test(git-sync): assert stable drawio markers, not the omittable center default (#351 review)

stabilize.test.ts used `data-align="center"` as proof the convergence pass
materialized the drawio node. Since this PR correctly stops emitting the
schema-default center align in the media builders, that marker is no longer a
reliable convergence proof. Assert on the stable canonical markers instead —
`data-type="drawio"` + `data-src="/d.drawio"` — which are always materialized
regardless of the align default. The fixpoint assertion (file2 === file1) is
unchanged; round-trip stays byte-stable.

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

.github/workflows/develop.yml

@@ -151,6 +151,12 @@ jobs:
       - name: Build editor-ext
         run: pnpm --filter @docmost/editor-ext build
 
+      # @docmost/prosemirror-markdown is an ESM workspace package the server
+      # imports at runtime; its build/ is gitignored and test:e2e has no pretest
+      # hook, so build it before the e2e run (mirrors the test.yml job).
+      - name: Build prosemirror-markdown
+        run: pnpm --filter @docmost/prosemirror-markdown build
+
       - name: Run migrations
         run: pnpm --filter ./apps/server migration:latest
 

.github/workflows/test.yml

@@ -13,6 +13,49 @@ permissions:
   contents: read
 
 jobs:
+  # Guard against a long-lived branch adding a migration whose timestamped
+  # filename sorts BEFORE migrations already applied on the target branch (and
+  # thus in prod). The Kysely startup migrator rejects that as "corrupted
+  # migrations" and crash-loops the app on boot (incident #361). This gate fails
+  # the PR so the migration is renamed to a current timestamp before merge. Only
+  # runs for pull_request events (needs a base branch to diff against).
+  migration-order:
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - name: Checkout (full history for the base-branch diff)
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - name: Added migrations must sort after the newest on the base branch
+        env:
+          TARGET_BRANCH: ${{ github.base_ref }}
+        run: |
+          set -euo pipefail
+          MIG_DIR="apps/server/src/database/migrations"
+          # checkout above already did fetch-depth:0 (full history). Fetch the base
+          # WITHOUT --depth (a shallow graft would truncate the base history and
+          # break the merge-base when the base has moved ahead of the PR merge —
+          # exactly the long-branch-vs-moving-base case this gate guards, #361).
+          git fetch --no-tags origin "$TARGET_BRANCH"
+          newest_on_target=$(git ls-tree -r --name-only "origin/${TARGET_BRANCH}" "$MIG_DIR" | sort | tail -1)
+          # NO `|| true`: a diff failure (e.g. an unresolved merge-base) must fail
+          # the job CLOSED — a gate whose job is to BLOCK must never pass on error.
+          # `set -e` above already aborts on a non-zero diff exit.
+          added=$(git diff --diff-filter=A --name-only "origin/${TARGET_BRANCH}...HEAD" -- "$MIG_DIR")
+          bad=0
+          for f in $added; do
+            if [[ "$f" < "$newest_on_target" || "$f" == "$newest_on_target" ]]; then
+              echo "::error::Migration $f sorts at or before the newest on ${TARGET_BRANCH} ($newest_on_target) — rename it with a CURRENT timestamp before merge (do not change its contents). See incident #361."
+              bad=1
+            fi
+          done
+          if [ "$bad" -eq 0 ]; then
+            echo "Migration order OK (added migrations all sort after $newest_on_target)."
+          fi
+          exit $bad
+
   test:
     runs-on: ubuntu-latest
     timeout-minutes: 20

AGENTS.md

@@ -201,7 +201,7 @@ pnpm workspace (`pnpm@10.4.0`) orchestrated by **Nx**. Four workspace packages:
 | `apps/client` | `client` | React 18 + Vite + Mantine 8 + TanStack Query + Jotai | SPA frontend |
 | `packages/editor-ext` | `@docmost/editor-ext` | Tiptap/ProseMirror | Shared Tiptap node/mark extensions, imported by both the client and the server |
 | `packages/mcp` | `@docmost/mcp` | MCP SDK, Tiptap, Yjs | Standalone MCP server, also bundled into the server at `/mcp`. Consumes the shared converter/schema from `@docmost/prosemirror-markdown` (#293) — it no longer carries its own vendored converter/schema copy |
-| `packages/prosemirror-markdown` | `@docmost/prosemirror-markdown` | Tiptap, marked, jsdom | The single, canonical ProseMirror↔Markdown converter + Docmost schema mirror (#293). Consumed by `mcp` and `git-sync`; there is exactly ONE copy of the converter now |
+| `packages/prosemirror-markdown` | `@docmost/prosemirror-markdown` | Tiptap, marked, jsdom | The single, canonical ProseMirror↔Markdown converter + Docmost schema mirror (#293). Consumed by `mcp`, `git-sync`, AND `apps/server` (server-side markdown import/export, #345); there is exactly ONE copy of the converter now |
 
 `build` targets are Nx-cached and dependency-ordered (`dependsOn: ["^build"]`), so `editor-ext` builds before the apps. `nx.json` sets `affected.defaultBase: main`.
 
@@ -214,6 +214,12 @@ Run from the repo root unless noted. The dev workflow needs **Postgres (with the
 > server, `APP_SECRET` mismatch between processes, a stale `editor-ext` white-
 > screening the client, LAN exposure. See **[docs/dev-stand.md](docs/dev-stand.md)**
 > for the step-by-step and the traps.
+>
+> **Testing the app against a stand** (browser E2E + out-of-band verification) has
+> its own non-obvious traps — the page has two ProseMirror editors (only the body is
+> collab-bound), a ~10s store debounce, and API-seeding the thing under test is a
+> silent no-test. See **[docs/how-to-test.md](docs/how-to-test.md)** before writing
+> UI tests.
 
 ```bash
 pnpm install                 # install all workspaces (uses pnpm patches; see package.json `pnpm.patchedDependencies`)
@@ -250,7 +256,10 @@ pnpm --filter server migration:codegen                   # regenerate src/databa
 ```
 Migration files live in `apps/server/src/database/migrations/` and are named `YYYYMMDDThhmmss-description.ts`. Fork-specific migrations only **add** tables (`page_embeddings`, `ai_chats`, `ai_chat_messages`, `ai_provider_credentials`, `ai_mcp_servers`, `page_template_references`) and columns (e.g. `pages.is_template`, a `NOT NULL DEFAULT false` boolean) — never drop/rewrite Docmost data.
 
-**Migration ordering — always check when merging branches/features.** Kysely runs migrations in **alphabetical (= timestamp) order** and refuses to start if a *new* migration sorts **before** one already applied to the DB (`corrupted migrations: ... must always have a name that comes alphabetically after the last executed migration`). When you merge a branch or land a feature, verify your migration's timestamp still sorts **after every migration that may already be applied on the target** (`/bin/ls -1 apps/server/src/database/migrations | sort | tail`). Branches developed in parallel routinely break this: a feature branch adds `…T130000-…`, `main` meanwhile ships and deploys `…T150000-…`, and after the merge the older-timestamped file is rejected at boot. **Fix = rename your migration to a timestamp after the latest one already in the target** (content unchanged — the filename is the ordering key), then rebuild so the compiled `dist/database/migrations/` picks up the new name.
+**Migration ordering — always check when merging branches/features.** Kysely runs migrations in **alphabetical (= timestamp) order**. A *new* migration that sorts **before** one already applied to the DB is a "back-dated" migration, which branches developed in parallel routinely produce: a feature branch adds `…T130000-…`, `develop` meanwhile ships and deploys `…T150000-…`, and after the merge the older-timestamped file has been skipped. Two layers guard this (both added for incident #361, where a back-dated migration crash-looped prod for ~11 min):
+
+- **CI gate (primary):** the `migration-order` job in `.github/workflows/test.yml` fails a PR whose added migration sorts at/before the newest on the base branch. **So the fix is to rename your migration to a timestamp after the latest one already in the target** (`/bin/ls -1 apps/server/src/database/migrations | sort | tail`; content unchanged — the filename is the ordering key), then rebuild so the compiled `dist/database/migrations/` picks up the new name.
+- **Runtime safety net:** both Migrators (`migration.service.ts` startup auto-migrate + `migrate.ts` CLI) set `allowUnorderedMigrations: true`, so the app does **not** refuse to start on an out-of-order migration — it applies the skipped older one instead of crash-looping. Kysely's `#ensureNoMissingMigrations` guard is still on (a *removed* applied migration is still an error). Because apply order can then differ from lexicographic across instances, migrations must stay **independent** (each creates its own objects) — the CI gate remains the primary line; this net only covers a gate bypass (manual push / hotfix branch).
 
 ## Architecture — the big picture
 
@@ -284,7 +293,7 @@ The API server is a Fastify app with a global `/api` prefix (`main.ts` excludes
 ### Client structure
 Vite SPA. Code is organized by feature under `apps/client/src/features/*` (mirrors the server domains: `page`, `space`, `comment`, `ai-chat`, `editor`, …). Conventions:
 - **TanStack Query** for server state (one `queries/` file per feature), **Jotai** atoms for local/shared UI state, **Mantine 8** + CSS modules (`*.module.css`) + `postcss-preset-mantine` for UI.
-- The editor is Tiptap; shared node/mark extensions live in `packages/editor-ext` and are imported by **both the client and the server** (collaboration, import/export) — editor schema changes often need to be made in `editor-ext`, not just the client. The ProseMirror↔Markdown converter and its Docmost schema mirror now live in a SINGLE package, `@docmost/prosemirror-markdown` (#293), consumed by both `mcp` and `git-sync` — do NOT reintroduce a per-package copy. `editor-ext` is the upstream source of the Tiptap schema; the package's `docmost-schema.ts` mirrors it and a serializer-contract test (`packages/prosemirror-markdown/test/serializer-contract.test.ts`) guards the boundary (every schema node must have a converter case), so a drift surfaces as a failing test rather than silent divergence.
+- The editor is Tiptap; shared node/mark extensions live in `packages/editor-ext` and are imported by **both the client and the server** (collaboration, schema, `canonicalizeFootnotes`) — editor schema changes often need to be made in `editor-ext`, not just the client. Server-side markdown import/export no longer lives in `editor-ext`: it goes through the canonical converter (#345, see below). The ProseMirror↔Markdown converter and its Docmost schema mirror now live in a SINGLE package, `@docmost/prosemirror-markdown` (#293), consumed by `mcp`, `git-sync`, and `apps/server` (#345) — do NOT reintroduce a per-package copy. `editor-ext` is the upstream source of the Tiptap schema; the package's `docmost-schema.ts` mirrors it and a serializer-contract test (`packages/prosemirror-markdown/test/serializer-contract.test.ts`) guards the boundary (every schema node must have a converter case), so a drift surfaces as a failing test rather than silent divergence.
 - API access goes through `apps/client/src/lib/api-client.ts` (axios). The `@` alias maps to `apps/client/src`.
 - Runtime config is injected at build time by `vite.config.ts` via `define` (`APP_URL`, `COLLAB_URL`, `APP_VERSION`, …) — these come from the root `.env`, not from `import.meta.env`.
 
@@ -294,7 +303,7 @@ Vite SPA. Code is organized by feature under `apps/client/src/features/*` (mirro
 - **Errors must never be swallowed or shown as generic messages.** Every caught error MUST (1) be logged in full to the console/logger — error name, message, stack, `cause`, and (for HTTP/provider failures) the status code and response body — and (2) be surfaced to the user with a *specific, human-readable explanation of what actually went wrong*, never a bare generic string like "Something went wrong" / "Could not start recording" / "Transcription failed". Include the real reason (the underlying error/provider message) in the user-facing text. On the server, wrap third-party/provider failures with `describeProviderError` (or equivalent) and rethrow as a meaningful HTTP status + message — never let them collapse into an opaque 500. On the client, `console.error(<context>, err)` the raw error AND show the extracted reason (e.g. `err.response?.data?.message`, or the error `name: message`) in the notification.
 - The version string shown in the UI comes from `APP_VERSION` (CI/Docker) or `git describe --tags --always` (local), resolved in `vite.config.ts` — not from `package.json`.
 - Server TS config is permissive (`noImplicitAny: false`, `strictNullChecks: false`, `no-explicit-any` lint disabled). Follow the existing relaxed style rather than tightening types broadly.
-- Dependency versions are heavily pinned via `pnpm.overrides` and `pnpm.patchedDependencies` (`scimmy`, `yjs`) in the root `package.json`. Don't bump pinned/patched deps casually; the patches and overrides exist for compatibility/security reasons.
+- Dependency versions are heavily pinned via `pnpm.overrides` and `pnpm.patchedDependencies` (`scimmy`, `yjs`, `ai`) in the root `package.json`. Don't bump pinned/patched deps casually; the patches and overrides exist for compatibility/security reasons. The `ai@6.0.134` patch disables the SDK's O(n²) cumulative `partialOutput` accumulation when no output strategy is requested (server heap OOM on long agent runs, #184; tripwire test: `apps/server/src/integrations/ai/ai-sdk-partial-output.patch.spec.ts`) — it MUST be re-created via `pnpm patch` when bumping `ai`.
 - **Adding/renaming/removing an MCP tool requires updating `SERVER_INSTRUCTIONS`** in `packages/mcp/src/index.ts` — the intent-routing guide MCP clients receive on initialize. This applies both to inline `server.registerTool(...)` calls in `index.ts` and to specs in `packages/mcp/src/tool-specs.ts`. Enforced by `packages/mcp/test/unit/server-instructions.test.mjs`, which fails when a registered tool is not mentioned in the guide (deliberate opt-outs go into its `EXCEPTIONS` list). `packages/mcp/build/` is gitignored and rebuilt in CI/Docker via `pnpm build` (same convention as `git-sync`/`prosemirror-markdown`) — never commit it; rebuild locally after editing to run the tests.
 
 ## CI / release

CHANGELOG.md

@@ -169,6 +169,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- **The server no longer runs out of heap during long autonomous agent runs.** A
+  new pnpm patch on `ai@6.0.134` stops the SDK from building a cumulative
+  snapshot of the ENTIRE turn text on every streamed text-delta when no output
+  strategy was requested (our server never requests one). Unpatched, those
+  O(n²) `partialOutput` snapshots piled up in a never-consumed internal
+  `tee()` branch of the stream result — a ~20-step, ~28k-chunk agent run
+  retained ~1.7 GB and OOM'd the 2 GB JS heap. Streaming granularity is
+  unchanged; the patch must be re-created if `ai` is ever bumped. (#184)
 - **Internal links in exported Markdown no longer lose their visible text.** A
   link whose target page name had no file extension (e.g. a bare title) was
   collapsed to empty text during export, producing an unclickable, label-less

apps/client/public/locales/en-US/translation.json

@@ -1373,6 +1373,39 @@
   "The role catalog is unavailable": "The role catalog is unavailable",
   "Please try again later.": "Please try again later.",
   "No bundles available": "No bundles available",
+  "Content": "Content",
+  "Content language of the roles": "Content language of the roles",
+  "{{count}} updates available in {{bundles}} bundles": "{{count}} updates available in {{bundles}} bundles",
+  "Update all ({{count}})": "Update all ({{count}})",
+  "Updating {{current}}/{{total}}…": "Updating {{current}}/{{total}}…",
+  "{{count}} roles are installed in another language. A different language installs separately and appears as new.": "{{count}} roles are installed in another language. A different language installs separately and appears as new.",
+  "{{count}} roles": "{{count}} roles",
+  "{{count}} new — none installed": "{{count}} new — none installed",
+  "All installed · up to date": "All installed · up to date",
+  "{{count}} updates · {{installed}} up to date": "{{count}} updates · {{installed}} up to date",
+  "{{count}} new": "{{count}} new",
+  "{{count}} installed": "{{count}} installed",
+  "{{count}} updates": "{{count}} updates",
+  "Install bundle": "Install bundle",
+  "Install {{count}} selected": "Install {{count}} selected",
+  "Install bundle ({{count}})": "Install bundle ({{count}})",
+  "{{selected}} of {{total}} selected": "{{selected}} of {{total}} selected",
+  "Select all": "Select all",
+  "Deselect all": "Deselect all",
+  "Skipped": "Skipped",
+  "v{{version}}": "v{{version}}",
+  "{{count}} roles installed": "{{count}} roles installed",
+  "{{count}} roles installed · {{renamed}} renamed": "{{count}} roles installed · {{renamed}} renamed",
+  "{{count}} roles updated": "{{count}} roles updated",
+  "Installed {{installed}} · {{skipped}} skipped": "Installed {{installed}} · {{skipped}} skipped",
+  "A role named \"{{name}}\" already exists in this workspace.": "A role named \"{{name}}\" already exists in this workspace.",
+  "\"{{name}}\" is already installed.": "\"{{name}}\" is already installed.",
+  "Rename & install": "Rename & install",
+  "Couldn’t load the catalog": "Couldn’t load the catalog",
+  "Check your connection and try again. Installed roles are not affected.": "Check your connection and try again. Installed roles are not affected.",
+  "Retry": "Retry",
+  "The catalog is empty": "The catalog is empty",
+  "No role bundles are published for this language yet. Try switching the content language.": "No role bundles are published for this language yet. Try switching the content language.",
   "Already up to date": "Already up to date",
   "Updated to the latest version": "Updated to the latest version",
   "This role is no longer in the catalog": "This role is no longer in the catalog",

apps/client/public/locales/ru-RU/translation.json

@@ -1235,6 +1235,39 @@
   "The role catalog is unavailable": "Каталог ролей недоступен",
   "Please try again later.": "Попробуйте позже.",
   "No bundles available": "Наборы недоступны",
+  "Content": "Язык контента",
+  "Content language of the roles": "Язык контента ролей",
+  "{{count}} updates available in {{bundles}} bundles": "Доступно обновлений: {{count}} в наборах: {{bundles}}",
+  "Update all ({{count}})": "Обновить все ({{count}})",
+  "Updating {{current}}/{{total}}…": "Обновление {{current}}/{{total}}…",
+  "{{count}} roles are installed in another language. A different language installs separately and appears as new.": "Ролей установлено на другом языке: {{count}}. Другой язык устанавливается отдельно и отображается как новый.",
+  "{{count}} roles": "ролей: {{count}}",
+  "{{count}} new — none installed": "новых: {{count}} — ничего не установлено",
+  "All installed · up to date": "Все установлены · актуальны",
+  "{{count}} updates · {{installed}} up to date": "обновлений: {{count}} · актуальны: {{installed}}",
+  "{{count}} new": "новых: {{count}}",
+  "{{count}} installed": "установлено: {{count}}",
+  "{{count}} updates": "обновлений: {{count}}",
+  "Install bundle": "Установить набор",
+  "Install {{count}} selected": "Установить выбранные ({{count}})",
+  "Install bundle ({{count}})": "Установить набор ({{count}})",
+  "{{selected}} of {{total}} selected": "выбрано {{selected}} из {{total}}",
+  "Select all": "Выбрать все",
+  "Deselect all": "Снять выбор",
+  "Skipped": "Пропущено",
+  "v{{version}}": "v{{version}}",
+  "{{count}} roles installed": "Установлено ролей: {{count}}",
+  "{{count}} roles installed · {{renamed}} renamed": "Установлено ролей: {{count}} · переименовано: {{renamed}}",
+  "{{count}} roles updated": "Обновлено ролей: {{count}}",
+  "Installed {{installed}} · {{skipped}} skipped": "Установлено: {{installed}} · пропущено: {{skipped}}",
+  "A role named \"{{name}}\" already exists in this workspace.": "Роль с именем «{{name}}» уже существует в этом рабочем пространстве.",
+  "\"{{name}}\" is already installed.": "«{{name}}» уже установлена.",
+  "Rename & install": "Переименовать и установить",
+  "Couldn’t load the catalog": "Не удалось загрузить каталог",
+  "Check your connection and try again. Installed roles are not affected.": "Проверьте подключение и попробуйте снова. Установленные роли не затронуты.",
+  "Retry": "Повторить",
+  "The catalog is empty": "Каталог пуст",
+  "No role bundles are published for this language yet. Try switching the content language.": "Для этого языка ещё не опубликовано ни одного набора ролей. Попробуйте сменить язык контента.",
   "No roles configured": "Роли не настроены",
   "Already up to date": "Уже актуальна",
   "Updated to the latest version": "Обновлено до последней версии",

apps/client/src/features/ai-chat/queries/ai-chat-query.ts

@@ -1,6 +1,7 @@
 import {
   useInfiniteQuery,
   useMutation,
+  useQueries,
   useQuery,
   useQueryClient,
 } from "@tanstack/react-query";
@@ -307,6 +308,29 @@ export function useAiRoleCatalogBundleQuery(
   });
 }
 
+/**
+ * Eagerly open EVERY listed bundle's content in parallel for one language. The
+ * redesigned catalog shows each bundle's status summary in its COLLAPSED header,
+ * which needs every role's install state up front — so contents can no longer be
+ * lazy-loaded on expand. The catalog is small, so a fan-out of `useQueries` (one
+ * cached read per bundle, sharing the same cache keys as
+ * `useAiRoleCatalogBundleQuery`) is cheap. Gated by `enabled` (modal open + a
+ * resolved language) so nothing fetches while the modal is closed.
+ */
+export function useAiRoleCatalogBundlesQueries(
+  bundleIds: string[],
+  language: string,
+  enabled: boolean,
+) {
+  return useQueries({
+    queries: bundleIds.map((bundleId) => ({
+      queryKey: AI_ROLE_CATALOG_BUNDLE_RQ_KEY(bundleId, language),
+      queryFn: () => getAiRoleCatalogBundle(bundleId, language),
+      enabled: enabled && !!language,
+    })),
+  });
+}
+
 export function useImportAiRolesFromCatalogMutation() {
   const queryClient = useQueryClient();
   const { t } = useTranslation();

apps/client/src/features/ai-chat/queries/import-from-catalog-message.test.tsx

@@ -77,7 +77,14 @@ describe("useImportAiRolesFromCatalogMutation — success notifications", () =>
   });
 
   it("errors:[] -> only the summary notification (counts interpolated)", async () => {
-    await runMutation({ created: 3, renamed: 1, skipped: 2, errors: [] });
+    await runMutation({
+      created: 3,
+      renamed: 1,
+      skipped: 2,
+      errors: [],
+      createdRoles: [],
+      skippedRoles: [],
+    });
     expect(notificationsShowMock).toHaveBeenCalledTimes(1);
     expect(notificationsShowMock).toHaveBeenCalledWith({
       message: "Imported 3, renamed 1, skipped 2",
@@ -93,6 +100,8 @@ describe("useImportAiRolesFromCatalogMutation — success notifications", () =>
         { slug: "a", message: "name taken" },
         { slug: "b", message: "name taken" },
       ],
+      createdRoles: [{ slug: "ok", name: "Ok" }],
+      skippedRoles: [],
     });
     expect(notificationsShowMock).toHaveBeenCalledTimes(2);
     expect(notificationsShowMock).toHaveBeenNthCalledWith(1, {

apps/client/src/features/ai-chat/types/ai-chat.types.ts

@@ -108,12 +108,25 @@ export interface IAiRoleImportPayload {
   conflict: "skip" | "rename";
 }
 
-/** Import result counts (mirrors `importFromCatalog()`). */
+/**
+ * Import result (mirrors `importFromCatalog()`). The counters (`created`,
+ * `skipped`, `renamed`) drive the summary notification; the per-role lists
+ * (`createdRoles`, `skippedRoles`) drive the redesigned catalog modal's inline
+ * result plaque — which roles were installed (and any rename) and which were
+ * skipped and why (so the plaque can name the conflicting role and offer
+ * "Rename & install").
+ */
 export interface IAiRoleImportResult {
   created: number;
   skipped: number;
   renamed: number;
   errors: { slug: string; message: string }[];
+  createdRoles: { slug: string; name: string; renamedTo?: string }[];
+  skippedRoles: {
+    slug: string;
+    name: string;
+    reason: "name-conflict" | "already-installed";
+  }[];
 }
 
 /**

apps/client/src/features/ai-chat/utils/catalog-bundle-model.test.ts

@@ -0,0 +1,234 @@
+import { describe, it, expect } from "vitest";
+import {
+  bundleCounts,
+  bundlePhase,
+  installedLangForRole,
+  mapBundleRolesToView,
+  mapCatalogRoleToView,
+  nameConflictSlugs,
+  partialOffersRename,
+  type CatalogViewRole,
+} from "./catalog-bundle-model.ts";
+import type {
+  IAiRole,
+  IAiRoleCatalogRole,
+} from "@/features/ai-chat/types/ai-chat.types.ts";
+
+function installedRole(
+  source: { slug: string; language: string; version: number },
+  overrides: Partial<IAiRole> = {},
+): IAiRole {
+  return {
+    id: `role-${source.slug}-${source.language}`,
+    name: source.slug,
+    emoji: null,
+    description: null,
+    enabled: true,
+    autoStart: true,
+    launchMessage: null,
+    source,
+    ...overrides,
+  };
+}
+
+function catalogRole(
+  overrides: Partial<IAiRoleCatalogRole> = {},
+): IAiRoleCatalogRole {
+  return {
+    slug: "writer",
+    emoji: "✍️",
+    name: "Writer",
+    description: "Drafts copy.",
+    instructions: "be a writer",
+    autoStart: true,
+    launchMessage: null,
+    version: 3,
+    ...overrides,
+  };
+}
+
+// Build a minimal view role for bundlePhase tests.
+function viewRole(status: CatalogViewRole["status"]): CatalogViewRole {
+  return { slug: `s-${status}`, name: status, description: "", version: 1, status };
+}
+
+describe("bundlePhase", () => {
+  it("empty bundle -> empty", () => {
+    expect(bundlePhase([])).toBe("empty");
+  });
+
+  it("all importable, none installed -> allNew", () => {
+    expect(bundlePhase([viewRole("import"), viewRole("import")])).toBe(
+      "allNew",
+    );
+  });
+
+  it("nothing to import or update -> allInstalled", () => {
+    expect(bundlePhase([viewRole("installed"), viewRole("installed")])).toBe(
+      "allInstalled",
+    );
+  });
+
+  it("updates present, nothing to import -> updates", () => {
+    expect(bundlePhase([viewRole("update"), viewRole("installed")])).toBe(
+      "updates",
+    );
+  });
+
+  it("import + installed (no updates) -> mixed", () => {
+    expect(bundlePhase([viewRole("import"), viewRole("installed")])).toBe(
+      "mixed",
+    );
+  });
+
+  it("import + update -> mixed", () => {
+    expect(bundlePhase([viewRole("import"), viewRole("update")])).toBe("mixed");
+  });
+
+  it("a skipped role with nothing installed -> mixed (NOT allInstalled)", () => {
+    // F1: a bundle whose only non-installed role was skipped has 0 installed for
+    // it, so the collapsed 'All installed · up to date' header would contradict
+    // the open 'Installed 0 · 1 skipped' plaque. It must be mixed until resolved.
+    expect(bundlePhase([viewRole("skipped")])).toBe("mixed");
+  });
+
+  it("installed + a skipped role -> mixed (partial success is not allInstalled)", () => {
+    expect(bundlePhase([viewRole("installed"), viewRole("skipped")])).toBe(
+      "mixed",
+    );
+  });
+});
+
+describe("bundleCounts", () => {
+  it("tallies each status once", () => {
+    expect(
+      bundleCounts([
+        viewRole("import"),
+        viewRole("import"),
+        viewRole("installed"),
+        viewRole("update"),
+        viewRole("skipped"),
+      ]),
+    ).toEqual({ importable: 2, installed: 1, update: 1, skipped: 1 });
+  });
+});
+
+describe("nameConflictSlugs / partialOffersRename (reason -> action)", () => {
+  it("only name-conflict skips become the transient overlay / offer rename", () => {
+    const skipped = [
+      { slug: "writer", name: "Writer", reason: "name-conflict" as const },
+      { slug: "editor", name: "Editor", reason: "already-installed" as const },
+    ];
+    expect(nameConflictSlugs(skipped)).toEqual(["writer"]);
+    expect(partialOffersRename(skipped)).toBe(true);
+  });
+
+  it("an already-installed-only skip is informational: no overlay, no rename", () => {
+    const skipped = [
+      { slug: "editor", name: "Editor", reason: "already-installed" as const },
+    ];
+    expect(nameConflictSlugs(skipped)).toEqual([]);
+    expect(partialOffersRename(skipped)).toBe(false);
+  });
+});
+
+describe("installedLangForRole", () => {
+  it("returns the other language when the same slug is installed elsewhere", () => {
+    const roles = [installedRole({ slug: "writer", language: "ru", version: 2 })];
+    expect(installedLangForRole("writer", roles, "en")).toBe("ru");
+  });
+
+  it("returns undefined when the same slug is installed in the SAME language", () => {
+    const roles = [installedRole({ slug: "writer", language: "en", version: 2 })];
+    expect(installedLangForRole("writer", roles, "en")).toBeUndefined();
+  });
+
+  it("returns undefined when no install of the slug exists", () => {
+    expect(installedLangForRole("writer", [], "en")).toBeUndefined();
+  });
+
+  it("ignores manually-created roles (no source)", () => {
+    const roles = [
+      installedRole({ slug: "writer", language: "ru", version: 2 }, {
+        source: null,
+      }),
+    ];
+    expect(installedLangForRole("writer", roles, "en")).toBeUndefined();
+  });
+});
+
+describe("mapCatalogRoleToView", () => {
+  it("no install -> import status, catalog version, emoji preserved", () => {
+    const view = mapCatalogRoleToView(catalogRole(), [], "en");
+    expect(view).toMatchObject({
+      slug: "writer",
+      emoji: "✍️",
+      name: "Writer",
+      description: "Drafts copy.",
+      status: "import",
+      version: 3,
+    });
+    expect(view.installedRoleId).toBeUndefined();
+    expect(view.installedLang).toBeUndefined();
+  });
+
+  it("import with the slug installed in another language -> installedLang set", () => {
+    const roles = [installedRole({ slug: "writer", language: "ru", version: 9 })];
+    const view = mapCatalogRoleToView(catalogRole(), roles, "en");
+    expect(view.status).toBe("import");
+    expect(view.installedLang).toBe("ru");
+  });
+
+  it("installed (up to date) -> installed status, catalog version, installedRoleId", () => {
+    const installed = installedRole({
+      slug: "writer",
+      language: "en",
+      version: 3,
+    });
+    const view = mapCatalogRoleToView(catalogRole(), [installed], "en");
+    expect(view).toMatchObject({
+      status: "installed",
+      version: 3,
+      installedRoleId: installed.id,
+    });
+  });
+
+  it("update -> version=from, newVersion=to, installedRoleId", () => {
+    const installed = installedRole({
+      slug: "writer",
+      language: "en",
+      version: 1,
+    });
+    const view = mapCatalogRoleToView(catalogRole(), [installed], "en");
+    expect(view).toMatchObject({
+      status: "update",
+      version: 1,
+      newVersion: 3,
+      installedRoleId: installed.id,
+    });
+  });
+
+  it("missing emoji -> emoji undefined; null description -> empty string", () => {
+    const view = mapCatalogRoleToView(
+      catalogRole({ emoji: null, description: null }),
+      [],
+      "en",
+    );
+    expect(view.emoji).toBeUndefined();
+    expect(view.description).toBe("");
+  });
+});
+
+describe("mapBundleRolesToView", () => {
+  it("maps a bundle's roles preserving order", () => {
+    const roles = [
+      catalogRole({ slug: "a", name: "A", version: 1 }),
+      catalogRole({ slug: "b", name: "B", version: 1 }),
+    ];
+    const installed = [installedRole({ slug: "a", language: "en", version: 1 })];
+    const view = mapBundleRolesToView(roles, installed, "en");
+    expect(view.map((r) => r.slug)).toEqual(["a", "b"]);
+    expect(view[0].status).toBe("installed");
+    expect(view[1].status).toBe("import");
+  });
+});

apps/client/src/features/ai-chat/utils/catalog-bundle-model.ts

@@ -0,0 +1,206 @@
+import type {
+  IAiRole,
+  IAiRoleCatalogRole,
+} from "@/features/ai-chat/types/ai-chat.types.ts";
+import { catalogRoleInstallState } from "@/features/ai-chat/utils/catalog-role-install-state.ts";
+
+/**
+ * The redesigned catalog modal renders bundles as cards with a summary status
+ * (readable without expanding) and a single primary action. The per-role and
+ * per-bundle view model that drives that UI is derived here as PURE functions so
+ * the mapping, the "installed in another language" hint, and the bundle-phase
+ * computation are unit-testable without mounting the component (mirrors the
+ * `catalogRoleInstallState` precedent).
+ */
+
+/**
+ * A role's status in the catalog view model.
+ *  - `import`    — not installed in the current content language.
+ *  - `installed` — installed and up to date.
+ *  - `update`    — installed, but the catalog ships a newer version.
+ *  - `skipped`   — TRANSIENT client-only status set after a conflicted import
+ *                  (a name collision under `conflict:'skip'`); never from the
+ *                  backend.
+ */
+export type RoleStatus = "import" | "installed" | "update" | "skipped";
+
+/** A catalog role mapped into the modal's view model. */
+export interface CatalogViewRole {
+  // Slug is the stable identity within a bundle; used as the row key and as the
+  // `slugs[]` payload for import.
+  slug: string;
+  // Optional in the catalog — the row reserves space and renders nothing when
+  // absent.
+  emoji?: string;
+  name: string;
+  description: string;
+  // For `installed`/`import`: the catalog version. For `update`: the installed
+  // (from) version, with `newVersion` holding the catalog (to) version.
+  version: number;
+  newVersion?: number;
+  status: RoleStatus;
+  // The language a same-slug role is installed under, when it differs from the
+  // current content language (drives the Р5 hint). Only set for `import` roles.
+  installedLang?: string;
+  // The workspace role id, present for `installed`/`update` — needed to call the
+  // update-from-catalog mutation.
+  installedRoleId?: string;
+}
+
+/**
+ * The summary phase of a bundle, derived from its roles' statuses. Determines
+ * the collapsed-header summary and the bundle's single primary action.
+ *  - `empty`        — the bundle has no roles.
+ *  - `allNew`       — everything is importable, nothing installed.
+ *  - `allInstalled` — everything installed & up to date; nothing else pending.
+ *  - `updates`      — updates available and nothing left to import.
+ *  - `mixed`        — any other combination.
+ */
+export type BundlePhase =
+  | "empty"
+  | "allNew"
+  | "allInstalled"
+  | "updates"
+  | "mixed";
+
+/** Per-status tallies for a bundle's roles (the single source of truth). */
+export interface BundleCounts {
+  importable: number;
+  installed: number;
+  update: number;
+  skipped: number;
+}
+
+/**
+ * Count a bundle's roles by status ONCE. Both `bundlePhase` and the panel derive
+ * from this, so the tally logic lives in exactly one place (no rescans / drift).
+ */
+export function bundleCounts(roles: CatalogViewRole[]): BundleCounts {
+  const counts: BundleCounts = {
+    importable: 0,
+    installed: 0,
+    update: 0,
+    skipped: 0,
+  };
+  for (const r of roles) {
+    if (r.status === "import") counts.importable += 1;
+    else if (r.status === "installed") counts.installed += 1;
+    else if (r.status === "update") counts.update += 1;
+    else if (r.status === "skipped") counts.skipped += 1;
+  }
+  return counts;
+}
+
+export function bundlePhase(roles: CatalogViewRole[]): BundlePhase {
+  if (roles.length === 0) return "empty";
+  const { importable, installed, update, skipped } = bundleCounts(roles);
+  // A `skipped` role is a pending post-import conflict (0 installed for it), so a
+  // bundle that has ANY skipped role is NOT "all installed & up to date" — that
+  // would make the collapsed green "up to date" header contradict the open
+  // panel's "Installed 0 · 1 skipped" plaque. It is `mixed` until resolved.
+  if (importable === 0 && update === 0 && skipped === 0) return "allInstalled";
+  if (update > 0 && importable === 0 && skipped === 0) return "updates";
+  if (importable > 0 && installed === 0 && update === 0 && skipped === 0)
+    return "allNew";
+  return "mixed";
+}
+
+/**
+ * The subset of a skip result that should be shown as a TRANSIENT `skipped`
+ * overlay in the bundle (so the row offers a re-import path). Only NAME-CONFLICT
+ * skips qualify: an `already-installed` skip (a concurrent-import race) has
+ * nothing to act on — re-importing the same slug would just skip again — so it
+ * must NOT be overlaid (else the row shows a misleading "Rename & install" that
+ * self-heals into a false "installed"). Pure so both reason branches are tested.
+ */
+export function nameConflictSlugs(
+  skipped: { slug: string; reason: "name-conflict" | "already-installed" }[],
+): string[] {
+  return skipped
+    .filter((s) => s.reason === "name-conflict")
+    .map((s) => s.slug);
+}
+
+/**
+ * Whether a partial-import result should offer the "Rename & install" action:
+ * only when at least one skip is a name conflict (renameable). An
+ * `already-installed`-only partial is informational.
+ */
+export function partialOffersRename(
+  skipped: { reason: "name-conflict" | "already-installed" }[],
+): boolean {
+  return skipped.some((s) => s.reason === "name-conflict");
+}
+
+/**
+ * For a role NOT installed in the current `language`, find a workspace role with
+ * the same catalog `slug` installed under a DIFFERENT language, and return that
+ * language. Drives the "installed in another language" hint (Р5): a different
+ * language of the same slug is a separate install and appears as `import`.
+ */
+export function installedLangForRole(
+  slug: string,
+  workspaceRoles: IAiRole[],
+  language: string,
+): string | undefined {
+  const other = workspaceRoles.find(
+    (r) =>
+      r.source?.slug === slug &&
+      !!r.source?.language &&
+      r.source.language !== language,
+  );
+  return other?.source?.language;
+}
+
+/**
+ * Map one catalog role to the view model, computing its install status against
+ * the workspace roles (via `catalogRoleInstallState`) and, for importable roles,
+ * the other-language hint.
+ */
+export function mapCatalogRoleToView(
+  role: IAiRoleCatalogRole,
+  workspaceRoles: IAiRole[],
+  language: string,
+): CatalogViewRole {
+  const state = catalogRoleInstallState(role, workspaceRoles, language);
+  const base = {
+    slug: role.slug,
+    emoji: role.emoji ?? undefined,
+    name: role.name,
+    description: role.description ?? "",
+  };
+  if (state.state === "update") {
+    return {
+      ...base,
+      status: "update",
+      version: state.fromVersion,
+      newVersion: state.toVersion,
+      installedRoleId: state.installed.id,
+    };
+  }
+  if (state.state === "installed") {
+    return {
+      ...base,
+      status: "installed",
+      version: role.version,
+      installedRoleId: state.installed.id,
+    };
+  }
+  return {
+    ...base,
+    status: "import",
+    version: role.version,
+    installedLang: installedLangForRole(role.slug, workspaceRoles, language),
+  };
+}
+
+/**
+ * Map a whole bundle's catalog roles to the view model, preserving order.
+ */
+export function mapBundleRolesToView(
+  roles: IAiRoleCatalogRole[],
+  workspaceRoles: IAiRole[],
+  language: string,
+): CatalogViewRole[] {
+  return roles.map((r) => mapCatalogRoleToView(r, workspaceRoles, language));
+}

