
+   */
+  describe('deferred tool loading ON — per-turn activation, no leak (#332)', () => {
+    // A stub deferred (non-core) tool the agent can activate. Its execute is never
+    // called — the model only needs to SEE it become active — but it must be a
+    // valid AI-SDK tool so the SDK includes it in a step's tool list once active.
+    const createPageStub = tool({
+      description: 'create a new page',
+      inputSchema: z.object({ title: z.string() }),
+      execute: async () => ({ id: 'p-stub' }),
+    });
+
+    // A CORE tool in the toolset, so a cold step shows CORE tools ARE active while
+    // the deferred createPage is not. `searchPages` is in CORE_TOOL_SET.
+    const searchPagesStub = tool({
+      description: 'search the wiki',
+      inputSchema: z.object({ query: z.string() }),
+      execute: async () => [],
+    });
+
+    // Same lifecycle harness as buildService() above, but with deferred loading ON
+    // and a toolset that exposes exactly one deferred tool (createPage) so it is
+    // catalogued + loadable-by-name. Kept separate so the OFF scenarios are
+    // untouched.
+    function buildDeferredService(): AiChatService {
+      return new AiChatService(
+        { getChatModel: async () => null } as any,
+        aiChatRepo,
+        msgRepo,
+        {} as any,
+        { resolve: async () => null } as any,
+        {
+          forUser: async () => ({
+            searchPages: searchPagesStub,
+            createPage: createPageStub,
+          }),
+          getInAppDeferredCatalog: async () => [
+            { name: 'createPage', catalogLine: 'createPage — create a new page.' },
+          ],
+        } as any,
+        mcpClients as any,
+        {} as any,
+        {} as any,
+        {} as any,
+        // #332: deferred tool loading ON — the property under test.
+        { isAiChatDeferredToolsEnabled: () => true } as any,
+      );
+    }
+
+    // Drive ONE stream() turn against `model` and wait for the assistant row to
+    // settle (mirrors runStream, but builds the deferred-ON service).
+    async function runDeferredTurn(
+      model: MockLanguageModelV3,
+      chatId: string,
+      body: any,
+    ): Promise<void> {
+      closeCalls = 0;
+      const service = buildDeferredService();
+      const { res, cleanup } = await makeRealResponse();
+      try {
+        await service.stream({
+          user: { id: userId, workspaceId } as any,
+          workspace: { id: workspaceId, name: 'WS' } as any,
+          sessionId: 'sess-1',
+          body,
+          res: { raw: res } as any,
+          signal: new AbortController().signal,
+          model: model as any,
+          role: null,
+        } as any);
+        await waitFor(async () => {
+          const rows = await msgRepo.findAllByChat(chatId, workspaceId);
+          return rows.some(
+            (r) =>
+              r.role === 'assistant' &&
+              ['completed', 'error', 'aborted'].includes(r.status as string),
+          );
+        });
+        await waitFor(() => closeCalls > 0, { timeoutMs: 5_000 });
+      } finally {
+        await cleanup();
+      }
+    }
+
+    // Tool names the provider actually received for a recorded step (activeTools
+    // filters this list, so it reflects what was active that step).
+    const toolNames = (call: any): string[] =>
+      ((call?.tools ?? []) as any[]).map((t) => t?.name).filter(Boolean);
+
+    // A model that, on step 1, calls loadTools(['createPage']); on step 2, answers.
+    function loadThenAnswerModel(): MockLanguageModelV3 {
+      let step = 0;
+      return new MockLanguageModelV3({
+        doStream: async () => {
+          const n = step++;
+          if (n === 0) {
+            return {
+              stream: convertArrayToReadableStream([
+                { type: 'stream-start', warnings: [] },
+                {
+                  type: 'tool-call',
+                  toolCallId: 'lt1',
+                  toolName: 'loadTools',
+                  input: JSON.stringify({ names: ['createPage'] }),
+                },
+                {
+                  type: 'finish',
+                  finishReason: 'tool-calls',
+                  usage: { inputTokens: 5, outputTokens: 3, totalTokens: 8 },
+                },
+              ] as any),
+            };
+          }
+          return { stream: successStream() };
+        },
+      } as any);
+    }
+
+    it('activates a deferred tool for the SAME turn, and a NEW turn starts cold (no leak)', async () => {
+      const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+
+      // --- Turn 1: loadTools(createPage) on step 1, then answer on step 2. ---
+      const model1 = loadThenAnswerModel();
+      await runDeferredTurn(model1, chatId, {
+        chatId,
+        messages: [userUiMessage('Make me a page')],
+      });
+
+      // The turn ran at least two steps (the load round-trip + the answer).
+      expect(model1.doStreamCalls.length).toBeGreaterThanOrEqual(2);
+      const step1Tools = toolNames(model1.doStreamCalls[0]);
+      const step2Tools = toolNames(model1.doStreamCalls[1]);
+
+      // Step 1 starts cold: CORE tools + the loadTools meta-tool are active, but
+      // the deferred createPage is NOT yet.
+      expect(step1Tools).toContain('loadTools');
+      expect(step1Tools).toContain('searchPages'); // a CORE tool, always active
+      expect(step1Tools).not.toContain('createPage');
+      // Step 2 of the SAME turn sees the just-activated deferred tool.
+      expect(step2Tools).toContain('createPage');
+
+      // --- Turn 2 on the SAME chat: must start cold again. ---
+      const model2 = new MockLanguageModelV3({
+        doStream: async () => ({ stream: successStream() }),
+      } as any);
+      await runDeferredTurn(model2, chatId, {
+        chatId,
+        messages: [userUiMessage('And another thing')],
+      });
+
+      const nextTurnFirstStep = toolNames(model2.doStreamCalls[0]);
+      expect(nextTurnFirstStep).toContain('loadTools');
+      // The activated set is per-turn: the prior turn's createPage did NOT leak,
+      // so the fresh turn's first step sees it deferred again.