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

@@ -24,6 +24,7 @@ import {
   GitmostListPagesResult,
   GitmostListSpacesResult,
   gitmostDecodePayloadToFile,
+  gitmostInsertTranscriptIntoEditor,
   gitmostUploadFileToEditor,
 } from "@/features/editor/gitmost/gitmost-recording.ts";
 
@@ -281,6 +282,18 @@ export default function GitmostGlobalBridge() {
             pageId: page.id,
           };
         }
+
+        // Best-effort: append the transcript (heading + one paragraph per line)
+        // below the just-inserted audio node. The audio insert already
+        // succeeded, so a transcript failure must NOT turn this into an error —
+        // wrap it and, on any throw, log and still return ok. A missing/empty/
+        // non-string transcript is a no-op inside the helper (audio only).
+        try {
+          gitmostInsertTranscriptIntoEditor(editor, payload?.transcript);
+        } catch (err) {
+          console.error("[gitmost] transcript insert failed", err);
+        }
+
         return { ok: true, pageId: page.id };
       } catch (err: any) {
         console.error("[gitmost] createPageWithRecording failed", err);

apps/client/src/features/editor/gitmost/gitmost-recording.test.ts

@@ -0,0 +1,150 @@
+import { describe, it, expect } from "vitest";
+import { Editor } from "@tiptap/core";
+import { Document } from "@tiptap/extension-document";
+import { Paragraph } from "@tiptap/extension-paragraph";
+import { Text } from "@tiptap/extension-text";
+import { Heading } from "@tiptap/extension-heading";
+import { Bold } from "@tiptap/extension-bold";
+import { Italic } from "@tiptap/extension-italic";
+import { Link } from "@tiptap/extension-link";
+import { gitmostInsertTranscriptIntoEditor } from "./gitmost-recording.ts";
+
+const ZWSP = "​"; // U+200B, the helper's block-trigger neutralizer
+
+/**
+ * #377 — the web-side bridge must append the native host's transcript below the
+ * recording. These exercise the pure insert helper through a REAL Tiptap editor
+ * (Document/Paragraph/Text/Heading + Bold/Italic/Link marks so an HTML-parsing
+ * regression would be caught), asserting the resulting document rather than
+ * mocking the editor: transcript present -> "Transcript" heading + one paragraph
+ * per non-empty line; content is inserted as LITERAL TEXT (no HTML/markdown
+ * parsing); col-0 markdown block triggers are neutralized so git-sync keeps them
+ * paragraphs; absent/empty/non-string -> no-op.
+ */
+describe("gitmostInsertTranscriptIntoEditor", () => {
+  const makeEditor = () =>
+    new Editor({
+      // Bold/Italic/Link are registered specifically so that IF the helper ever
+      // regressed to inserting an HTML/markdown string (instead of a text node),
+      // TipTap would parse `<b>`/`*..*`/`[..](..)` into marks and the literal-
+      // text assertions below would fail.
+      extensions: [Document, Paragraph, Text, Heading, Bold, Italic, Link],
+      // Start from a single empty paragraph (a fresh page's baseline). The
+      // helper appends at the end of the doc, i.e. below existing content.
+      content: { type: "doc", content: [{ type: "paragraph" }] },
+    });
+
+  it("inserts a Transcript heading + one paragraph per non-empty line, verbatim", () => {
+    const editor = makeEditor();
+
+    const inserted = gitmostInsertTranscriptIntoEditor(
+      editor,
+      "You: hello there\nSpeaker 1: hi\n\nYou: bye",
+    );
+
+    expect(inserted).toBe(true);
+
+    const nodes = (editor.getJSON().content ?? []) as any[];
+    // A level-2 "Transcript" heading is present.
+    const heading = nodes.find((n) => n.type === "heading");
+    expect(heading?.attrs?.level).toBe(2);
+    expect(heading?.content?.[0]?.text).toBe("Transcript");
+
+    // Every non-empty transcript line becomes a paragraph, in order, verbatim;
+    // the blank line between them is dropped.
+    const texts = nodes
+      .filter((n) => n.type === "paragraph")
+      .map((n) => n.content?.[0]?.text)
+      .filter((t) => typeof t === "string");
+    expect(texts).toEqual(["You: hello there", "Speaker 1: hi", "You: bye"]);
+
+    editor.destroy();
+  });
+
+  it("inserts HTML + markdown metacharacters as LITERAL text (no injection / no mark parsing)", () => {
+    const editor = makeEditor();
+    const line =
+      "You: <b>bold</b> <script>alert(1)</script> and *stars* and [link](x)";
+
+    const inserted = gitmostInsertTranscriptIntoEditor(editor, line);
+    expect(inserted).toBe(true);
+
+    const paras = (editor.getJSON().content ?? []).filter(
+      (n: any) => n.type === "paragraph",
+    ) as any[];
+    // The transcript line is exactly ONE paragraph holding a SINGLE text node
+    // whose text is the verbatim string — not split into bold/link/other nodes,
+    // not carrying any marks, not raw HTML. This FAILS if the helper switched to
+    // insertContent(htmlString): TipTap would then parse <b>/[link](x)/*stars*.
+    const content = paras[paras.length - 1].content;
+    expect(content).toHaveLength(1);
+    expect(content[0].type).toBe("text");
+    expect(content[0].marks ?? []).toEqual([]);
+    expect(content[0].text).toBe(line);
+    // And no bold/italic/link mark exists anywhere in the document.
+    const html = editor.getHTML();
+    expect(html).not.toMatch(/<(strong|b|em|i|a)\b/);
+    // The angle brackets survived as escaped entities (literal text), not a live
+    // <script>/<b> element.
+    expect(html).not.toMatch(/<script/i);
+
+    editor.destroy();
+  });
+
+  it("neutralizes col-0 markdown block triggers with a leading ZWSP (git-sync safety)", () => {
+    const editor = makeEditor();
+    // Trigger lines (some with a leaked indent) + a normal prefixed line.
+    const inserted = gitmostInsertTranscriptIntoEditor(
+      editor,
+      [
+        "- dash",
+        "  > quote", // leading indent must be trimmed then neutralized
+        "# hash",
+        "1. one",
+        "> [!info] note",
+        "```js",
+        "---", // solid thematic break -> horizontalRule (text-losing) if unneutralized
+        "***",
+        "___",
+        "You: normal line",
+      ].join("\n"),
+    );
+    expect(inserted).toBe(true);
+
+    const texts = (editor.getJSON().content ?? [])
+      .filter((n: any) => n.type === "paragraph")
+      .map((n: any) => n.content?.[0]?.text)
+      .filter((t: any) => typeof t === "string") as string[];
+
+    // Every block-trigger line is prefixed with the invisible ZWSP (indent
+    // trimmed first); the normal `You:` line is left byte-exact.
+    expect(texts).toEqual([
+      ZWSP + "- dash",
+      ZWSP + "> quote",
+      ZWSP + "# hash",
+      ZWSP + "1. one",
+      ZWSP + "> [!info] note",
+      ZWSP + "```js",
+      ZWSP + "---",
+      ZWSP + "***",
+      ZWSP + "___",
+      "You: normal line",
+    ]);
+
+    editor.destroy();
+  });
+
+  it("is a no-op for undefined / empty / whitespace-only / non-string transcripts", () => {
+    for (const value of [undefined, "", "   \n  \n", 42, {}, null]) {
+      const editor = makeEditor();
+      const before = JSON.stringify(editor.getJSON());
+
+      const inserted = gitmostInsertTranscriptIntoEditor(editor, value as any);
+
+      expect(inserted).toBe(false);
+      // Document is untouched (audio-only behavior preserved).
+      expect(JSON.stringify(editor.getJSON())).toBe(before);
+      editor.destroy();
+    }
+  });
+});

apps/client/src/features/editor/gitmost/gitmost-recording.ts

@@ -65,6 +65,11 @@ export interface GitmostCreatePagePayload {
   base64: string;
   filename: string;
   mimeType: string;
+  // Optional transcript for the recording: plain text, `\n`-separated, each
+  // line already formatted as `You: ...` / `Speaker N: ...` by the native host
+  // (ready to insert, no parsing needed). Omitted (no speech / no models) ->
+  // audio only.
+  transcript?: string;
 }
 
 export interface GitmostCreatePageResult {
@@ -235,6 +240,83 @@ export async function gitmostUploadFileToEditor(
   }
 }
 
+// Zero-width space (U+200B). Prepended to a transcript line that begins with a
+// markdown BLOCK trigger: it is invisible in the rendered doc but shifts the
+// trigger off column 0, so the git-sync doc->markdown->doc round-trip keeps the
+// line a plain paragraph (see GITMOST_MD_BLOCK_TRIGGER_RE).
+const GITMOST_ZWSP = "​";
+
+// A markdown BLOCK-level construct that, sitting at column 0 of a paragraph
+// line, the git-sync markdown serializer (packages/prosemirror-markdown
+// markdown-converter.ts, `case "paragraph"`) would re-parse into a NON-paragraph
+// block on the doc->markdown->doc cycle. That serializer emits paragraph text
+// verbatim with NO block-escape (the pre-existing root cause), so a leading
+// `#`/`-`/`*`/`+`/`>`, an ordered-list `N.`/`N)`, a code fence ```/~~~, a table
+// `|`, or a `> [!info]` callout opener would silently become a heading / list /
+// quote / code block / table / callout. The final alternative matches a WHOLE-
+// LINE thematic break — solid `---`/`***`/`___` or spaced `- - -`/`_ _ _` (3+ of
+// the same `-`/`*`/`_`) — which round-trips into a `horizontalRule`; because
+// that node carries NO text, an un-neutralized separator line would LOSE its
+// text entirely (worse than the list/quote case). This matches a TRIMMED line's
+// start; the transcript's own `You:` / `Speaker N:` prefix begins with a letter
+// and never matches, so prefixed lines are left byte-exact.
+const GITMOST_MD_BLOCK_TRIGGER_RE =
+  /^(?:#{1,6}(?:\s|$)|[-*+](?:\s|$)|>|\d+[.)](?:\s|$)|```|~~~|\||([-*_])(?:\s*\1){2,}\s*$)/;
+
+// Append a transcript block BELOW the recording's audio node in a live editor:
+// a "Transcript" heading followed by one paragraph per non-empty transcript
+// line. The transcript is plain text, `\n`-separated, each line already
+// formatted as `You: ...` / `Speaker N: ...` by the native host — line text is
+// inserted as a TEXT node (never HTML/markdown), so there is no injection or
+// mark-parsing surface. Each kept line is trimmed (drops an indent that would
+// both leak into the display and, at col 0, form a markdown block trigger) and,
+// if it still begins with a col-0 markdown block trigger, gets an invisible
+// zero-width space prepended so the git-sync round-trip cannot turn it into a
+// list/quote/heading/callout/code/table (defensive boundary against the
+// serializer's missing block-escape). This is best-effort and meant to run
+// AFTER the audio has already been inserted; the caller must guard against a
+// throw so a transcript failure never fails the (already successful) recording.
+// Returns true when a block was inserted, false when there was nothing to
+// insert (transcript undefined/empty/not-a-string). A non-string value is a
+// no-op, not an error.
+export function gitmostInsertTranscriptIntoEditor(
+  editor: Editor,
+  transcript: unknown,
+): boolean {
+  if (typeof transcript !== "string") return false;
+  const lines = transcript
+    .split("\n")
+    // Trim each line and drop blank (whitespace-only) ones.
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0)
+    // Neutralize a col-0 markdown block trigger with an invisible ZWSP so the
+    // git-sync round-trip keeps the line a paragraph. Host lines (`You:` /
+    // `Speaker N:`) never match and stay byte-exact.
+    .map((line) =>
+      GITMOST_MD_BLOCK_TRIGGER_RE.test(line) ? GITMOST_ZWSP + line : line,
+    );
+  if (lines.length === 0) return false;
+
+  const content = [
+    {
+      type: "heading",
+      attrs: { level: 2 },
+      content: [{ type: "text", text: "Transcript" }],
+    },
+    ...lines.map((line) => ({
+      type: "paragraph",
+      content: [{ type: "text", text: line }],
+    })),
+  ];
+
+  // Append at the end of the document. On a freshly-created recording page the
+  // audio node is the last block, so the end position places the transcript
+  // directly below it.
+  const endPos = editor.state.doc.content.size;
+  editor.chain().focus().insertContentAt(endPos, content).run();
+  return true;
+}
+
 // Full insert path used by the open-page bridge (insertRecording): guard the
 // editor, validate/decode the payload, then upload. Never throws — resolves to
 // a result code.

apps/server/package.json

@@ -23,7 +23,7 @@
     "migration:reset": "tsx src/database/migrate.ts down-to NO_MIGRATIONS",
     "migration:codegen": "kysely-codegen --dialect=postgres --camel-case --env-file=../../.env --out-file=./src/database/types/db.d.ts",
     "lint": "eslint \"{src,apps,libs,test}/**/*.ts\" --fix",
-    "pretest": "pnpm --filter @docmost/editor-ext build",
+    "pretest": "pnpm --filter @docmost/editor-ext build && pnpm --filter @docmost/prosemirror-markdown build",
     "test": "jest",
     "test:int": "jest --config test/jest-integration.json",
     "test:watch": "jest --watch",
@@ -43,6 +43,7 @@
     "@clickhouse/client": "^1.18.2",
     "@docmost/mcp": "workspace:*",
     "@docmost/pdf-inspector": "1.9.6",
+    "@docmost/prosemirror-markdown": "workspace:*",
     "@fastify/cookie": "^11.0.2",
     "@fastify/multipart": "^10.0.0",
     "@fastify/static": "^9.1.3",
@@ -175,7 +176,7 @@
       "/node_modules/"
     ],
     "transform": {
-      "happy-dom.+\\.js$": [
+      "(happy-dom.+|prosemirror-markdown/build/.+)\\.js$": [
         "babel-jest",
         {
           "presets": [
@@ -193,7 +194,7 @@
       "^.+\\.(t|j)sx?$": "ts-jest"
     },
     "transformIgnorePatterns": [
-      "/node_modules/(?!(\\.pnpm/)?(nanoid|uuid|image-dimensions|marked|happy-dom|lib0)(@|/))"
+      "/node_modules/(?!(\\.pnpm/)?(nanoid|uuid|image-dimensions|marked|happy-dom|lib0|@docmost/prosemirror-markdown)(@|/))"
     ],
     "collectCoverageFrom": [
       "**/*.(t|j)s"
@@ -204,7 +205,8 @@
       "^@docmost/db/(.*)$": "<rootDir>/database/$1",
       "^@docmost/transactional/(.*)$": "<rootDir>/integrations/transactional/$1",
       "^@docmost/ee/(.*)$": "<rootDir>/ee/$1",
-      "^src/(.*)$": "<rootDir>/$1"
+      "^src/(.*)$": "<rootDir>/$1",
+      "^@tiptap/react$": "<rootDir>/../test/stubs/tiptap-react.js"
     }
   }
 }

apps/server/src/collaboration/collaboration.util.ts

@@ -43,7 +43,6 @@ import {
   Column,
   Status,
   addUniqueIdsToDoc,
-  htmlToMarkdown,
   TransclusionSource,
   TransclusionReference,
   FootnoteReference,
@@ -51,6 +50,7 @@ import {
   FootnoteDefinition,
   PageEmbed,
 } from '@docmost/editor-ext';
+import { convertProseMirrorToMarkdown } from '@docmost/prosemirror-markdown';
 import { generateText, getSchema, JSONContent } from '@tiptap/core';
 import { generateHTML, generateJSON } from '../common/helpers/prosemirror/html';
 // @tiptap/html library works best for generating prosemirror json state but not HTML
@@ -239,6 +239,10 @@ export function prosemirrorNodeToYElement(node: any): Y.XmlElement | Y.XmlText {
 }
 
 export function jsonToMarkdown(tiptapJson: any): string {
-  const html = jsonToHtml(tiptapJson);
-  return htmlToMarkdown(html);
+  // Direct ProseMirror JSON -> Markdown via the canonical converter
+  // (`@docmost/prosemirror-markdown`) — no HTML intermediate, no second
+  // editor-ext markdown layer. Same serializer as the page/space export and the
+  // git-sync vault writer, so every server PM->MD path emits identical canonical
+  // markdown (issue #345).
+  return convertProseMirrorToMarkdown(tiptapJson);
 }

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

@@ -610,6 +610,63 @@ describe('AiAgentRolesService guards', () => {
       expect(repo.insert.mock.calls[0][0].name).toBe('Researcher (2)');
     });
 
+    it('createdRoles lists the installed role (no renamedTo when not renamed)', async () => {
+      const { service } = makeImportService({});
+      const res = await service.importFromCatalog('ws-1', 'u1', dto());
+      expect(res.createdRoles).toEqual([
+        { slug: 'researcher', name: 'Researcher' },
+      ]);
+      expect(res.skippedRoles).toEqual([]);
+    });
+
+    it('createdRoles carries renamedTo on a rename', async () => {
+      const existing = [makeRow({ id: 'r-x', name: 'Researcher' })];
+      const { service } = makeImportService({ existing });
+      const res = await service.importFromCatalog(
+        'ws-1',
+        'u1',
+        dto({ conflict: 'rename' }),
+      );
+      expect(res.createdRoles).toEqual([
+        { slug: 'researcher', name: 'Researcher', renamedTo: 'Researcher (2)' },
+      ]);
+      expect(res.skippedRoles).toEqual([]);
+    });
+
+    it('skippedRoles: already-installed slug carries reason "already-installed"', async () => {
+      const existing = [
+        makeRow({
+          id: 'r-existing',
+          name: 'Old researcher',
+          source: { slug: 'researcher', language: 'en', version: 1 } as never,
+        }),
+      ];
+      const { service } = makeImportService({ existing });
+      const res = await service.importFromCatalog('ws-1', 'u1', dto());
+      expect(res.skippedRoles).toEqual([
+        {
+          slug: 'researcher',
+          name: 'Researcher',
+          reason: 'already-installed',
+        },
+      ]);
+      expect(res.createdRoles).toEqual([]);
+    });
+
+    it('skippedRoles: a name collision under conflict:skip carries reason "name-conflict"', async () => {
+      const existing = [makeRow({ id: 'r-x', name: 'Researcher' })];
+      const { service } = makeImportService({ existing });
+      const res = await service.importFromCatalog(
+        'ws-1',
+        'u1',
+        dto({ conflict: 'skip' }),
+      );
+      expect(res.skippedRoles).toEqual([
+        { slug: 'researcher', name: 'Researcher', reason: 'name-conflict' },
+      ]);
+      expect(res.createdRoles).toEqual([]);
+    });
+
     it('dto.slugs filters; an unknown slug becomes an error entry', async () => {
       const { service, repo } = makeImportService({
         bundleRoles: [catalogRole()],
@@ -677,6 +734,15 @@ describe('AiAgentRolesService guards', () => {
       // 'a' converged on the concurrent install (skip); 'b' imported; no errors.
       expect(res).toMatchObject({ created: 1, skipped: 1, renamed: 0 });
       expect(res.errors).toEqual([]);
+      // The per-role list records 'a' as an already-installed skip (the UI reads
+      // skippedRoles, not the counter, to render its plaque — assert the array,
+      // not just the count).
+      expect(res.skippedRoles).toContainEqual({
+        slug: 'a',
+        name: 'A',
+        reason: 'already-installed',
+      });
+      expect(res.createdRoles.map((r) => r.slug)).toEqual(['b']);
       // Both inserts were attempted (the batch did not abort on the 23505).
       expect(repo.insert).toHaveBeenCalledTimes(2);
     });

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

@@ -305,6 +305,16 @@ export class AiAgentRolesService {
     skipped: number;
     renamed: number;
     errors: { slug: string; message: string }[];
+    // Per-role lists alongside the counters (kept for back-compat). The redesigned
+    // catalog UI needs the actual roles — which were created (and any rename) and
+    // which were skipped and why — to render an inline result plaque with the
+    // conflicting role's name and a "Rename & install" affordance.
+    createdRoles: { slug: string; name: string; renamedTo?: string }[];
+    skippedRoles: {
+      slug: string;
+      name: string;
+      reason: 'name-conflict' | 'already-installed';
+    }[];
   }> {
     const { file, versions } = await this.loadBundleById(
       dto.bundleId,
@@ -312,6 +322,13 @@ export class AiAgentRolesService {
     );
 
     const errors: { slug: string; message: string }[] = [];
+    const createdRoles: { slug: string; name: string; renamedTo?: string }[] =
+      [];
+    const skippedRoles: {
+      slug: string;
+      name: string;
+      reason: 'name-conflict' | 'already-installed';
+    }[] = [];
 
     // Resolve the selected catalog roles (honor dto.slugs; flag unknown ones).
     let selected = file.roles;
@@ -351,16 +368,27 @@ export class AiAgentRolesService {
       // Already installed from the catalog in THIS language => skip (use
       // update-from-catalog). A different language of the same slug still imports.
       const installKey = `${role.slug}:${dto.language}`;
+      const originalName = role.name.trim();
       if (installedKeys.has(installKey)) {
         skipped++;
+        skippedRoles.push({
+          slug: role.slug,
+          name: originalName,
+          reason: 'already-installed',
+        });
         continue;
       }
 
-      let name = role.name.trim();
+      let name = originalName;
       let didRename = false;
       if (takenNames.has(name.toLowerCase())) {
         if (dto.conflict === 'skip') {
           skipped++;
+          skippedRoles.push({
+            slug: role.slug,
+            name: originalName,
+            reason: 'name-conflict',
+          });
           continue;
         }
         // conflict === 'rename': find a free " (N)" suffix.
@@ -380,6 +408,11 @@ export class AiAgentRolesService {
         });
         created++;
         if (didRename) renamed++;
+        createdRoles.push({
+          slug: role.slug,
+          name: originalName,
+          ...(didRename ? { renamedTo: name } : {}),
+        });
         takenNames.add(name.toLowerCase());
         installedKeys.add(installKey);
       } catch (err) {
@@ -391,6 +424,11 @@ export class AiAgentRolesService {
         // skipped (already installed) and continue; do NOT abort or error.
         if (isSourceUniqueViolation(err)) {
           skipped++;
+          skippedRoles.push({
+            slug: role.slug,
+            name: originalName,
+            reason: 'already-installed',
+          });
           installedKeys.add(installKey);
           continue;
         }
@@ -407,7 +445,7 @@ export class AiAgentRolesService {
       }
     }
 
-    return { created, skipped, renamed, errors };
+    return { created, skipped, renamed, errors, createdRoles, skippedRoles };
   }
 
   /**

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

@@ -539,3 +539,115 @@ describe('AiChatToolsService model-friendly input validation (#190)', () => {
     expect(result.error?.message).toContain('parameter "pageId": missing (required)');
   });
 });
+
+/**
+ * #294 F1 — the contract-parity test introspects only the ADVERTISED schema keys
+ * (buildShape), not the execute bodies. Most execs are unchanged pass-throughs,
+ * but two wirings actually CHANGED in the migration and are otherwise untested:
+ *   - movePage now forwards the newly-added optional `position` field to the
+ *     client (client.movePage(pageId, parentPageId, position));
+ *   - the table trio unified its `tableRef` param to `table` and must forward it
+ *     positionally. A field destructured under the wrong name would silently pass
+ *     `undefined` to the client (execute is `any`-cast, so tsc won't catch it).
+ */
+describe('AiChatToolsService #294 changed execute wirings', () => {
+  const calls: Record<string, unknown[][]> = {
+    movePage: [],
+    tableInsertRow: [],
+    tableDeleteRow: [],
+    tableUpdateCell: [],
+  };
+  const fakeClient: Partial<DocmostClientLike> = {
+    movePage: (...args: unknown[]) => {
+      calls.movePage.push(args);
+      return Promise.resolve({ success: true });
+    },
+    tableInsertRow: (...args: unknown[]) => {
+      calls.tableInsertRow.push(args);
+      return Promise.resolve({ ok: true });
+    },
+    tableDeleteRow: (...args: unknown[]) => {
+      calls.tableDeleteRow.push(args);
+      return Promise.resolve({ ok: true });
+    },
+    tableUpdateCell: (...args: unknown[]) => {
+      calls.tableUpdateCell.push(args);
+      return Promise.resolve({ ok: true });
+    },
+  };
+  const tokenServiceStub = {
+    generateAccessToken: jest.fn().mockResolvedValue('access-token'),
+    generateCollabToken: jest.fn().mockResolvedValue('collab-token'),
+  };
+  let service: AiChatToolsService;
+
+  beforeEach(() => {
+    for (const k of Object.keys(calls)) calls[k].length = 0;
+    jest.spyOn(loader, 'loadDocmostMcp').mockResolvedValue(
+      mockLoaded(function () {
+        return fakeClient as DocmostClientLike;
+      } as unknown as loader.DocmostClientCtor),
+    );
+    service = new AiChatToolsService(
+      tokenServiceStub as never,
+      {} as never,
+      {} as never,
+      {} as never,
+      {} as never,
+      {
+        asSink: () => ({ put: jest.fn(), has: jest.fn(), evict: jest.fn() }),
+      } as never,
+    );
+  });
+  afterEach(() => jest.restoreAllMocks());
+
+  const buildTools = () =>
+    service.forUser(
+      { id: 'user-1', email: 'u@example.com', workspaceId: 'ws-1' } as never,
+      'session-1',
+      'ws-1',
+      'chat-1',
+    );
+
+  it('movePage forwards the optional position to the client', async () => {
+    const tools = await buildTools();
+    await tools.movePage.execute(
+      { pageId: 'p1', parentPageId: 'parent1', position: 'a5' } as never,
+      {} as never,
+    );
+    expect(calls.movePage).toEqual([['p1', 'parent1', 'a5']]);
+  });
+
+  it('movePage passes undefined position and null parent when omitted (unchanged behavior)', async () => {
+    const tools = await buildTools();
+    await tools.movePage.execute({ pageId: 'p2' } as never, {} as never);
+    expect(calls.movePage).toEqual([['p2', null, undefined]]);
+  });
+
+  it('tableInsertRow forwards the unified `table` param positionally', async () => {
+    const tools = await buildTools();
+    await tools.tableInsertRow.execute(
+      { pageId: 'p1', table: '#0', cells: ['a', 'b'], index: 2 } as never,
+      {} as never,
+    );
+    expect(calls.tableInsertRow).toEqual([['p1', '#0', ['a', 'b'], 2]]);
+  });
+
+  it('tableDeleteRow forwards `table` positionally', async () => {
+    const tools = await buildTools();
+    await tools.tableDeleteRow.execute(
+      { pageId: 'p1', table: '#0', index: 1 } as never,
+      {} as never,
+    );
+    expect(calls.tableDeleteRow).toEqual([['p1', '#0', 1]]);
+  });
+
+  it('tableUpdateCell forwards `table` positionally', async () => {
+    const tools = await buildTools();
+    await tools.tableUpdateCell.execute(
+      { pageId: 'p1', table: '#0', row: 1, col: 2, text: 'x' } as never,
+      {} as never,
+    );
+    expect(calls.tableUpdateCell).toEqual([['p1', '#0', 1, 2, 'x']]);
+  });
+});

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

@@ -316,50 +316,27 @@ export class AiChatToolsService {
         execute: async () => resolveCurrentPageResult(openedPage),
       }),
 
-      getPage: tool({
-        description:
-          'Fetch a single page as Markdown by its page id. Returns the page ' +
-          'title and its Markdown content. Inline <span data-comment-id> tags ' +
-          'in the markdown are comment highlight anchors (also present for ' +
-          'RESOLVED threads) — treat them as markup, not page text.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id (or slugId) of the page.'),
-        }),
-        execute: async ({ pageId }) => {
-          // getPage(pageId) -> { data: filterPage(page, markdown), success }.
-          const result = await client.getPage(pageId);
-          const data = (result?.data ?? {}) as {
-            title?: string;
-            content?: string;
-          };
-          return {
-            title: data.title ?? '',
-            markdown: typeof data.content === 'string' ? data.content : '',
-          };
-        },
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      // The execute body keeps this layer's { title, markdown } projection.
+      getPage: sharedTool(sharedToolSpecs.getPage, async ({ pageId }) => {
+        // getPage(pageId) -> { data: filterPage(page, markdown), success }.
+        const result = await client.getPage(pageId);
+        const data = (result?.data ?? {}) as {
+          title?: string;
+          content?: string;
+        };
+        return {
+          title: data.title ?? '',
+          markdown: typeof data.content === 'string' ? data.content : '',
+        };
       }),
 
       // --- WRITE tools (all reversible — history/trash; §6.5 / D3) ---
 
-      createPage: tool({
-        description:
-          'Create a new page with a Markdown body in a space, optionally under ' +
-          'a parent page. Returns the new page id and title. Reversible: a page ' +
-          'can be moved to trash later.',
-        inputSchema: modelFriendlyInput({
-          title: z.string().describe('The title of the new page.'),
-          content: z
-            .string()
-            .describe('The page body as Markdown (may be empty).'),
-          spaceId: z
-            .string()
-            .describe('The id of the space to create the page in.'),
-          parentPageId: z
-            .string()
-            .optional()
-            .describe('Optional parent page id to nest the new page under.'),
-        }),
-        execute: async ({ title, content, spaceId, parentPageId }) => {
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      createPage: sharedTool(
+        sharedToolSpecs.createPage,
+        async ({ title, content, spaceId, parentPageId }) => {
           // createPage(title, content, spaceId, parentPageId?) ->
           // { data: filterPage(page, markdown), success }.
           const result = await client.createPage(
@@ -375,7 +352,7 @@ export class AiChatToolsService {
           };
           return { id: data.id ?? data.slugId, title: data.title ?? title };
         },
-      }),
+      ),
 
       updatePageContent: tool({
         description:
@@ -399,115 +376,46 @@ export class AiChatToolsService {
         },
       }),
 
-      renamePage: tool({
-        description:
-          "Rename a page (change its title only; the body is untouched). " +
-          'Reversible: rename back at any time.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page to rename.'),
-          title: z.string().describe('The new title.'),
-        }),
-        execute: async ({ pageId, title }) => {
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      renamePage: sharedTool(
+        sharedToolSpecs.renamePage,
+        async ({ pageId, title }) => {
           // renamePage(pageId, title) -> { success, pageId, title }.
           await client.renamePage(pageId, title);
           return { pageId, title };
         },
-      }),
+      ),
 
-      movePage: tool({
-        description:
-          'Move a page under a new parent page, or to the space root when no ' +
-          'parent is given. Reversible: move it back at any time.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page to move.'),
-          parentPageId: z
-            .string()
-            .nullable()
-            .optional()
-            .describe(
-              'Target parent page id. Null/omitted moves the page to the ' +
-                'space root.',
-            ),
-        }),
-        execute: async ({ pageId, parentPageId }) => {
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      // The shared schema adds the optional `position` field this layer lacked
+      // before; the execute now forwards it (the client already accepted it).
+      movePage: sharedTool(
+        sharedToolSpecs.movePage,
+        async ({ pageId, parentPageId, position }) => {
           // movePage(pageId, parentPageId, position?) -> raw move response.
-          await client.movePage(pageId, parentPageId ?? null);
+          await client.movePage(pageId, parentPageId ?? null, position);
           return { pageId, parentPageId: parentPageId ?? null, moved: true };
         },
+      ),
+
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      // GUARDRAIL (§14 H4) preserved: the shared schema exposes ONLY pageId, so
+      // permanentlyDelete/forceDelete are never part of the input and can never
+      // be forwarded — the agent physically cannot permanently delete a page.
+      deletePage: sharedTool(sharedToolSpecs.deletePage, async ({ pageId }) => {
+        // deletePage(pageId) hits POST /pages/delete with { pageId } only,
+        // which is the soft-delete (trash) path on the server.
+        await client.deletePage(pageId);
+        return { pageId, trashed: true };
       }),
 
-      deletePage: tool({
-        description:
-          'Move a page to the trash (SOFT delete only — fully reversible; the ' +
-          'page can be restored from trash). This NEVER permanently deletes.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page to move to trash.'),
-        }),
-        // GUARDRAIL (§14 H4): the only field ever passed to the client is
-        // pageId. permanentlyDelete/forceDelete are not part of the schema and
-        // are never forwarded, so the agent physically cannot permanently
-        // delete a page through this tool.
-        execute: async ({ pageId }) => {
-          // deletePage(pageId) hits POST /pages/delete with { pageId } only,
-          // which is the soft-delete (trash) path on the server.
-          await client.deletePage(pageId);
-          return { pageId, trashed: true };
-        },
-      }),
-
-      // INTENTIONAL per-transport divergence (not shared): the description is
-      // tuned for the in-app agent (e.g. "retry with a corrected EXACT selection"
-      // and "Reversible via the comment UI"); the standalone MCP `create_comment`
-      // keeps its own wording. Kept per-layer.
-      createComment: tool({
-        description:
-          'Add an INLINE comment to a page, or reply to an existing top-level ' +
-          'comment (one level only — the backend rejects replies to replies). ' +
-          'The comment is anchored inline to the given exact `selection` text ' +
-          '(which gets highlighted); page-level comments are NOT supported. A ' +
-          "new top-level comment REQUIRES a `selection`. Replies inherit the " +
-          "parent's anchor and take no selection. If the call fails with a " +
-          '"selection not found" error, retry with a corrected EXACT selection ' +
-          'copied verbatim from a single paragraph/block. You may also attach a ' +
-          '`suggestedText` proposing a replacement for the `selection` (a human ' +
-          'applies it from the UI); when set, the `selection` must occur exactly ' +
-          'once in the page. Reversible via the comment UI.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page to comment on.'),
-          content: z.string().describe('The comment body as Markdown.'),
-          selection: z
-            .string()
-            .min(1)
-            .max(250)
-            .optional()
-            .describe(
-              'EXACT contiguous text from a SINGLE paragraph/block to anchor ' +
-                '(highlight) the comment on (<=250 chars, avoid spanning across ' +
-                'formatting boundaries). Required for a new top-level comment; ' +
-                'omit only when replying via parentCommentId.',
-            ),
-          parentCommentId: z
-            .string()
-            .optional()
-            .describe(
-              'Optional id of a TOP-LEVEL comment to reply to (one level ' +
-                'of replies only).',
-            ),
-          suggestedText: z
-            .string()
-            .min(1)
-            .max(2000)
-            .optional()
-            .describe(
-              'Optional proposed replacement (PLAIN TEXT) for the `selection`, ' +
-                'applied by a human via the UI (never auto-applied). REQUIRES a ' +
-                '`selection`; NOT allowed on a reply. When set, the `selection` ' +
-                'must be UNIQUE in the page — expand it with surrounding context ' +
-                '(still <=250 chars) if it occurs more than once, or the call is ' +
-                'refused.',
-            ),
-        }),
-        execute: async ({
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      // This layer keeps only its own execute-side guards (require a selection
+      // for a top-level comment; reject suggestedText on a reply / without a
+      // selection) — the schema+description are shared.
+      createComment: sharedTool(
+        sharedToolSpecs.createComment,
+        async ({
           pageId,
           content,
           selection,
@@ -548,26 +456,17 @@ export class AiChatToolsService {
           const data = (result?.data ?? {}) as { id?: string };
           return { commentId: data.id, pageId };
         },
-      }),
+      ),
 