+      expect(nextTurnFirstStep).not.toContain('createPage');
+    });
+  });
 });

packages/mcp/src/tool-specs.ts

@@ -31,6 +31,22 @@ export interface SharedToolSpec {
   inAppKey: string;
   /** Single canonical model-facing description used by both layers. */
   description: string;
+  /**
+   * Deferred-tool tier for the IN-APP agent (#332). 'core' tools are always
+   * active; 'deferred' tools are hidden behind the <tool_catalog> and loaded on
+   * demand via the loadTools meta-tool. This is an IN-APP concern only: the
+   * standalone /mcp server ignores this field and registers every tool normally
+   * (registerShared in index.ts reads mcpName/description/buildShape only).
+   */
+  tier: 'core' | 'deferred';
+  /**
+   * Hand-written one-liner "name — purpose" shown in the in-app agent's
+   * <tool_catalog> for a DEFERRED tool (#332). Deliberately NOT derived from the
+   * description's first sentence — a concise, accurate purpose line. Present on
+   * every spec (core tools too) for uniformity; only deferred ones are rendered.
+   * Inert for the external /mcp server.
+   */
+  catalogLine: string;
   /**
    * Builds the tool's input schema as a plain object of zod fields (a
    * ZodRawShape). Called with the consumer's own zod namespace. Omitted for
@@ -47,6 +63,8 @@ export const SHARED_TOOL_SPECS = {
     mcpName: 'get_workspace',
     inAppKey: 'getWorkspace',
     description: 'Fetch metadata about the current workspace (name, settings).',
+    tier: 'core',
+    catalogLine: 'getWorkspace — fetch current workspace metadata (name, settings).',
   },
 
   listSpaces: {
@@ -55,6 +73,8 @@ export const SHARED_TOOL_SPECS = {
     description:
       'List the spaces the current user can access. Returns the array of ' +
       'spaces (id, name, slug, ...).',
+    tier: 'core',
+    catalogLine: 'listSpaces — list the spaces the user can access (id, name, slug).',
   },
 
   listShares: {
@@ -62,6 +82,8 @@ export const SHARED_TOOL_SPECS = {
     inAppKey: 'listShares',
     description:
       'List all public shares in the workspace with page titles and public URLs.',
+    tier: 'deferred',
+    catalogLine: 'listShares — list all public shares in the workspace with their URLs.',
   },
 
   // --- single-pageId read tools ---
@@ -74,6 +96,9 @@ export const SHARED_TOOL_SPECS = {
       'includes block ids, callouts, tables, link/image attributes) plus the ' +
       'slugId used in URLs. Use the block ids it returns to make precise ' +
       'structural edits or surgical text edits without resending the page.',
+    tier: 'deferred',
+    catalogLine:
+      "getPageJson — get a page's raw ProseMirror JSON (lossless, with block ids).",
     buildShape: (z) => ({
       pageId: z.string().min(1),
     }),
@@ -88,6 +113,9 @@ export const SHARED_TOOL_SPECS = {
       'count) WITHOUT the full document body. Use it to locate sections/tables ' +
       'and grab block ids cheaply before fetching, patching or inserting ' +
       'individual blocks.',
+    tier: 'core',
+    catalogLine:
+      "getOutline — compact outline of a page's top-level blocks with their ids.",
     buildShape: (z) => ({
       pageId: z.string().min(1),
     }),
@@ -104,6 +132,9 @@ export const SHARED_TOOL_SPECS = {
       'outline or page-JSON view (works for headings/paragraphs/callouts/images), OR ' +
       '`#<index>` to fetch a top-level block by its outline index — use the ' +
       '`#<index>` form for tables/rows/cells, which carry no id.',
+    tier: 'core',
+    catalogLine:
+      "getNode — fetch one block's ProseMirror subtree by block id or #index.",
     buildShape: (z) => ({
       pageId: z.string().min(1),
       nodeId: z.string().min(1),
@@ -137,6 +168,9 @@ export const SHARED_TOOL_SPECS = {
       'caseSensitive:true to match case. Ideal for systematic ' +
       'editorial sweeps (unquoted "ё", straight quotes, "т.е.", stray units). An ' +
       'invalid regex or an empty query returns a clear error to fix.',
+    tier: 'core',
+    catalogLine:
+      'searchInPage — find every occurrence of a string/regex inside one page, with locations.',
     buildShape: (z) => ({
       pageId: z.string().min(1).describe('ID of the page to search'),
       query: z
@@ -172,6 +206,8 @@ export const SHARED_TOOL_SPECS = {
     description:
       'Remove a single block by its attrs.id (from the page outline or ' +
       'page-JSON view) WITHOUT resending the whole document.',
+    tier: 'deferred',
+    catalogLine: 'deleteNode — remove a single content block by its block id.',
     buildShape: (z) => ({
       pageId: z.string().min(1),
       nodeId: z.string().min(1),
@@ -203,6 +239,9 @@ export const SHARED_TOOL_SPECS = {
       'JSON object or a JSON string (both accepted). Cheaper and safer than ' +
       'replacing the whole document for one-block structural edits. Reversible: ' +
       'the previous version is kept in page history.',
+    tier: 'deferred',
+    catalogLine:
+      'patchNode — replace one block with a new ProseMirror node, keeping its id.',
     buildShape: (z) => ({
       pageId: z.string().min(1).describe('ID of the page containing the block'),
       nodeId: z
@@ -245,6 +284,9 @@ export const SHARED_TOOL_SPECS = {
       '[{"type":"text","text":"Title"}]}. Bold is a mark: ' +
       '{"type":"text","text":"x","marks":[{"type":"bold"}]}. The node may be a ' +
       'JSON object or a JSON string (both accepted). Reversible via page history.',
+    tier: 'deferred',
+    catalogLine:
+      'insertNode — insert a block before/after an anchor, or append at the end.',
     buildShape: (z) => ({
       pageId: z.string().min(1),
       node: z
@@ -278,6 +320,8 @@ export const SHARED_TOOL_SPECS = {
     mcpName: 'unshare_page',
     inAppKey: 'unsharePage',
     description: 'Remove the public share of a page (revokes the public URL).',
+    tier: 'deferred',
+    catalogLine: "unsharePage — revoke a page's public share (removes the public URL).",
     buildShape: (z) => ({
       pageId: z.string().min(1).describe('ID of the page to unshare'),
     }),
@@ -295,6 +339,9 @@ export const SHARED_TOOL_SPECS = {
       "`from`/`to` each accept a historyId, or null/'current' for the page's " +
       'current content (defaults: from=current, to=current — pass a historyId ' +
       'from the page-history list to compare against the live page).',
+    tier: 'deferred',
+    catalogLine:
+      'diffPageVersions — diff two page versions and return the change set + summary.',
     buildShape: (z) => ({
       pageId: z.string().min(1),
       from: z
@@ -315,6 +362,9 @@ export const SHARED_TOOL_SPECS = {
       "List a page's saved versions (Docmost auto-snapshots on every save), " +
       'newest first, cursor-paginated. Returns { items, nextCursor }; each ' +
       "item's id is the historyId to pass to the page diff or restore tools.",
+    tier: 'deferred',
+    catalogLine:
+      "listPageHistory — list a page's saved versions (newest first, paginated).",
     buildShape: (z) => ({
       pageId: z.string().min(1),
       cursor: z
@@ -332,6 +382,9 @@ export const SHARED_TOOL_SPECS = {
       'as the page\'s current content (Docmost has no restore endpoint, so ' +
       'this creates a NEW history snapshot — the restore is itself revertible). ' +
       'Get the historyId from the page-history list.',
+    tier: 'deferred',
+    catalogLine:
+      'restorePageVersion — restore a page to a saved history version (revertible).',
     buildShape: (z) => ({
       historyId: z.string().min(1),
     }),
@@ -349,6 +402,9 @@ export const SHARED_TOOL_SPECS = {
       'thread records are NOT created/updated/deleted on the server by this ' +
       'tool — only the page body + inline comment marks are written; manage ' +
       'comment threads via the comment tools/UI.',
+    tier: 'deferred',
+    catalogLine:
+      "importPageMarkdown — replace a page's content from exported Docmost Markdown.",
     buildShape: (z) => ({
       pageId: z.string().min(1),
       markdown: z.string().min(1),
@@ -365,6 +421,9 @@ export const SHARED_TOOL_SPECS = {
       'entirely server-side — the document is NOT sent through the model. The ' +
       'target keeps its own title and slug; only its body is replaced. Ideal ' +
       "for 'make page A's content equal to B' or 'replace A with B but keep A's URL'.",
+    tier: 'deferred',
+    catalogLine:
+      "copyPageContent — replace one page's body with a copy of another page's body.",
     buildShape: (z) => ({
       sourcePageId: z.string().min(1).describe('Page to copy content FROM'),
       targetPageId: z
@@ -402,6 +461,9 @@ export const SHARED_TOOL_SPECS = {
       'page JSON and use a structural node patch/update to set its marks. ' +
       'Examples: edits:[{find:"teh",replace:"the"}]; edits:[{find:"Hello ' +
       'world",replace:"Hello there"}] (crosses a bold boundary).',
+    tier: 'core',
+    catalogLine:
+      "editPageText — surgical find/replace of plain text in a page, preserving ids/marks.",
     buildShape: (z) => ({
       pageId: z.string().describe('ID of the page to edit'),
       edits: z
@@ -440,6 +502,9 @@ export const SHARED_TOOL_SPECS = {
       'server instance that created it: in a multi-replica deployment without ' +
       'sticky sessions a blob stored on one instance is not retrievable via the ' +
       'sandbox URL on another (it 404s like an expired one).',
+    tier: 'deferred',
+    catalogLine:
+      'stashPage — serialize a whole page to a short anonymous URL without loading its body.',
     buildShape: (z) => ({
       pageId: z.string().min(1),
     }),

packages/prosemirror-markdown/src/lib/docmost-schema.ts

@@ -635,13 +635,17 @@ const Attachment = Node.create({
       },
       name: {
         default: null,
-        parseHTML: (el: HTMLElement) => el.getAttribute("data-attachment-name"),
+        // Empty-string-vs-absent idempotency (GS-EDIT-REVERT class): "" -> default.
+        parseHTML: (el: HTMLElement) =>
+          el.getAttribute("data-attachment-name") || null,
         renderHTML: (attrs: Record<string, any>) =>
           attrs.name ? { "data-attachment-name": attrs.name } : {},
       },
       mime: {
         default: null,
-        parseHTML: (el: HTMLElement) => el.getAttribute("data-attachment-mime"),
+        // Empty-string-vs-absent idempotency (GS-EDIT-REVERT class): "" -> default.
+        parseHTML: (el: HTMLElement) =>
+          el.getAttribute("data-attachment-mime") || null,
         renderHTML: (attrs: Record<string, any>) =>
           attrs.mime ? { "data-attachment-mime": attrs.mime } : {},
       },
@@ -689,7 +693,10 @@ const Video = Node.create({
       },
       alt: {
         default: null,
-        parseHTML: (el: HTMLElement) => el.getAttribute("aria-label"),
+        // Empty-string-vs-absent idempotency: coerce "" back to the default so a
+        // stray empty `aria-label` never materializes `alt: ""` on a video stored
+        // with no alt (same GS-EDIT-REVERT class as the image `alt` fix).
+        parseHTML: (el: HTMLElement) => el.getAttribute("aria-label") || null,
         renderHTML: (attrs: Record<string, any>) =>
           attrs.alt ? { "aria-label": attrs.alt } : {},
       },
@@ -864,13 +871,15 @@ const diagramAttributes = () => ({
   },
   title: {
     default: null,
-    parseHTML: (el: HTMLElement) => el.getAttribute("data-title"),
+    // Empty-string-vs-absent idempotency (GS-EDIT-REVERT class): "" -> default.
+    parseHTML: (el: HTMLElement) => el.getAttribute("data-title") || null,
     renderHTML: (attrs: Record<string, any>) =>
       attrs.title ? { "data-title": attrs.title } : {},
   },
   alt: {
     default: null,
-    parseHTML: (el: HTMLElement) => el.getAttribute("data-alt"),
+    // Empty-string-vs-absent idempotency (GS-EDIT-REVERT class): "" -> default.
+    parseHTML: (el: HTMLElement) => el.getAttribute("data-alt") || null,
     renderHTML: (attrs: Record<string, any>) =>
       attrs.alt ? { "data-alt": attrs.alt } : {},
   },
@@ -1106,7 +1115,8 @@ const Pdf = Node.create({
       },
       name: {
         default: null,
-        parseHTML: (el: HTMLElement) => el.getAttribute("data-name"),
+        // Empty-string-vs-absent idempotency (GS-EDIT-REVERT class): "" -> default.