-      resolveComment: tool({
-        description:
-          'Resolve or reopen a top-level comment thread (reversible — toggle ' +
-          'the resolved flag). Only top-level comments can be resolved.',
-        inputSchema: modelFriendlyInput({
-          commentId: z
-            .string()
-            .describe('The id of the top-level comment to resolve/reopen.'),
-          resolved: z
-            .boolean()
-            .describe('true to resolve the thread, false to reopen it.'),
-        }),
-        execute: async ({ commentId, resolved }) => {
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      resolveComment: sharedTool(
+        sharedToolSpecs.resolveComment,
+        async ({ commentId, resolved }) => {
           // resolveComment(commentId, resolved) -> { success, commentId, resolved }.
           await client.resolveComment(commentId, resolved);
           return { commentId, resolved };
         },
-      }),
+      ),
 
       // --- READ tools (added) ---
 
@@ -585,33 +484,12 @@ export class AiChatToolsService {
       // hierarchy mode but is worded for the in-app agent; the standalone MCP
       // `list_pages` carries its own wording. Kept per-layer so each side tunes
       // its own guidance.
-      listPages: tool({
-        description:
-          'List the most recent pages, optionally scoped to a single space. ' +
-          'Returns a bounded list (default 50, max 100). Pass tree:true (with ' +
-          "spaceId) to instead get the space's full page hierarchy as a nested tree.",
-        inputSchema: modelFriendlyInput({
-          spaceId: z
-            .string()
-            .optional()
-            .describe('Optional space id to scope the listing to.'),
-          limit: z
-            .number()
-            .int()
-            .min(1)
-            .max(100)
-            .optional()
-            .describe('Maximum number of pages (1-100).'),
-          tree: z
-            .boolean()
-            .optional()
-            .describe(
-              'When true, return the full page hierarchy of the given space as a nested tree (children arrays) instead of the recent-pages flat list. Requires spaceId; ignores limit.',
-            ),
-        }),
-        execute: async ({ spaceId, limit, tree }) =>
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      listPages: sharedTool(
+        sharedToolSpecs.listPages,
+        async ({ spaceId, limit, tree }) =>
           await client.listPages(spaceId, limit, tree),
-      }),
+      ),
 
       listSidebarPages: tool({
         description:
@@ -656,41 +534,34 @@ export class AiChatToolsService {
           }),
       ),
 
+      // NOT shared (kept inline): the MCP tool name `table_get` is noun-first
+      // while this key is `getTable` (verb-first), breaking the
+      // snake_case(inAppKey) convention the shared registry enforces. Its
+      // reference parameter is still named `table` (was `tableRef`) so it matches
+      // the migrated table row/cell tools below.
       getTable: tool({
         description:
           'Read a table as a matrix of cell texts (plus a parallel cellIds ' +
           'matrix so cells can be addressed for rich edits).',
         inputSchema: modelFriendlyInput({
           pageId: z.string().describe('The id of the page.'),
-          tableRef: z
+          table: z
             .string()
             .describe(
-              '"#<index>" from getOutline, or a block id of any node inside ' +
-                'the table.',
+              '"#<index>" from the page outline, or a block id of any node ' +
+                'inside the table.',
             ),
         }),
-        execute: async ({ pageId, tableRef }) =>
-          await client.getTable(pageId, tableRef),
+        execute: async ({ pageId, table }) =>
+          await client.getTable(pageId, table),
       }),
 
-      listComments: tool({
-        description:
-          'List comments on a page in one call. By DEFAULT only ACTIVE ' +
-          'threads are returned; resolved threads (a resolved top-level ' +
-          'comment and all its replies) are hidden and their count reported ' +
-          'as `resolvedThreadsHidden` so you can re-query with ' +
-          '`includeResolved: true` to see everything. Returns ' +
-          '`{ items, resolvedThreadsHidden }`. Content is returned as Markdown.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page.'),
-          includeResolved: z
-            .boolean()
-            .optional()
-            .describe('default only active threads; true — include resolved'),
-        }),
-        execute: async ({ pageId, includeResolved }) =>
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      listComments: sharedTool(
+        sharedToolSpecs.listComments,
+        async ({ pageId, includeResolved }) =>
           await client.listComments(pageId, includeResolved),
-      }),
+      ),
 
       getComment: tool({
         description: 'Fetch a single comment by id (content as Markdown).',
@@ -700,26 +571,12 @@ export class AiChatToolsService {
         execute: async ({ commentId }) => await client.getComment(commentId),
       }),
 
-      checkNewComments: tool({
-        description:
-          'Find new comments across a space (optionally scoped to a subtree) ' +
-          'created after a given timestamp.',
-        inputSchema: modelFriendlyInput({
-          spaceId: z.string().describe('The id of the space to scan.'),
-          since: z
-            .string()
-            .describe('An ISO-8601 timestamp; only comments created after it.'),
-          parentPageId: z
-            .string()
-            .optional()
-            .describe(
-              'Optional page id to scope the scan to that page and its ' +
-                'descendants.',
-            ),
-        }),
-        execute: async ({ spaceId, since, parentPageId }) =>
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      checkNewComments: sharedTool(
+        sharedToolSpecs.checkNewComments,
+        async ({ spaceId, since, parentPageId }) =>
           await client.checkNewComments(spaceId, since, parentPageId),
-      }),
+      ),
 
       listShares: sharedTool(
         sharedToolSpecs.listShares,
@@ -749,19 +606,14 @@ export class AiChatToolsService {
           await client.diffPageVersions(pageId, from, to),
       ),
 
-      exportPageMarkdown: tool({
-        description:
-          'Export a page to a single self-contained Docmost-flavoured ' +
-          'Markdown file (meta + body + comment threads). Lossless round-trip ' +
-          'with importPageMarkdown.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page to export.'),
-        }),
-        execute: async ({ pageId }) => {
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      exportPageMarkdown: sharedTool(
+        sharedToolSpecs.exportPageMarkdown,
+        async ({ pageId }) => {
           const markdown = await client.exportPageMarkdown(pageId);
           return { markdown };
         },
-      }),
+      ),
 
       // --- WRITE tools (added; reversible via page history/trash) ---
 
@@ -811,28 +663,12 @@ export class AiChatToolsService {
         async ({ pageId, nodeId }) => await client.deleteNode(pageId, nodeId),
       ),
 
-      updatePageJson: tool({
-        description:
-          "Replace a page's body with a full ProseMirror document — a full " +
-          'overwrite — and/or update its title. Minimal example content: ' +
-          '{"type":"doc","content":[{"type":"paragraph","content":' +
-          '[{"type":"text","text":"Hi"}]}]}. The content arg may be a JSON ' +
-          'object or a JSON string (both accepted). Omit content for a ' +
-          'title-only update. Reversible: the previous version is kept in page ' +
-          'history.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page to update.'),
-          content: z
-            .any()
-            .optional()
-            .describe(
-              'Full ProseMirror doc {"type":"doc","content":[...]} (JSON ' +
-                'object or JSON string); omit for a title-only update.',
-            ),
-          title: z.string().optional().describe('Optional new title.'),
-        }),
-        execute: async ({ pageId, content, title }) => {
-          // Parity with the standalone MCP server (index.ts update_page_json):
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      // The execute body keeps this layer's content normalization (parity with
+      // the standalone MCP server, index.ts update_page_json).
+      updatePageJson: sharedTool(
+        sharedToolSpecs.updatePageJson,
+        async ({ pageId, content, title }) => {
           // undefined/null pass through as undefined (title-only / no-op); any
           // string is JSON.parsed (so an empty string "" throws, matching the
           // MCP server); an object is passed through unchanged.
@@ -845,66 +681,29 @@ export class AiChatToolsService {
           }
           return await client.updatePageJson(pageId, doc, title);
         },
-      }),
+      ),
 
-      // NOT in the shared registry: this layer names the table argument
-      // `tableRef`, while the standalone MCP tool names it `table` (index.ts).
-      // Sharing one buildShape would rename a model-facing parameter on one
-      // transport, so the table row/cell tools stay per-layer by design.
-      tableInsertRow: tool({
-        description:
-          'Insert a row of plain-text cells into a table. Reversible via ' +
-          'page history.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page.'),
-          tableRef: z
-            .string()
-            .describe('"#<index>" from getOutline, or a block id in the table.'),
-          cells: z.array(z.string()).describe('The cell texts for the row.'),
-          index: z
-            .number()
-            .int()
-            .optional()
-            .describe('0-based insert position (omit/out-of-range to append).'),
-        }),
-        execute: async ({ pageId, tableRef, cells, index }) =>
-          await client.tableInsertRow(pageId, tableRef, cells, index),
-      }),
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      // The table reference parameter was unified to `table` (was `tableRef`).
+      tableInsertRow: sharedTool(
+        sharedToolSpecs.tableInsertRow,
+        async ({ pageId, table, cells, index }) =>
+          await client.tableInsertRow(pageId, table, cells, index),
+      ),
 
-      // NOT shared — same `tableRef` (here) vs `table` (MCP) parameter-name
-      // divergence as tableInsertRow.
-      tableDeleteRow: tool({
-        description:
-          'Delete a table row at a 0-based index. Reversible via page history.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page.'),
-          tableRef: z
-            .string()
-            .describe('"#<index>" from getOutline, or a block id in the table.'),
-          index: z.number().int().describe('0-based row index to delete.'),
-        }),
-        execute: async ({ pageId, tableRef, index }) =>
-          await client.tableDeleteRow(pageId, tableRef, index),
-      }),
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      tableDeleteRow: sharedTool(
+        sharedToolSpecs.tableDeleteRow,
+        async ({ pageId, table, index }) =>
+          await client.tableDeleteRow(pageId, table, index),
+      ),
 
-      // NOT shared — same `tableRef` (here) vs `table` (MCP) parameter-name
-      // divergence as tableInsertRow.
-      tableUpdateCell: tool({
-        description:
-          'Set the plain-text content of a table cell at [row, col] (0-based). ' +
-          'Reversible via page history.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page.'),
-          tableRef: z
-            .string()
-            .describe('"#<index>" from getOutline, or a block id in the table.'),
-          row: z.number().int().describe('0-based row index.'),
-          col: z.number().int().describe('0-based column index.'),
-          text: z.string().describe('The new cell text.'),
-        }),
-        execute: async ({ pageId, tableRef, row, col, text }) =>
-          await client.tableUpdateCell(pageId, tableRef, row, col, text),
-      }),
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      tableUpdateCell: sharedTool(
+        sharedToolSpecs.tableUpdateCell,
+        async ({ pageId, table, row, col, text }) =>
+          await client.tableUpdateCell(pageId, table, row, col, text),
+      ),
 
       copyPageContent: sharedTool(
         sharedToolSpecs.copyPageContent,
@@ -918,25 +717,14 @@ export class AiChatToolsService {
           await client.importPageMarkdown(pageId, markdown),
       ),
 
-      // INTENTIONAL per-transport divergence (not shared): adds a security
-      // confirmation framing ("Only share when the user explicitly asked, since
-      // this exposes the page to anyone with the link") for the in-app agent; the
-      // standalone MCP `share_page` keeps the plain public-URL wording.
-      sharePage: tool({
-        description:
-          'Make a page PUBLICLY accessible and return its public URL. ' +
-          'Reversible via unsharePage. Only share when the user explicitly ' +
-          'asked, since this exposes the page to anyone with the link.',
-        inputSchema: modelFriendlyInput({
-          pageId: z.string().describe('The id of the page to share.'),
-          searchIndexing: z
-            .boolean()
-            .optional()
-            .describe('Allow public search engines to index it (default true).'),
-        }),
-        execute: async ({ pageId, searchIndexing }) =>
+      // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
+      // Both layers already carried the security-confirmation framing, so there
+      // was no real divergence to preserve — only wording drift.
+      sharePage: sharedTool(
+        sharedToolSpecs.sharePage,
+        async ({ pageId, searchIndexing }) =>
           await client.sharePage(pageId, searchIndexing),
-      }),
+      ),
 
       unsharePage: sharedTool(
         sharedToolSpecs.unsharePage,

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

@@ -100,54 +100,26 @@ export const INLINE_TOOL_TIERS: Record<
     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).',
-  },
+  // NOTE: getPage and listPages moved to @docmost/mcp's SHARED_TOOL_SPECS
+  // (#294); they carry their own tier ('core') + catalogLine there.
+  // NOTE: createComment, listComments and resolveComment moved to
+  // @docmost/mcp's SHARED_TOOL_SPECS (#294); they carry their own tier +
+  // catalogLine there. getComment stays inline (MCP-only shape divergence is
+  // n/a — it simply has no shared spec).
   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.',
-  },
+  // NOTE: createPage, renamePage, movePage, deletePage, updatePageJson and
+  // exportPageMarkdown moved to @docmost/mcp's SHARED_TOOL_SPECS (#294); they
+  // carry their own deferred tier + catalogLine there.
   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:
@@ -157,42 +129,21 @@ export const INLINE_TOOL_TIERS: Record<
     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.',
-  },
+  // NOTE: tableInsertRow, tableDeleteRow and tableUpdateCell moved to
+  // @docmost/mcp's SHARED_TOOL_SPECS (#294); they carry their own deferred tier +
+  // catalogLine there. getTable stays inline (its MCP name table_get breaks the
+  // snake_case(inAppKey) convention, so it has no shared spec).
+  // NOTE: checkNewComments moved to @docmost/mcp's SHARED_TOOL_SPECS (#294);
+  // it carries its own deferred tier + catalogLine there.
   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.',
-  },
+  // NOTE: sharePage moved to @docmost/mcp's SHARED_TOOL_SPECS (#294); it carries
+  // its own deferred tier + catalogLine there. transformPage stays inline (its
+  // schema deliberately diverges — it omits the deleteComments field the MCP
+  // docmost_transform exposes, a comment-deletion guardrail).
   transformPage: {
     tier: 'deferred',
     catalogLine: "transformPage — run a sandboxed JS transform over a page's document.",

apps/server/src/core/page/services/page.service.ts

@@ -52,7 +52,9 @@ import {
   INTERNAL_LINK_REGEX,
   extractPageSlugId,
 } from '../../../integrations/export/utils';
-import { markdownToHtml, canonicalizeFootnotes } from '@docmost/editor-ext';
+import { canonicalizeFootnotes } from '@docmost/editor-ext';
+import { markdownToProseMirror } from '@docmost/prosemirror-markdown';
+import { normalizeForeignMarkdown } from '../../../integrations/import/utils/foreign-markdown';
 import { WatcherService } from '../../watcher/watcher.service';
 import { sql } from 'kysely';
 import { TransclusionService } from '../transclusion/transclusion.service';
@@ -1301,8 +1303,14 @@ export class PageService {
 
     switch (format) {
       case 'markdown': {
-        const html = await markdownToHtml(content as string);
-        prosemirrorJson = htmlToJson(html as string);
+        // Canonical markdown -> ProseMirror JSON directly via
+        // `@docmost/prosemirror-markdown` (issue #345) — no HTML intermediate,
+        // no editor-ext markdown layer. Foreign markdown surfaces the strict
+        // parser rejects (GFM `[^id]` reference footnotes) are normalized to the
+        // canonical inline form first.
+        prosemirrorJson = await markdownToProseMirror(
+          normalizeForeignMarkdown(content as string),
+        );
         break;
       }
       case 'html': {

apps/server/src/database/listeners/page.listener.ts

@@ -4,7 +4,6 @@ import { EventName } from '../../common/events/event.contants';
 import { InjectQueue } from '@nestjs/bullmq';
 import { QueueJob, QueueName } from '../../integrations/queue/constants';
 import { Queue } from 'bullmq';
-import { EnvironmentService } from '../../integrations/environment/environment.service';
 
 /**
  * Thin snapshot of a page node carried inside domain events so the WebSocket
@@ -112,48 +111,24 @@ export class PageListener {
   private readonly logger = new Logger(PageListener.name);
 
   constructor(
-    private readonly environmentService: EnvironmentService,
-    @InjectQueue(QueueName.SEARCH_QUEUE) private searchQueue: Queue,
     @InjectQueue(QueueName.AI_QUEUE) private aiQueue: Queue,
   ) {}
 
   @OnEvent(EventName.PAGE_CREATED)
   async handlePageCreated(event: PageEvent) {
     const { pageIds, workspaceId } = event;
-    if (this.isTypesense()) {
-      await this.searchQueue.add(QueueJob.PAGE_CREATED, {
-        pageIds,
-      });
-    }
-
     await this.aiQueue.add(QueueJob.PAGE_CREATED, { pageIds, workspaceId });
   }
 
-  @OnEvent(EventName.PAGE_UPDATED)
-  async handlePageUpdated(event: PageEvent) {
-    const { pageIds } = event;
-
-    await this.searchQueue.add(QueueJob.PAGE_UPDATED, { pageIds });
-  }
-
   @OnEvent(EventName.PAGE_DELETED)
   async handlePageDeleted(event: PageEvent) {
     const { pageIds, workspaceId } = event;
-    if (this.isTypesense()) {
-      await this.searchQueue.add(QueueJob.PAGE_DELETED, { pageIds });
-    }
-
     await this.aiQueue.add(QueueJob.PAGE_DELETED, { pageIds, workspaceId });
   }
 
   @OnEvent(EventName.PAGE_SOFT_DELETED)
   async handlePageSoftDeleted(event: PageEvent) {
     const { pageIds, workspaceId } = event;
-
-    if (this.isTypesense()) {
-      await this.searchQueue.add(QueueJob.PAGE_SOFT_DELETED, { pageIds });
-    }
-
     await this.aiQueue.add(QueueJob.PAGE_SOFT_DELETED, {
       pageIds,
       workspaceId,
@@ -163,14 +138,6 @@ export class PageListener {
   @OnEvent(EventName.PAGE_RESTORED)
   async handlePageRestored(event: PageEvent) {
     const { pageIds, workspaceId } = event;
-    if (this.isTypesense()) {
-      await this.searchQueue.add(QueueJob.PAGE_RESTORED, { pageIds });
-    }
-
     await this.aiQueue.add(QueueJob.PAGE_RESTORED, { pageIds, workspaceId });
   }
-
-  isTypesense(): boolean {
-    return this.environmentService.getSearchDriver() === 'typesense';
-  }
 }

apps/server/src/database/listeners/space.listener.ts

@@ -4,7 +4,6 @@ import { EventName } from '../../common/events/event.contants';
 import { InjectQueue } from '@nestjs/bullmq';
 import { QueueJob, QueueName } from '../../integrations/queue/constants';
 import { Queue } from 'bullmq';
-import { EnvironmentService } from '../../integrations/environment/environment.service';
 
 export class SpaceEvent {
   spaceId: string;
@@ -15,22 +14,12 @@ export class SpaceListener {
   private readonly logger = new Logger(SpaceListener.name);
 
   constructor(
-    private readonly environmentService: EnvironmentService,
-    @InjectQueue(QueueName.SEARCH_QUEUE) private searchQueue: Queue,
     @InjectQueue(QueueName.AI_QUEUE) private aiQueue: Queue,
   ) {}
 
   @OnEvent(EventName.SPACE_DELETED)
   async handleSpaceDeleted(event: SpaceEvent) {
     const { spaceId } = event;
-    if (this.isTypesense()) {
-      await this.searchQueue.add(QueueJob.SPACE_DELETED, { spaceId });
-    }
-
     await this.aiQueue.add(QueueJob.SPACE_DELETED, { spaceId });
   }
-
-  isTypesense(): boolean {
-    return this.environmentService.getSearchDriver() === 'typesense';
-  }
 }

apps/server/src/database/listeners/workspace.listener.ts

@@ -4,7 +4,6 @@ import { EventName } from '../../common/events/event.contants';
 import { InjectQueue } from '@nestjs/bullmq';
 import { QueueJob, QueueName } from '../../integrations/queue/constants';
 import { Queue } from 'bullmq';
-import { EnvironmentService } from '../../integrations/environment/environment.service';
 
 export class WorkspaceEvent {
   workspaceId: string;
@@ -15,22 +14,12 @@ export class WorkspaceListener {
   private readonly logger = new Logger(WorkspaceListener.name);
 
   constructor(
-    private readonly environmentService: EnvironmentService,
-    @InjectQueue(QueueName.SEARCH_QUEUE) private searchQueue: Queue,
     @InjectQueue(QueueName.AI_QUEUE) private aiQueue: Queue,
   ) {}
 
   @OnEvent(EventName.WORKSPACE_DELETED)
   async handlePageDeleted(event: WorkspaceEvent) {
     const { workspaceId } = event;
-    if (this.isTypesense()) {
-      await this.searchQueue.add(QueueJob.WORKSPACE_DELETED, { workspaceId });
-    }
-
     await this.aiQueue.add(QueueJob.WORKSPACE_DELETED, { workspaceId });
   }
-
-  isTypesense(): boolean {
-    return this.environmentService.getSearchDriver() === 'typesense';
-  }
 }

apps/server/src/database/migrate.ts

@@ -24,6 +24,10 @@ const migrator = new Migrator({
     path,
     migrationFolder,
   }),
+  // Match the startup auto-migrator (migration.service.ts): a back-dated
+  // migration from a long-lived branch must be applied, not rejected as
+  // "corrupted migrations" (incident #361). See that file for the full rationale.
+  allowUnorderedMigrations: true,
 });
 
 run(db, migrator, migrationFolder);

apps/server/src/database/services/migration.service.ts

@@ -19,6 +19,16 @@ export class MigrationService {
         path,
         migrationFolder: path.join(__dirname, '..', 'migrations'),
       }),
+      // A long-lived branch can add a migration whose timestamped filename sorts
+      // BEFORE migrations already applied in prod (e.g. #234's 20260627 landing
+      // after 20260704 was live). With the default (ordered) setting the startup
+      // migrator then sees "corrupted migrations" — the applied set is no longer a
+      // prefix of the sorted list — throws, and the app crash-loops on boot
+      // (incident #361: 502s for ~11 min). allowUnorderedMigrations runs any
+      // not-yet-applied migration regardless of filename order, so a back-dated
+      // migration is applied instead of bricking startup. A CI order-gate still
+      // discourages back-dating; this is the runtime safety net.
+      allowUnorderedMigrations: true,
     });
 
     const { error, results } = await migrator.migrateToLatest();

apps/server/src/integrations/ai/ai-sdk-partial-output.patch.spec.ts

@@ -0,0 +1,124 @@
+import { readFileSync } from 'fs';
+import { streamText, Output } from 'ai';
+import { MockLanguageModelV3, simulateReadableStream } from 'ai/test';
+
+/**
+ * Regression tests for patches/ai@6.0.134.patch (server heap OOM on long
+ * autonomous agent runs, #184).
+ *
+ * Unpatched ai@6.0.134 substitutes the default text() output strategy even
+ * when the caller passes NO `output` option. Its createOutputTransformStream
+ * then accumulates the ENTIRE turn text and, on EVERY text-delta, enqueues a
+ * flat snapshot of all text so far as `partialOutput` (O(n^2) memory). Those
+ * snapshots pile up in the never-consumed leftover tee() branch of
+ * DefaultStreamTextResult.baseStream, which is what OOM'd production during a
+ * ~28k-chunk agent turn. The pnpm patch skips partialOutput production
+ * entirely when no output strategy was requested, while keeping per-delta
+ * streaming granularity.
+ */
+describe('ai@6.0.134 pnpm patch: no partialOutput accumulation without an output strategy', () => {
+  const makeModel = () =>
+    new MockLanguageModelV3({
+      doStream: async () => ({
+        stream: simulateReadableStream({
+          chunks: [
+            { type: 'stream-start' as const, warnings: [] },
+            { type: 'text-start' as const, id: '1' },
+            { type: 'text-delta' as const, id: '1', delta: 'Hello' },
+            { type: 'text-delta' as const, id: '1', delta: ', ' },
+            { type: 'text-delta' as const, id: '1', delta: 'world!' },
+            { type: 'text-end' as const, id: '1' },
+            {
+              type: 'finish' as const,
+              finishReason: { unified: 'stop' as const, raw: 'stop' },
+              usage: {
+                inputTokens: {
+                  total: 1,
+                  noCache: undefined,
+                  cacheRead: undefined,
+                  cacheWrite: undefined,
+                },
+                outputTokens: { total: 1, text: 1, reasoning: undefined },
+              },
+            },
+          ],
+        }),
+      }),
+    });
+
+  it('preserves per-delta streaming granularity in textStream', async () => {
+    const result = streamText({ model: makeModel(), prompt: 'hi' });
+
+    const deltas: string[] = [];
+    for await (const delta of result.textStream) {
+      deltas.push(delta);
+    }
+
+    // The patch must NOT coalesce or drop deltas: three model deltas arrive
+    // as three separate textStream chunks.
+    expect(deltas).toEqual(['Hello', ', ', 'world!']);
+  });
+
+  it('emits NO partialOutput values when the caller did not request an output strategy', async () => {
+    const result = streamText({ model: makeModel(), prompt: 'hi' });
+
+    // Fully consume the primary stream first (mirrors production usage).
+    for await (const _ of result.textStream) {
+      // drain
+    }
+
+    const partials: unknown[] = [];
+    for await (const partial of result.experimental_partialOutputStream) {
+      partials.push(partial);
+    }
+
+    // TRIPWIRE: on unpatched ai@6.0.134 the default text() output strategy
+    // yields one cumulative partial per text-delta here (['Hello', 'Hello, ',
+    // 'Hello, world!']). An empty stream proves the patch is applied and no
+    // cumulative snapshots are being produced (and thus none can pile up in
+    // the leftover internal tee branch).
+    expect(partials).toEqual([]);
+  });
+
+  it('preserves cumulative partialOutput when the caller DOES request an output strategy', async () => {
+    // PRESERVE-BRANCH GUARD: the patch only short-circuits partialOutput when
+    // `output == null`. When an output strategy IS set (here Output.text()),
+    // createOutputTransformStream must fall through to the ORIGINAL code path
+    // and keep publishing cumulative snapshots, so object/text-output consumers
+    // behave byte-identically to unpatched ai. A careless re-port that routed
+    // output-set calls into the skip branch would leave partialOutput empty and
+    // silently break those consumers — this test is the tripwire for that.
+    const result = streamText({
+      model: makeModel(),
+      prompt: 'hi',
+      experimental_output: Output.text(),
+    });
+
+    // Drain the primary stream fully and accumulate the complete output text.
+    let fullText = '';
+    for await (const delta of result.textStream) {
+      fullText += delta;
+    }
+
+    const partials: string[] = [];
+    for await (const partial of result.experimental_partialOutputStream) {
+      partials.push(partial);
+    }
+
+    // With a strategy set, partialOutput must be PRESERVED (non-empty) and
+    // cumulative: the last emitted partial equals the full accumulated text.
+    expect(partials.length).toBeGreaterThan(0);
+    expect(partials[partials.length - 1]).toBe(fullText);
+    expect(fullText).toBe('Hello, world!');
+  });
+
+  it('both installed dist builds (CJS and ESM) carry the patch marker', () => {
+    // Secondary guard: pins the patch to BOTH bundles the SDK ships, since
+    // the NestJS server consumes CJS while other tooling may load ESM.
+    const cjsPath = require.resolve('ai');
+    const mjsPath = cjsPath.replace(/index\.js$/, 'index.mjs');
+    expect(cjsPath).toMatch(/index\.js$/);
+    expect(readFileSync(cjsPath, 'utf8')).toContain('PATCH(docmost');
+    expect(readFileSync(mjsPath, 'utf8')).toContain('PATCH(docmost');
+  });
+});

apps/server/src/integrations/export/export-markdown.spec.ts

@@ -0,0 +1,145 @@
+// export.service.ts imports the ESM-only @sindresorhus/slugify (not in jest's
+// transform allowlist). It is irrelevant to the markdown-serialization path under
+// test (only used for page-mention link slugs on the DB path), so it is mocked
+// out to keep the module graph loadable under ts-jest (mirrors the import specs).
+jest.mock('@sindresorhus/slugify', () => ({
+  __esModule: true,
+  default: (input: string) => String(input),
+}));
+
+import { convertProseMirrorToMarkdown } from '@docmost/prosemirror-markdown';
+import { ExportService } from './export.service';
+import { ExportFormat } from './dto/export-dto';
+
+/**
+ * STEP 1 golden test for issue #345: server MARKDOWN export runs DIRECTLY through
+ * the canonical converter (`convertProseMirrorToMarkdown`) — no HTML intermediate
+ * and no `@docmost/editor-ext` markdown layer — so the emitted markdown is in the
+ * canonical package forms and is byte-identical to the git-sync vault body.
+ *
+ * These are the goldens the swap has to satisfy: they assert the CANONICAL
+ * surface (callout `> [!type]`, inline footnote `^[…]`, lossless image
+ * `<!--img …-->`) rather than the old editor-ext forms (`:::type`, `[^id]`,
+ * lossy `![alt](src)`).
+ *
+ * `exportPage(..., singlePage=false)` takes no DB path (no mention rewriting), so
+ * the service is constructed with null collaborators and only the pure
+ * PM -> Markdown path is exercised.
+ */
+
+function makeService(): ExportService {
+  return new ExportService(
+    null as any, // pageRepo
+    null as any, // pagePermissionRepo
+    null as any, // db
+    null as any, // storageService
+    null as any, // environmentService
+    null as any, // domainService
+  );
+}
+
+// A representative page exercising the node types whose canonical markdown form
+// changed with the move off the editor-ext layer: callout, inline footnote, and a
+// lossless image carrying width/align attrs that the old layer dropped.
+const REPRESENTATIVE_DOC = {
+  type: 'doc',
+  content: [
+    {
+      type: 'paragraph',
+      content: [
+        { type: 'text', text: 'Body ' },
+        { type: 'footnoteReference', attrs: { id: 'fn-1' } },
+        { type: 'text', text: ' end.' },
+      ],
+    },
+    {
+      type: 'callout',
+      attrs: { type: 'info', icon: null },
+      content: [
+        {
+          type: 'paragraph',
+          content: [{ type: 'text', text: 'Heads up' }],
+        },
+      ],
+    },
+    {
+      type: 'image',
+      attrs: {
+        src: '/files/pic.png',
+        alt: 'Pic',
+        width: 320,
+        align: 'left',
+      },
+    },
+    {
+      type: 'footnotesList',
+      content: [
+        {
+          type: 'footnoteDefinition',
+          attrs: { id: 'fn-1' },
+          content: [
+            {
+              type: 'paragraph',
+              content: [{ type: 'text', text: 'the note' }],
+            },
+          ],
+        },
+      ],
+    },
+  ],
+};
+
+describe('ExportService — markdown export via the canonical converter (#345)', () => {
+  it('emits canonical callout, inline footnote and lossless image forms', async () => {
+    const service = makeService();
+    const md = (await service.exportPage(ExportFormat.Markdown, {
+      title: '',
+      content: REPRESENTATIVE_DOC,
+    } as any)) as string;
+
+    // Callout: Obsidian `> [!type]`, NOT the legacy `:::type`.
+    expect(md).toContain('> [!info]');
+    expect(md).not.toContain(':::');
+
+    // Inline footnote: `^[…]`, NOT the reference `[^id]` form.
+    expect(md).toContain('^[the note]');
+    expect(md).not.toMatch(/\[\^/);
+
+    // Lossless image: trailing `<!--img …-->` carrying the dropped attrs.
+    expect(md).toContain('![Pic](/files/pic.png)');
+    expect(md).toContain('<!--img');
+    expect(md).toContain('"width":"320"');
+    expect(md).toContain('"align":"left"');
+  });
+
+  it('export body is byte-identical to the git-sync vault serializer (export == vault)', async () => {
+    const service = makeService();
+    // A title-less page: exportPage prepends NO heading, so the whole output is
+    // the page BODY — exactly what git-sync serializes (git-sync stores the title
+    // in frontmatter / the filename, never as an in-body H1).
+    const exported = (await service.exportPage(ExportFormat.Markdown, {
+      title: '',
+      content: REPRESENTATIVE_DOC,
+    } as any)) as string;
+
+    // The git-sync vault writer feeds this SAME converter (git-sync
+    // `stabilizePageBody` = convertProseMirrorToMarkdown(content) at the
+    // fixpoint). For an already-stable doc the single pass IS the fixpoint, so
+    // the two are byte-identical by construction — assert it.
+    const vaultBody = convertProseMirrorToMarkdown(REPRESENTATIVE_DOC);
+    expect(exported).toBe(vaultBody);
+  });
+
+  it('prepends the page title as an H1 heading (the one documented export/vault delta)', async () => {
+    const service = makeService();
+    const md = (await service.exportPage(ExportFormat.Markdown, {
+      title: 'My Page',
+      content: { type: 'doc', content: [] },
+    } as any)) as string;
+
+    // Export makes standalone files, so it prepends the title as an H1. This is
+    // the ONE deliberate difference from the vault body (which carries the title
+    // in frontmatter). The body below the heading still serializes canonically.
+    expect(md.startsWith('# My Page')).toBe(true);
+  });
+});

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

@@ -37,7 +37,7 @@ import {
   getAttachmentIds,
   getProsemirrorContent,
 } from '../../common/helpers/prosemirror/utils';
-import { htmlToMarkdown } from '@docmost/editor-ext';
+import { convertProseMirrorToMarkdown } from '@docmost/prosemirror-markdown';
 
 type AllowedAttachment = { id: string; fileName: string; filePath: string };
 
@@ -79,9 +79,8 @@ export class ExportService {
       prosemirrorJson.content.unshift(titleNode);
     }
 
-    const pageHtml = jsonToHtml(prosemirrorJson);
-
     if (format === ExportFormat.HTML) {
+      const pageHtml = jsonToHtml(prosemirrorJson);
       return `<!DOCTYPE html>
       <html>
         <head>
@@ -92,11 +91,14 @@ export class ExportService {
     }
 
     if (format === ExportFormat.Markdown) {
-      const newPageHtml = pageHtml.replace(
-        /<colgroup[^>]*>[\s\S]*?<\/colgroup>/gim,
-        '',
-      );
-      return htmlToMarkdown(newPageHtml);
+      // Direct ProseMirror JSON -> Markdown via the canonical converter
+      // (`@docmost/prosemirror-markdown`). This is the SAME serializer the
+      // git-sync vault writer feeds (see git-sync `stabilizePageBody`), so an
+      // exported page body is byte-identical to its vault representation — no
+      // HTML intermediate, no second markdown layer, no format drift (issue
+      // #345). The old `<colgroup>` scrub is gone with the HTML step: the
+      // converter emits GFM tables directly and never produces `<colgroup>`.
+      return convertProseMirrorToMarkdown(prosemirrorJson);
     }
 
     return;

apps/server/src/integrations/import/services/file-import-task.service.footnote-canonicalize.spec.ts

@@ -17,6 +17,22 @@ jest.mock('image-dimensions', () => ({
   __esModule: true,
   imageDimensionsFromData: () => undefined,
 }));
+// FileImportTaskService -> PageService -> collaboration.gateway ->
+// metrics.registry imports `prom-client`, which is not resolvable in this
+// workspace's node_modules (types-only stub, no runtime entry). Metrics are
+// disabled on this path, so a virtual no-op mock keeps the module graph loadable.
+jest.mock(
+  'prom-client',
+  () => ({
+    collectDefaultMetrics: () => undefined,
+    Registry: class {},
+    Histogram: class {},
+    Gauge: class {},
+    Counter: class {},
+    Summary: class {},
+  }),
+  { virtual: true },
+);
 
 import { promises as fs } from 'fs';
 import * as os from 'os';
@@ -26,14 +42,17 @@ import { ImportService } from './import.service';
 
 /**
  * Binding test for issue #228 / review #5: FileImportTaskService.processGenericImport
- * is a NON-editor write path (markdownToHtml -> processHTML -> JSON, never runs
- * footnoteSyncPlugin), so it canonicalizes footnotes before persisting. This pins
- * that binding — the same one import.service has a spec for — which previously had
- * NO spec at all.
+ * is a NON-editor write path, so a zip-imported `.md` page ends up with canonical
+ * footnotes before persisting: ordered by first reference, reused refs deduped,
+ * orphan definitions dropped.
  *
- * The markdown -> HTML -> ProseMirror conversion is REAL (a real ImportService,
- * its createYdoc stubbed); the filesystem is a real temp dir with one .md file;
- * the DB transaction is stubbed to capture the persisted page content.
+ * Since #345 the `.md` parse runs `normalizeForeignMarkdown` ->
+ * `markdownToProseMirror` -> `jsonToHtml` (feeding the shared HTML attachment /
+ * link pipeline) -> `processHTML` -> `canonicalizeFootnotes`. The parser assigns
+ * fresh `fn-*` ids, so we assert by definition BODY order rather than the source
+ * labels. The conversion is REAL (a real ImportService, its createYdoc stubbed);
+ * the filesystem is a real temp dir with one .md file; the DB transaction is
+ * stubbed to capture the persisted page content.
  */
 
 // Out-of-order references (c, a, b), a REUSED reference ([^a] twice), and an
@@ -49,13 +68,14 @@ const MARKDOWN = [
   '[^z]: orphan note',
 ].join('\n');
 
-function footnoteListIds(content: any): string[] {
+/** Definition body texts of the (single) footnotesList, in list order. */
+function footnoteListBodies(content: any): string[] {
   const list = (content?.content ?? []).find(
     (n: any) => n.type === 'footnotesList',
   );
   return (list?.content ?? [])
     .filter((n: any) => n.type === 'footnoteDefinition')
-    .map((n: any) => n.attrs?.id);
+    .map((n: any) => n.content?.[0]?.content?.[0]?.text);
 }
 
 // A permissive chainable stub for the spaces lookup (selectFrom(...).select(...)
@@ -71,80 +91,127 @@ function chainable(result: any): any {
   return proxy;
 }
 
+/**
+ * Run one markdown file through the REAL zip-import pipeline
+ * (`processGenericImport` -> `markdownToProseMirror` -> `jsonToHtml` ->
+ * `processHTML`/`htmlToJson`) and return the persisted page `content`. This is
+ * the server-specific PM->HTML->PM hop that the package's own PM<->MD tests do
+ * NOT cover.
+ */
+async function runZipImport(markdown: string): Promise<any> {
+  const extractDir = await fs.mkdtemp(path.join(os.tmpdir(), 'fit-canon-'));
+  await fs.writeFile(path.join(extractDir, 'note.md'), markdown, 'utf-8');
+
+  const importService = new ImportService(
+    {} as any,
+    {} as any,
+    {} as any,
+    {} as any,
+  );
+  jest
+    .spyOn(importService as any, 'createYdoc')
+    .mockResolvedValue(Buffer.from([]) as any);
+
+  let captured: any = null;
+  const trx = {
+    insertInto: (table: string) => ({
+      values: (v: any) => {
+        if (table === 'pages') captured = v;
+        return { execute: async () => {} };
+      },
+    }),
+  };
+  const db: any = {
+    selectFrom: () => chainable({ slug: 'space-slug' }),
+    transaction: () => ({ execute: (fn: any) => fn(trx) }),
+  };
+
+  const importAttachmentService = {
+    processAttachments: async ({ html }: any) => html,
+  };
+  const service = new FileImportTaskService(
+    {} as any, // storageService
+    importService as any,
+    { nextPagePosition: async () => 'a0' } as any,
+    { insertBacklink: jest.fn() } as any,
+    db,
+    importAttachmentService as any,
+    { emit: jest.fn() } as any,
+    { logBatchWithContext: jest.fn() } as any,
+  );
+
+  const fileTask: any = {
+    id: 'task-1',
+    source: 'generic',
+    spaceId: 'space-1',
+    workspaceId: 'ws-1',
+    creatorId: 'user-1',
+  };
+
+  try {
+    await service.processGenericImport({ extractDir, fileTask });
+    expect(captured).toBeTruthy();
+    return captured.content;
+  } finally {
+    await fs.rm(extractDir, { recursive: true, force: true });
+  }
+}
+
+/** Find the first node of a given type anywhere in a PM content tree. */
+function findFirst(node: any, type: string): any {
+  if (!node || typeof node !== 'object') return null;
+  if (node.type === type) return node;
+  for (const child of node.content ?? []) {
+    const hit = findFirst(child, type);
+    if (hit) return hit;
+  }
+  return null;
+}
+
 describe('FileImportTaskService.processGenericImport — footnote canonicalization (#228)', () => {
   it('orders footnotes by first reference, dedupes reuse, and drops orphans on zip import', async () => {
-    const extractDir = await fs.mkdtemp(path.join(os.tmpdir(), 'fit-canon-'));
-    await fs.writeFile(path.join(extractDir, 'note.md'), MARKDOWN, 'utf-8');
-
-    // Real ImportService for the html -> JSON conversion; stub the yjs encode.
-    const importService = new ImportService(
-      {} as any,
-      {} as any,
-      {} as any,
-      {} as any,
+    const content = await runZipImport(MARKDOWN);
+    // Definitions ordered by FIRST REFERENCE (C, A, B), NOT the markdown
+    // definition order (A, B, C). Ids are the parser's fresh `fn-*`, so pin
+    // the BODIES.
+    expect(footnoteListBodies(content)).toEqual(['note C', 'note A', 'note B']);
+    // Orphan [^z] dropped; reused [^a] collapses to one definition; one list.
+    expect(footnoteListBodies(content)).not.toContain('orphan note');
+    const lists = (content.content ?? []).filter(
+      (n: any) => n.type === 'footnotesList',
     );
-    jest
-      .spyOn(importService as any, 'createYdoc')
-      .mockResolvedValue(Buffer.from([]) as any);
+    expect(lists).toHaveLength(1);
+    expect(
+      footnoteListBodies(content).filter((b) => b === 'note A'),
+    ).toHaveLength(1);
+  });
 
-    let captured: any = null;
-    const trx = {
-      insertInto: (table: string) => ({
-        values: (v: any) => {
-          if (table === 'pages') captured = v;
-          return { execute: async () => {} };
-        },
-      }),
-    };
-    const db: any = {
-      selectFrom: () => chainable({ slug: 'space-slug' }),
-      transaction: () => ({ execute: (fn: any) => fn(trx) }),
-    };
+  // #345 F4: the zip path routes markdown through jsonToHtml -> processHTML ->
+  // htmlToJson (the shared HTML attachment pipeline). #345's headline is LOSSLESS
+  // image width/align via the `<!--img {...}-->` comment; a callout carries its
+  // `type`. This asserts those survive the PM->HTML->PM hop — the one hop the
+  // package's PM<->MD suite does not exercise.
+  it('preserves image width/align and callout type through the PM->HTML->PM hop', async () => {
+    const md = [
+      '# Doc',
+      '',
+      '![a picture](https://example.com/i.png) <!--img {"width":"320","align":"left"}-->',
+      '',
+      ':::warning',
+      'Careful now.',
+      ':::',
+    ].join('\n');
 
-    const importAttachmentService = {
-      processAttachments: async ({ html }: any) => html,
-    };
-    const backlinkRepo = { insertBacklink: jest.fn() };
-    const eventEmitter = { emit: jest.fn() };
-    const auditService = { logBatchWithContext: jest.fn() };
+    const content = await runZipImport(md);
 
-    const pageService = { nextPagePosition: async () => 'a0' };
+    const image = findFirst(content, 'image');
+    expect(image).toBeTruthy();
+    // The lossless sizing/alignment must survive the HTML hop.
+    expect(String(image.attrs?.width)).toBe('320');
+    expect(image.attrs?.align).toBe('left');
 
-    const service = new FileImportTaskService(
-      {} as any, // storageService
-      importService as any,
-      pageService as any,
-      backlinkRepo as any,
-      db,
-      importAttachmentService as any,
-      eventEmitter as any,
-      auditService as any,
-    );
-
-    const fileTask: any = {
-      id: 'task-1',
-      source: 'generic',
-      spaceId: 'space-1',
-      workspaceId: 'ws-1',
-      creatorId: 'user-1',
-    };
-
-    try {
-      await service.processGenericImport({ extractDir, fileTask });
-
-      expect(captured).toBeTruthy();
-      const content = captured.content;
-      // Reference order is c, a, b (NOT the markdown definition order a, b, c).
-      expect(footnoteListIds(content)).toEqual(['c', 'a', 'b']);
-      // Orphan [^z] dropped; reused [^a] collapses to one definition; one list.
-      expect(footnoteListIds(content)).not.toContain('z');
-      const lists = (content.content ?? []).filter(
-        (n: any) => n.type === 'footnotesList',
-      );
-      expect(lists).toHaveLength(1);
-      expect(footnoteListIds(content).filter((id) => id === 'a')).toHaveLength(1);
-    } finally {
-      await fs.rm(extractDir, { recursive: true, force: true });
-    }
+    const callout = findFirst(content, 'callout');
+    expect(callout).toBeTruthy();
+    expect(callout.attrs?.type).toBe('warning');
   });
 });

apps/server/src/integrations/import/services/file-import-task.service.ts

@@ -1,6 +1,9 @@
 import { Inject, Injectable, Logger } from '@nestjs/common';
 import * as path from 'path';
-import { jsonToText } from '../../../collaboration/collaboration.util';
+import {
+  jsonToHtml,
+  jsonToText,
+} from '../../../collaboration/collaboration.util';
 import { InjectKysely } from 'nestjs-kysely';
 import { KyselyDB } from '@docmost/db/types/kysely.types';
 import {
@@ -18,9 +21,11 @@ import { generateSlugId } from '../../../common/helpers';
 import { v7 } from 'uuid';
 import { generateJitteredKeyBetween } from 'fractional-indexing-jittered';
 import { FileTask, InsertablePage } from '@docmost/db/types/entity.types';
-import { markdownToHtml, canonicalizeFootnotes } from '@docmost/editor-ext';
+import { canonicalizeFootnotes } from '@docmost/editor-ext';
+import { markdownToProseMirror } from '@docmost/prosemirror-markdown';
 import { getProsemirrorContent } from '../../../common/helpers/prosemirror/utils';
 import { formatImportHtml } from '../utils/import-formatter';
+import { normalizeForeignMarkdown } from '../utils/foreign-markdown';
 import {
   buildAttachmentCandidates,
   collectMarkdownAndHtmlFiles,
@@ -461,7 +466,18 @@ export class FileImportTaskService {
               content = await fs.readFile(absPath, 'utf-8');
 
               if (page.fileExtension.toLowerCase() === '.md') {
-                content = await markdownToHtml(content);
+                // Parse markdown with the single canonical converter
+                // (`@docmost/prosemirror-markdown`), after normalizing foreign
+                // reference footnotes, then serialize to HTML so the shared HTML
+                // pipeline below (processAttachments + formatImportHtml +
+                // processHTML) keeps handling `.md` and `.html` imports
+                // uniformly. The markdown PARSE no longer goes through the
+                // editor-ext markdown layer (issue #345) — the drift source is
+                // gone. The PM -> HTML -> PM hop that follows is lossless
+                // plumbing for attachment/link resolution, NOT a second parse.
+                content = jsonToHtml(
+                  await markdownToProseMirror(normalizeForeignMarkdown(content)),
+                );
               }
             } catch (err: any) {
               if (err?.code === 'ENOENT') {
@@ -500,10 +516,12 @@ export class FileImportTaskService {
               this.importService.extractTitleAndRemoveHeading(pmState);
 
             // Canonicalize footnote topology on this non-editor write path
-            // (markdownToHtml/processHTML never runs footnoteSyncPlugin), so a
-            // zip-imported page's footnotes are reference-ordered, deduped, and
+            // (the HTML pipeline's processHTML never runs footnoteSyncPlugin), so
+            // a zip-imported page's footnotes are reference-ordered, deduped, and
             // orphan-free like the editor's invariant (issue #228). Pure +
-            // idempotent + shape-safe; a footnote-free doc is unchanged.
+            // idempotent + shape-safe; a footnote-free doc is unchanged. (For a
+            // `.md` file the package parser already yields canonical footnotes,
+            // so this is a no-op there.)
             // (Future consolidation, architecture B: like import.service, this
             // path persists directly rather than via PageService — a shared
             // "prepare JSON for persist" helper would centralize this call.)

apps/server/src/integrations/import/services/import.service.footnote-canonicalize.spec.ts

@@ -12,13 +12,19 @@ import { canonicalizeFootnotes } from '@docmost/editor-ext';
 
 /**
  * Integration-ish test for the USER-FACING markdown import path
- * (`ImportService.importPage`). It exercises the REAL markdown -> HTML -> JSON
- * conversion and asserts that the stored page content has its footnotes
- * canonicalized — the gap that issue #228 fixes: the import path builds
- * ProseMirror JSON directly (never running the editor's footnoteSyncPlugin), so
- * before this wiring the stored footnotes kept the markdown's physical
- * definition order (out of order vs. references), retained orphan definitions,
- * and did not collapse reused references.
+ * (`ImportService.importPage`). It exercises the REAL markdown -> ProseMirror
+ * conversion and asserts the stored page's footnotes are canonical: ordered by
+ * FIRST REFERENCE (not markdown definition order), reused references deduped to a
+ * single definition, and orphan definitions dropped.
+ *
+ * Since #345 the markdown parse runs through the canonical package
+ * (`normalizeForeignMarkdown` -> `markdownToProseMirror`), which owns this
+ * canonicalization: the input's GFM `[^id]` reference footnotes are normalized to
+ * inline `^[…]`, and the parser assigns fresh sequential ids (`fn-*`) in
+ * reference order while merging identical bodies — so we assert by definition
+ * BODY order, not by the source labels. `canonicalizeFootnotes` remains wired as
+ * an idempotent safety net (issue #228) and is a no-op on this already-canonical
+ * output.
  *
  * The DB/ydoc side-effects are stubbed: `getNewPagePosition` (DB query) and
  * `createYdoc` (Yjs encode) are spied, and `pageRepo.insertPage` captures the
@@ -67,24 +73,14 @@ function makeService() {
 }
 
 /** List the footnote-definition ids of the (single) footnotesList, in order. */
-function footnoteListIds(content: any): string[] {
+/** Definition body texts of the (single) footnotesList, in list order. */
+function footnoteListBodies(content: any): string[] {
   const list = (content.content ?? []).find(
     (n: any) => n.type === 'footnotesList',
   );
-  if (!list) return [];
-  return (list.content ?? [])
+  return (list?.content ?? [])
     .filter((n: any) => n.type === 'footnoteDefinition')
-    .map((n: any) => n.attrs?.id);
-}
-
-function definitionText(content: any, id: string): string | undefined {
-  const list = (content.content ?? []).find(
-    (n: any) => n.type === 'footnotesList',
-  );
-  const def = (list?.content ?? []).find(
-    (n: any) => n.type === 'footnoteDefinition' && n.attrs?.id === id,
-  );
-  return def?.content?.[0]?.content?.[0]?.text;
+    .map((n: any) => n.content?.[0]?.content?.[0]?.text);
 }
 
 describe('ImportService.importPage — footnote canonicalization (#228)', () => {
@@ -101,23 +97,23 @@ describe('ImportService.importPage — footnote canonicalization (#228)', () =>
     const content = getCaptured().content;
     expect(content).toBeTruthy();
 
-    // Reference order is c, a, b (NOT the markdown definition order a, b, c).
-    expect(footnoteListIds(content)).toEqual(['c', 'a', 'b']);
-
-    // Definitions preserved and attached to the right ids.
-    expect(definitionText(content, 'c')).toBe('note C');
-    expect(definitionText(content, 'a')).toBe('note A');
-    expect(definitionText(content, 'b')).toBe('note B');
+    // Definitions ordered by FIRST REFERENCE (C, A, B) — NOT the markdown
+    // definition order (A, B, C) — with the orphan [^z] dropped and the reused
+    // [^a] collapsed to a single definition. (Ids are the parser's fresh `fn-*`,
+    // so we pin the BODIES.)
+    expect(footnoteListBodies(content)).toEqual(['note C', 'note A', 'note B']);
 
     // Orphan definition [^z] is dropped.
-    expect(footnoteListIds(content)).not.toContain('z');
+    expect(footnoteListBodies(content)).not.toContain('orphan note');
 
     // Reused [^a] yields exactly ONE definition, and exactly one list.
     const lists = (content.content ?? []).filter(
       (n: any) => n.type === 'footnotesList',
     );
     expect(lists).toHaveLength(1);
-    expect(footnoteListIds(content).filter((id) => id === 'a')).toHaveLength(1);
+    expect(
+      footnoteListBodies(content).filter((b) => b === 'note A'),
+    ).toHaveLength(1);
   });
 
   it('is idempotent: canonicalizing the stored output again is a no-op', async () => {
@@ -134,6 +130,6 @@ describe('ImportService.importPage — footnote canonicalization (#228)', () =>
     // time must not change it (safe to wire into every write path).
     const second = canonicalizeFootnotes(stored);
     expect(second).toEqual(stored);
-    expect(footnoteListIds(second)).toEqual(['c', 'a', 'b']);
+    expect(footnoteListBodies(second)).toEqual(['note C', 'note A', 'note B']);
   });
 });

apps/server/src/integrations/import/services/import.service.ts

@@ -17,7 +17,9 @@ import {
 import { generateJitteredKeyBetween } from 'fractional-indexing-jittered';
 import { TiptapTransformer } from '@hocuspocus/transformer';
 import * as Y from 'yjs';
-import { markdownToHtml, canonicalizeFootnotes } from '@docmost/editor-ext';
+import { canonicalizeFootnotes } from '@docmost/editor-ext';
+import { markdownToProseMirror } from '@docmost/prosemirror-markdown';
+import { normalizeForeignMarkdown } from '../utils/foreign-markdown';
 import {
   FileTaskStatus,
   FileTaskType,
@@ -85,11 +87,13 @@ export class ImportService {
 
     const extracted = this.extractTitleAndRemoveHeading(prosemirrorState);
     const title = extracted.title;
-    // Imported markdown/HTML is built via markdownToHtml -> htmlToJson, which
-    // never runs the editor's footnoteSyncPlugin, so the footnote topology keeps
-    // the source's PHYSICAL definition order (out of order vs. references),
-    // retains orphan definitions, and is not deduped. Canonicalize before
-    // persisting so the stored page matches the editor's invariant (issue #228).
+    // The markdown path now canonicalizes footnotes itself (the package parser),
+    // but the HTML path (processHTML -> htmlToJson) does NOT run the editor's
+    // footnoteSyncPlugin, so an imported HTML doc can keep its source's PHYSICAL
+    // definition order (out of order vs. references), retain orphan definitions,
+    // and not be deduped. Canonicalize before persisting so the stored page
+    // matches the editor's invariant (issue #228); it is an idempotent no-op on
+    // the already-canonical markdown output.
     // Pure + idempotent + shape-safe: a doc with no footnotes is unchanged.
     // (Future consolidation, architecture B: this import path persists directly
     // via pageRepo.insertPage rather than through PageService.createPage, so the
@@ -133,12 +137,15 @@ export class ImportService {
   }
 
   async processMarkdown(markdownInput: string): Promise<any> {
-    try {
-      const html = await markdownToHtml(markdownInput);
-      return this.processHTML(html);
-    } catch (err) {
-      throw err;
-    }
+    // Canonical markdown -> ProseMirror JSON directly via
+    // `@docmost/prosemirror-markdown` (issue #345) — no HTML intermediate and no
+    // second editor-ext markdown layer. Foreign markdown surfaces the strict
+    // canonical parser does not accept (GFM `[^id]` reference footnotes) are
+    // rewritten to the canonical inline form by `normalizeForeignMarkdown` first.
+    // The HTML-cleanup pass (`normalizeImportHtml`) is intentionally skipped here:
+    // it targets foreign *HTML* (Notion/XWiki), which only ever arrives on the
+    // `.html` path (`processHTML`), never as canonical markdown.
+    return markdownToProseMirror(normalizeForeignMarkdown(markdownInput));
   }
 
   async processHTML(htmlInput: string): Promise<any> {

apps/server/src/integrations/import/utils/foreign-markdown.spec.ts

@@ -0,0 +1,218 @@
+import {
+  convertProseMirrorToMarkdown,
+  markdownToProseMirror,
+} from '@docmost/prosemirror-markdown';
+import { normalizeForeignMarkdown } from './foreign-markdown';
+
+/**
+ * STEP 2 goldens for issue #345: the foreign-markdown normalizer that runs at the
+ * import boundary BEFORE the strict canonical parser (`markdownToProseMirror`).
+ *
+ * Two layers:
+ *  1. PURE string→string cases pinning the normalizer's own behavior (GFM
+ *     reference footnotes → inline `^[…]`).
+ *  2. END-TO-END acceptance: for a foreign corpus, `normalizeForeignMarkdown`
+ *     then `markdownToProseMirror` then `convertProseMirrorToMarkdown` must leave
+ *     NO literal `[^id]` / `:::` garbage in the document and must re-export in the
+ *     canonical forms.
+ */
+
+describe('normalizeForeignMarkdown — GFM reference footnotes', () => {
+  it('inlines a single-line reference footnote and drops its definition', () => {
+    const out = normalizeForeignMarkdown(
+      'A note[^1] here.\n\n[^1]: The definition.',
+    );
+    expect(out).toBe('A note^[The definition.] here.\n');
+  });
+
+  it('inlines every reference to a reused id (downstream dedups)', () => {
+    const out = normalizeForeignMarkdown(
+      'X[^a] and Y[^a].\n\n[^a]: shared.',
+    );
+    expect(out).toBe('X^[shared.] and Y^[shared.].\n');
+  });
+
+  it('joins indented continuation lines of a definition with a space', () => {
+    const out = normalizeForeignMarkdown(
+      'See[^n].\n\n[^n]: line one\n    line two',
+    );
+    expect(out).toBe('See^[line one line two].\n');
+  });
+
+  it('never rewrites a reference inside a fenced code block', () => {
+    const out = normalizeForeignMarkdown(
+      '```\ncode[^1] here\n```\n\n[^1]: def.',
+    );
+    expect(out).toContain('code[^1] here');
+    // The (now orphaned) definition line is still removed.
+    expect(out).not.toContain('[^1]: def.');
+  });
+
+  it('never rewrites a reference inside an INLINE-code span (backticks)', () => {
+    // The `[^1]` inside backticks is literal code and must survive verbatim;
+    // the one outside is rewritten. (Bug #1: only fenced blocks were protected.)
+    const out = normalizeForeignMarkdown(
+      'Use `arr[^1]` in code but note[^1] in prose.\n\n[^1]: def.',
+    );
+    expect(out).toBe('Use `arr[^1]` in code but note^[def.] in prose.\n');
+  });
+
+  it('escapes brackets in a body so an unbalanced ] cannot truncate the footnote', () => {
+    // A foreign definition body with a stray `]` would, unescaped, close the
+    // canonical `^[...]` early and leak the tail as text (bug #2). The body's
+    // brackets are backslash-escaped so the footnote stays whole.
+    const out = normalizeForeignMarkdown(
+      'Ref[^1] here.\n\n[^1]: see item ] and [more] later',
+    );
+    expect(out).toBe('Ref^[see item \\] and \\[more\\] later] here.\n');
+    // The tokenizer must see exactly one unescaped closing bracket (our own).
+    expect(out.match(/(?<!\\)\]/g)).toHaveLength(1);
+  });
+
+  it('leaves a reference with no matching definition literal (no body to inline)', () => {
+    const out = normalizeForeignMarkdown('Dangling[^x] ref.');
+    expect(out).toBe('Dangling[^x] ref.');
+  });
+
+  it('returns the input unchanged when there are no reference footnotes', () => {
+    const md = '# Title\n\nJust text with `inline code` and a [link](/x).';
+    expect(normalizeForeignMarkdown(md)).toBe(md);
+  });
+
+  it('does NOT touch callout surfaces — the canonical parser handles them', () => {
+    const callouts = ':::info\nHi\n:::\n\n> [!warning]\n> Careful';
+    expect(normalizeForeignMarkdown(callouts)).toBe(callouts);
+  });
+
+  it('strips a leading YAML front-matter block (Obsidian/Hugo/git-sync files)', () => {
+    const out = normalizeForeignMarkdown(
+      '---\ntitle: My Page\ntags: [a, b]\n---\n\n# Heading\n\nBody.',
+    );
+    expect(out).toBe('# Heading\n\nBody.');
+    // The front-matter must not leak into the body as a setext heading.
+    expect(out).not.toContain('title: My Page');
+    expect(out).not.toContain('---');
+  });
+
+  it('does not strip a horizontal rule that is not leading front-matter', () => {
+    const md = 'Intro paragraph.\n\n---\n\nAfter the rule.';
+    expect(normalizeForeignMarkdown(md)).toBe(md);
+  });
+
+  it('is linear on a document with thousands of definitions (no quadratic blowup)', () => {
+    // F2(a): the pass-2 rewrite must be O(text), not O(text × defs). Build a
+    // pathological doc (many defs + many plain text lines) and assert it
+    // completes well under a second — a quadratic implementation took ~14s.
+    const N = 4000;
+    const refs = Array.from({ length: N }, (_, i) => `line ${i} plain text`).join('\n');
+    const defs = Array.from({ length: N }, (_, i) => `[^n${i}]: def ${i}`).join('\n');
+    const doc = `start[^n0] and[^n${N - 1}] end\n\n${refs}\n\n${defs}`;
+    const t0 = Date.now();
+    const out = normalizeForeignMarkdown(doc);
+    const elapsed = Date.now() - t0;
+    expect(elapsed).toBeLessThan(2000);
+    // Sanity: the two real references were still inlined.
+    expect(out).toContain('^[def 0]');
+    expect(out).toContain(`^[def ${N - 1}]`);
+  });
+
+  it('is bounded on a long unclosed backtick run (no inline-split ReDoS)', () => {
+    // F2(b): a huge unterminated backtick run must not cause quadratic
+    // backtracking in the inline-code split. Oversized lines skip the split
+    // entirely (left untouched), so this returns promptly.
+    const line = 'x' + '`'.repeat(200000);
+    const doc = `${line}\n\n[^1]: def`;
+    const t0 = Date.now();
+    normalizeForeignMarkdown(doc);
+    expect(Date.now() - t0).toBeLessThan(2000);
+  });
+
+  it('does not crash or slow down on thousands of prefix-chain definition ids', () => {
+    // F7: the rewrite must use a FIXED generic scanner, not an alternation built
+    // from the ids. A `(a|aa|aaa|…)` alternation over prefix-chain ids blows the
+    // V8 regex compiler (FATAL RegExpCompiler Allocation failed — uncatchable,
+    // kills the process). A fixed scanner has no id-dependent compilation cost.
+    const N = 4000;
+    const ids = Array.from({ length: N }, (_, i) => 'a'.repeat(i + 1));
+    const defs = ids.map((id) => `[^${id}]: body ${id.length}`).join('\n');
+    const doc = `ref[^${ids[0]}] and[^${ids[N - 1]}] end\n\n${defs}`;
+    const t0 = Date.now();
+    const out = normalizeForeignMarkdown(doc);
+    expect(Date.now() - t0).toBeLessThan(2000);
+    // Prefix disambiguation is correct: [^a] and [^aaaa...] inline their OWN body.
+    expect(out).toContain('^[body 1]');
+    expect(out).toContain(`^[body ${N}]`);
+  });
+
+  it('strips a CRLF (Windows) front-matter block, not just LF', () => {
+    // F9: the line-anchored regex needs LF after the opening `---`, so a Windows
+    // file (`---\r\n…`) would slip past the strip and leak the front-matter into
+    // the body. normalizeForeignMarkdown normalizes CRLF -> LF first.
+    const out = normalizeForeignMarkdown(
+      '---\r\ntitle: Foo\r\ntags: [a]\r\n---\r\n\r\n# Heading\r\n\r\nBody.',
+    );
+    expect(out).toBe('# Heading\n\nBody.');
+    expect(out).not.toContain('title: Foo');
+    expect(out).not.toContain('---');
+  });
+
+  it('strips front-matter whose value contains a triple-dash (line-anchored)', () => {
+    // F8: the block must close only on a `\n---` LINE, not the first inline
+    // `---`. A value like `title: Q1 --- Q2` must not truncate the front-matter
+    // and leak the rest (author/closing ---) into the body.
+    const out = normalizeForeignMarkdown(
+      '---\ntitle: Q1 --- Q2 results\nauthor: bob\n---\n\nReal body.',
+    );
+    expect(out).toBe('Real body.');
+    expect(out).not.toContain('author: bob');
+    expect(out).not.toContain('Q2 results');
+  });
+});
+
+describe('foreign markdown import acceptance (normalizer + canonical parser)', () => {
+  const FOREIGN = [
+    '# Doc',
+    '',
+    'Body refs [^c] and [^a] and [^b] and again [^a].',
+    '',
+    ':::info',
+    'A legacy callout.',
+    ':::',
+    '',
+    '| h1 | h2 |',
+    '| --- | --- |',
+    '| 1 | 2 |',
+    '',
+    '[^a]: note A',
+    '[^b]: note B',
+    '[^c]: note C',
+    '[^z]: orphan note',
+  ].join('\n');
+
+  it('leaves no literal [^id] or ::: in the imported doc and re-exports canonically', async () => {
+    const normalized = normalizeForeignMarkdown(FOREIGN);
+    const doc = await markdownToProseMirror(normalized);
+    const reexport = convertProseMirrorToMarkdown(doc);
+
+    // No foreign garbage leaks into the document.
+    expect(reexport).not.toMatch(/\[\^/); // no reference footnote refs/defs
+    expect(reexport).not.toContain(':::'); // no legacy callout fences
+
+    // Canonical forms are present.
+    expect(reexport).toContain('^[note C]');
+    expect(reexport).toContain('> [!info]');
+    expect(reexport).toContain('| h1 | h2 |');
+
+    // Footnotes: ordered by first reference (C, A, B), reused [^a] deduped to one,
+    // orphan [^z] dropped (it had no reference after normalization).
+    const list = doc.content.find((n: any) => n.type === 'footnotesList');
+    const bodies = list.content.map(
+      (d: any) => d.content[0].content[0].text,
+    );
+    expect(bodies).toEqual(['note C', 'note A', 'note B']);
+    expect(bodies).not.toContain('orphan note');
+    expect(
+      doc.content.filter((n: any) => n.type === 'footnotesList'),
+    ).toHaveLength(1);
+  });
+});