+        parseHTML: (el: HTMLElement) => el.getAttribute("data-name") || null,
         renderHTML: (attrs: Record<string, any>) =>
           attrs.name ? { "data-name": attrs.name } : {},
       },
@@ -1491,6 +1501,29 @@ export const docmostExtensions = [
           ...parent.height,
           parseHTML: (el: HTMLElement) => el.getAttribute("height"),
         },
+        // Empty-string-vs-absent idempotency (GS-EDIT-REVERT class). `marked`
+        // renders `![](src)` as `<img alt="">`, so the stock Image `alt`
+        // parseHTML (`getAttribute("alt")`) materializes `alt: ""` on an image
+        // that was stored with NO alt (attr absent). That is a false diff against
+        // the editor-stored form (a no-alt image has alt ABSENT, not ""), so a
+        // git-sync / ai-chat touch of a page with a plain image produced phantom
+        // churn. Coerce an empty string back to the attr's default (null) so the
+        // import is idempotent. A real alt survives verbatim (`|| undefined` keeps
+        // the truthy value; the default fills the empty case). `title` is coerced
+        // the same way for the whole class, even though `marked` does not
+        // currently emit `title=""` — defence in depth against any path that does.
+        // NOTE: this DIVERGES from editor-ext's literal image `alt` parseHTML
+        // (`getAttribute("alt")`, which returns "" verbatim), but CONVERGES on
+        // editor-ext's real STORED shape: an editor image inserted without alt
+        // renders with no `alt` attribute and re-parses as absent, never "".
+        alt: {
+          ...parent.alt,
+          parseHTML: (el: HTMLElement) => el.getAttribute("alt") || null,
+        },
+        title: {
+          ...parent.title,
+          parseHTML: (el: HTMLElement) => el.getAttribute("title") || null,
+        },
       };
     },
   }).configure({ inline: false }),

packages/prosemirror-markdown/test/roundtrip-stability.helper.ts

@@ -0,0 +1,443 @@
+/**
+ * Reusable round-trip-STABILITY matrix helper (fixtures-first).
+ *
+ * A single stored node authored WITHOUT a given string attribute (attr
+ * absent / undefined) must not gain a phantom EMPTY-STRING value after a
+ * markdown round-trip — the "empty-string-vs-absent" churn class. This helper,
+ * given a node spec, drives a matrix of attribute combinations through the REAL
+ * converter (`convertProseMirrorToMarkdown` -> `markdownToProseMirror`) and
+ * asserts byte-stability on two contours:
+ *
+ *   1. RAW round-trip: for the node under test, every attribute the round-trip
+ *      materializes must equal what the INPUT authored — an authored attr keeps
+ *      its value, an ABSENT attr may only reappear at its SCHEMA DEFAULT. If an
+ *      absent attr comes back as a NON-default value (e.g. `alt: ""` where the
+ *      default is `null`), that is an instability and is reported precisely as
+ *      `type.attr: absent -> "<got>"`. This is the contour git-sync / stored
+ *      JSON diffs on, so masking it only in `canonicalize` would leave the noise.
+ *
+ *   2. CANONICAL round-trip: `canonicalizeContent(original)` must deep-equal
+ *      `canonicalizeContent(roundtrip)` (a second, semantic contour).
+ *
+ * The ONLY normalization the helper treats as allowed (not an instability) is
+ * the DOCUMENTED numeric width/height/size/aspectRatio -> string coercion the
+ * converter performs on purpose (a stored numeric `640` re-parses via
+ * `getAttribute` as the string `"640"`). It is encoded here as an explicit
+ * per-spec `numericStringAttrs` set applied to BOTH contours, NOT a silent skip.
+ *
+ * The helper is node-type agnostic: image and the whole media family share the
+ * `align !== "center"` predicate + `<!--name {…}-->` comment machinery, so one
+ * matrix guards the shared class.
+ */
+import { getSchema } from "@tiptap/core";
+import {
+  convertProseMirrorToMarkdown,
+  markdownToProseMirror,
+  canonicalizeContent,
+  docmostExtensions,
+} from "../src/lib/index.js";
+import { firstDivergence } from "./roundtrip-helpers.js";
+
+/** One attribute's two probe values. */
+export interface AttrMatrixEntry {
+  /** Attribute name on the node. */
+  attr: string;
+  /**
+   * The "default" pick. `undefined` means the attribute is OMITTED entirely
+   * (the absent case — the one that can materialize an empty string on import).
+   * A concrete value is authored verbatim.
+   */
+  default: unknown;
+  /** A representative NON-default value to exercise (must survive verbatim). */
+  nonDefault: unknown;
+  /**
+   * Marks the attr as a member of the EMPTY-STRING class the fix targets: a
+   * string attr whose schema default is `null`/absent and whose parseHTML
+   * coerces `"" -> default` (image/drawio `alt`+`title`, video `alt` via
+   * aria-label, pdf/attachment `name`, attachment `mime`). Set true to also
+   * drive the THIRD-STATE convergence case (see runConvergenceCase) for this
+   * attr. Attrs whose default is NOT null (e.g. embed `provider`, default "")
+   * or that are not `""`-coerced (control attrs) are left unset.
+   */
+  emptyStringClass?: boolean;
+}
+
+/** A node type + the attribute matrix to sweep for it. */
+export interface NodeStabilitySpec {
+  /** Node type (e.g. "image", "video"). */
+  type: string;
+  /** Attributes always present on the node (e.g. `{ src: "/i.png" }`). */
+  baseAttrs?: Record<string, unknown>;
+  /** Attributes to sweep at default and non-default. */
+  attrMatrix: AttrMatrixEntry[];
+  /**
+   * Attributes whose numeric -> string coercion on round-trip is DOCUMENTED and
+   * intentional; compared modulo `String(x)` on both sides. Defaults to the
+   * converter's known sizing set.
+   */
+  numericStringAttrs?: string[];
+}
+
+/** A single unstable finding, legible enough to tie a gate-lock to. */
+export interface Instability {
+  type: string;
+  attr: string;
+  /** What the input authored: the literal value, or the ABSENT sentinel. */
+  authored: unknown | typeof ABSENT;
+  /** What the round-trip produced. */
+  got: unknown;
+  /** What a stable round-trip should have produced (authored value or default). */
+  expected: unknown;
+}
+
+/** One matrix cell's result. */
+export interface ComboResult {
+  label: string;
+  authored: Record<string, unknown>;
+  /** RAW-contour instabilities on the node under test. */
+  raw: Instability[];
+  /** CANONICAL-contour divergence (path + values) or null when equal. */
+  canonical: { path: string; a: unknown; b: unknown } | null;
+  /** True when the node type failed to round-trip at all (structural loss). */
+  missing: boolean;
+  md: string;
+}
+
+/** Whole-matrix report for one node spec. */
+export interface MatrixReport {
+  type: string;
+  combos: ComboResult[];
+}
+
+/** Sentinel marking an attribute the input did NOT author. */
+export const ABSENT = Symbol("ABSENT");
+
+const DEFAULT_NUMERIC_STRING_ATTRS = [
+  "width",
+  "height",
+  "size",
+  "aspectRatio",
+];
+
+// The ProseMirror schema the converter targets — its attribute `default`s are
+// the authoritative "what an absent attr should re-materialize as" oracle.
+const schema = getSchema(docmostExtensions);
+
+/** Read the schema default for every attribute of a node type. */
+function schemaDefaults(type: string): Record<string, unknown> {
+  const specAttrs = (schema.nodes[type]?.spec?.attrs ?? {}) as Record<
+    string,
+    { default: unknown }
+  >;
+  const out: Record<string, unknown> = {};
+  for (const [k, v] of Object.entries(specAttrs)) out[k] = v.default;
+  return out;
+}
+
+/** Find the first node of a given type anywhere in a PM doc tree. */
+function findFirst(node: any, type: string): any {
+  if (node && node.type === type) return node;
+  for (const child of node?.content ?? []) {
+    const hit = findFirst(child, type);
+    if (hit) return hit;
+  }
+  return null;
+}
+
+/** Coerce a scalar for the documented numeric->string comparison. */
+const numStr = (x: unknown): unknown => (x == null ? x : String(x));
+
+/**
+ * Enumerate the cartesian product of the matrix: every attribute independently
+ * at its default (index 0) or non-default (index 1) pick. The all-default
+ * corner is included (the baseline). Small by construction (2^N over a handful
+ * of at-risk string attrs).
+ */
+function enumerateCombos(matrix: AttrMatrixEntry[]): number[][] {
+  let combos: number[][] = [[]];
+  for (let i = 0; i < matrix.length; i++) {
+    const next: number[][] = [];
+    for (const c of combos) {
+      next.push([...c, 0]);
+      next.push([...c, 1]);
+    }
+    combos = next;
+  }
+  return combos;
+}
+
+/** Build the authored attrs for one combo pick vector. */
+function authoredAttrs(
+  spec: NodeStabilitySpec,
+  picks: number[],
+): Record<string, unknown> {
+  const attrs: Record<string, unknown> = { ...(spec.baseAttrs ?? {}) };
+  spec.attrMatrix.forEach((entry, i) => {
+    if (picks[i] === 1) {
+      attrs[entry.attr] = entry.nonDefault;
+    } else if (entry.default !== undefined) {
+      attrs[entry.attr] = entry.default;
+    }
+    // default === undefined -> OMIT the attr entirely (the absent case).
+  });
+  return attrs;
+}
+
+/** Human-readable label for a combo (which attrs are at non-default). */
+function comboLabel(spec: NodeStabilitySpec, picks: number[]): string {
+  const on = spec.attrMatrix
+    .filter((_, i) => picks[i] === 1)
+    .map((e) => e.attr);
+  return on.length === 0 ? "<all-default>" : on.join("+");
+}
+
+/**
+ * Run the full stability matrix for one node spec and return a structured
+ * report (does NOT throw — the caller asserts, so a failure can print the whole
+ * report). Every combo runs the real export->import pipeline once.
+ */
+export async function runStabilityMatrix(
+  spec: NodeStabilitySpec,
+): Promise<MatrixReport> {
+  const numericStringAttrs = new Set(
+    spec.numericStringAttrs ?? DEFAULT_NUMERIC_STRING_ATTRS,
+  );
+  const defaults = schemaDefaults(spec.type);
+  const combos: ComboResult[] = [];
+
+  for (const picks of enumerateCombos(spec.attrMatrix)) {
+    const authored = authoredAttrs(spec, picks);
+    const doc = { type: "doc", content: [{ type: spec.type, attrs: authored }] };
+    const md = convertProseMirrorToMarkdown(doc);
+    const rt = await markdownToProseMirror(md);
+    const node = findFirst(rt, spec.type);
+
+    const result: ComboResult = {
+      label: comboLabel(spec, picks),
+      authored,
+      raw: [],
+      canonical: null,
+      missing: node == null,
+      md,
+    };
+
+    if (node != null) {
+      // RAW contour: every materialized attr must equal the authored value, or
+      // (for an absent attr) the schema default — modulo the documented numeric
+      // string coercion.
+      const rtAttrs = (node.attrs ?? {}) as Record<string, unknown>;
+      for (const key of Object.keys(rtAttrs)) {
+        const authoredHas = Object.prototype.hasOwnProperty.call(authored, key);
+        const expected = authoredHas ? authored[key] : defaults[key];
+        let got = rtAttrs[key];
+        let exp = expected;
+        if (numericStringAttrs.has(key)) {
+          got = numStr(got);
+          exp = numStr(exp);
+        }
+        if (firstDivergence(got, exp) !== null) {
+          result.raw.push({
+            type: spec.type,
+            attr: key,
+            authored: authoredHas ? authored[key] : ABSENT,
+            got: rtAttrs[key],
+            expected,
+          });
+        }
+      }
+
+      // CANONICAL contour: canonical forms deep-equal, modulo the same numeric
+      // string coercion (applied to both trees so a documented coercion is not
+      // counted as a divergence).
+      const ca = normalizeNumeric(canonicalizeContent(doc), numericStringAttrs);
+      const cb = normalizeNumeric(canonicalizeContent(rt), numericStringAttrs);
+      result.canonical = firstDivergence(ca, cb);
+    }
+
+    combos.push(result);
+  }
+
+  return { type: spec.type, combos };
+}
+
+/**
+ * Deep-copy a canonical tree, coercing the documented numeric->string attrs to
+ * their string form so an intentional `640 -> "640"` coercion is not reported
+ * as a canonical divergence. Only touches the listed attribute keys.