apps/server/src/integrations/import/utils/foreign-markdown.ts

@@ -0,0 +1,265 @@
+/**
+ * Foreign-markdown normalizer — an input-liberal / output-canonical adapter that
+ * runs at the IMPORT boundary, BEFORE the canonical parser
+ * (`markdownToProseMirror` from `@docmost/prosemirror-markdown`).
+ *
+ * The canonical parser is deliberately STRICT: it only understands Docmost's
+ * canonical markdown surface (Obsidian-style `> [!type]` callouts, Pandoc/Obsidian
+ * inline footnotes `^[body]`, lossless `![alt](src) <!--img {...}-->` images, …).
+ * Import, however, ingests FOREIGN files (GitHub/GFM, Notion, old Docmost
+ * exports). Those use surfaces the canonical parser does not accept, most notably
+ * GitHub-flavoured *reference* footnotes:
+ *
+ *     Text with a note[^1] and another[^long].
+ *
+ *     [^1]: The first definition.
+ *     [^long]: A second one.
+ *
+ * Left untouched, the parser does NOT recognise `[^id]` (it only parses `^[body]`),
+ * so the reference leaks as literal text — and worse, the trailing `[^id]: def`
+ * line is a valid CommonMark *link-reference definition*, so `[^id]` is silently
+ * rendered as a bogus link. This normalizer rewrites reference footnotes into the
+ * canonical inline form so the parser materialises real footnote nodes.
+ *
+ * This is a TEXT pre-pass, NOT a second parser fork: it does not re-implement any
+ * converter logic. Callout surfaces (`:::type` and `> [!type]`) are intentionally
+ * NOT touched here — the canonical parser already accepts BOTH natively (its
+ * `preprocessCallouts` pass), so normalizing them would be redundant and would
+ * only risk degrading the parser's nesting/code-fence-aware handling.
+ */
+
+/** Matches a fenced code block delimiter (``` or ~~~), capturing the marker run. */
+const CODE_FENCE_RE = /^(\s*)(`{3,}|~{3,})/;
+
+/**
+ * Matches a GFM footnote DEFINITION line: `[^id]: body`. The id is any run of
+ * non-`]` characters; the body is the remainder of the line (possibly empty).
+ */
+const FOOTNOTE_DEF_RE = /^\[\^([^\]]+)\]:[ \t]?(.*)$/;
+
+/** True when a line is a code-fence delimiter that toggles fenced-code state. */
+function fenceMarker(line: string): string | null {
+  const m = line.match(CODE_FENCE_RE);
+  return m ? m[2] : null;
+}
+
+/** True when a line is indented (leading space/tab) and not blank — a continuation. */
+function isIndentedContinuation(line: string): boolean {
+  return /^[ \t]+\S/.test(line);
+}
+
+function escapeRegExp(value: string): string {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+}
+
+/**
+ * Backslash-escape any square bracket in a footnote body before it is wrapped in
+ * `^[...]`. The canonical inline-footnote tokenizer scans the body with bracket
+ * balancing and closes on the first UNMATCHED `]`, so an unbalanced bracket in a
+ * foreign definition (e.g. `[^1]: see item ] later`) would otherwise truncate the
+ * footnote and leak the tail as literal text. Escaping every `[`/`]` makes the
+ * body an inert run of characters — the tokenizer then closes only on our own
+ * closing `]`. (A balanced `[link](url)` inside a body still round-trips because
+ * the escaped form renders the literal brackets, which is the safe reading for a
+ * footnote body; the alternative — brittle balance tracking — risks worse.)
+ */
+function escapeFootnoteBody(body: string): string {
+  return body.replace(/[[\]]/g, '\\$&');
+}
+
+/**
+ * Rewrite every `[^id]` reference on a line to its `^[body]` form, but ONLY in the
+ * text OUTSIDE inline-code spans. A `[^id]` inside backticks is literal code
+ * content and must be preserved verbatim (a footnote ref never lives inside code).
+ * We split the line on inline-code spans (paired backtick runs) and rewrite only
+ * the non-code segments.
+ */
+// Above this length a single line is not split into inline-code spans (see
+// below). A genuine markdown line carrying a footnote reference is never tens of
+// KB; the cap only bypasses the inline-code protection for pathological lines.
+const INLINE_SPLIT_MAX_LINE = 8192;
+
+function rewriteRefsOutsideInlineCode(
+  line: string,
+  replace: (text: string) => string,
+): string {
+  // The inline-code split alternation `(`+)(?:(?!\1)[\s\S])*\1` backtracks
+  // quadratically on a long UNCLOSED backtick run (its middle can consume the
+  // rest of the line, then fail to find a closing run and retry from each
+  // position). On an untrusted import this is a request-thread ReDoS. A real
+  // footnote line is short, so for an oversized line we skip the inline-code
+  // protection entirely and leave the line UNTOUCHED (rewriting it wholesale
+  // could corrupt a `[^id]` that legitimately lives inside inline code). This is
+  // a conservative bypass: an over-8KB line simply does not get its reference
+  // footnotes inlined — acceptable for a pathological input.
+  if (line.length > INLINE_SPLIT_MAX_LINE) return line;
+
+  // Alternation: an inline-code span (one or more backticks, then anything up to
+  // the SAME run of backticks) OR a run of non-backtick text. Unterminated
+  // backticks fall through as ordinary text (matched by the second branch on the
+  // leftover), so a stray backtick never swallows the rest of the line.
+  const parts = line.match(/(`+)(?:(?!\1)[\s\S])*\1|[^`]+|`+/g);
+  if (!parts) return line;
+  return parts
+    .map((seg) => (seg.startsWith('`') ? seg : replace(seg)))
+    .join('');
+}
+
+/**
+ * Convert GFM reference footnotes (`[^id]` + `[^id]: def`) into canonical inline
+ * footnotes (`^[def]`).
+ *
+ * - Definitions are collected first (a leading `[^id]: text` line plus any
+ *   immediately-following indented continuation lines, joined with a space) and
+ *   removed from the output.
+ * - Each in-text reference `[^id]` for which a definition was found is replaced by
+ *   `^[def]`. References with no matching definition are left literal (there is no
+ *   body to inline; the parser fails them open the same way).
+ * - Code is respected on both passes: `[^id]` inside a fenced ``` / ~~~ block is
+ *   never rewritten and a `[^id]:` line inside a fence is never a definition; and
+ *   on the rewrite pass a `[^id]` inside an INLINE-code span (backticks) is left
+ *   literal too.
+ * - The inlined body is bracket-escaped so an unbalanced `[`/`]` in a foreign
+ *   definition cannot truncate the resulting `^[...]` footnote.
+ *
+ * Deduplication / reference-ordering / orphan-dropping of the resulting footnotes
+ * is handled downstream by the canonical parser (`assembleFootnotes`); this pass
+ * only changes the surface syntax.
+ */
+function convertReferenceFootnotes(markdown: string): string {
+  const lines = markdown.split('\n');
+
+  // Pass 1: collect definitions and mark their lines for removal.
+  const defs = new Map<string, string>();
+  const dropped = new Array<boolean>(lines.length).fill(false);
+  let inFence = false;
+  let fence = '';
+
+  for (let i = 0; i < lines.length; i++) {
+    const line = lines[i];
+    const marker = fenceMarker(line);
+    if (inFence) {
+      if (marker && marker[0] === fence[0] && marker.length >= fence.length) {
+        inFence = false;
+        fence = '';
+      }
+      continue;
+    }
+    if (marker) {
+      inFence = true;
+      fence = marker;
+      continue;
+    }
+
+    const def = line.match(FOOTNOTE_DEF_RE);
+    if (!def) continue;
+
+    const id = def[1];
+    const body: string[] = [def[2].trim()];
+    dropped[i] = true;
+
+    // Consume immediately-following indented continuation lines (GFM lazy
+    // continuation is not supported by design — keep it simple and predictable).
+    let j = i + 1;
+    while (j < lines.length && isIndentedContinuation(lines[j])) {
+      body.push(lines[j].trim());
+      dropped[j] = true;
+      j++;
+    }
+    i = j - 1;
+
+    // Last definition wins for a duplicated id (matches CommonMark link-ref
+    // semantics closely enough for a foreign-input adapter).
+    defs.set(id, body.filter((s) => s.length > 0).join(' '));
+  }
+
+  if (defs.size === 0) {
+    return markdown;
+  }
+
+  // ONE fixed, generic scanner regex — NOT one built from the definition ids.
+  // It matches ANY `[^id]` shape, and the replacer decides per match via a map
+  // lookup whether that id is a real definition (replace) or not (leave as-is).
+  // This is genuinely O(total text) with no per-document regex compilation.
+  //
+  // Do NOT rebuild this as an alternation over `[...defs.keys()]`: a giant
+  // `(id1|id2|...)` alternation over thousands of ids can blow the V8 regex
+  // compiler's stack — a fatal, UNCATCHABLE "RegExpCompiler Allocation failed"
+  // on prefix-chain ids (`a`, `aa`, `aaa`, ...) that kills the whole process
+  // (worse than the earlier per-def thread-hang). A fixed scanner has no
+  // id-dependent compilation cost and cannot blow up.
+  const refRe = /\[\^([^\]]+)\]/g;
+  const rewriteSegment = (segment: string): string =>
+    segment.replace(refRe, (whole, id: string) => {
+      const body = defs.get(id);
+      // Only real definitions are inlined; an unknown id is left literal (same as
+      // the old per-def loop, which simply never matched it).
+      return body === undefined ? whole : `^[${escapeFootnoteBody(body)}]`;
+    });
+
+  // Pass 2: rewrite in-text references, skipping fenced code and dropped lines.
+  const out: string[] = [];
+  inFence = false;
+  fence = '';
+  for (let i = 0; i < lines.length; i++) {
+    if (dropped[i]) continue;
+    let line = lines[i];
+
+    const marker = fenceMarker(line);
+    if (inFence) {
+      out.push(line);
+      if (marker && marker[0] === fence[0] && marker.length >= fence.length) {
+        inFence = false;
+        fence = '';
+      }
+      continue;
+    }
+    if (marker) {
+      inFence = true;
+      fence = marker;
+      out.push(line);
+      continue;
+    }
+
+    line = rewriteRefsOutsideInlineCode(line, rewriteSegment);
+    out.push(line);
+  }
+
+  return out.join('\n');
+}
+
+/**
+ * Strip a single leading YAML front-matter block (`---\n…\n---`). Foreign files
+ * from Obsidian / Hugo / Jekyll / Notion — and Docmost's OWN git-sync page files
+ * — open with front-matter that the canonical parser does not consume, so
+ * without this it leaks into the body (and `title: Foo` above the closing `---`
+ * renders as a setext `<h2>` that `extractTitleAndRemoveHeading` can hijack as
+ * the page title). It is a no-op for front-matter-free input.
+ *
+ * LINE-ANCHORED (the same shape the canonical parser uses in
+ * prosemirror-markdown/page-file.ts): the block opens only on `---\n` at the
+ * very start and closes only on a `\n---` line. The retired `markdownToHtml`
+ * strip closed on the FIRST `---` ANYWHERE (an unanchored close), so a value
+ * containing a triple-dash (e.g. `title: Q1 --- Q2`) truncated the front-matter
+ * and leaked the rest into the body. An optional leading BOM is tolerated.
+ */
+const YAML_FRONT_MATTER_RE = /^\uFEFF?---\n[\s\S]*?\n---\n?/;
+
+/**
+ * Normalize a foreign markdown string into Docmost's canonical markdown surface
+ * so the strict canonical parser accepts it losslessly: normalize line endings,
+ * strip a leading YAML front-matter block, then rewrite GFM reference footnotes
+ * into inline footnotes. Add further fixture-driven foreign-surface cases here as
+ * they are found.
+ */
+export function normalizeForeignMarkdown(markdown: string): string {
+  if (!markdown) return markdown;
+  // Normalize CRLF -> LF FIRST. The line-anchored front-matter regex requires a
+  // bare `\n` after the opening `---`, and convertReferenceFootnotes splits on
+  // `\n`; a Windows/CRLF foreign file (`---\r\n…`) would otherwise slip past the
+  // front-matter strip and leak into the body. The canonical parser
+  // (page-file.ts parsePageFile) normalizes the same way before its FRONTMATTER_RE.
+  const src = markdown.replace(/\r\n/g, '\n');
+  const withoutFrontMatter = src.replace(YAML_FRONT_MATTER_RE, '').trimStart();
+  return convertReferenceFootnotes(withoutFrontMatter);
+}

apps/server/src/integrations/metrics/metrics-bull.service.ts

@@ -46,7 +46,6 @@ export class MetricsBullService implements OnModuleInit, OnModuleDestroy {
     @InjectQueue(QueueName.GENERAL_QUEUE) generalQueue: Queue,
     @InjectQueue(QueueName.BILLING_QUEUE) billingQueue: Queue,
     @InjectQueue(QueueName.FILE_TASK_QUEUE) fileTaskQueue: Queue,
-    @InjectQueue(QueueName.SEARCH_QUEUE) searchQueue: Queue,
     @InjectQueue(QueueName.AI_QUEUE) aiQueue: Queue,
     @InjectQueue(QueueName.HISTORY_QUEUE) historyQueue: Queue,
     @InjectQueue(QueueName.NOTIFICATION_QUEUE) notificationQueue: Queue,
@@ -58,7 +57,6 @@ export class MetricsBullService implements OnModuleInit, OnModuleDestroy {
       { label: 'general', queue: generalQueue },
       { label: 'billing', queue: billingQueue },
       { label: 'file-task', queue: fileTaskQueue },
-      { label: 'search', queue: searchQueue },
       { label: 'ai', queue: aiQueue },
       { label: 'history', queue: historyQueue },
       { label: 'notification', queue: notificationQueue },

apps/server/src/integrations/queue/constants/queue.constants.ts

@@ -4,7 +4,6 @@ export enum QueueName {
   GENERAL_QUEUE = '{general-queue}',
   BILLING_QUEUE = '{billing-queue}',
   FILE_TASK_QUEUE = '{file-task-queue}',
-  SEARCH_QUEUE = '{search-queue}',
   AI_QUEUE = '{ai-queue}',
   HISTORY_QUEUE = '{history-queue}',
   NOTIFICATION_QUEUE = '{notification-queue}',
@@ -32,12 +31,6 @@ export enum QueueJob {
   IMPORT_TASK = 'import-task',
   EXPORT_TASK = 'export-task',
 
-  SEARCH_INDEX_PAGE = 'search-index-page',
-  SEARCH_INDEX_PAGES = 'search-index-pages',
-  SEARCH_INDEX_COMMENT = 'search-index-comment',
-  SEARCH_INDEX_COMMENTS = 'search-index-comments',
-  SEARCH_INDEX_ATTACHMENT = 'search-index-attachment',
-  SEARCH_INDEX_ATTACHMENTS = 'search-index-attachments',
   SEARCH_REMOVE_PAGE = 'search-remove-page',
   SEARCH_REMOVE_ASSET = 'search-remove-attachment',
   SEARCH_REMOVE_FACE = 'search-remove-comment',

apps/server/src/integrations/queue/queue.module.ts

@@ -57,14 +57,6 @@ import { GeneralQueueProcessor } from './processors/general-queue.processor';
         attempts: 1,
       },
     }),
-    BullModule.registerQueue({
-      name: QueueName.SEARCH_QUEUE,
-      defaultJobOptions: {
-        removeOnComplete: true,
-        removeOnFail: true,
-        attempts: 2,
-      },
-    }),
     BullModule.registerQueue({
       name: QueueName.AI_QUEUE,
       defaultJobOptions: {

apps/server/test/integration/db.ts

@@ -38,6 +38,24 @@ export const TEST_DATABASE_URL =
   process.env.TEST_DATABASE_URL ??
   'postgresql://docmost:docmost_dev_pw@localhost:5432/docmost_test';
 
+// Build the raw postgres.js client (mirrors database.module.ts: max pool,
+// silenced notices, bigint-as-number parsing). Kept separate so the singleton
+// can hold a reference to bound its shutdown in destroyTestDb.
+function buildTestSql(url: string = TEST_DATABASE_URL) {
+  return postgres(url, {
+    max: 5,
+    onnotice: () => {},
+    types: {
+      bigint: {
+        to: 20,
+        from: [20, 1700],
+        serialize: (value: number) => value.toString(),
+        parse: (value: string) => Number.parseInt(value),
+      },
+    },
+  });
+}
+
 /**
  * Build a Kysely instance that MIRRORS the app's setup in database.module.ts:
  * PostgresJSDialect over postgres(), CamelCasePlugin, and the bigint type
@@ -47,38 +65,40 @@ export const TEST_DATABASE_URL =
  */
 export function buildTestDb(url: string = TEST_DATABASE_URL): Kysely<any> {
   return new Kysely<any>({
-    dialect: new PostgresJSDialect({
-      postgres: postgres(url, {
-        max: 5,
-        onnotice: () => {},
-        types: {
-          bigint: {
-            to: 20,
-            from: [20, 1700],
-            serialize: (value: number) => value.toString(),
-            parse: (value: string) => Number.parseInt(value),
-          },
-        },
-      }),
-    }),
+    dialect: new PostgresJSDialect({ postgres: buildTestSql(url) }),
     plugins: [new CamelCasePlugin()],
   });
 }
 
 let singleton: Kysely<any> | undefined;
+let singletonSql: ReturnType<typeof buildTestSql> | undefined;
 
 /** Lazily-built shared Kysely for the test suite (one per worker; maxWorkers=1). */
 export function getTestDb(): Kysely<any> {
   if (!singleton) {
-    singleton = buildTestDb();
+    singletonSql = buildTestSql();
+    singleton = new Kysely<any>({
+      dialect: new PostgresJSDialect({ postgres: singletonSql }),
+      plugins: [new CamelCasePlugin()],
+    });
   }
   return singleton;
 }
 
 export async function destroyTestDb(): Promise<void> {
-  if (singleton) {
-    await singleton.destroy();
-    singleton = undefined;
+  if (!singleton) return;
+  const sql = singletonSql;
+  // Clear the refs first so a hung end() cannot leave a half-closed singleton.
+  singleton = undefined;
+  singletonSql = undefined;
+  // postgres.js .end() waits indefinitely for in-flight queries by default; a
+  // leaked/stuck pooled connection would hang the afterAll hook (a 60s hook
+  // timeout in CI). Bound the shutdown: the { timeout } grace period lets
+  // active queries drain, then force-closes lingering sockets so teardown
+  // always completes. We close the pool directly instead of Kysely.destroy()
+  // (which would call sql.end() again with no timeout).
+  if (sql) {
+    await sql.end({ timeout: 5 });
   }
 }
 

apps/server/test/jest-e2e.json

@@ -4,10 +4,14 @@
   "testEnvironment": "node",
   "testRegex": ".e2e-spec.ts$",
   "transform": {
+    "prosemirror-markdown/build/.+\\.js$": [
+      "babel-jest",
+      { "presets": [["@babel/preset-env", { "targets": { "node": "current" } }]] }
+    ],
     "^.+\\.(t|j)sx?$": ["ts-jest", { "tsconfig": { "allowJs": true } }]
   },
   "transformIgnorePatterns": [
-    "/node_modules/(?!(\\.pnpm/)?(nanoid|uuid|image-dimensions|marked|happy-dom|lib0|@sindresorhus[+/][a-z0-9-]+|escape-string-regexp|p-limit|yocto-queue)(@|/))"
+    "/node_modules/(?!(\\.pnpm/)?(nanoid|uuid|image-dimensions|marked|happy-dom|lib0|@sindresorhus[+/][a-z0-9-]+|escape-string-regexp|p-limit|yocto-queue|@docmost/prosemirror-markdown)(@|/))"
   ],
   "moduleNameMapper": {
     "^@docmost/db/(.*)$": "<rootDir>/../src/database/$1",

apps/server/test/jest-integration.json

@@ -4,14 +4,19 @@
   "testRegex": ".*\\.int-spec\\.ts$",
   "testPathIgnorePatterns": ["/node_modules/"],
   "transform": {
+    "prosemirror-markdown/build/.+\\.js$": [
+      "babel-jest",
+      { "presets": [["@babel/preset-env", { "targets": { "node": "current" } }]] }
+    ],
     "^.+\\.(t|j)sx?$": "ts-jest"
   },
   "transformIgnorePatterns": [
-    "/node_modules/(?!(\\.pnpm/)?(nanoid|uuid|image-dimensions|marked|happy-dom|lib0)(@|/))"
+    "/node_modules/(?!(\\.pnpm/)?(nanoid|uuid|image-dimensions|marked|happy-dom|lib0|@docmost/prosemirror-markdown)(@|/))"
   ],
   "testEnvironment": "node",
   "testTimeout": 60000,
   "maxWorkers": 1,
+  "forceExit": true,
   "globalSetup": "<rootDir>/test/integration/global-setup.ts",
   "globalTeardown": "<rootDir>/test/integration/global-teardown.ts",
   "moduleNameMapper": {

apps/server/test/stubs/tiptap-react.js

@@ -0,0 +1,23 @@
+// Jest stub for @tiptap/react.
+//
+// The server export/import code paths transitively import editor-ext, whose node
+// extensions import from `@tiptap/react`. The real module re-exports all of
+// `@tiptap/core` (headless, safe under node) AND adds React view helpers
+// (`ReactNodeViewRenderer`, …) that eagerly pull in react-dom — which throws
+// `navigator is not defined` under jest's node environment.
+//
+// So this stub DELEGATES to the real `@tiptap/core` (keeping `mergeAttributes`,
+// `Node`, `Mark`, `nodeInputRule`, … working — they are used by
+// `jsonToHtml`/`htmlToJson` on the server) and overrides ONLY the React view
+// helpers with no-ops. Those helpers are referenced solely inside `addNodeView()`
+// — code that runs only in a live browser editor, never on the server; if any
+// were actually invoked here it would (correctly) surface as a test failure.
+const core = require('@tiptap/core');
+
+module.exports = {
+  ...core,
+  ReactNodeViewRenderer: () => () => ({}),
+  NodeViewWrapper: () => null,
+  NodeViewContent: () => null,
+  ReactRenderer: class {},
+};

docs/dev-stand.md

@@ -131,5 +131,14 @@ const { Client } = require("pg");
 7. **Migrations don't auto-run in dev** — run `migration:latest` after every pull
    or branch switch.
 
+8. **Automation (Playwright): type into the BODY editor, not the title.** A page has
+   two `.ProseMirror` editors — `[aria-label='Page title']` (non-collab) and
+   `[aria-label='Page content']` (the collab body). `document.querySelector('.ProseMirror')`
+   returns the TITLE editor, so typing there never changes body content and `mod+S`
+   versions nothing. Target `[aria-label='Page content']`, confirm it's collab-bound
+   (`el.editor.extensionManager.extensions.some(e=>e.name==='collaboration')`), and
+   wait ~10-12s for the store debounce before asserting `pages.content` changed. Full
+   testing methodology + traps: **[how-to-test.md](how-to-test.md)**.
+
 See also the **Commands** and **Architecture → Two server processes** sections in
 [`AGENTS.md`](../AGENTS.md).

docs/how-to-test.md

@@ -0,0 +1,108 @@
+# How to test the application (browser E2E + out-of-band)
+
+How to actually verify a feature end-to-end against a running stand — driving the
+**real app in a browser** and confirming results **out-of-band** in the DB/git, not
+through the same API you're supposed to be testing. Written from real false-positives
+that wasted hours (see **Traps** — read them before you write a test).
+
+Prereq: a running stand — see **[dev-stand.md](dev-stand.md)**. Automation uses
+Playwright (`pip install playwright && python -m playwright install chromium`).
+
+## Principles
+
+1. **Drive the behaviour under test through the browser.** The stand exists so you
+   exercise the real UI + realtime-collab + server path. Using `POST /api/pages/*` to
+   perform the action you're validating tests the API, not the app — an e2e suite can
+   do that. API calls are fine ONLY for one-time setup/fixtures, never for the
+   interaction you're asserting on.
+2. **Evidence before claim.** Nothing "passes" without an artifact: a DB row, a git
+   diff, a screenshot looked at as an image. If you can't show it, you didn't verify it.
+3. **Verify out-of-band.** Judge results from a source independent of the UI: `psql`
+   against the DB, a fresh `git clone` of a synced repo, a hard reload. Optimistic UI
+   lies about persistence.
+4. **Disconfirm by default.** For each feature, actively try to prove it's broken
+   before concluding it works. Reload after every create/edit/save.
+5. **Recon actuatability FIRST.** Before building editor tests, confirm the
+   interaction even works in your harness (does a typed edit reach the DB?). Skipping
+   this is how you ship a pile of tests that all silently exercised the wrong thing.
+
+## The editor: two ProseMirror instances (READ THIS)
+
+A page has **two** `.ProseMirror` editors:
+
+| index | selector | role | collab? |
+|---|---|---|---|
+| 0 | `[aria-label='Page title']` | title field | **NO** (16 exts, no `collaboration`) |
+| 1 | `[aria-label='Page content']` | body | **YES** (95 exts, has `collaboration`) |
+
+`document.querySelector('.ProseMirror')` returns the **title** editor (first match).
+Type there and you edit the title only — body page content never changes, so `mod+S`
+"versions" unchanged content and every content test silently no-ops.
+
+**Always target the body editor** and confirm it's collab-bound before typing:
+
+```js
+const el = document.querySelector("[aria-label='Page content']");
+el.editor.extensionManager.extensions.some(e => e.name === 'collaboration'); // must be true
+```
+
+Body edits emit ~20 `/collab` websocket frames while typing and land in
+`pages.content` after the **hocuspocus store debounce (~10s)** — so **wait ~12s**
+before asserting persistence (checking at 6–8s is a false negative). `mod+S` (the
+`save-version` stateless message) flushes immediately, so a version created right
+after a settled body edit holds the typed text.
+
+## A known-good browser flow
+
+```
+1. goto /s/<space-slug>            # the "Create page" button lives in the space sidebar, not /home
+2. click button[aria-label='Create page']   # fully UI-driven page creation
+3. type into [aria-label='Page title']       # optional title
+4. click [aria-label='Page content'] → type body text
+5. wait ~12s (store debounce)
+6. assert pages.content changed (psql)       # out-of-band
+7. mod+S / menu Save → assert page_history row (psql)
+8. reload / fresh context → re-assert (persistence round-trip)
+```
+
+Auth: log in ONCE, save `storage_state.json`, reuse it across pages/agents (re-login
+per run trips shared rate-limits). Cookie-based session authorizes both REST and the
+collab websocket.
+
+## Judging out-of-band
+
+```bash
+# page content / history
+docker exec <db> psql -U docmost -d docmost -tAc \
+  "select coalesce(kind,'null'), content::text from page_history where page_id='<id>' order by created_at;"
+# git-sync round-trip: clone the space repo and diff against what you pushed
+git clone http://<user>:<pass>@127.0.0.1:3000/git/<spaceId>.git /tmp/x
+```
+
+`page_history.content` is full JSON — parse it, don't truncate the snippet, or a
+marker check misses. For sync/async features (autosave, git-sync, idle-flush) use an
+active probe: write a unique marker, wait past the debounce/poll window, re-read
+out-of-band, ≥2 iterations — never conclude "broken" from a single snapshot.
+
+## Traps (each of these produced a false result in a real run)
+
+- **Wrong editor.** Typed into `.ProseMirror` (= title). Edits never touched body
+  content. → target `[aria-label='Page content']`.
+- **Checked persistence too early.** Store debounce ~10s; a 6–8s check reads stale.
+- **Truncated the DB snapshot** below where the test marker sits → false "content
+  missing".
+- **API-seeded the content under test**, then "verified" the feature — that validated
+  the API, not the app.
+- **Reused a fixed marker on a non-rebooted stand** → title/row collisions inflate
+  counts (`count==2`). Use a unique per-run marker (timestamp).
+- **Idle/async read once** and called it "permanently broken" — it was mid-debounce.
+- **Concluded env-limitation without a cross-build control.** If unsure whether a
+  failure is your harness or the product, run the SAME harness against a known-good
+  build; a divergence localizes it.
+
+## Scope note
+
+Some paths genuinely need a human in a real browser (rich drag-drop, native file
+pickers, clipboard, and anything the harness can't actuate). Label those UNTESTED in
+the report — "handled gracefully" is not "works". Keep four states distinct:
+verified-working, defect, untested, env-limitation.

package.json

@@ -96,7 +96,8 @@
   "pnpm": {
     "patchedDependencies": {
       "scimmy@1.3.5": "patches/scimmy@1.3.5.patch",
-      "yjs@13.6.30": "patches/yjs@13.6.30.patch"
+      "yjs@13.6.30": "patches/yjs@13.6.30.patch",
+      "ai@6.0.134": "patches/ai@6.0.134.patch"
     },
     "overrides": {
       "prosemirror-changeset": "2.4.0",

packages/git-sync/test/stabilize.test.ts

@@ -55,10 +55,15 @@ describe('stabilizePageFile — normalize-on-write fixpoint (SPEC §11)', () =>
     const file2 = await stabilizePageFile(doc2, meta);
     expect(file2).toBe(file1);
 
-    // The materialized diagram default is present in the stabilized body (proof
-    // that the convergence pass actually ran, not just that two naive exports
-    // happened to match).
-    expect(body1).toContain('data-align="center"');
+    // The drawio node was materialized to its canonical HTML form by the
+    // convergence pass — a bare `{ src }` doc node becomes the full
+    // `<div data-type="drawio" data-src=...>` — proof the pass actually ran, not
+    // just two naive exports happening to match. Assert on the stable canonical
+    // markers rather than `data-align="center"`: center is a schema default the
+    // converter may omit (see prosemirror-markdown media-html.ts), so it is not
+    // a reliable convergence proof.
+    expect(body1).toContain('data-type="drawio"');
+    expect(body1).toContain('data-src="/d.drawio"');
   });
 
   it('already-stable content is unchanged by the pass (idempotent)', async () => {

packages/mcp/src/index.ts

@@ -118,56 +118,19 @@ export function createDocmostMcpServer(config: DocmostMcpConfig): McpServer {
 // transport exposes a `tree:true` mode that returns the full nested hierarchy;
 // the in-app copy keeps the same tree option but is worded for the in-app agent.
 // Kept per-layer so each side can tune its own guidance.
-server.registerTool(
-  "list_pages",
-  {
-    description:
-      "List most recent pages in a space ordered by updatedAt (descending). " +
-      "Returns a bounded list (default 50, max 100) — use search for lookups " +
-      "in large spaces. Pass tree:true (with spaceId) to instead get the " +
-      "space's full page hierarchy as a nested tree.",
-    inputSchema: {
-      spaceId: z.string().optional(),
-      limit: z
-        .number()
-        .int()
-        .min(1)
-        .max(100)
-        .optional()
-        .describe("Max pages to return (default 50, max 100)"),
-      tree: z
-        .boolean()
-        .optional()
-        .describe(
-          "When true, return the space's full page hierarchy as a nested tree (each node has a children array) instead of the recent-by-updatedAt flat list. Requires spaceId; ignores limit.",
-        ),
-    },
-  },
-  async ({ spaceId, limit, tree }) => {
-    const result = await docmostClient.listPages(spaceId, limit ?? 50, tree ?? false);
-    return jsonContent(result);
-  },
-);
+// Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294). This
+// transport keeps applying its own defaults (limit=50, tree=false) in execute.
+registerShared(SHARED_TOOL_SPECS.listPages, async ({ spaceId, limit, tree }) => {
+  const result = await docmostClient.listPages(spaceId, limit ?? 50, tree ?? false);
+  return jsonContent(result);
+});
 
 // Tool: get_page
-server.registerTool(
-  "get_page",
-  {
-    description:
-      "Get page details with content converted to Markdown. The conversion is " +
-      "LOSSY (block ids, exact table/callout structure are approximated); for a " +
-      "lossless representation use get_page_json. Inline <span data-comment-id> " +
-      "tags in the markdown are comment highlight anchors (also present for " +
-      "RESOLVED threads) — treat them as markup, not page text.",
-    inputSchema: {
-      pageId: z.string().min(1),
-    },
-  },
-  async ({ pageId }) => {
-    const page = await docmostClient.getPage(pageId);
-    return jsonContent(page);
-  },
-);
+// Schema + description now live in the shared registry (#294).
+registerShared(SHARED_TOOL_SPECS.getPage, async ({ pageId }) => {
+  const page = await docmostClient.getPage(pageId);
+  return jsonContent(page);
+});
 
 // Tool: get_page_json
 registerShared(SHARED_TOOL_SPECS.getPageJson, async ({ pageId }) => {
@@ -201,6 +164,10 @@ registerShared(
 );
 
 // Tool: table_get
+// NOT in the shared registry: the MCP tool name `table_get` is noun-first while
+// the in-app key is `getTable` (verb-first), breaking the snake_case(inAppKey)
+// convention the shared registry enforces (shared-tool-specs.contract.spec.ts).
+// Renaming the public MCP tool would break external clients, so it stays inline.
 server.registerTool(
   "table_get",
   {
@@ -223,25 +190,10 @@ server.registerTool(
 );
 
 // Tool: table_insert_row
-// NOT in the shared registry: this transport names the table argument `table`,
-// while the in-app tool names it `tableRef` (ai-chat-tools.service.ts). Sharing
-// one buildShape would rename a public MCP parameter, so the table row/cell
-// tools stay per-transport by design.
-server.registerTool(
-  "table_insert_row",
-  {
-    description:
-      "Insert a row of plain-text cells into a table. `table` = `#<index>` or " +
-      "a block id inside it. `cells` = text per column (padded to the table's " +
-      "column count; error if more cells than columns). `index` = 0-based " +
-      "insert position (0 inserts before the header); omit to append at the end.",
-    inputSchema: {
-      pageId: z.string().min(1),
-      table: z.string().min(1),
-      cells: z.array(z.string()),
-      index: z.number().int().optional(),
-    },
-  },
+// Schema + description now live in the shared registry (#294); the `table`
+// parameter name is the canonical one (the in-app layer was unified to it).
+registerShared(
+  SHARED_TOOL_SPECS.tableInsertRow,
   async ({ pageId, table, cells, index }) => {
     const result = await docmostClient.tableInsertRow(
       pageId,
@@ -254,22 +206,9 @@ server.registerTool(
 );
 
 // Tool: table_delete_row
-// NOT shared — same `table` (here) vs `tableRef` (in-app) parameter-name
-// divergence as table_insert_row.
-server.registerTool(
-  "table_delete_row",
-  {
-    description:
-      "Delete the row at 0-based `index` from a table (`table` = `#<index>` or " +
-      "a block id inside it). Refuses to delete the table's only row. An " +
-      "out-of-range `index` throws. Deleting `index` 0 removes the header row, " +
-      "and the next row becomes the new header.",
-    inputSchema: {
-      pageId: z.string().min(1),
-      table: z.string().min(1),
-      index: z.number().int(),
-    },
-  },
+// Schema + description now live in the shared registry (#294).
+registerShared(
+  SHARED_TOOL_SPECS.tableDeleteRow,
   async ({ pageId, table, index }) => {
     const result = await docmostClient.tableDeleteRow(pageId, table, index);
     return jsonContent(result);
@@ -277,24 +216,9 @@ server.registerTool(
 );
 
 // Tool: table_update_cell
-// NOT shared — same `table` (here) vs `tableRef` (in-app) parameter-name
-// divergence as table_insert_row.
-server.registerTool(
-  "table_update_cell",
-  {
-    description:
-      "Set the plain-text content of cell [row,col] (0-based) in a table " +
-      "(`table` = `#<index>` or a block id inside it). Replaces the cell's " +
-      "content with a single text paragraph; for rich formatting use patch_node " +
-      "on the cell's paragraph id from table_get.",
-    inputSchema: {
-      pageId: z.string().min(1),
-      table: z.string().min(1),
-      row: z.number().int(),
-      col: z.number().int(),
-      text: z.string(),
-    },
-  },
+// Schema + description now live in the shared registry (#294).
+registerShared(
+  SHARED_TOOL_SPECS.tableUpdateCell,
   async ({ pageId, table, row, col, text }) => {
     const result = await docmostClient.tableUpdateCell(
       pageId,
@@ -308,22 +232,9 @@ server.registerTool(
 );
 
 // Tool: create_page
-server.registerTool(
-  "create_page",
-  {
-    description:
-      "Create a new page from Markdown in a space. Pass parentPageId to nest " +
-      "it under a parent; omit it to create at the space root.",
-    inputSchema: {
-      title: z.string().min(1).describe("Title of the page"),
-      content: z.string().min(1).describe("Markdown content"),
-      spaceId: z.string().min(1),
-      parentPageId: z
-        .string()
-        .optional()
-        .describe("Optional parent page ID to nest under"),
-    },
-  },
+// Schema + description now live in the shared registry (#294).
+registerShared(
+  SHARED_TOOL_SPECS.createPage,
   async ({ title, content, spaceId, parentPageId }) => {
     const result = await docmostClient.createPage(
       title,
@@ -336,32 +247,11 @@ server.registerTool(
 );
 
 // Tool: update_page_json
-server.registerTool(
-  "update_page_json",
-  {
-    description:
-      "Replace a page's content with a raw ProseMirror JSON document " +
-      "(lossless write: preserves the block ids, callouts, tables and " +
-      "attributes you pass in). Typical flow: get_page_json -> modify the " +
-      "JSON -> update_page_json. Keep existing node ids intact so heading " +
-      "anchors and history stay stable. Minimal full-doc example: " +
-      '{"type":"doc","content":[{"type":"paragraph","content":' +
-      '[{"type":"text","text":"Hi"}]}]}. `content` may be a JSON object or a ' +
-      "JSON string (both accepted), and is OPTIONAL: omit it to update only " +
-      "the title (though prefer rename_page for a title-only change). " +
-      "Supplying neither content nor title is an error.",
-    inputSchema: {
-      pageId: z.string().min(1).describe("ID of the page to update"),
-      content: z
-        .any()
-        .optional()
-        .describe(
-          'ProseMirror document {"type":"doc","content":[...]} (JSON object or ' +
-            "JSON string). Omit to rename only.",
-        ),
-      title: z.string().optional().describe("Optional new title"),
-    },
-  },
+// Schema + description now live in the shared registry (#294). The execute body
+// keeps this transport's content normalization (parse a JSON-string content,
+// pass undefined/null through for a title-only/no-op update).
+registerShared(
+  SHARED_TOOL_SPECS.updatePageJson,
   async ({ pageId, content, title }) => {
     // Only parse/validate the document when it was actually supplied; when it
     // is omitted, pass it straight through so the client performs a title-only
@@ -379,26 +269,11 @@ server.registerTool(
 );
 
 // Tool: export_page_markdown
-server.registerTool(
-  "export_page_markdown",
-  {
-    description:
-      "Export a page to a single self-contained, lossless Docmost-flavoured " +
-      "Markdown file (custom extensions): YAML-free meta header, body with " +
-      "inline comment anchors and diagrams, and a trailing comments-thread " +
-      "block. Designed for a download -> edit body -> import_page_markdown " +
-      "round-trip that preserves everything, including comment highlights. " +
-      "Comment THREADS are preserved in the file but are not re-pushed to the " +
-      "server on import.",
-    inputSchema: {
-      pageId: z.string().min(1),
-    },
-  },
-  async ({ pageId }) => {
-    const md = await docmostClient.exportPageMarkdown(pageId);
-    return { content: [{ type: "text" as const, text: md }] };
-  },
-);
+// Schema + description now live in the shared registry (#294).
+registerShared(SHARED_TOOL_SPECS.exportPageMarkdown, async ({ pageId }) => {
+  const md = await docmostClient.exportPageMarkdown(pageId);
+  return { content: [{ type: "text" as const, text: md }] };
+});
 
 // Tool: import_page_markdown
 registerShared(
@@ -422,22 +297,11 @@ registerShared(
 );
 
 // Tool: rename_page
-server.registerTool(
-  "rename_page",
-  {
-    description:
-      "Rename a page (change its title only) without touching or resending " +
-      "its content.",
-    inputSchema: {
-      pageId: z.string().min(1).describe("ID of the page to rename"),
-      title: z.string().min(1).describe("New title"),
-    },
-  },
-  async ({ pageId, title }) => {
-    const result = await docmostClient.renamePage(pageId, title);
-    return jsonContent(result);
-  },
-);
+// Schema + description now live in the shared registry (#294).
+registerShared(SHARED_TOOL_SPECS.renamePage, async ({ pageId, title }) => {
+  const result = await docmostClient.renamePage(pageId, title);
+  return jsonContent(result);
+});
 
 // Tool: edit_page_text
 registerShared(SHARED_TOOL_SPECS.editPageText, async ({ pageId, edits }) => {
@@ -516,6 +380,10 @@ registerShared(SHARED_TOOL_SPECS.deleteNode, async ({ pageId, nodeId }) => {
 });
 
 // Tool: insert_image
+// MCP-only by design (NOT in the shared registry): the in-app AI-chat agent
+// exposes no image tools (insert/replace), so there is no second layer to unify
+// — a SHARED_TOOL_SPECS entry's tier/catalogLine are in-app metadata and the
+// catalog-partition test forbids a spec without a live in-app tool (#294).
 server.registerTool(
   "insert_image",
   {
@@ -561,6 +429,7 @@ server.registerTool(
 );
 
 // Tool: replace_image
+// MCP-only by design (see insert_image): no in-app equivalent, stays inline.
 server.registerTool(
   "replace_image",
   {
@@ -603,25 +472,10 @@ server.registerTool(
 );
 
 // Tool: share_page
-// INTENTIONAL per-transport divergence (not shared): the in-app copy adds a
-// security-confirmation framing ("only share when the user explicitly asked,
-// since this exposes the page to anyone with the link") tuned for the in-app
-// agent; this transport keeps the plain public-URL wording.
-server.registerTool(
-  "share_page",
-  {
-    description:
-      "Make a page publicly accessible (idempotent) and return its public " +
-      "URL. The URL format is <app>/share/<key>/p/<slugId>. This exposes the " +
-      "page content to ANYONE with the URL — do it only when explicitly asked.",
-    inputSchema: {
-      pageId: z.string().min(1).describe("ID of the page to share"),
-      searchIndexing: z
-        .boolean()
-        .optional()
-        .describe("Allow search engines to index the page (default true)"),
-    },
-  },
+// Schema + description now live in the shared registry (#294). The execute body
+// keeps this transport's own `searchIndexing ?? true` default.
+registerShared(
+  SHARED_TOOL_SPECS.sharePage,
   async ({ pageId, searchIndexing }) => {
     const result = await docmostClient.sharePage(pageId, searchIndexing ?? true);
     return jsonContent(result);
@@ -641,29 +495,11 @@ registerShared(SHARED_TOOL_SPECS.listShares, async () => {
 });
 