+ */
+function normalizeNumeric(node: any, attrs: Set<string>): any {
+  if (Array.isArray(node)) return node.map((n) => normalizeNumeric(n, attrs));
+  if (node === null || typeof node !== "object") return node;
+  const out: Record<string, unknown> = {};
+  for (const key of Object.keys(node)) {
+    if (key === "attrs" && node.attrs && typeof node.attrs === "object") {
+      const a: Record<string, unknown> = {};
+      for (const [k, v] of Object.entries(node.attrs)) {
+        a[k] = attrs.has(k) ? numStr(v) : v;
+      }
+      out.attrs = a;
+    } else {
+      out[key] = normalizeNumeric(node[key], attrs);
+    }
+  }
+  return out;
+}
+
+/** Flatten a report to just its unstable combos (for a terse assertion). */
+export function unstableCombos(report: MatrixReport): ComboResult[] {
+  return report.combos.filter(
+    (c) => c.missing || c.raw.length > 0 || c.canonical !== null,
+  );
+}
+
+// ---------------------------------------------------------------------------
+// THIRD STATE: an EXPLICITLY-STORED empty string on a string attr.
+//
+// The matrix above sweeps TWO states per string attr: absent/default and a
+// non-default value — and asserts FIRST-pass byte-stability for both. There is
+// a third, degenerate state the matrix does NOT cover: the attr stored as a
+// LITERAL `""`. This is DISTINCT from "the node never had the attr": a user
+// types an alt in the editor, then deletes it, and Tiptap's
+// `updateAttributes({ alt: "" })` persists a literal `alt: ""` in the stored
+// JSON. There is no absent-vs-"" distinction in the DOM once serialized, so the
+// fix's `getAttribute("alt") || null` coercion canonicalizes BOTH to the
+// default (`null`).
+//
+// Consequence — and this is CORRECT, not a bug: a doc carrying an explicit `""`
+// converges to the default on the FIRST round-trip (a ONE-TIME diff: `"" ->
+// null`), then is byte-stable from the SECOND round-trip on (idempotent). So
+// this state must be pinned with a DIFFERENT contract than the matrix's:
+//   - do NOT assert first-pass byte-stability (the first pass legitimately
+//     changes `""` -> default), and
+//   - DO assert the first pass converges to the default AND the second pass is
+//     idempotent (rt2 deep-equals rt1).
+//
+// A future sync/QA pass diffing stored pages will see this one-time `"" -> null`
+// normalization exactly once per affected node; it is the converter canon, not
+// corruption, and must not be flagged as data loss.
+// ---------------------------------------------------------------------------
+
+/** Result of the third-state ("explicit empty string") convergence probe. */
+export interface ConvergenceResult {
+  type: string;
+  attr: string;
+  /** The schema default the attr must converge to on pass 1 (null / absent). */
+  expectedDefault: unknown;
+  /** rt1's materialized value for the attr — must equal `expectedDefault`. */
+  firstPassValue: unknown;
+  /** True when the node round-tripped AND rt1 converged the attr to default. */
+  convergedToDefault: boolean;
+  /** rt1-vs-rt2 divergence; MUST be null (idempotent from pass 2 on). */
+  secondPassDivergence: { path: string; a: unknown; b: unknown } | null;
+  /** True when the node type failed to round-trip at all (structural loss). */
+  missing: boolean;
+}
+
+/** Round-trip a full PM doc through the real converter once. */
+async function roundtripDoc(doc: any): Promise<any> {
+  return markdownToProseMirror(convertProseMirrorToMarkdown(doc));
+}
+
+/**
+ * Third-state convergence probe for one string attr of the empty-string class.
+ *
+ * (a) builds a doc with the attr EXPLICITLY set to `""` (baseAttrs + `""`),
+ * (b) rt1 = roundtrip(doc); asserts rt1's attr equals the schema default — the
+ *     documented ONE-TIME `"" -> default` normalization (NOT byte-stable vs the
+ *     `""` input, so first-pass stability is deliberately NOT asserted here),
+ * (c) rt2 = roundtrip(rt1); asserts rt2 deep-equals rt1 — idempotent from the
+ *     second round-trip on.
+ *
+ * Returns a structured result (does NOT throw) so the caller can assert and
+ * print. Reusable across the whole node family: drive it for every attr flagged
+ * `emptyStringClass` on every spec (see convergenceCasesFor / the test driver).
+ */
+export async function runConvergenceCase(
+  spec: NodeStabilitySpec,
+  attr: string,
+): Promise<ConvergenceResult> {
+  const expectedDefault = schemaDefaults(spec.type)[attr];
+
+  // (a) The degenerate third state: attr persisted as a LITERAL "".
+  const authored = { ...(spec.baseAttrs ?? {}), [attr]: "" };
+  const doc = { type: "doc", content: [{ type: spec.type, attrs: authored }] };
+
+  // (b) First round-trip: "" must normalize to the default (a one-time diff).
+  const rt1 = await roundtripDoc(doc);
+  const node1 = findFirst(rt1, spec.type);
+  const firstPassValue = node1?.attrs?.[attr];
+  const convergedToDefault =
+    node1 != null && firstDivergence(firstPassValue, expectedDefault) === null;
+
+  // (c) Second round-trip: must be byte-stable (rt2 deep-equals rt1). We compare
+  // the WHOLE docs — both are converter OUTPUTS already in the same materialized
+  // form (numeric attrs are strings on both sides), so no numeric normalization
+  // is needed here, unlike the raw/canonical contours above.
+  const rt2 = node1 != null ? await roundtripDoc(rt1) : rt1;
+  const secondPassDivergence =
+    node1 != null ? firstDivergence(rt1, rt2) : null;
+
+  return {
+    type: spec.type,
+    attr,
+    expectedDefault,
+    firstPassValue,
+    convergedToDefault,
+    secondPassDivergence,
+    missing: node1 == null,
+  };
+}
+
+/** The attrs of a spec flagged as members of the empty-string class. */
+export function convergenceCasesFor(spec: NodeStabilitySpec): string[] {
+  return spec.attrMatrix
+    .filter((e) => e.emptyStringClass)
+    .map((e) => e.attr);
+}
+
+/** True when a convergence result honours the "converges once, then stable" contract. */
+export function convergenceOk(r: ConvergenceResult): boolean {
+  return !r.missing && r.convergedToDefault && r.secondPassDivergence === null;
+}
+
+/** Render a convergence result as a legible one-liner for a failed assertion. */
+export function formatConvergence(r: ConvergenceResult): string {
+  if (r.missing) return `${r.type}.${r.attr}: DID-NOT-ROUND-TRIP`;
+  const parts: string[] = [];
+  if (!r.convergedToDefault) {
+    parts.push(
+      `pass1 did NOT converge: got ${JSON.stringify(r.firstPassValue)} (expected default ${JSON.stringify(r.expectedDefault)})`,
+    );
+  }
+  if (r.secondPassDivergence) {
+    parts.push(
+      `pass2 NOT idempotent @ ${r.secondPassDivergence.path}: ${JSON.stringify(r.secondPassDivergence.a)} vs ${JSON.stringify(r.secondPassDivergence.b)}`,
+    );
+  }
+  const status = parts.length === 0 ? "converges-once-then-stable" : parts.join("; ");
+  return `${r.type}.${r.attr}: ${status}`;
+}
+
+/** Render a report as a legible multi-line string for a failed assertion. */
+export function formatReport(report: MatrixReport): string {
+  const lines: string[] = [`node "${report.type}":`];
+  for (const c of report.combos) {
+    const flags: string[] = [];
+    if (c.missing) flags.push("DID-NOT-ROUND-TRIP");
+    for (const i of c.raw) {
+      const authored =
+        i.authored === ABSENT ? "absent" : JSON.stringify(i.authored);
+      flags.push(
+        `RAW ${i.type}.${i.attr}: ${authored} -> ${JSON.stringify(i.got)} (expected ${JSON.stringify(i.expected)})`,
+      );
+    }
+    if (c.canonical) {
+      flags.push(
+        `CANON @ ${c.canonical.path}: ${JSON.stringify(c.canonical.a)} vs ${JSON.stringify(c.canonical.b)}`,
+      );
+    }
+    const status = flags.length === 0 ? "stable" : flags.join("; ");
+    lines.push(`  [${c.label}] ${status}`);
+  }
+  return lines.join("\n");
+}

packages/prosemirror-markdown/test/roundtrip-stability.test.ts

@@ -0,0 +1,164 @@
+import { describe, expect, it } from "vitest";
+import {
+  runStabilityMatrix,
+  unstableCombos,
+  formatReport,
+  runConvergenceCase,
+  convergenceCasesFor,
+  convergenceOk,
+  formatConvergence,
+  type NodeStabilitySpec,
+} from "./roundtrip-stability.helper.js";
+
+// ---------------------------------------------------------------------------
+// Round-trip STABILITY matrix for image + the media family.
+//
+// Guards the "empty-string-vs-absent" churn class (GS-EDIT-REVERT family): a
+// stored node authored WITHOUT a string attr (alt/title/caption/aria-label/...)
+// must not gain a phantom `attr: ""` after `markdownToProseMirror(convert…)`.
+// Each spec sweeps the at-risk string attrs at DEFAULT (absent) and at a real
+// NON-default value; the helper asserts both the RAW round-trip (attrs equal the
+// input's, modulo the documented numeric width/height/size/aspectRatio -> string
+// coercion) and the CANONICAL round-trip (canonical forms deep-equal).
+//
+// The image + media family share the `align !== "center"` predicate and the
+// `<!--name {…}-->` comment machinery, so one matrix guards the shared class.
+// align is NOT part of this class (it round-trips correctly) and is not swept.
+// ---------------------------------------------------------------------------
+
+const SPECS: NodeStabilitySpec[] = [
+  {
+    // Image carries the most at-risk string attrs. `alt` is the one marked
+    // materializes as `<img alt="">` on `![](src)` import (the real bug); title
+    // and caption are covered as the same class. attachmentId is a string attr
+    // that must stay absent when unset (control).
+    type: "image",
+    baseAttrs: { src: "/i.png" },
+    attrMatrix: [
+      { attr: "alt", default: undefined, nonDefault: "a real alt text", emptyStringClass: true },
+      { attr: "title", default: undefined, nonDefault: "a real title", emptyStringClass: true },
+      { attr: "caption", default: undefined, nonDefault: "a real caption" },
+      { attr: "attachmentId", default: undefined, nonDefault: "att-42" },
+    ],
+  },
+  {
+    // Video's `alt` rides the `aria-label` attribute (media aria-label at risk).
+    type: "video",
+    baseAttrs: { src: "/v.mp4" },
+    attrMatrix: [
+      { attr: "alt", default: undefined, nonDefault: "a clip", emptyStringClass: true },
+      { attr: "attachmentId", default: undefined, nonDefault: "att-1" },
+    ],
+  },
+  {
+    // Audio carries no alt/title; attachmentId is its only optional string attr.
+    type: "audio",
+    baseAttrs: { src: "/a.mp3" },
+    attrMatrix: [
+      { attr: "attachmentId", default: undefined, nonDefault: "att-2" },
+    ],
+  },
+  {
+    // pdf: link-form media. `name` (filename) is its at-risk string attr.
+    type: "pdf",
+    baseAttrs: { src: "/d.pdf" },
+    attrMatrix: [
+      { attr: "name", default: undefined, nonDefault: "report.pdf", emptyStringClass: true },
+      { attr: "attachmentId", default: undefined, nonDefault: "att-3" },
+    ],
+  },
+  {
+    // attachment: link-form media (file card). `name` + `mime` string attrs.
+    type: "attachment",
+    baseAttrs: { url: "/f.zip" },
+    attrMatrix: [
+      { attr: "name", default: undefined, nonDefault: "bundle.zip", emptyStringClass: true },
+      { attr: "mime", default: undefined, nonDefault: "application/zip", emptyStringClass: true },
+      { attr: "attachmentId", default: undefined, nonDefault: "att-4" },
+    ],
+  },
+  {
+    // embed: link-form media. `provider` is its at-risk string attr (schema
+    // default ""). embed's numeric width/height defaults (800/600) are a SEPARATE,
+    // documented limitation OUTSIDE the empty-string class: they are not in
+    // canonicalize's KNOWN_DEFAULTS, so an ABSENT width/height re-imports as the
+    // 800/600 default and diverges canonically (see the note in canonicalize.ts).