 // Tool: move_page
-server.registerTool(
-  "move_page",
-  {
-    description:
-      "Move a page under a new parent (nesting) or to the space root.",
-    inputSchema: {
-      pageId: z.string().min(1),
-      parentPageId: z
-        .string()
-        .nullable()
-        .optional()
-        .describe(
-          "Target parent page ID. Pass 'null' or empty string to move to root.",
-        ),
-      position: z
-        .string()
-        .min(5)
-        .optional()
-        .describe(
-          "fractional-index position key; min 5 chars; omit to append at the end.",
-        ),
-    },
-  },
+// Schema + description now live in the shared registry (#294). The execute body
+// keeps this transport's cycle guard, its 'null'/'' -> null string coercion, and
+// its positive-confirmation check on the move response.
+registerShared(
+  SHARED_TOOL_SPECS.movePage,
   async ({ pageId, parentPageId, position }) => {
     const finalParentId =
       parentPageId === "" || parentPageId === "null" ? null : parentPageId;
@@ -698,49 +534,22 @@ server.registerTool(
 );
 
 // Tool: delete_page
-server.registerTool(
-  "delete_page",
-  {
-    description:
-      "Delete a single page by ID. SOFT delete only: the page is moved to " +
-      "trash and can be restored; nothing is permanently deleted.",
-    inputSchema: {
-      pageId: z.string().min(1),
-    },
-  },
-  async ({ pageId }) => {
-    await docmostClient.deletePage(pageId);
-    return {
-      content: [
-        { type: "text" as const, text: `Successfully deleted page ${pageId}` },
-      ],
-    };
-  },
-);
+// Schema + description now live in the shared registry (#294). The shared schema
+// exposes ONLY pageId, so no permanent/force-delete flag can reach the client.
+registerShared(SHARED_TOOL_SPECS.deletePage, async ({ pageId }) => {
+  await docmostClient.deletePage(pageId);
+  return {
+    content: [
+      { type: "text" as const, text: `Successfully deleted page ${pageId}` },
+    ],
+  };
+});
 
 // --- Comment tools (ported from upstream PR #3 by Max Nikitin) ---
 
 // Tool: list_comments
-server.registerTool(
-  "list_comments",
-  {
-    description:
-      "List comments on a page in one call (pagination is handled " +
-      "internally). By DEFAULT only ACTIVE threads are returned; resolved " +
-      "threads (a resolved top-level comment and all its replies) are hidden " +
-      "and their count reported as `resolvedThreadsHidden` so you can re-query " +
-      "with `includeResolved: true` to see everything. Returns " +
-      "`{ items, resolvedThreadsHidden }`. Content is returned as Markdown.",
-    inputSchema: {
-      pageId: z.string().describe("ID of the page"),
-      includeResolved: z
-        .boolean()
-        .optional()
-        .describe(
-          "default only active threads; true — include resolved",
-        ),
-    },
-  },
+registerShared(
+  SHARED_TOOL_SPECS.listComments,
   async ({ pageId, includeResolved }) => {
     const comments = await docmostClient.listComments(pageId, includeResolved);
     return jsonContent(comments);
@@ -748,55 +557,11 @@ server.registerTool(
 );
 
 // Tool: create_comment
-// INTENTIONAL per-transport divergence (not shared): the in-app copy tunes the
-// guidance for the in-app agent (e.g. "retry with a corrected EXACT selection"
-// and "Reversible via the comment UI"); this transport keeps its own wording.
-server.registerTool(
-  "create_comment",
-  {
-    description:
-      "Create a new comment on a page. The comment is ALWAYS inline and is " +
-      "anchored to (highlights) its `selection` text — there are no page-level " +
-      "comments. Content is provided as Markdown and automatically converted. " +
-      "A top-level comment REQUIRES an exact `selection`; if the selection " +
-      "cannot be found in the page the call fails (no orphan comment is left). " +
-      "Replies (parentCommentId set) inherit the parent's anchor and take no " +
-      "selection. You may also attach a `suggestedText` proposing a replacement " +
-      "for the `selection`; a human applies (or rejects) it from the UI. When " +
-      "`suggestedText` is set the `selection` MUST occur exactly once in the " +
-      "page — expand it with surrounding context if it is ambiguous.",
-    inputSchema: {
-      pageId: z.string().describe("ID of the page to comment on"),
-      content: z.string().min(1).describe("Comment content in Markdown format"),
-      selection: z
-        .string()
-        .min(1)
-        // Enforce the documented 250-char cap to match the description above.
-        .max(250)
-        .optional()
-        .describe(
-          "EXACT contiguous text from a single paragraph/block to anchor the " +
-            "comment on (<=250 chars). Required for a top-level comment; omit " +
-            "only when replying via parentCommentId.",
-        ),
-      parentCommentId: z
-        .string()
-        .optional()
-        .describe("Parent comment ID to create a reply (max 2 nesting levels)"),
-      suggestedText: z
-        .string()
-        .min(1)
-        .max(2000)
-        .optional()
-        .describe(
-          "Optional proposed replacement (PLAIN TEXT) for the `selection`, " +
-            "applied by a human via the UI (never auto-applied). REQUIRES a " +
-            "`selection`; NOT allowed on a reply. When set, the `selection` must " +
-            "be UNIQUE in the page — expand it with surrounding context (still " +
-            "<=250 chars) if it occurs more than once, or the call is refused.",
-        ),
-    },
-  },
+// Schema + description now live in the shared registry (#294). The execute body
+// keeps this transport's own guards (require a selection for a top-level
+// comment; reject suggestedText on a reply / without a selection).
+registerShared(
+  SHARED_TOOL_SPECS.createComment,
   async ({ pageId, content, selection, parentCommentId, suggestedText }) => {
     if (!parentCommentId && (!selection || !selection.trim())) {
       throw new Error(
@@ -872,28 +637,9 @@ server.registerTool(
 );
 
 // Tool: resolve_comment
-server.registerTool(
-  "resolve_comment",
-  {
-    description:
-      "Resolve (close) or reopen a comment thread. Only top-level comments can " +
-      "be resolved — the server rejects resolving a reply. Reversible: pass " +
-      "resolved=false to reopen. Resolving keeps the thread and its replies " +
-      "(unlike delete_comment, which permanently removes them).",
-    inputSchema: {
-      commentId: z
-        .string()
-        .min(1)
-        .describe("ID of the top-level comment thread to resolve or reopen"),
-      resolved: z
-        .boolean()
-        .optional()
-        .default(true)
-        .describe(
-          "true (default) marks the thread resolved/closed; false reopens it",
-        ),
-    },
-  },
+// Schema + description now live in the shared registry (#294).
+registerShared(
+  SHARED_TOOL_SPECS.resolveComment,
   async ({ commentId, resolved }) => {
     const result = await docmostClient.resolveComment(commentId, resolved);
     return jsonContent(result);
@@ -901,30 +647,10 @@ server.registerTool(
 );
 
 // Tool: check_new_comments
-server.registerTool(
-  "check_new_comments",
-  {
-    description:
-      "Check for new comments across pages in a space since a given timestamp. " +
-      "Optionally scope to a page subtree (folder). Returns only comments " +
-      "created after the specified time.",
-    inputSchema: {
-      spaceId: z.string().describe("Space ID to check for new comments"),
-      since: z
-        .string()
-        .min(1)
-        .describe(
-          "ISO 8601 timestamp — only return comments created after this time (e.g. '2026-03-10T00:00:00Z')",
-        ),
-      parentPageId: z
-        .string()
-        .optional()
-        .describe(
-          "Optional root page ID to scope the check to a subtree (folder). " +
-            "Only pages under this parent will be checked.",
-        ),
-    },
-  },
+// Schema + description now live in the shared registry (#294). The execute body
+// keeps this transport's own guard rejecting an unparseable `since` timestamp.
+registerShared(
+  SHARED_TOOL_SPECS.checkNewComments,
   async ({ spaceId, since, parentPageId }) => {
     // Reject an unparseable timestamp up front: otherwise the comparison
     // against NaN silently treats every comment as "not new" and the tool
@@ -1053,6 +779,8 @@ server.registerTool(
 );
 
 // Tool: insert_footnote
+// MCP-only by design (see insert_image): the in-app AI-chat agent exposes no
+// footnote tool, so there is no second layer to unify — stays inline (#294).
 server.registerTool(
   "insert_footnote",
   {

packages/mcp/src/tool-specs.ts

@@ -316,6 +316,34 @@ export const SHARED_TOOL_SPECS = {
 
   // --- share management ---
 
+  // Unified from the per-layer inline definitions (#294). Both layers already
+  // carried the "only share when explicitly asked" security framing (the
+  // "per-transport divergence" note on the old inline copies was stale), so
+  // there was no real behavioral divergence to preserve — only wording drift.
+  sharePage: {
+    mcpName: 'share_page',
+    inAppKey: 'sharePage',
+    // CANONICAL: merges the MCP copy's URL-format + idempotency detail with the
+    // in-app copy's reversibility note; keeps the security framing both had.
+    description:
+      'Make a page PUBLICLY accessible (idempotent) and return its public URL ' +
+      '(format: <app>/share/<key>/p/<slugId>). This exposes the page content ' +
+      'to ANYONE with the URL — only share when the user explicitly asked. ' +
+      'Reversible: unshare it later to revoke the public URL.',
+    tier: 'deferred',
+    catalogLine: 'sharePage — make a page publicly accessible and return its URL.',
+    // Reconciled: MCP's stricter .min(1) on pageId kept; field descriptions from
+    // the in-app copy. The MCP execute keeps its own `searchIndexing ?? true`
+    // default (a per-layer concern, not part of the shared schema).
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page to share.'),
+      searchIndexing: z
+        .boolean()
+        .optional()
+        .describe('Allow public search engines to index it (default true).'),
+    }),
+  },
+
   unsharePage: {
     mcpName: 'unshare_page',
     inAppKey: 'unsharePage',
@@ -509,4 +537,470 @@ export const SHARED_TOOL_SPECS = {
       pageId: z.string().min(1),
     }),
   },
+
+  // --- page tools (unified from the per-layer inline definitions, #294) ---
+  //
+  // Descriptions merge both layers (the MCP copy's richer structural notes + the
+  // in-app copy's "Reversible via history/trash" framing where it added one).
+  // Field constraints keep the MCP copy's stricter .min(1) EXCEPT where the
+  // in-app layer deliberately allowed a looser value (documented per field).
+
+  getPage: {
+    mcpName: 'get_page',
+    inAppKey: 'getPage',
+    description:
+      'Fetch a single page as Markdown by its id. Returns the page title and ' +
+      'its Markdown content. The Markdown conversion is LOSSY (block ids, exact ' +
+      'table/callout structure are approximated); for a lossless representation ' +
+      'use the lossless page-JSON read tool. Inline <span data-comment-id> tags in the markdown ' +
+      'are comment highlight anchors (also present for RESOLVED threads) — ' +
+      'treat them as markup, not page text.',
+    tier: 'core',
+    catalogLine: 'getPage — fetch a page as Markdown by its id.',
+    // Reconciled: MCP's stricter .min(1) kept; in-app's more-informative
+    // "(or slugId)" describe kept.
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id (or slugId) of the page.'),
+    }),
+  },
+
+  listPages: {
+    mcpName: 'list_pages',
+    inAppKey: 'listPages',
+    description:
+      'List the most recent pages (ordered by updatedAt, descending), ' +
+      'optionally scoped to a single space. Returns a bounded list (default ' +
+      '50, max 100) — use search for lookups in large spaces. Pass tree:true ' +
+      "(with spaceId) to instead get the space's full page hierarchy as a " +
+      'nested tree.',
+    tier: 'core',
+    catalogLine: "listPages — list recent pages, or a space's full page tree.",
+    buildShape: (z) => ({
+      spaceId: z
+        .string()
+        .optional()
+        .describe('Optional space id to scope the listing to.'),
+      limit: z
+        .number()
+        .int()
+        .min(1)
+        .max(100)
+        .optional()
+        .describe('Maximum number of pages (default 50, max 100).'),
+      tree: z
+        .boolean()
+        .optional()
+        .describe(
+          "When true, return the space's full page hierarchy as a nested tree " +
+            '(children arrays) instead of the recent-by-updatedAt flat list. ' +
+            'Requires spaceId; ignores limit.',
+        ),
+    }),
+  },
+
+  createPage: {
+    mcpName: 'create_page',
+    inAppKey: 'createPage',
+    description:
+      'Create a new page with a Markdown body in a space, optionally under a ' +
+      'parent page (omit parentPageId to create at the space root). Returns ' +
+      'the new page id and title. Reversible: a page can be moved to trash ' +
+      'later.',
+    tier: 'deferred',
+    catalogLine: 'createPage — create a new page with a Markdown body in a space.',
+    // Reconciled schema DRIFT: the MCP copy pinned `content` to .min(1) while
+    // the in-app copy left it unbounded and DOCUMENTS an empty body as valid
+    // ("may be empty") — creating an empty page to fill in later is a real use
+    // case. The looser (no-min) form is kept, so create_page now also accepts an
+    // empty body (harmless — it creates an empty page) and no previously-valid
+    // in-app input is ever rejected. `title`/`spaceId` keep the MCP .min(1)
+    // (an empty title or space is never valid).
+    buildShape: (z) => ({
+      title: z.string().min(1).describe('The title of the new page.'),
+      content: z.string().describe('The page body as Markdown (may be empty).'),
+      spaceId: z.string().min(1).describe('The id of the space to create the page in.'),
+      parentPageId: z
+        .string()
+        .optional()
+        .describe('Optional parent page id to nest the new page under.'),
+    }),
+  },
+
+  movePage: {
+    mcpName: 'move_page',
+    inAppKey: 'movePage',
+    description:
+      'Move a page under a new parent page, or to the space root when no ' +
+      'parent is given. Reversible: move it back at any time.',
+    tier: 'deferred',
+    catalogLine: 'movePage — move a page under a new parent or to the space root.',
+    // Reconciled schema DRIFT: the MCP copy exposed a `position` field
+    // (fractional-index ordering) that the in-app copy lacked. Unified by
+    // KEEPING position (the in-app client already accepts an optional position
+    // arg, so the in-app execute now forwards it) — it is optional, so no
+    // previously-valid in-app call is rejected. `parentPageId` is `.nullable()`
+    // on both, so a real JSON null moves to root on either transport; the MCP
+    // execute additionally coerces the strings 'null'/'' to null as a robustness
+    // fallback (kept in its execute body, not in the shared schema).
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page to move.'),
+      parentPageId: z
+        .string()
+        .nullable()
+        .optional()
+        .describe(
+          'Target parent page id. Null or omitted moves the page to the space ' +
+            'root.',
+        ),
+      position: z
+        .string()
+        .min(5)
+        .optional()
+        .describe(
+          'Optional fractional-index position key (min 5 chars); omit to ' +
+            'append at the end.',
+        ),
+    }),
+  },
+
+  renamePage: {
+    mcpName: 'rename_page',
+    inAppKey: 'renamePage',
+    description:
+      'Rename a page (change its title only; the body is untouched, never ' +
+      'resent). Reversible: rename back at any time.',
+    tier: 'deferred',
+    catalogLine: "renamePage — change a page's title only (body untouched).",
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page to rename.'),
+      title: z.string().min(1).describe('The new title.'),
+    }),
+  },
+
+  deletePage: {
+    mcpName: 'delete_page',
+    inAppKey: 'deletePage',
+    description:
+      'Move a page to the trash — SOFT delete only: the page can be restored ' +
+      'from trash and nothing is ever permanently deleted.',
+    tier: 'deferred',
+    catalogLine: 'deletePage — move a page to trash (soft delete, reversible).',
+    // GUARDRAIL preserved (§14 H4): the schema exposes ONLY pageId, so a
+    // permanentlyDelete/forceDelete flag can never reach the client through this
+    // tool (asserted by ai-chat-tools.service.spec.ts).
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page to move to trash.'),
+    }),
+  },
+
+  updatePageJson: {
+    mcpName: 'update_page_json',
+    inAppKey: 'updatePageJson',
+    description:
+      "Replace a page's content with a raw ProseMirror JSON document (lossless " +
+      'write: preserves the block ids, callouts, tables and attributes you pass ' +
+      'in). Typical flow: read the page-JSON view -> modify the JSON -> write it back. ' +
+      'Keep existing node ids intact so heading anchors and history stay ' +
+      'stable. Minimal full-doc example: {"type":"doc","content":[{"type":' +
+      '"paragraph","content":[{"type":"text","text":"Hi"}]}]}. `content` may be ' +
+      'a JSON object or a JSON string (both accepted), and is OPTIONAL: omit it ' +
+      'to update only the title (though prefer the rename-page tool for a title-only ' +
+      'change). Supplying neither content nor title is an error. Reversible: ' +
+      'the previous version is kept in page history.',
+    tier: 'deferred',
+    catalogLine:
+      "updatePageJson — overwrite a page's body with a full ProseMirror document.",
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('ID of the page to update'),
+      content: z
+        .any()
+        .optional()
+        .describe(
+          'ProseMirror document {"type":"doc","content":[...]} (JSON object or ' +
+            'JSON string). Omit to update only the title.',
+        ),
+      title: z.string().optional().describe('Optional new title'),
+    }),
+  },
+
+  exportPageMarkdown: {
+    mcpName: 'export_page_markdown',
+    inAppKey: 'exportPageMarkdown',
+    // CANONICAL: the MCP copy (a strict superset of the terse in-app wording).
+    description:
+      'Export a page to a single self-contained, lossless Docmost-flavoured ' +
+      'Markdown file (custom extensions): YAML-free meta header, body with ' +
+      'inline comment anchors and diagrams, and a trailing comments-thread ' +
+      'block. Designed for a download -> edit body -> page-Markdown import ' +
+      'round-trip that preserves everything, including comment highlights. ' +
+      'Comment THREADS are preserved in the file but are not re-pushed to the ' +
+      'server on import.',
+    tier: 'deferred',
+    catalogLine:
+      'exportPageMarkdown — export a page to self-contained Markdown (body + comments).',
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page to export.'),
+    }),
+  },
+
+  // --- comment tools (unified from the per-layer inline definitions, #294) ---
+  //
+  // create_comment and resolve_comment previously carried a "per-transport
+  // divergence" note in BOTH layers; #294 unifies their schema + description
+  // here. Only the four tools that genuinely exist in BOTH layers live in the
+  // registry: create/list/resolve comment and check_new_comments.
+  //
+  // update_comment and delete_comment are intentionally NOT here: they exist
+  // ONLY on the standalone MCP server. The in-app agent deliberately exposes no
+  // hard comment edit/delete tool (comment edits are irreversible / not
+  // version-tracked; see the guardrail tests in ai-chat-tools.service.spec.ts),
+  // so there is nothing to unify — they stay inline in index.ts.
+
+  createComment: {
+    mcpName: 'create_comment',
+    inAppKey: 'createComment',
+    // CANONICAL: the in-app copy (the more-maintained one). It keeps the same
+    // rules as the MCP copy — inline-only, top-level requires a `selection`, no
+    // page-level comments, replies inherit the anchor, suggestedText must be
+    // unique — and adds the "retry with a corrected EXACT selection" and reply-
+    // to-reply-rejected guidance the MCP copy lacked. Execute-side validation
+    // (reject suggestedText on a reply, require a selection) stays per-layer.
+    description:
+      'Add an INLINE comment to a page, or reply to an existing top-level ' +
+      'comment (one level only — the backend rejects replies to replies). ' +
+      'The comment is anchored inline to the given exact `selection` text ' +
+      '(which gets highlighted); page-level comments are NOT supported. A ' +
+      'new top-level comment REQUIRES a `selection`. Replies inherit the ' +
+      "parent's anchor and take no selection. If the call fails with a " +
+      '"selection not found" error, retry with a corrected EXACT selection ' +
+      'copied verbatim from a single paragraph/block. You may also attach a ' +
+      '`suggestedText` proposing a replacement for the `selection` (a human ' +
+      'applies it from the UI); when set, the `selection` must occur exactly ' +
+      'once in the page. Reversible via the comment UI.',
+    tier: 'core',
+    catalogLine:
+      'createComment — add an inline comment (optionally with a suggested edit).',
+    // Reconciled schema: the field set is identical across both layers; the
+    // only constraint drift is `content`, which the MCP copy pinned to
+    // .min(1) while the in-app copy left unbounded — the stricter MCP form is
+    // kept (an empty comment body is never valid).
+    buildShape: (z) => ({
+      pageId: z.string().describe('The id of the page to comment on.'),
+      content: z.string().min(1).describe('The comment body as Markdown.'),
+      selection: z
+        .string()
+        .min(1)
+        .max(250)
+        .optional()
+        .describe(
+          'EXACT contiguous text from a SINGLE paragraph/block to anchor ' +
+            '(highlight) the comment on (<=250 chars, avoid spanning across ' +
+            'formatting boundaries). Required for a new top-level comment; ' +
+            'omit only when replying via parentCommentId.',
+        ),
+      parentCommentId: z
+        .string()
+        .optional()
+        .describe(
+          'Optional id of a TOP-LEVEL comment to reply to (one level ' +
+            'of replies only).',
+        ),
+      suggestedText: z
+        .string()
+        .min(1)
+        .max(2000)
+        .optional()
+        .describe(
+          'Optional proposed replacement (PLAIN TEXT) for the `selection`, ' +
+            'applied by a human via the UI (never auto-applied). REQUIRES a ' +
+            '`selection`; NOT allowed on a reply. When set, the `selection` ' +
+            'must be UNIQUE in the page — expand it with surrounding context ' +
+            '(still <=250 chars) if it occurs more than once, or the call is ' +
+            'refused.',
+        ),
+    }),
+  },
+
+  listComments: {
+    mcpName: 'list_comments',
+    inAppKey: 'listComments',
+    // CANONICAL: the two copies are near-identical; the MCP copy is the
+    // superset (it keeps the "(pagination is handled internally)" note the
+    // in-app copy dropped), so it is used verbatim.
+    description:
+      'List comments on a page in one call (pagination is handled ' +
+      'internally). By DEFAULT only ACTIVE threads are returned; resolved ' +
+      'threads (a resolved top-level comment and all its replies) are hidden ' +
+      'and their count reported as `resolvedThreadsHidden` so you can re-query ' +
+      'with `includeResolved: true` to see everything. Returns ' +
+      '`{ items, resolvedThreadsHidden }`. Content is returned as Markdown.',
+    tier: 'core',
+    catalogLine:
+      'listComments — list all comments on a page (including resolved).',
+    buildShape: (z) => ({
+      pageId: z.string().describe('ID of the page'),
+      includeResolved: z
+        .boolean()
+        .optional()
+        .describe('default only active threads; true — include resolved'),
+    }),
+  },
+
+  resolveComment: {
+    mcpName: 'resolve_comment',
+    inAppKey: 'resolveComment',
+    // CANONICAL: the MCP copy's richer wording, minus its snake_case reference
+    // to `delete_comment` (a sibling tool that does NOT exist in the in-app
+    // layer) — rephrased transport-neutrally per the registry convention.
+    description:
+      'Resolve (close) or reopen a top-level comment thread (reversible — ' +
+      'pass resolved=false to reopen). Only top-level comments can be ' +
+      'resolved; the server rejects resolving a reply. Resolving keeps the ' +
+      'thread and its replies intact (it is not a deletion).',
+    tier: 'core',
+    catalogLine: 'resolveComment — resolve or reopen a comment thread.',
+    // Reconciled schema: `resolved` drifted — the MCP copy made it optional
+    // with .default(true) (resolve is the common case, documented), the in-app
+    // copy made it required. The MCP form is kept (a strict superset: it never
+    // rejects a previously-valid input and adds a sensible default), and
+    // commentId keeps the MCP copy's stricter .min(1).
+    buildShape: (z) => ({
+      commentId: z
+        .string()
+        .min(1)
+        .describe('ID of the top-level comment thread to resolve or reopen'),
+      resolved: z
+        .boolean()
+        .optional()
+        .default(true)
+        .describe(
+          'true (default) marks the thread resolved/closed; false reopens it',
+        ),
+    }),
+  },
+
+  checkNewComments: {
+    mcpName: 'check_new_comments',
+    inAppKey: 'checkNewComments',
+    // CANONICAL: the MCP copy (the more detailed of the two). The MCP layer's
+    // execute-side guard that rejects an unparseable `since` timestamp stays in
+    // its execute body (per-layer logic), not in the shared schema.
+    description:
+      'Check for new comments across pages in a space since a given ' +
+      'timestamp. Optionally scope to a page subtree (folder). Returns only ' +
+      'comments created after the specified time.',
+    tier: 'deferred',
+    catalogLine:
+      'checkNewComments — find comments in a space created after a timestamp.',
+    // Reconciled schema: `since` keeps the MCP copy's stricter .min(1) (the
+    // in-app copy left it unbounded); field descriptions use the MCP copy's
+    // more detailed wording (it carries an example timestamp).
+    buildShape: (z) => ({
+      spaceId: z.string().describe('Space ID to check for new comments'),
+      since: z
+        .string()
+        .min(1)
+        .describe(
+          "ISO 8601 timestamp — only return comments created after this time " +
+            "(e.g. '2026-03-10T00:00:00Z')",
+        ),
+      parentPageId: z
+        .string()
+        .optional()
+        .describe(
+          'Optional root page ID to scope the check to a subtree (folder). ' +
+            'Only pages under this parent will be checked.',
+        ),
+    }),
+  },
+
+  // --- table tools (unified from the per-layer inline definitions, #294) ---
+  //
+  // These tools carried a "NOT shared" note in BOTH layers because of a single
+  // parameter-NAME drift: the MCP layer named the table reference `table` while
+  // the in-app layer named it `tableRef`. #294 reconciles that drift by unifying
+  // on the MCP name `table` — renaming the MCP public parameter would break
+  // external MCP clients, whereas the in-app parameter is model-facing
+  // (prompt-only) and safe to rename. The in-app execute bodies now destructure
+  // `table` instead of `tableRef` (nothing else changes). Descriptions take the
+  // MCP copy's richer wording (it documented `#<index>`, padding, header-row
+  // behavior) plus the in-app copy's "Reversible via page history" note; sibling
+  // tool references are phrased transport-neutrally.
+  //
+  // NOT here (kept inline in index.ts): table_get / getTable. Its MCP tool name
+  // is noun-first (`table_get`) while the in-app key is verb-first (`getTable`),
+  // so it breaks the snake_case(inAppKey) naming convention the registry enforces
+  // (shared-tool-specs.contract.spec.ts). Renaming the public MCP tool would
+  // break external clients, so it stays per-transport (its in-app param was still
+  // aligned to `table` for consistency with the migrated trio below).
+
+  tableInsertRow: {
+    mcpName: 'table_insert_row',
+    inAppKey: 'tableInsertRow',
+    description:
+      'Insert a row of plain-text cells into a table. `table` is `#<index>` ' +
+      'from the page outline, or a block id inside it. `cells` is the text per ' +
+      "column (padded to the table's column count; an error if more cells than " +
+      'columns). `index` is the 0-based insert position (0 inserts before the ' +
+      'header); omit to append at the end. Reversible via page history.',
+    tier: 'deferred',
+    catalogLine: 'tableInsertRow — insert a row of plain-text cells into a table.',
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page.'),
+      table: z
+        .string()
+        .min(1)
+        .describe('"#<index>" from the page outline, or a block id in the table.'),
+      cells: z.array(z.string()).describe('The cell texts for the row (one per column).'),
+      index: z
+        .number()
+        .int()
+        .optional()
+        .describe('0-based insert position (0 inserts before the header); omit to append.'),
+    }),
+  },
+
+  tableDeleteRow: {
+    mcpName: 'table_delete_row',
+    inAppKey: 'tableDeleteRow',
+    description:
+      'Delete the row at 0-based `index` from a table (`table` is `#<index>` ' +
+      'from the page outline, or a block id inside it). Refuses to delete the ' +
+      "table's only row; an out-of-range `index` throws. Deleting `index` 0 " +
+      'removes the header row, and the next row becomes the new header. ' +
+      'Reversible via page history.',
+    tier: 'deferred',
+    catalogLine: 'tableDeleteRow — delete a table row at a 0-based index.',
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page.'),
+      table: z
+        .string()
+        .min(1)
+        .describe('"#<index>" from the page outline, or a block id in the table.'),
+      index: z.number().int().describe('0-based row index to delete.'),
+    }),
+  },
+
+  tableUpdateCell: {
+    mcpName: 'table_update_cell',
+    inAppKey: 'tableUpdateCell',
+    description:
+      'Set the plain-text content of cell [row, col] (0-based) in a table ' +
+      '(`table` is `#<index>` from the page outline, or a block id inside it). ' +
+      "Replaces the cell's content with a single text paragraph; for rich " +
+      "formatting, patch the cell's paragraph id (obtained from reading the " +
+      'table) instead. Reversible via page history.',
+    tier: 'deferred',
+    catalogLine: 'tableUpdateCell — set the text of a table cell at [row, col].',
+    buildShape: (z) => ({
+      pageId: z.string().min(1).describe('The id of the page.'),
+      table: z
+        .string()
+        .min(1)
+        .describe('"#<index>" from the page outline, or a block id in the table.'),
+      row: z.number().int().describe('0-based row index.'),
+      col: z.number().int().describe('0-based column index.'),
+      text: z.string().describe('The new cell text.'),
+    }),
+  },
 } satisfies Record<string, SharedToolSpec>;

packages/prosemirror-markdown/src/lib/markdown-converter.ts

@@ -48,6 +48,52 @@ export interface ConvertProseMirrorToMarkdownOptions {
   dropResolvedCommentAnchors?: boolean;
 }
 
+/**
+ * Adjacent sibling lists that share a markdown MARKER FAMILY re-parse as ONE
+ * merged list — bulletList and taskList both emit `- ` markers (→ a single
+ * `<ul>`), and two orderedLists both emit `1.` markers (→ a single `<ol>`). The
+ * cross-type case is real data loss the editor CAN produce (e.g. a taskList
+ * followed by a bulletList: the merged `<ul>` has a mix of checkbox and plain
+ * items, so `bridgeTaskLists` refuses to convert it and every taskItem loses its
+ * checkbox). Between two such adjacent list children we emit an empty HTML comment
+ * `<!-- -->`: marked renders it as its own HTML block that interrupts the list, so
+ * the two lists stay distinct; on import the comment is inert (parseAttachedComment
+ * → null) and dropped by generateJSON, and re-export re-inserts it, so the marker
+ * is byte-stable. It fires ONLY between two adjacent same-family list nodes — no
+ * separator is emitted for any other join, so non-list output is unchanged.
+ */
+const LIST_MARKER_SEPARATOR = "<!-- -->";
+function listMarkerFamily(type: string | undefined): "ul" | "ol" | null {
+  if (type === "bulletList" || type === "taskList") return "ul";
+  if (type === "orderedList") return "ol";
+  return null;
+}
+function adjacentListsMerge(
+  prevType: string | undefined,
+  curType: string | undefined,
+): boolean {
+  const a = listMarkerFamily(prevType);
+  return a !== null && a === listMarkerFamily(curType);
+}
+/**
+ * Render each block child, inserting a `<!-- -->` separator entry between any two
+ * adjacent same-marker-family list nodes (see LIST_MARKER_SEPARATOR). Callers
+ * join the returned strings with their own context separator.
+ */
+function renderBlockChildren(
+  children: any[],
+  render: (n: any) => string,
+): string[] {
+  const out: string[] = [];
+  let prevType: string | undefined;
+  for (const child of children) {
+    if (adjacentListsMerge(prevType, child?.type)) out.push(LIST_MARKER_SEPARATOR);
+    out.push(render(child));
+    prevType = child?.type;
+  }
+  return out;
+}
+
 /**
  * Convert ProseMirror/TipTap JSON content to Markdown
  * Supports all Docmost-specific node types and extensions
@@ -347,9 +393,15 @@ export function convertProseMirrorToMarkdown(
         // lossless (the body survives) and byte-stable (it re-exports identically),
         // so it is deliberately not treated as data loss.
         const parts: string[] = [];
+        let prevDocType: string | undefined;
         for (const child of nodeContent) {
           if (child?.type === "footnotesList") continue;
+          // Keep adjacent same-family sibling lists distinct (see renderBlockChildren).
+          if (adjacentListsMerge(prevDocType, child?.type)) {
+            parts.push(LIST_MARKER_SEPARATOR);
+          }
           parts.push(processNode(child));
+          prevDocType = child?.type;
         }
         for (const [id, def] of footnoteDefs) {
           if (!referencedFootnoteIds.has(id)) {
@@ -599,20 +651,34 @@ export function convertProseMirrorToMarkdown(
         return processTaskItem(node);
 
       case "listItem":
-        return nodeContent.map(processNode).join("\n");
+        // Direct-listItem path (lists normally render via processListItem, which
+        // handles the marker + indentation). Blank line between block children so
+        // multiple paragraphs do not merge on re-parse; a `<!-- -->` entry
+        // (renderBlockChildren) separates adjacent sibling lists.
+        return renderBlockChildren(nodeContent, processNode).join("\n\n");
 
-      case "blockquote":
+      case "blockquote": {
         // Prefix EVERY line of EVERY child with "> " and separate block-level
         // children with a blank ">" line so code blocks / multi-paragraph
-        // quotes round-trip correctly.
-        return nodeContent
-          .map((n: any) =>
+        // quotes round-trip correctly. A `> <!-- -->` separator line is inserted
+        // between two adjacent same-family sibling lists (renderBlockChildren)
+        // so they stay distinct inside the quote.
+        const bqParts: string[] = [];
+        let prevBqType: string | undefined;
+        for (const n of nodeContent) {
+          if (adjacentListsMerge(prevBqType, n?.type)) {
+            bqParts.push(`> ${LIST_MARKER_SEPARATOR}`);
+          }
+          bqParts.push(
             processNode(n)
               .split("\n")
               .map((line: string) => (line.length ? `> ${line}` : ">"))
               .join("\n"),
-          )
-          .join("\n>\n");
+          );
+          prevBqType = n?.type;
+        }
+        return bqParts.join("\n>\n");
+      }
 
       case "horizontalRule":
         return "---";
@@ -787,9 +853,12 @@ export function convertProseMirrorToMarkdown(
         // blockquote-prefixed; a blank line becomes a bare `>` so the callout is
         // not split.
         const calloutType = (node.attrs?.type || "info").toLowerCase();
-        const calloutBody = nodeContent
-          .map(processNode)
-          .join("\n")
+        const calloutBody = renderBlockChildren(nodeContent, processNode)
+          // Blank line between block children (rendered as a bare `>` after the
+          // prefix pass below) so multiple paragraphs stay separate nodes instead
+          // of merging on re-parse — same rule blockquote already uses. A
+          // `<!-- -->` entry (renderBlockChildren) separates adjacent sibling lists.
+          .join("\n\n")
           .split("\n")
           .map((l: string) => (l.length ? `> ${l}` : ">"))
           .join("\n");
@@ -809,7 +878,10 @@ export function convertProseMirrorToMarkdown(
         return `<summary>${renderInlineChildren(nodeContent)}</summary>\n\n`;
 
       case "detailsContent":
-        return `${nodeContent.map(processNode).join("\n")}\n`;
+        // Blank line between block children so multiple paragraphs in a details
+        // body survive as separate nodes (a single "\n" merges them on re-parse);
+        // a `<!-- -->` entry (renderBlockChildren) separates adjacent sibling lists.
+        return `${renderBlockChildren(nodeContent, processNode).join("\n\n")}\n`;
 
       case "mathInline": {
         // #293 canon #6: inline math serializes as Obsidian-native `$LaTeX$`
@@ -1338,11 +1410,19 @@ export function convertProseMirrorToMarkdown(
         // The code itself is element TEXT content (between <code> tags), so it
         // must escape < > & — NOT the attribute escaper. The language rides in
         // a class ATTRIBUTE, so it uses escapeAttr.
+        //
+        // Read the child text RAW (as `case "codeBlock"` does) and keep it
+        // VERBATIM — do NOT strip the trailing newline. Unlike the markdown fence
+        // path (which strips then relies on marked re-adding one `\n`), the schema
+        // codeBlock parseHTML reads the `<code>` text content back byte-for-byte,
+        // so stripping here would drop a trailing newline the node legitimately
+        // carries and break the round trip inside a column/cell.
         const code = escapeHtmlText(
           children
-            .map(processNode)
-            .join("")
-            .replace(/\n+$/, ""),
+            .map((child: any) =>
+              typeof child?.text === "string" ? child.text : "",
+            )
+            .join(""),
         );
         const cls = lang ? ` class="language-${escapeAttr(lang)}"` : "";
         return `<pre><code${cls}>${code}</code></pre>`;
@@ -1484,6 +1564,13 @@ export function convertProseMirrorToMarkdown(
     const indent = " ".repeat(indentWidth);
     const lines: string[] = [];
     childStrings.forEach((child, childIndex) => {
+      // Separate consecutive block children with a BLANK line so the item is a
+      // CommonMark "loose" list item and each block stays its own node. Without
+      // it, a second paragraph (`- a\n  b`) is re-parsed as a lazy continuation
+      // of the first and the two merge into one paragraph — silent data loss.
+      // The blank line still sits INSIDE the item (the following block keeps the
+      // continuation indent), so nested lists/code blocks remain nested.
+      if (childIndex > 0) lines.push("");
       child.split("\n").forEach((line, lineIndex) => {
         if (childIndex === 0 && lineIndex === 0) {
           // First physical line of the first block gets the marker.
@@ -1500,7 +1587,9 @@ export function convertProseMirrorToMarkdown(
 
   const processListItem = (item: any, prefix: string): string => {
     const itemContent = item.content || [];
-    const childStrings = itemContent.map(processNode);
+    // A `<!-- -->` entry separates two adjacent same-family sublists inside the
+    // item so they do not merge on re-parse (see renderBlockChildren).
+    const childStrings = renderBlockChildren(itemContent, processNode);
     if (childStrings.length === 0) return prefix;
     // The rendered marker is `${prefix} ` (prefix + one space), so its width —
     // and thus the continuation indent — is prefix.length + 1. This is correct
@@ -1514,7 +1603,9 @@ export function convertProseMirrorToMarkdown(
     const checkbox = checked ? "[x]" : "[ ]";
     const prefix = `- ${checkbox}`;
     const itemContent = item.content || [];
-    const childStrings = itemContent.map(processNode);
+    // A `<!-- -->` entry separates two adjacent same-family sublists inside the
+    // item so they do not merge on re-parse (see renderBlockChildren).
+    const childStrings = renderBlockChildren(itemContent, processNode);
     // An empty task item still needs its checkbox marker; without this guard
     // the indent below produces "" and the "- [ ]"/"- [x]" row disappears.
     if (childStrings.length === 0) return prefix;

packages/prosemirror-markdown/src/lib/markdown-to-prosemirror.ts

@@ -270,9 +270,10 @@ const CALLOUT_CLOSE_RE = /^:::\s*$/;
  * optional title after the type is allowed but ignored (the Docmost callout
  * schema has no title). The body is the following contiguous blockquote lines.
  */
-const CALLOUT_BQ_OPEN_RE = /^>\s*\[!(\w+)\]/;
-/** Matches any blockquote continuation line (`>` … ). */
-const BLOCKQUOTE_LINE_RE = /^>/;
+// The callout's own `>` marker may be preceded by an ENCLOSING container prefix:
+// list-item indentation (`  `) and/or blockquote markers (`> `). Group 1 captures
+// that prefix (lazily, so the LAST `>` before `[!type]` is the callout's own).
+const CALLOUT_BQ_OPEN_RE = /^([>\s]*?)>\s*\[!(\w+)\]/;
 /** Matches the start/end of a code fence (``` or ~~~), capturing the marker. */
 const CODE_FENCE_RE = /^(\s*)(`{3,}|~{3,})/;
 
@@ -402,20 +403,47 @@ async function preprocessCallouts(markdown: string): Promise<string> {
       // recurse so nested callouts (`> > [!type]`) are handled, then emit the same
       // callout div the `:::` path produces. A normal blockquote (no `[!type]` on
       // its first line) does not match and stays a blockquote.
+      //
+      // PREFIX-aware: a callout nested inside a list item and/or a blockquote is
+      // serialized with the enclosing container prefix in front of its own `>`
+      // marker — `  > [!type]` (list indent) or `> > [!type]` (blockquote). We
+      // capture that prefix, take only continuation lines carrying `prefix>`, strip
+      // it, and re-apply the prefix to the emitted HTML block so the callout div
+      // stays WITHIN its container (an unprefixed div would escape and re-parse as
+      // a top-level callout / plain blockquote — silent structure loss).
       const bqOpen = line.match(CALLOUT_BQ_OPEN_RE);
       if (bqOpen) {
-        const type = bqOpen[1].toLowerCase();
+        const prefix = bqOpen[1];
+        const type = bqOpen[2].toLowerCase();
+        const cont = prefix + ">"; // a body line = prefix + the callout's own `>`
         const bodyLines: string[] = [];
         let j = i + 1;
         for (; j < lines.length; j++) {
-          if (!BLOCKQUOTE_LINE_RE.test(lines[j])) break;
-          bodyLines.push(lines[j].replace(/^>\s?/, ""));
+          if (!lines[j].startsWith(cont)) break;
+          // Drop the prefix + `>` + one optional space, leaving the body content.
+          bodyLines.push(lines[j].slice(prefix.length).replace(/^>\s?/, ""));
         }
         const inner = await transform(bodyLines);
         const renderedInner = await markedInstance.parse(inner);
-        out.push(
-          `\n<div data-type="callout" data-callout-type="${type}">${renderedInner}</div>\n`,
-        );
+        const block = `<div data-type="callout" data-callout-type="${type}">${renderedInner}</div>`;
+        if (prefix.length === 0) {
+          // Top-level callout: blank lines isolate the HTML block.
+          out.push(`\n${block}\n`);
+        } else if (prefix.includes(">")) {
+          // Enclosing BLOCKQUOTE: prefix every line and add NO surrounding blank
+          // lines — a blank line would terminate the blockquote and split the
+          // callout out of it.
+          out.push(block.split("\n").map((l) => prefix + l).join("\n"));
+        } else {
+          // Pure LIST-ITEM indentation: re-indent and keep the blank-line
+          // separators (a loose list item), so the div sits at the marker column.
+          out.push(
+            `\n${block
+              .split("\n")
+              .map((l) => (l.length ? prefix + l : l))
+              .join("\n")}\n`,
+          );
+        }
         i = j;
         continue;
       }
@@ -597,6 +625,40 @@ function bridgeTaskLists(html: string): string {
  * (null from parseAttachedComment), an unknown name, a wrong-position comment, or
  * an unknown/empty attr value is ignored.
  */
+/**
+ * A directive comment is in ATTACHED position when it sits inside a `<p>`/`<hN>`
+ * textblock — bound to that block's text (the `attrs`/`img` conventions). Every
+ * other parent (body, document level, a block container like blockquote/details/
+ * li/column div) is STANDALONE position, where a lone-block directive
+ * (subpages/pagebreak/pageembed/transclusion) is materialized. Broadening
+ * standalone beyond body/document is what lets these nodes survive NESTED inside
+ * a blockquote/callout/details/list item (previously dropped -> silent data loss).
+ */
+function isAttachedPosition(tag: string): boolean {
+  return tag === "p" || /^h[1-6]$/.test(tag);
+}
+
+/**
+ * Place a materialized standalone-directive element in the DOM: replace the
+ * comment IN PLACE when it has a real element parent inside <body> (body itself
+ * or a nested block container), preserving document order; queue it as a leading
+ * div only when the comment is at document level (no parentElement) or directly
+ * under `<html>` (outside <body>, which `document.body.innerHTML` would drop).
+ */
+function placeStandalone(
+  comment: any,
+  el: any,
+  tag: string,
+  leadingDivs: any[],
+): void {
+  if (comment.parentElement && tag !== "html") {
+    comment.replaceWith(el);
+  } else {
+    comment.remove();
+    leadingDivs.push(el);
+  }
+}
+
 function applyCommentDirectives(html: string): string {
   // Cheap early-out: no comments at all -> nothing to intercept.
   if (!html.includes("<!--")) return html;
@@ -664,13 +726,13 @@ function applyCommentDirectives(html: string): string {
 
     if (parsed.name === "subpages" || parsed.name === "pagebreak") {
       // #293 canon #5 STANDALONE machinery. A lone comment line is rendered by
-      // marked as an HTML block; the parser places it either directly under
-      // <body> (when other content surrounds it) or at document level (when it
-      // leads the output). Both are STANDALONE position. A `subpages`/`pagebreak`
-      // comment sitting inside a `<p>`/`<hN>` (or any other element) is attached
-      // position -> INERT.
-      const standalone = tag === "" || tag === "body" || tag === "html";
-      if (!standalone) continue; // wrong position -> inert
+      // marked as its own HTML block; the parser places it under <body>, at
+      // document level (leading), or — when the directive is NESTED — inside a
+      // block CONTAINER (`<blockquote>` for blockquote/callout, `<details>`,
+      // `<li>`, a column `<div>`, …). All of those are STANDALONE position. Only a
+      // comment ATTACHED inside a `<p>`/`<hN>` (bound to that block's text) is
+      // attached position -> INERT.
+      if (isAttachedPosition(tag)) continue; // wrong position -> inert
       const div = document.createElement("div");
       if (parsed.name === "pagebreak") {
         div.setAttribute("data-type", "pageBreak");
@@ -680,26 +742,18 @@ function applyCommentDirectives(html: string): string {
           div.setAttribute("data-recursive", "true");
         }
       }
-      if (tag === "body") {
-        // In-body: replace in place so surrounding content keeps its order.
-        comment.replaceWith(div);
-      } else {
-        // Document-level (leading): drop the stray comment and queue the div to
-        // be prepended into body below.
-        comment.remove();
-        leadingDivs.push(div);
-      }
+      placeStandalone(comment, div, tag, leadingDivs);
       continue;
     }
 
     if (parsed.name === "pageembed" || parsed.name === "transclusion") {
       // #293 canon #8 STANDALONE media. Like subpages/pagebreak: a lone comment
-      // line placed under <body> or at document level (leading). An attached-
-      // position comment (inside a <p>/<hN> with a sibling) is INERT. We rebuild
-      // the schema div the raw-HTML path emits (media-html.ts) from the decoded
-      // attrs so serialize/parse stay in sync.
-      const standalone = tag === "" || tag === "body" || tag === "html";
-      if (!standalone) continue; // wrong position -> inert
+      // line placed under <body>, at document level (leading), or NESTED inside a
+      // block container (blockquote/callout/details/li/column). An ATTACHED-
+      // position comment (inside a `<p>`/`<hN>`) is INERT. We rebuild the schema
+      // div the raw-HTML path emits (media-html.ts) from the decoded attrs so
+      // serialize/parse stay in sync.
+      if (isAttachedPosition(tag)) continue; // wrong position -> inert
       const el = buildElement(
         parsed.name === "pageembed"
           ? pageEmbedToHtml({ sourcePageId: parsed.attrs.sourcePageId })
@@ -709,12 +763,7 @@ function applyCommentDirectives(html: string): string {
             }),
       );
       if (!el) continue; // defensive: builder always yields an element
-      if (tag === "body") {
-        comment.replaceWith(el);
-      } else {
-        comment.remove();
-        leadingDivs.push(el);
-      }
+      placeStandalone(comment, el, tag, leadingDivs);
       continue;
     }
 
@@ -806,18 +855,36 @@ function applyCommentDirectives(html: string): string {
 
     if (!parent) continue; // attrs comment must have an element parent
     if (parsed.name !== "attrs") continue; // unknown name -> inert
-    // #293 canon #9 ATTACHED attrs: honored only in attached position.
-    const isBlock = tag === "p" || /^h[1-6]$/.test(tag);
-    if (!isBlock) continue; // misplaced comment -> inert
     const align = parsed.attrs.textAlign;
-    if (typeof align === "string" && align) {
-      // Re-express as an inline style; the schema's textAlign parseHTML reads
-      // `el.style.textAlign` back onto the paragraph/heading node.
-      parent.style.textAlign = align;
+    // #293 canon #9 ATTACHED attrs: honored only in attached position.
+    if (tag === "p" || /^h[1-6]$/.test(tag)) {
+      // A real <p>/<hN> host (loose list item, top-level block, …): re-express as
+      // an inline style; the schema's textAlign parseHTML reads `el.style.textAlign`
+      // back onto the paragraph/heading node.
+      if (typeof align === "string" && align) parent.style.textAlign = align;
+      comment.remove();
+    } else if (tag === "li" || tag === "td" || tag === "th") {
+      // TIGHT list item / GFM table cell: marked emits the paragraph's inline
+      // content DIRECTLY inside the <li>/<td>/<th> with NO <p> wrapper, so there
+      // is no element to carry the style — generateJSON materializes the
+      // paragraph later. Wrap the host's LEADING inline content (everything up to
+      // the comment; any trailing block child such as a nested list stays put) in
+      // a <p> carrying the alignment, so the materialized paragraph re-reads it.
+      if (typeof align === "string" && align) {
+        const p = document.createElement("p");
+        p.style.textAlign = align;
+        while (parent.firstChild && parent.firstChild !== comment) {
+          p.appendChild(parent.firstChild);
+        }
+        parent.insertBefore(p, comment);
+      }
+      comment.remove();
+    } else {
+      // Misplaced `attrs` comment (not a textblock/li/cell host): inert. Consume
+      // it anyway so no attached marker ever survives into the parsed body
+      // (matches the pre-existing "consume regardless" behaviour).
+      comment.remove();
     }
-    // Consume the marker regardless (unknown keys are simply ignored) so no
-    // attached comment ever survives into the parsed body.
-    comment.remove();
   }
   // Prepend any document-level (leading) standalone divs into body, preserving
   // their document order relative to each other and ahead of existing content.

packages/prosemirror-markdown/src/lib/media-html.ts

@@ -46,7 +46,11 @@ export function videoToHtml(attrs: Record<string, any>): string {
   if (attrs.width != null) parts.push(`width="${escapeAttr(attrs.width)}"`);
   if (attrs.height != null) parts.push(`height="${escapeAttr(attrs.height)}"`);
   if (attrs.size != null) parts.push(`data-size="${escapeAttr(attrs.size)}"`);
-  if (attrs.align) parts.push(`data-align="${escapeAttr(attrs.align)}"`);
+  // align default is "center" (schema): OMIT it so a bare/center node stays
+  // clean and parse's re-materialized "center" default is not a P2 churn — only
+  // a genuinely non-default left/right emits data-align (mirrors imageToHtml).
+  if (attrs.align && attrs.align !== "center")
+    parts.push(`data-align="${escapeAttr(attrs.align)}"`);
   if (attrs.aspectRatio != null)
     parts.push(`data-aspect-ratio="${escapeAttr(attrs.aspectRatio)}"`);
   return `<div><video ${parts.join(" ")}></video></div>`;
@@ -62,7 +66,9 @@ export function youtubeToHtml(attrs: Record<string, any>): string {
     parts.push(`data-width="${escapeAttr(attrs.width)}"`);
   if (attrs.height != null)
     parts.push(`data-height="${escapeAttr(attrs.height)}"`);
-  if (attrs.align) parts.push(`data-align="${escapeAttr(attrs.align)}"`);
+  // "center" is the schema default -> omit (see videoToHtml rationale).
+  if (attrs.align && attrs.align !== "center")
+    parts.push(`data-align="${escapeAttr(attrs.align)}"`);
   return `<div ${parts.join(" ")}></div>`;
 }
 
@@ -97,7 +103,9 @@ export function diagramToHtml(
   if (attrs.size != null) parts.push(`data-size="${escapeAttr(attrs.size)}"`);
   if (attrs.aspectRatio != null)
     parts.push(`data-aspect-ratio="${escapeAttr(attrs.aspectRatio)}"`);
-  if (attrs.align) parts.push(`data-align="${escapeAttr(attrs.align)}"`);
+  // "center" is the schema default -> omit (see videoToHtml rationale).
+  if (attrs.align && attrs.align !== "center")
+    parts.push(`data-align="${escapeAttr(attrs.align)}"`);
   if (attrs.attachmentId)
     parts.push(`data-attachment-id="${escapeAttr(attrs.attachmentId)}"`);
   return `<div ${parts.join(" ")}></div>`;
@@ -110,10 +118,18 @@ export function embedToHtml(attrs: Record<string, any>): string {
     `data-src="${escapeAttr(attrs.src ?? "")}"`,
     `data-provider="${escapeAttr(attrs.provider ?? "")}"`,
   ];
-  if (attrs.align) parts.push(`data-align="${escapeAttr(attrs.align)}"`);
-  if (attrs.width != null)
+  // "center" is the schema default -> omit (see videoToHtml rationale).
+  if (attrs.align && attrs.align !== "center")
+    parts.push(`data-align="${escapeAttr(attrs.align)}"`);
+  // embed width/height default to the NUMBERS 800/600 (schema). Getting a data-
+  // attribute back always yields a STRING, so emitting the default here would
+  // round-trip 800 -> "800" (a number->string P1 divergence canonicalize does
+  // NOT normalize). OMIT the defaults so parse re-materializes the numeric
+  // default instead — mirrors the top-level embed path (markdown-converter.ts),
+  // which also emits width/height only when they differ from 800/600.
+  if (attrs.width != null && attrs.width !== 800)
     parts.push(`data-width="${escapeAttr(attrs.width)}"`);
-  if (attrs.height != null)
+  if (attrs.height != null && attrs.height !== 600)
     parts.push(`data-height="${escapeAttr(attrs.height)}"`);
   return `<div ${parts.join(" ")}></div>`;
 }

packages/prosemirror-markdown/test/fixtures/counterexamples/columns-column-width-percent.json

@@ -0,0 +1,16 @@
+{
+  "_bug": "BUG #351: a `column` whose `width` is a percentage string (e.g. \"50%\") is NOT byte-stable across export->import->export (violates P2). The `column` schema's parseHTML does `parseFloat(getAttribute('data-width'))`, which silently drops the '%' unit and returns the NUMBER 50. So the first export emits data-width=\"50%\" but the re-import stores width=50, and the second export emits data-width=\"50\": md2 !== md1, a permanent GS-EDIT-REVERT churn (every git-sync pull rewrites the column width). The editor authors column widths as percentages, so this is a real data/round-trip defect. Fix belongs in src/lib/docmost-schema.ts column.width parseHTML (preserve the unit / keep the string), which is OUT OF SCOPE for this test-only PR and must be a separate, maintainer-approved change. This flat generator therefore keeps `column.width` frozen (never generates a non-default width).",
+  "doc": {
+    "type": "doc",
+    "content": [
+      {
+        "type": "columns",
+        "attrs": { "layout": "two_equal", "widthMode": "normal" },
+        "content": [
+          { "type": "column", "attrs": { "width": "50%" }, "content": [{ "type": "paragraph", "content": [{ "type": "text", "text": "L" }] }] },
+          { "type": "column", "attrs": { "width": "50%" }, "content": [{ "type": "paragraph", "content": [{ "type": "text", "text": "R" }] }] }
+        ]
+      }
+    ]
+  }
+}

packages/prosemirror-markdown/test/fixtures/counterexamples/ordered-list-start.json

@@ -0,0 +1,19 @@
+{
+  "doc": {
+    "type": "doc",
+    "content": [
+      {
+        "type": "orderedList",
+        "attrs": { "type": null, "start": 5 },
+        "content": [
+          {
+            "type": "listItem",
+            "content": [
+              { "type": "paragraph", "content": [{ "type": "text", "text": "alpha" }] }
+            ]
+          }
+        ]
+      }
+    ]
+  }
+}

packages/prosemirror-markdown/test/generative/attr-arbitraries.ts

@@ -0,0 +1,325 @@
+/**
+ * Schema-DERIVED attribute-state fast-check arbitraries (#351, PR 1).
+ *
+ * This GENERALIZES the #350 stability-matrix helper (roundtrip-stability.helper.ts)
+ * to fast-check. Where that helper sweeps a HAND-WRITTEN 2-state matrix for one
+ * node spec, this module reads the attribute list straight from
+ * `schema.nodes[type].spec.attrs` (never a hand list) and, per attribute,
+ * generates over the FOUR states the issue calls for:
+ *
+ *   - `absent`      : the attribute is OMITTED entirely (the empty-string-vs-
+ *                     absent churn class the #350 fix targets).
+ *   - `default`     : the schema default value, authored explicitly.
+ *   - `nonDefault`  : a representative legal non-default value.
+ *   - `degenerate`  : `""` for strings, `0`/negative for numbers, the flipped
+ *                     value for booleans.
+ *
+ * ── Why a per-attribute override table ──────────────────────────────────────
+ * Everything that CAN be derived generically from the default's runtime type is
+ * (booleans flip; the degenerate value follows the runtime type). But two facts
+ * force a small, DOCUMENTED override table:
+ *
+ *  1. CONSTRAINED domains the schema does not encode. `image.align ∈
+ *     {left,center,right}`, `heading.level ∈ 1..6`, `callout.type ∈
+ *     {info,success,warning,danger}`, `columns.layout`, table-cell `align`,
+ *     `status.color`, `orderedList.start ≥ 1`, etc. A generic "default + 1"
+ *     would emit an ILLEGAL value, so these get an explicit legal domain.
+ *  2. ROUND-TRIP-safety, established EMPIRICALLY by probing the live converter
+ *     (the classification captured in flat-roundtrip.property.test.ts). A frozen
+ *     attribute falls into ONE of TWO explicitly-distinguished classes — never a
+ *     silent "it just doesn't round-trip":
+ *
+ *       (a) ACCEPTED LIMITATION — the attribute has NO markdown representation,
+ *           so the loss is inherent to targeting markdown, not a converter
+ *           defect. These: `paragraph`/`heading` `indent`, `callout.icon`,
+ *           `orderedList.type` (a/A/i markers), table `colwidth` /
+ *           `backgroundColor(Name)` (dropped by the raw-<table> fallback). Each is
+ *           tagged `// ACCEPTED:` inline. Freezing them is correct — there is
+ *           nothing to preserve in the target format.
+ *
+ *       (b) PINNED BUG — the attribute IS representable in markdown but the
+ *           converter drops it anyway (a real defect). These are NOT silently
+ *           frozen: each is captured as a LOUD `it.fails` counterexample in
+ *           test/fixtures/counterexamples/ + counterexamples.test.ts, and the
+ *           freeze here only keeps the P1/P2 union green until a MAINTAINER rules
+ *           on accept-vs-fix (the epic guardrail reserves that call). These:
+ *           `column.width` (parseFloat drops `%`), `orderedList.start` (non-1
+ *           start renders as `1.`). Tagged `// PINNED-BUG:` inline.
+ *
+ *       (c) DEFERRED-BUG — representable AND round-trips, frozen only because the
+ *           flat generator can't yet build a valid instance. Table
+ *           `colspan`/`rowspan` round-trip via the raw-<table> fallback, but a
+ *           geometrically-valid spanned table is PR-2 structural work; the flat
+ *           generator hardcodes span = 1. Tagged `// DEFERRED-BUG:` inline so a
+ *           maintainer does not read them as an inherent limitation.
+ *       - Several non-null-default attrs are MATERIALIZED on import but are not
+ *         in canonicalize's KNOWN_DEFAULTS (`callout.type`, `status.color`,
+ *         table `colspan`/`rowspan`, `columns.layout`/`widthMode`,
+ *         `embed.width`/`height`, `heading.level`, `taskItem.checked`,
+ *         `details.open`, `subpages.recursive`, `orderedList.start`). If left
+ *         `absent` they re-materialize as a non-canonical default and diverge
+ *         under P1. We mark them `always` so they are authored explicitly.
+ *       - The documented numeric→string coercion set (`width height size
+ *         aspectRatio`) is generated as STRINGS for the media family (a stored
+ *         number re-parses as a string), EXCEPT `embed.width/height` which the
+ *         embed schema keeps numeric — handled per-attr.
+ *
+ * Both PINNED-BUG attrs (`column.width` P2 churn, `orderedList.start` P1 loss)
+ * are captured as committed `it.fails` counterexamples — NOT hidden here.
+ */
+import fc from 'fast-check';
+import { getSchema } from '@tiptap/core';
+import { docmostExtensions } from '../../src/lib/index.js';
+import { phraseArb, letterPhraseArb, urlArb } from './text-arbitraries.js';
+
+/** The exact ProseMirror schema the converter targets. */
+export const schema = getSchema(docmostExtensions as any);
+
+/** Sentinel: this attribute is OMITTED (the `absent` state). */
+export const ABSENT = Symbol('ABSENT');
+
+/** The documented numeric→string coercion set (issue + roundtrip-stability.helper). */
+export const NUMERIC_STRING_ATTRS = ['width', 'height', 'size', 'aspectRatio'];
+
+/** Read the schema default for every attribute of a node type. */
+export function schemaAttrDefaults(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;
+}
+
+/** Attribute names for a node type, straight from the schema (never hand-listed). */
+export function schemaAttrNames(type: string): string[] {
+  return Object.keys((schema.nodes[type]?.spec?.attrs ?? {}) as object);
+}
+
+/**
+ * Per-attribute policy. Everything unlisted falls back to a generic policy:
+ *   - a BOOLEAN default is fuzzable (its non-default is the flipped value);
+ *   - any other default is `frozen` (only `absent`/`default` are generated) so
+ *     we never invent an unverified non-default that might not round-trip.
+ * Listed attrs override this with a legal `arb` domain and/or flags.
+ */
+interface AttrPolicy {
+  /** Arbitrary for the `nonDefault` state's value. */
+  arb?: fc.Arbitrary<unknown>;
+  /** Value for the `degenerate` state (fuzz mode only). Omit to skip degenerate. */
+  degen?: unknown;
+  /** Never emit `absent` — the attr must be authored (materialized default class). */
+  always?: boolean;
+  /** Never emit the schema default value (required-ish attrs like `src`). Implies always. */
+  noDefault?: boolean;
+  /** Never emit non-default/degenerate — attr has no md representation or churns. */
+  frozen?: boolean;
+}
+
+const num = (...xs: number[]) => fc.constantFrom(...xs);
+const str = (...xs: string[]) => fc.constantFrom(...xs);
+const widthStr = str('120', '320', '640');
+
+// The documented override table, keyed `type.attr`. Every entry is grounded in
+// the empirical converter probe (see flat-roundtrip.property.test.ts header).
+const OVERRIDES: Record<string, AttrPolicy> = {
+  // ── block text containers ────────────────────────────────────────────────
+  // 'left' is the IMPLICIT default alignment: the converter drops it on export
+  // (empirically confirmed), so it never round-trips. Only center/right/justify
+  // carry through the `<!--attrs {textAlign}-->` comment.
+  'paragraph.textAlign': { arb: str('center', 'right', 'justify') },
+  'paragraph.indent': { frozen: true }, // ACCEPTED: no md representation
+  'heading.level': { always: true, arb: num(2, 3, 4, 5, 6) },
+  'heading.textAlign': { arb: str('center', 'right', 'justify') },
+  'heading.indent': { frozen: true }, // ACCEPTED: no md representation
+  // ── lists ────────────────────────────────────────────────────────────────
+  // PINNED-BUG: markdown CAN express a non-1 start ("5."), but the converter
+  // renders "1." and drops it -> P1 loss. See counterexamples.test.ts
+  // (ordered-list-start.json). Frozen only until the maintainer rules accept-vs-fix.
+  'orderedList.start': { always: true, frozen: true },
+  'orderedList.type': { frozen: true }, // ACCEPTED: a/A/i markers not expressible in GFM
+  'taskItem.checked': { always: true, arb: fc.constant(true) }, // boolean, default false
+  // ── codeBlock ────────────────────────────────────────────────────────────
+  'codeBlock.language': { arb: str('js', 'ts', 'python', 'go', 'rust', 'bash') },
+  // ── image / media (numeric→string width family) ──────────────────────────
+  'image.src': { noDefault: true, arb: urlArb, degen: '' },
+  'image.align': { arb: str('left', 'right') },
+  'image.alt': { arb: letterPhraseArb, degen: '' },
+  'image.title': { arb: letterPhraseArb },
+  'image.width': { arb: widthStr, degen: '' },
+  'image.height': { arb: widthStr, degen: '' },
+  'video.src': { noDefault: true, arb: urlArb, degen: '' },
+  'video.alt': { arb: letterPhraseArb },
+  'video.width': { arb: widthStr },
+  'video.height': { arb: widthStr },
+  'audio.src': { noDefault: true, arb: urlArb, degen: '' },
+  'youtube.src': { noDefault: true, arb: urlArb },
+  'pdf.src': { noDefault: true, arb: urlArb },
+  'pdf.name': { arb: phraseArb },
+  'drawio.src': { noDefault: true, arb: urlArb },
+  'excalidraw.src': { noDefault: true, arb: urlArb },
+  'attachment.url': { noDefault: true, arb: urlArb },
+  'attachment.name': { arb: phraseArb },
+  // ── callout / status ─────────────────────────────────────────────────────
+  'callout.type': { always: true, arb: str('success', 'warning', 'danger') },
+  'callout.icon': { frozen: true }, // ACCEPTED: no md representation (dropped on export)
+  'status.text': { noDefault: true, arb: phraseArb, degen: '' },
+  'status.color': { always: true, arb: str('green', 'orange', 'red', 'blue', 'yellow', 'purple') },
+  // ── table cells ────────────────────────────────────────────────────────────
+  // DEFERRED-BUG (not ACCEPTED): colspan/rowspan ARE representable and round-trip
+  // — a spanned cell makes the converter emit the whole table as a raw <table>
+  // with colspan/rowspan attrs (markdown-converter.ts tableToHtml), which the
+  // tiptap parser reads back. They are frozen only because generating a
+  // geometrically-valid spanned table is deferred STRUCTURAL work (the flat
+  // generator hardcodes colspan/rowspan = 1), NOT a markdown limitation.
+  'tableCell.colspan': { always: true, frozen: true },
+  'tableCell.rowspan': { always: true, frozen: true },
+  // ACCEPTED: colwidth / backgroundColor(Name) have no representation — the
+  // raw-<table> fallback (tableToHtml) drops them, so there is nothing to preserve.
+  'tableCell.colwidth': { frozen: true },
+  'tableCell.backgroundColor': { frozen: true },
+  'tableCell.backgroundColorName': { frozen: true },
+  'tableCell.align': { arb: str('left', 'center', 'right') },
+  'tableHeader.colspan': { always: true, frozen: true }, // DEFERRED-BUG (see tableCell.colspan)
+  'tableHeader.rowspan': { always: true, frozen: true }, // DEFERRED-BUG (see tableCell.rowspan)
+  'tableHeader.colwidth': { frozen: true }, // ACCEPTED: no representation
+  'tableHeader.backgroundColor': { frozen: true }, // ACCEPTED: no representation
+  'tableHeader.backgroundColorName': { frozen: true }, // ACCEPTED: no representation
+  'tableHeader.align': { arb: str('left', 'center', 'right') },
+  // ── details ──────────────────────────────────────────────────────────────
+  'details.open': { always: true, arb: fc.constant(true) }, // boolean, default false
+  // ── columns ──────────────────────────────────────────────────────────────
+  'columns.layout': { always: true, arb: str('three_equal', 'left_sidebar', 'right_sidebar') },
+  // widthMode round-trips via the `data-width-mode` attribute (verified P1+P2),
+  // so it is fuzzed, not frozen.
+  'columns.widthMode': { always: true, arb: str('custom') },
+  // PINNED-BUG: parseFloat import drops the `%` unit -> P2 churn. See
+  // counterexamples.test.ts (columns-column-width-percent.json).
+  'column.width': { frozen: true },
+  // ── embed (schema keeps width/height NUMERIC, not string-coerced) ─────────
+  'embed.src': { noDefault: true, arb: urlArb, degen: '' },
+  'embed.provider': { noDefault: true, arb: str('iframe', 'youtube', 'vimeo') },
+  'embed.width': { always: true, frozen: true },
+  'embed.height': { always: true, frozen: true },
+  // ── subpages / math / htmlEmbed ──────────────────────────────────────────
+  'subpages.recursive': { always: true, arb: fc.constant(true) }, // boolean, default false
+  'mathBlock.text': { noDefault: true, arb: str('x^2', 'a < b', '\\frac{1}{2}'), degen: '' },
+  'mathInline.text': { noDefault: true, arb: str('x^2', 'a < b', '\\frac{1}{2}'), degen: '' },
+  'htmlEmbed.source': { noDefault: true, arb: str('<b>hi</b>', '<i>x</i>', '<span>y</span>'), degen: '' },
+  'htmlEmbed.height': { arb: num(200, 300, 400) },
+  // ── footnotes / transclusion / pageEmbed / mention ───────────────────────
+  'footnoteDefinition.id': { noDefault: true, arb: str('fn1', 'fn2', 'note') },
+  'footnoteReference.id': { noDefault: true, arb: str('fn1', 'fn2', 'note') },
+  'pageEmbed.sourcePageId': { noDefault: true, arb: fc.uuid() },
+  'transclusionSource.id': { noDefault: true, arb: str('src1', 'src2') },
+  'transclusionReference.sourcePageId': { noDefault: true, arb: fc.uuid() },
+  'transclusionReference.transclusionId': { noDefault: true, arb: str('tr1', 'tr2') },
+  'mention.id': { noDefault: true, arb: fc.uuid() },
+  'mention.label': { noDefault: true, arb: phraseArb },
+  'mention.entityType': { noDefault: true, arb: str('user') },
+  'mention.entityId': { noDefault: true, arb: fc.uuid() },
+};
+
+/** Resolve the effective policy for one attribute (override merged over generic). */
+function policyFor(type: string, attr: string, def: unknown): AttrPolicy {
+  const override = OVERRIDES[`${type}.${attr}`];
+  if (override) return override;
+  // Generic: booleans are fuzzable via their flipped value; everything else is
+  // frozen (only absent/default) so no unverified non-default is invented.
+  if (typeof def === 'boolean') return { arb: fc.constant(!def) };
+  return { frozen: true };
+}
+
+/**
+ * Whether an attribute is actually exercised at a NON-DEFAULT value (i.e. its
+ * policy has an `arb`, which the generic fallback does not). Used by the
+ * attribute-coverage snapshot test to make the generic-frozen space VISIBLE: any
+ * string/number attr not in OVERRIDES is silently only tested at absent/default,
+ * so the snapshot pins exactly which attrs are NOT value-fuzzed and forces a
+ * reviewer to look when a new attr lands in that invisible bucket.
+ */
+export function attrIsValueFuzzed(type: string, attr: string): boolean {
+  const def = schemaAttrDefaults(type)[attr];
+  return !!policyFor(type, attr, def).arb;
+}
+
+/** Every node `type.attr` in the schema (excluding the auto `id`), sorted. */
+export function allSchemaAttrKeys(): string[] {
+  const keys: string[] = [];
+  for (const type of Object.keys(schema.nodes)) {
+    for (const attr of schemaAttrNames(type)) {
+      if (attr === 'id') continue;
+      keys.push(`${type}.${attr}`);
+    }
+  }
+  return keys.sort();
+}
+
+/**
+ * Every MARK attribute in the schema, keyed `mark:<name>.<attr>`, sorted. Marks
+ * are not driven by the node OVERRIDES table (they are fuzzed by the text
+ * generator, text-arbitraries.ts), so their value-fuzz coverage is tracked with a
+ * separate snapshot (see flat-roundtrip.property.test.ts) — without this the
+ * "no invisible coverage hole" guarantee would hold for node attrs only, letting a
+ * new mark attr slip through unfuzzed and unallowlisted.
+ */
+export function allSchemaMarkAttrKeys(): string[] {
+  const keys: string[] = [];
+  for (const [name, mark] of Object.entries(schema.marks)) {
+    const attrs = (mark.spec?.attrs ?? {}) as Record<string, unknown>;
+    for (const attr of Object.keys(attrs)) keys.push(`mark:${name}.${attr}`);
+  }
+  return keys.sort();
+}
+
+export type AttrMode = 'p1' | 'fuzz';
+
+/**
+ * Build an arbitrary for ONE attribute's value (or the ABSENT sentinel) across
+ * the states legal for `mode`:
+ *   - p1   : absent / default / nonDefault (the round-trip-safe space).
+ *   - fuzz : the above PLUS degenerate (P2 tolerates the one-time
+ *            normalization; P3 only needs totality).
+ */
+export function attrValueArb(
+  type: string,
+  attr: string,
+  mode: AttrMode,
+): fc.Arbitrary<unknown | typeof ABSENT> {
+  const def = schemaAttrDefaults(type)[attr];
+  const p = policyFor(type, attr, def);
+
+  const states: fc.Arbitrary<unknown | typeof ABSENT>[] = [];
+  if (!p.always && !p.noDefault) states.push(fc.constant(ABSENT));
+  if (!p.noDefault) states.push(fc.constant(def));
+  if (!p.frozen && p.arb) states.push(p.arb);
+  if (mode === 'fuzz' && !p.frozen && p.degen !== undefined) {
+    states.push(fc.constant(p.degen));
+  }
+  if (states.length === 0) states.push(fc.constant(def));
+  return fc.oneof(...states);
+}
+
+/**
+ * Build an arbitrary for a node's full `attrs` object over all schema attrs.
+ * `base` pins caller-required attrs (e.g. a concrete `src`) verbatim; any attr
+ * present in `base` is NOT re-generated. Omitted (ABSENT) attrs are dropped.
+ */
+export function nodeAttrsArb(
+  type: string,
+  mode: AttrMode,
+  base: Record<string, unknown> = {},
+): fc.Arbitrary<Record<string, unknown>> {
+  const names = schemaAttrNames(type).filter((n) => !(n in base) && n !== 'id');
+  if (names.length === 0) return fc.constant({ ...base });
+  return fc
+    .tuple(...names.map((n) => attrValueArb(type, n, mode)))
+    .map((vals) => {
+      const attrs: Record<string, unknown> = { ...base };
+      names.forEach((n, i) => {
+        if (vals[i] !== ABSENT) attrs[n] = vals[i];
+      });
+      return attrs;
+    });
+}

packages/prosemirror-markdown/test/generative/counterexamples.test.ts

@@ -0,0 +1,74 @@
+import { describe, expect, it } from 'vitest';
+import { readFileSync } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+import path from 'node:path';
+import { convertProseMirrorToMarkdown } from '../../src/lib/markdown-converter.js';
+import { markdownToProseMirror } from '../../src/lib/markdown-to-prosemirror.js';
+import { docsCanonicallyEqual } from '../../src/lib/canonicalize.js';
+
+// ---------------------------------------------------------------------------
+// #351 committed counterexamples — REAL round-trip bugs surfaced by the flat
+// generative probing (attribute level). Each is pinned here as an `it.fails`
+// (vitest passes ONLY WHILE the assertion still fails), so that the day the
+// underlying src/ bug is fixed, the `it.fails` starts PASSING and vitest turns
+// this test RED — forcing us to delete the counterexample and (per the epic
+// guardrail) tighten the generator. A bare `it.fails` would ship silent
+// corruption, so every case below carries a loud `// BUG #351:` explanation.
+//
+// These bugs are NOT worked around by weakening any property: the offending
+// attribute is kept OUT of the P1/P2 generators (documented in
+// attr-arbitraries.ts), and the exact failing document lives here as the
+// regression pin. FIXING the bug is a separate, maintainer-approved src/ change.
+// ---------------------------------------------------------------------------
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const fixtureDir = path.resolve(here, '../fixtures/counterexamples');
+
+function loadDoc(file: string): any {
+  return JSON.parse(readFileSync(path.join(fixtureDir, file), 'utf8')).doc;
+}
+
+describe('#351 counterexamples (known round-trip bugs, pinned as it.fails)', () => {
+  // BUG #351: a `column` with a PERCENTAGE width ("50%") is not byte-stable.
+  // The column schema parses `data-width` with parseFloat, dropping the '%':
+  //   md1 = '...data-width="50%"...'   (first export)
+  //   re-import stores width = 50 (number)
+  //   md2 = '...data-width="50"...'    (second export)  => md2 !== md1
+  // A permanent GS-EDIT-REVERT churn on every git-sync pull. The editor stores
+  // column widths as percentages, so this is a genuine defect. The fix is in
+  // src/lib/docmost-schema.ts (column.width parseHTML must preserve the unit)
+  // and is out of scope for this test-only PR.
+  it.fails('column percentage width is byte-stable (P2)', async () => {
+    const doc = loadDoc('columns-column-width-percent.json');
+    const md1 = convertProseMirrorToMarkdown(doc);
+    const doc2 = await markdownToProseMirror(md1);
+    const md2 = convertProseMirrorToMarkdown(doc2);
+    // This assertion currently FAILS (md2 drops the '%'), which is exactly what
+    // `it.fails` expects. When the schema is fixed, it will PASS and flip this
+    // test red — our cue to remove the pin.
+    expect(md2).toBe(md1);
+  });
+
+  // BUG #351: an `orderedList` with a non-1 `start` loses its start number.
+  // CommonMark CAN express this ("5." starts the list at 5), but the converter
+  // always emits "1." and ignores `attrs.start` (markdown-converter.ts renders
+  // `${index + 1}.`; the <ol> HTML path also omits `start`):
+  //   doc.start = 5   ->   md1 = "1. alpha"   (start dropped on export)
+  //   re-import stores start = 1   =>   docsCanonicallyEqual(rt, doc) === false
+  // This is a P1 (semantic round-trip) loss of the SAME class as column.width:
+  // representable in markdown, silently dropped by the converter. It is pinned
+  // here as the LOUD counterexample rather than being masked as an "accepted
+  // normalization" in the generator — per the epic guardrail, deciding
+  // accept-vs-fix for a markdown-representable loss is a MAINTAINER call, so this
+  // stays a visible known-bug until the maintainer rules on it. The fix would be
+  // in src/lib/markdown-converter.ts (emit the start number on the first item)
+  // and is out of scope for this test-only PR.
+  it.fails('ordered list start number is preserved (P1)', async () => {
+    const doc = loadDoc('ordered-list-start.json');
+    const md1 = convertProseMirrorToMarkdown(doc);
+    const doc2 = await markdownToProseMirror(md1);
+    // Currently FAILS: doc2.start === 1 while doc.start === 5. When the converter
+    // preserves `start`, this PASSES and flips the test red — remove the pin then.
+    expect(docsCanonicallyEqual(doc2, doc)).toBe(true);
+  });
+});

packages/prosemirror-markdown/test/generative/doc-generator.ts

@@ -0,0 +1,392 @@
+/**
+ * Nested whole-document generator (#351, PR 2) — "variant A": a random walk over
+ * the schema's ContentMatch automaton.
+ *
+ * Where the FLAT generator (node-generators.ts) emits `{doc:[ <one target> ]}`,
+ * this module produces arbitrarily DEEP, valid ProseMirror documents. Validity is
+ * NOT hand-asserted: it comes straight from the schema. To fill a container's
+ * block content we start at `nodeType.contentMatch` and walk the automaton —
+ * enumerate the legal next node types (`match.edge(i).type`), let fast-check pick
+ * one (or STOP once `match.validEnd`), generate that child RECURSIVELY, then
+ * advance the automaton via `match.matchType(childType)`. A bad walk therefore
+ * cannot emit a structurally-invalid doc (the `generator validity` test guards
+ * this with `schema.nodeFromJSON(json).check()`).
+ *
+ * ── What the walk drives, and what it delegates ──────────────────────────────
+ * The walk owns BLOCK STRUCTURE (which blocks nest inside which containers, in
+ * what order and how deep). Two things it deliberately delegates, to avoid
+ * REGRESSING the byte-stable space the flat suite already proved empirically:
+ *
+ *   - INLINE content of textblocks (paragraph/heading/codeBlock/detailsSummary)
+ *     is filled from text-arbitraries.ts — the exact hostile-but-byte-stable
+ *     inline corpus the flat suite established. Walking the inline ContentMatch
+ *     instead would re-derive (and re-fail) the text-space limitations the flat
+ *     suite already pins, which is not this generator's job.
+ *   - Node ATTRS come from nodeAttrsArb(type, 'p1') — the round-trip-safe
+ *     attribute space. Attribute-degenerate fuzzing (P2/P3 over 'fuzz') is the
+ *     flat suite's concern; the nested suite isolates STRUCTURAL round-trip, so
+ *     it stays in 'p1' and does not re-litigate frozen/pinned attributes.
+ *
+ * ── Coordination the automaton cannot express ────────────────────────────────
+ * A ContentMatch guarantees a child SEQUENCE is legal, but not cross-sibling
+ * invariants. Two nodes need coordination the walk injects by hand:
+ *   - `table`: GFM needs a RECTANGULAR grid with column-consistent alignment.
+ *     The automaton happily allows ragged rows / per-cell align, which are not a
+ *     converter bug — just a malformed table. So a table is generated ATOMICALLY
+ *     (same shape the flat suite proved) rather than walked.
+ *   - `columns`: the `layout` attr must agree with the column COUNT. We pick the
+ *     layout, derive the count, then walk each column's block body normally — so
+ *     columns still gain real nested content, only the count is coordinated.
+ *
+ * ── Excluded from the nested walk ────────────────────────────────────────────
+ * Footnote nodes (footnoteReference / footnotesList / footnoteDefinition) need a
+ * DOCUMENT-GLOBAL id match between a reference and its definition. That
+ * coordination is owned by the flat suite's `footnotes` generator; placing them
+ * independently here would fabricate id mismatches that look like converter bugs
+ * but are generator defects. They are filtered out of the walk (documented in
+ * EXCLUDED). The completeness contract lives in the FLAT suite and is unaffected.
+ *
+ * ── Termination / budgets ────────────────────────────────────────────────────
+ * Two bounds keep every doc finite and the suite fast:
+ *   - MAX_DEPTH — a hard cap on block-nesting depth. A precomputed `minDepth`
+ *     fixpoint (the minimum extra nesting a subtree of each type needs to be
+ *     valid) lets the walk pick a CONTAINER child only when there is depth
+ *     headroom to complete it — so the walk can never paint itself into a corner
+ *     where a required child cannot fit (no invalid docs, guaranteed termination).
+ *   - NODE_BUDGET — a soft cap on total nodes; as it runs low the walk biases
+ *     toward STOP (when validEnd) or toward cheap terminating children.
+ */
+import fc from 'fast-check';
+import { getSchema } from '@tiptap/core';
+import { docmostExtensions } from '../../src/lib/docmost-schema.js';
+import { nodeAttrsArb } from './attr-arbitraries.js';
+import {
+  inlineContentArb,
+  headingInlineContentArb,
+  plainInlineContentArb,
+  phraseArb,
+} from './text-arbitraries.js';
+
+/** The exact ProseMirror schema the converter targets (built per the issue). */
+export const schema = getSchema(docmostExtensions as never);
+
+/** Hard cap on block-nesting depth (doc = depth 0). Kept in the issue's 4–5 band. */
+export const MAX_DEPTH = 4;
+/**
+ * Soft cap on total node count per generated document. Kept moderate: every P1/P2
+ * run parses the emitted markdown through jsdom (heavy), so 100+ node docs across
+ * hundreds of runs exhaust the worker heap. 60 still yields deeply-nested docs
+ * (depth 4) while keeping the suite within memory.
+ */
+export const NODE_BUDGET = 60;
+
+/**
+ * Nodes kept OUT of the nested walk: footnote nodes need a doc-global id match a
+ * local walk cannot coordinate (owned by the flat suite's `footnotes` generator).
+ */
+const EXCLUDED = new Set<string>([
+  'footnoteReference',
+  'footnotesList',
+  'footnoteDefinition',
+]);
+
+/**
+ * Structural-only children that carry `group: "block"` in the schema and so leak
+ * into EVERY block container's ContentMatch, even though the editor only ever
+ * places them inside their one true parent. Choosing them freely (e.g. a bare
+ * `column` at the document root) fabricates documents no editor produces and that
+ * the converter is not designed to round-trip — a GENERATOR artifact, not a
+ * converter bug. They are admitted ONLY when the container being filled is their
+ * dedicated parent. (`column` is in fact always built inside columnsArb, so this
+ * just double-guards it.)
+ */
+const DEDICATED_PARENT: Record<string, string> = {
+  column: 'columns',
+  detailsSummary: 'details',
+  detailsContent: 'details',
+};
+
+/** Is child type `t` legal as a freely-chosen child of container `parentType`? */
+function childAllowedUnder(t: string, parentType: string): boolean {
+  const dedicated = DEDICATED_PARENT[t];
+  return dedicated === undefined || dedicated === parentType;
+}
+
+/** Textblock (inlineContent) types — filled from the proven inline corpus. */
+function isTextblock(typeName: string): boolean {
+  return !!schema.nodes[typeName]?.isTextblock;
+}
+/** Leaf/atom types — no content, only generated attrs. */
+function isLeaf(typeName: string): boolean {
+  return !!schema.nodes[typeName]?.isLeaf;
+}
+
+// ---------------------------------------------------------------------------
+// minDepth fixpoint: the minimum EXTRA block-nesting depth a valid subtree
+// rooted at each node type requires. Leaves and textblocks need 0 (a textblock
+// is satisfied by inline content, no block recursion). A container needs
+// 1 + the cheapest way to satisfy its ContentMatch. Computed as a min–max path
+// to `validEnd` over the automaton, iterated to a fixpoint over node types.
+// ---------------------------------------------------------------------------
+
+/** Cheapest (min over reachable validEnd of max child minDepth) to complete a match. */
+function minCompletion(
+  match: any,
+  md: Record<string, number>,
+  seen: Set<any>,
+  parentType: string,
+): number {
+  let best = match.validEnd ? 0 : Infinity;
+  if (seen.has(match)) return best; // a cycle never completes more cheaply
+  seen.add(match);
+  for (let i = 0; i < match.edgeCount; i++) {
+    const edge = match.edge(i);
+    const t = edge.type.name;
+    if (EXCLUDED.has(t) || t === 'text') continue;
+    if (!childAllowedUnder(t, parentType)) continue;
+    const childCost = md[t];
+    if (childCost === undefined || childCost === Infinity) continue;
+    const rest = minCompletion(edge.next, md, seen, parentType);
+    if (rest === Infinity) continue;
+    best = Math.min(best, Math.max(childCost, rest));
+  }
+  seen.delete(match);
+  return best;
+}
+
+function computeMinDepth(): Record<string, number> {
+  const md: Record<string, number> = {};
+  for (const name of Object.keys(schema.nodes)) {
+    md[name] = isLeaf(name) || isTextblock(name) ? 0 : Infinity;
+  }
+  let changed = true;
+  while (changed) {
+    changed = false;
+    for (const name of Object.keys(schema.nodes)) {
+      if (md[name] === 0) continue; // leaves/textblocks fixed at 0
+      const nt: any = schema.nodes[name];
+      const completion = minCompletion(nt.contentMatch, md, new Set(), name);
+      const next = completion === Infinity ? Infinity : 1 + completion;
+      if (next < md[name]) {
+        md[name] = next;
+        changed = true;
+      }
+    }
+  }
+  return md;
+}
+
+const MIN_DEPTH = computeMinDepth();
+
+// ---------------------------------------------------------------------------
+// Leaf / textblock builders (attrs from 'p1', inline from the proven corpus).
+// ---------------------------------------------------------------------------
+
+function attachAttrs(typeName: string, base: Record<string, unknown> = {}) {
+  return nodeAttrsArb(typeName, 'p1', base).map((attrs) => {
+    const node: any = { type: typeName };
+    if (Object.keys(attrs).length) node.attrs = attrs;
+    return node;
+  });
+}
+
+/** A leaf/atom block: attrs only, no content. */
+function leafArb(typeName: string): fc.Arbitrary<any> {
+  return attachAttrs(typeName);
+}
+
+/** A textblock, inline content taken from the byte-stable flat corpus. */
+function textblockArb(typeName: string): fc.Arbitrary<any> {
+  if (typeName === 'codeBlock') {
+    return fc
+      .tuple(
+        nodeAttrsArb('codeBlock', 'p1'),
+        // Fenced code re-imports with a TRAILING NEWLINE (flat suite finding);
+        // author it so the doc is already at the round-trip fixpoint.
+        fc.array(phraseArb, { minLength: 1, maxLength: 3 }).map((l) => l.join('\n') + '\n'),
+      )
+      .map(([attrs, code]) => ({
+        type: 'codeBlock',
+        ...(Object.keys(attrs).length ? { attrs } : {}),
+        content: [{ type: 'text', text: code }],
+      }));
+  }
+  const inline =
+    typeName === 'heading'
+      ? headingInlineContentArb
+      : typeName === 'detailsSummary'
+        ? plainInlineContentArb
+        : inlineContentArb;
+  return fc
+    .tuple(nodeAttrsArb(typeName, 'p1'), inline)
+    .map(([attrs, content]) => ({
+      type: typeName,
+      ...(Object.keys(attrs).length ? { attrs } : {}),
+      content,
+    }));
+}
+
+// ---------------------------------------------------------------------------
+// Coordinated builders: table (atomic, rectangular, column-consistent align)
+// and columns (layout coupled to count, bodies walked).
+// ---------------------------------------------------------------------------
+
+/** A rectangular GFM-safe table (mirrors the flat suite's proven shape). */
+function tableArb(): fc.Arbitrary<any> {
+  return fc.integer({ min: 1, max: 3 }).chain((cols) => {
+    // One alignment per COLUMN, identical on header + every body cell, so the
+    // second export cannot re-align and churn.
+    const alignsArb = fc.array(fc.constantFrom(undefined, 'left', 'center', 'right'), {
+      minLength: cols,
+      maxLength: cols,
+    });
+    const cell = (header: boolean, align?: string) =>
+      phraseArb.map((t) => ({
+        type: header ? 'tableHeader' : 'tableCell',
+        attrs: { colspan: 1, rowspan: 1, ...(align ? { align } : {}) },
+        content: [{ type: 'paragraph', content: [{ type: 'text', text: t }] }],
+      }));
+    return alignsArb.chain((aligns) => {
+      const headerRow = fc
+        .tuple(...aligns.map((a) => cell(true, a)))
+        .map((cells) => ({ type: 'tableRow', content: cells }));
+      const bodyRow = fc
+        .tuple(...aligns.map((a) => cell(false, a)))
+        .map((cells) => ({ type: 'tableRow', content: cells }));
+      return fc
+        .tuple(headerRow, fc.array(bodyRow, { minLength: 1, maxLength: 2 }))
+        .map(([h, body]) => ({ type: 'table', content: [h, ...body] }));
+    });
+  });
+}
+
+/** A columns block: layout ↔ count coupled, each column body walked as blocks. */
+function columnsArb(depth: number, budget: number): fc.Arbitrary<any> {
+  return fc
+    .constantFrom('two_equal', 'three_equal', 'left_sidebar', 'right_sidebar')
+    .chain((layout) => {
+      const count = layout === 'three_equal' ? 3 : 2;
+      const columnType: any = schema.nodes.column;
+      // Split the remaining budget across the fixed number of columns.
+      const per = Math.max(2, Math.floor((budget - 1) / count));
+      return nodeAttrsArb('columns', 'p1', { layout, widthMode: 'normal' }).chain((attrs) =>
+        fc
+          .tuple(
+            ...Array.from({ length: count }, () =>
+              fillMatch(columnType.contentMatch, depth + 1, per, 'column').map(({ children }) => ({
+                type: 'column',
+                content: children,
+              })),
+            ),
+          )
+          .map((cols) => ({ type: 'columns', attrs, content: cols })),
+      );
+    });
+}
+
+// ---------------------------------------------------------------------------
+// The ContentMatch walk.
+// ---------------------------------------------------------------------------
+
+/** Count every node in a subtree (block + inline), for budget accounting. */
+function countNodes(node: any): number {
+  let n = 1;
+  for (const c of node.content ?? []) n += countNodes(c);
+  return n;
+}
+
+/** Build a single child node of a given type at `depth`, within `budget`. */
+function blockNode(typeName: string, depth: number, budget: number): fc.Arbitrary<any> {
+  if (typeName === 'table') return tableArb();
+  if (typeName === 'columns') return columnsArb(depth, budget);
+  if (isTextblock(typeName)) return textblockArb(typeName);
+  if (isLeaf(typeName)) return leafArb(typeName);
+  // Generic container: attrs from 'p1', block content from the automaton walk.
+  const nt: any = schema.nodes[typeName];
+  return nodeAttrsArb(typeName, 'p1').chain((attrs) =>
+    fillMatch(nt.contentMatch, depth, budget - 1, typeName).map(({ children }) => {
+      const node: any = { type: typeName };
+      if (Object.keys(attrs).length) node.attrs = attrs;
+      if (children.length) node.content = children;
+      return node;
+    }),
+  );
+}
+
+/**
+ * Fill a container's block content by walking its ContentMatch automaton from
+ * `match`. Returns the children array plus the budget left after them.
+ */
+function fillMatch(
+  match: any,
+  depth: number,
+  budget: number,
+  parentType: string,
+): fc.Arbitrary<{ children: any[]; budget: number }> {
+  const canStop = match.validEnd;
+  // A child lives at depth+1; only pick it if its subtree can complete within
+  // MAX_DEPTH. This headroom rule is what makes the walk deadlock-free.
+  const headroom = MAX_DEPTH - (depth + 1);
+  const edges: { t: string; next: any }[] = [];
+  if (headroom >= 0) {
+    for (let i = 0; i < match.edgeCount; i++) {
+      const edge = match.edge(i);
+      const t = edge.type.name;
+      if (EXCLUDED.has(t) || t === 'text') continue;
+      if (!childAllowedUnder(t, parentType)) continue;
+      if ((MIN_DEPTH[t] ?? Infinity) > headroom) continue;
+      edges.push({ t, next: edge.next });
+    }
+  }
+
+  // Decide the next action: STOP (if allowed) or extend with one more child.
+  // Bias toward stopping when the budget is spent; force a child only when the
+  // match is not yet at a valid end.
+  const pool: { weight: number; arbitrary: fc.Arbitrary<{ t: string; next: any } | null> }[] = [];
+  const canGo = edges.length > 0 && (budget > 0 || !canStop);
+  if (canStop) {
+    // Stop is weighted higher when the budget is low so docs stay bounded.
+    pool.push({ weight: budget > 0 ? 2 : 5, arbitrary: fc.constant(null) });
+  }
+  if (canGo && !(canStop && budget <= 0)) {
+    pool.push({ weight: 3, arbitrary: fc.constantFrom(...edges) });
+  }
+  // Forced continuation: not a valid end yet and (budget exhausted) — must place
+  // a mandatory child regardless of budget.
+  if (pool.length === 0) {
+    if (edges.length > 0) {
+      pool.push({ weight: 1, arbitrary: fc.constantFrom(...edges) });
+    } else {
+      // No legal child and not required to place one: stop with what we have.
+      return fc.constant({ children: [], budget });
+    }
+  }
+
+  return fc.oneof(...pool).chain((choice) => {
+    if (choice === null) return fc.constant({ children: [], budget });
+    return blockNode(choice.t, depth + 1, budget).chain((node) => {
+      const cost = countNodes(node);
+      return fillMatch(choice.next, depth, budget - cost, parentType).map(
+        ({ children, budget: left }) => ({
+          children: [node, ...children],
+          budget: left,
+        }),
+      );
+    });
+  });
+}
+
+/**
+ * The nested-document arbitrary: a valid, arbitrarily-deep ProseMirror doc built
+ * by walking the schema from the document root. Attrs stay in the round-trip-safe
+ * 'p1' space; inline content reuses the byte-stable flat corpus.
+ */
+export const docArb: fc.Arbitrary<any> = fillMatch(
+  schema.nodes.doc.contentMatch,
+  0,
+  NODE_BUDGET,
+  'doc',
+).map(({ children }) => ({ type: 'doc', content: children }));
+
+/** The precomputed minDepth table, exported for inspection/debugging. */
+export { MIN_DEPTH };

packages/prosemirror-markdown/test/generative/flat-roundtrip.property.test.ts

@@ -0,0 +1,284 @@
+import { describe, expect, it, vi } from 'vitest';
+import fc from 'fast-check';
+// Real converter, imported the same way the sibling property test does.
+import { convertProseMirrorToMarkdown } from '../../src/lib/markdown-converter.js';
+// Importing markdownToProseMirror mutates the global DOM via jsdom at module
+// load (expected, required for @tiptap/html's generateJSON under Node).
+import { markdownToProseMirror } from '../../src/lib/markdown-to-prosemirror.js';
+import { docsCanonicallyEqual, canonicalizeContent } from '../../src/lib/index.js';
+import { firstDivergence } from '../roundtrip-helpers.js';
+import {
+  schema,
+  allSchemaAttrKeys,
+  allSchemaMarkAttrKeys,
+  attrIsValueFuzzed,
+} from './attr-arbitraries.js';
+import {
+  buildGenerators,
+  coveredTypes,
+  KNOWN_UNCOVERED,
+} from './node-generators.js';
+
+// ── Attribute-value coverage allowlist ──────────────────────────────────────
+// The node/mark completeness contract guarantees every TYPE is generated, but
+// NOT that every attribute is exercised at a NON-DEFAULT value. An attribute
+// with no `arb` in attr-arbitraries.ts is only ever tested at absent/default —
+// an INVISIBLE coverage hole (the reviewer's concern). This allowlist makes that
+// hole EXPLICIT: it is the exact set of attrs deliberately not value-fuzzed, so
+// a NEW attribute (or a newly-frozen one) that lands in this bucket flips the
+// snapshot test red and forces a reviewer to classify it. Each belongs to one of:
+//   - internal/opaque ids & placeholders (attachmentId, slugId, placeholder,
+//     creatorId, anchorId) — no meaningful non-default to assert;
+//   - dimensions/among the media family with no standalone md form here
+//     (aspectRatio, size, caption, drawio/excalidraw/pdf/video/youtube w/h/align)
+//     — round-trip candidates deferred to a later PR, not silently dropped;
+//   - ACCEPTED limitations with no md representation (indent, callout.icon,
+//     orderedList.type, table spans/bg/colwidth);
+//   - PINNED bugs (column.width, orderedList.start) tracked in
+//     counterexamples.test.ts.
+const ATTR_VALUE_FUZZ_ALLOWLIST = new Set<string>([
+  'attachment.attachmentId', 'attachment.mime', 'attachment.placeholder', 'attachment.size',
+  'audio.attachmentId', 'audio.placeholder', 'audio.size',
+  'callout.icon', 'column.width',
+  'drawio.align', 'drawio.alt', 'drawio.aspectRatio', 'drawio.attachmentId',
+  'drawio.height', 'drawio.size', 'drawio.title', 'drawio.width',
+  'embed.align', 'embed.height', 'embed.width',
+  'excalidraw.align', 'excalidraw.alt', 'excalidraw.aspectRatio', 'excalidraw.attachmentId',
+  'excalidraw.height', 'excalidraw.size', 'excalidraw.title', 'excalidraw.width',
+  'heading.indent',
+  'image.aspectRatio', 'image.attachmentId', 'image.caption', 'image.placeholder', 'image.size',
+  'mention.anchorId', 'mention.creatorId', 'mention.slugId',
+  'orderedList.start', 'orderedList.type', 'paragraph.indent',
+  'pdf.attachmentId', 'pdf.height', 'pdf.placeholder', 'pdf.size', 'pdf.width',
+  'tableCell.backgroundColor', 'tableCell.backgroundColorName', 'tableCell.colspan',
+  'tableCell.colwidth', 'tableCell.rowspan',
+  'tableHeader.backgroundColor', 'tableHeader.backgroundColorName', 'tableHeader.colspan',
+  'tableHeader.colwidth', 'tableHeader.rowspan',
+  'video.align', 'video.aspectRatio', 'video.attachmentId', 'video.placeholder', 'video.size',
+  'youtube.align', 'youtube.height', 'youtube.width',
+]);
+
+// ── MARK attribute-value coverage ───────────────────────────────────────────
+// Marks are fuzzed by the text generator (text-arbitraries.ts markedTextRunArb),
+// not the node OVERRIDES table, so their value-fuzz coverage is tracked with this
+// separate registry — otherwise the "no invisible coverage hole" guarantee would
+// hold for node attrs only, and a new mark attr (or a new attributed mark) would
+// silently escape the fuzz set. Every schema mark attr must be in exactly one of:
+//   MARK_ATTR_FUZZED    — actually driven at a non-default value by the generator;
+//   MARK_ATTR_ALLOWLIST — deliberately not value-fuzzed, with a reason.
+const MARK_ATTR_FUZZED = new Set<string>([
+  'mark:link.href', // markedTextRunArb sets a random webUrl href
+  'mark:link.title', // ...and an optional letter-bearing title
+  'mark:highlight.color', // highlight mark carries a generated color
+  'mark:textStyle.color', // textStyle mark carries a generated color
+  'mark:comment.commentId', // comment anchor id (alphanumeric token)
+  'mark:comment.resolved', // comment resolved flag (rides only when true)
+]);
+const MARK_ATTR_ALLOWLIST = new Set<string>([
+  // link presentational/routing attrs: not part of the markdown link surface the
+  // converter emits (it round-trips href + title only), so there is no
+  // non-default value to assert here — a deferred concern for a link-specific
+  // fixture, not the flat generative pass.
+  'mark:link.internal',
+  'mark:link.target',
+  'mark:link.rel',
+  'mark:link.class',
+]);
+
+// Each run does a real convert + marked + jsdom parse (~ms). Give ample headroom
+// so the suite is deterministic regardless of parallel worker load (like the
+// sibling property file).
+vi.setConfig({ testTimeout: 30000 });
+
+// ---------------------------------------------------------------------------
+// #351 PR 1 — GENERATIVE (property-based) round-trip over FLAT (single-node)
+// documents at the ATTRIBUTE level.
+//
+// We assert three invariants for ANY generated valid flat document `d`
+// (pmToMd = convertProseMirrorToMarkdown, mdToPm = markdownToProseMirror):
+//
+//   P1 — semantic round-trip (nothing lost):
+//        docsCanonicallyEqual(await mdToPm(pmToMd(d)), d) === true
+//   P2 — byte fixpoint (anti "GS-EDIT-REVERT" churn):
+//        pmToMd(await mdToPm(pmToMd(d))) === pmToMd(d)
+//   P3 — totality: neither converter throws; bounded.
+//
+// The generators are schema-DERIVED (attribute lists come from
+// schema.nodes[type].spec.attrs) and stay inside the round-trip-supported space
+// proven empirically by probing the live converter (see attr-arbitraries.ts and
+// text-arbitraries.ts). P1 runs over the safe attribute space; P2/P3 run over
+// the wider 'fuzz' space that also injects degenerate attribute states, which
+// P2 tolerates via a one-time first-pass normalization and P3 via totality only.
+// ---------------------------------------------------------------------------
+
+// Fixed seed so every failure is reproducible; fast-check also prints the
+// shrunk counterexample. numRuns starts modest to keep CI under budget — the
+// issue's CI target is ~300-500 per property; the nightly / PR 3 will crank
+// this up further. Each property runs over the UNION (fc.oneof) of all flat
+// node generators, so the runs are shared across node types (one test per
+// property keeps the jsdom import cost and memory bounded — a per-generator ×
+// per-property matrix is ~200 heavy tests that OOMs the worker).
+const SEED = 20250705;
+const NUM_RUNS = 300;
+
+const P1_GENERATORS = buildGenerators('p1');
+const FUZZ_GENERATORS = buildGenerators('fuzz');
+
+// Union arbitraries: a single draw picks one node generator, then a document
+// from it. On failure fast-check prints the shrunk counterexample doc, which
+// names the offending node type directly.
+const p1Union = fc.oneof(...P1_GENERATORS.map((g) => g.arb));
+const fuzzUnion = fc.oneof(...FUZZ_GENERATORS.map((g) => g.arb));
+
+async function roundTrip(doc: unknown): Promise<{ md1: string; md2: string; doc2: any }> {
+  const md1 = convertProseMirrorToMarkdown(doc);
+  const doc2 = await markdownToProseMirror(md1);
+  const md2 = convertProseMirrorToMarkdown(doc2);
+  return { md1, md2, doc2 };
+}
+
+describe('#351 flat generative round-trip — completeness contract', () => {
+  it('every schema node and mark is covered by a generator or explicitly allowlisted', () => {
+    const covered = coveredTypes();
+    const uncovered: string[] = [];
+
+    for (const nodeType of Object.keys(schema.nodes)) {
+      if (covered.has(nodeType)) continue;
+      if (nodeType in KNOWN_UNCOVERED) continue;
+      uncovered.push(`node:${nodeType}`);
+    }
+    for (const markType of Object.keys(schema.marks)) {
+      if (covered.has(`mark:${markType}`)) continue;
+      if (markType in KNOWN_UNCOVERED) continue;
+      uncovered.push(`mark:${markType}`);
+    }
+
+    // A new node/mark added to the schema with no generator AND no allowlist
+    // entry MUST turn this test red — that is the whole point (no silent blind
+    // spots).
+    expect(
+      uncovered,
+      `these schema types have no generator and no KNOWN_UNCOVERED reason:\n  ${uncovered.join(
+        '\n  ',
+      )}`,
+    ).toEqual([]);
+  });
+
+  it('every KNOWN_UNCOVERED entry is a real schema type (no stale allowlist rows)', () => {
+    const all = new Set([...Object.keys(schema.nodes), ...Object.keys(schema.marks)]);
+    for (const t of Object.keys(KNOWN_UNCOVERED)) {
+      expect(all.has(t), `stale KNOWN_UNCOVERED entry: ${t}`).toBe(true);
+    }
+  });
+
+  it('every attribute is value-fuzzed OR explicitly allowlisted (no invisible hole)', () => {
+    // Makes the "generic-frozen" coverage hole VISIBLE: any schema attr not
+    // exercised at a non-default value must be a KNOWN entry in the allowlist.
+    // A new attr (or one that loses its `arb`) that falls into the not-fuzzed
+    // bucket without an allowlist row turns this red — no silent blind spots.
+    const unaccounted: string[] = [];
+    for (const key of allSchemaAttrKeys()) {
+      const i = key.indexOf('.');
+      const fuzzed = attrIsValueFuzzed(key.slice(0, i), key.slice(i + 1));
+      if (!fuzzed && !ATTR_VALUE_FUZZ_ALLOWLIST.has(key)) unaccounted.push(key);
+    }
+    expect(
+      unaccounted,
+      `these attrs are not value-fuzzed and not in ATTR_VALUE_FUZZ_ALLOWLIST:\n  ${unaccounted.join(
+        '\n  ',
+      )}`,
+    ).toEqual([]);
+  });
+
+  it('the attribute allowlist has no stale rows (every entry is really not-fuzzed)', () => {
+    const notFuzzed = new Set(
+      allSchemaAttrKeys().filter((key) => {
+        const i = key.indexOf('.');
+        return !attrIsValueFuzzed(key.slice(0, i), key.slice(i + 1));
+      }),
+    );
+    for (const key of ATTR_VALUE_FUZZ_ALLOWLIST) {
+      expect(
+        notFuzzed.has(key),
+        `stale allowlist row (attr is now value-fuzzed, remove it): ${key}`,
+      ).toBe(true);
+    }
+  });
+
+  it('every MARK attribute is value-fuzzed OR allowlisted (no invisible hole)', () => {
+    // The node guard above covers node attrs; marks are fuzzed by the text
+    // generator, so their coverage is tracked separately. A new mark attr (or a
+    // newly-attributed mark) that lands in neither set turns this red.
+    const unaccounted: string[] = [];
+    for (const key of allSchemaMarkAttrKeys()) {
+      if (!MARK_ATTR_FUZZED.has(key) && !MARK_ATTR_ALLOWLIST.has(key)) {
+        unaccounted.push(key);
+      }
+    }
+    expect(
+      unaccounted,
+      `these mark attrs are neither in MARK_ATTR_FUZZED nor MARK_ATTR_ALLOWLIST:\n  ${unaccounted.join(
+        '\n  ',
+      )}`,
+    ).toEqual([]);
+  });
+
+  it('the MARK fuzz/allowlist sets have no stale rows (every entry is a real schema mark attr)', () => {
+    const all = new Set(allSchemaMarkAttrKeys());
+    for (const key of [...MARK_ATTR_FUZZED, ...MARK_ATTR_ALLOWLIST]) {
+      expect(all.has(key), `stale mark-attr registry row: ${key}`).toBe(true);
+    }
+  });
+});
+
+describe('#351 flat generative round-trip — properties', () => {
+  it('generator validity: every generated doc passes schema.check()', () => {
+    // A generator that emits an invalid ProseMirror document is a GENERATOR bug.
+    fc.assert(
+      fc.property(fuzzUnion, (doc) => {
+        schema.nodeFromJSON(doc).check(); // throws on an invalid doc
+        return true;
+      }),
+      { numRuns: NUM_RUNS, seed: SEED },
+    );
+  });
+
+  it('P1 — semantic round-trip: docsCanonicallyEqual(mdToPm(pmToMd(d)), d)', async () => {
+    await fc.assert(
+      fc.asyncProperty(p1Union, async (doc) => {
+        const { doc2 } = await roundTrip(doc);
+        if (!docsCanonicallyEqual(doc2, doc)) {
+          // Surface the precise divergence in the failure message.
+          const div = firstDivergence(
+            JSON.parse(JSON.stringify(canonicalizeContent(doc2))),
+            JSON.parse(JSON.stringify(canonicalizeContent(doc))),
+          );
+          throw new Error(
+            `P1 divergence @ ${div?.path}: got=${JSON.stringify(div?.a)} want=${JSON.stringify(