+    // That is canonicalize-owned and out of scope here, so we author the
+    // dimensions at their defaults (as real editor embeds carry them) to keep this
+    // guard focused on the empty-string/provider class.
+    // provider's schema default is "" (NOT null), so a re-imported "" is the
+    // correct value, not a phantom — it is outside the null-default empty-string
+    // class. We author it at its "" default (the default pick) so the sweep still
+    // asserts a non-default provider ("youtube") round-trips, without tripping the
+    // canonicalize KNOWN_DEFAULTS gap for embed's non-null defaults.
+    type: "embed",
+    baseAttrs: { src: "https://example.com/x", width: 800, height: 600 },
+    attrMatrix: [
+      { attr: "provider", default: "", nonDefault: "youtube" },
+    ],
+  },
+  {
+    // drawio: image-form diagram. `title` + `alt` string attrs (data-title/-alt).
+    type: "drawio",
+    baseAttrs: { src: "blob:drawio" },
+    attrMatrix: [
+      { attr: "title", default: undefined, nonDefault: "flow chart", emptyStringClass: true },
+      { attr: "alt", default: undefined, nonDefault: "an alt", emptyStringClass: true },
+      { attr: "attachmentId", default: undefined, nonDefault: "att-5" },
+    ],
+  },
+  {
+    // excalidraw: image-form diagram, same shared diagramAttributes set.
+    type: "excalidraw",
+    baseAttrs: { src: "blob:excalidraw" },
+    attrMatrix: [
+      { attr: "title", default: undefined, nonDefault: "sketch", emptyStringClass: true },
+      { attr: "alt", default: undefined, nonDefault: "an alt", emptyStringClass: true },
+      { attr: "attachmentId", default: undefined, nonDefault: "att-6" },
+    ],
+  },
+];
+
+describe("round-trip stability matrix (image + media family)", () => {
+  for (const spec of SPECS) {
+    it(`${spec.type}: no attr materializes an empty-string / phantom value`, async () => {
+      const report = await runStabilityMatrix(spec);
+      const unstable = unstableCombos(report);
+      // On failure, print the WHOLE matrix so which (attr, value) combos are
+      // unstable is legible.
+      expect(unstable, `\n${formatReport(report)}\n`).toEqual([]);
+    });
+  }
+});
+
+// ---------------------------------------------------------------------------
+// THIRD STATE: an attr EXPLICITLY stored as a literal "" (GS-EDIT-REVERT: a user
+// typed alt/title/name/... then deleted it, so Tiptap persisted `attr: ""` — a
+// value DISTINCT from "attr was never set"). Unlike the absent case above, this
+// state is NOT first-pass byte-stable: the fix's `"" -> default` coercion is a
+// deliberate ONE-TIME normalization on the FIRST sync round-trip, stable
+// thereafter. We therefore assert a DIFFERENT contract — "converges to default
+// on pass 1, then idempotent from pass 2 on" — for every empty-string-class attr
+// across the whole node family (image/video/pdf/attachment/drawio/excalidraw).
+//
+// IMPORTANT for a future sync/QA pass: the pass-1 `"" -> null` diff is the
+// converter canon, not corruption. It appears at most once per affected node and
+// must NOT be flagged as "the converter is losing/corrupting page data".
+// ---------------------------------------------------------------------------
+describe("round-trip third state: explicit empty string converges once, then idempotent", () => {
+  for (const spec of SPECS) {
+    for (const attr of convergenceCasesFor(spec)) {
+      it(`${spec.type}.${attr}: "" normalizes to default on pass 1, byte-stable from pass 2`, async () => {
+        const r = await runConvergenceCase(spec, attr);
+        // Pass 1 must converge "" -> the schema default (the one-time diff) and
+        // pass 2 (roundtrip of pass-1 output) must be byte-stable. formatConvergence
+        // prints exactly which half failed.
+        expect(convergenceOk(r), `\n${formatConvergence(r)}\n`).toBe(true);
+        // Spell the contract out explicitly so the intent is legible in the test:
+        expect(r.convergedToDefault, `\n${formatConvergence(r)}\n`).toBe(true);
+        expect(r.firstPassValue).toEqual(r.expectedDefault);
+        expect(r.secondPassDivergence, `\n${formatConvergence(r)}\n`).toBeNull();
+      });
+    }
+  }
+});

pnpm-lock.yaml

@@ -269,6 +269,9 @@ importers:
       '@atlaskit/pragmatic-drag-and-drop-live-region':
         specifier: 1.3.4
         version: 1.3.4
+      '@braintree/sanitize-url':
+        specifier: 7.1.2
+        version: 7.1.2
       '@casl/react':
         specifier: 5.0.1
         version: 5.0.1(@casl/ability@6.8.0)(react@18.3.1)
@@ -16153,7 +16156,7 @@ snapshots:
       obug: 2.1.1
       std-env: 4.1.0
       tinyrainbow: 3.1.0
-      vitest: 4.1.6(@opentelemetry/api@1.9.0)(@types/node@25.5.0)(@vitest/coverage-v8@4.1.6)(happy-dom@20.8.9)(jsdom@27.4.0(@noble/hashes@2.0.1))(vite@8.0.5(@types/node@25.5.0)(esbuild@0.28.0)(jiti@2.4.2)(less@4.2.0)(sugarss@5.0.1(postcss@8.5.14))(terser@5.39.0)(tsx@4.21.0)(yaml@2.8.3))
+      vitest: 4.1.6(@opentelemetry/api@1.9.0)(@types/node@22.19.1)(@vitest/coverage-v8@4.1.6)(happy-dom@20.8.9)(jsdom@25.0.0)(vite@8.0.5(@types/node@22.19.1)(esbuild@0.28.0)(jiti@2.4.2)(less@4.2.0)(sugarss@5.0.1(postcss@8.5.14))(terser@5.39.0)(tsx@4.21.0)(yaml@2.8.3))
 
   '@vitest/expect@4.1.6':
     dependencies: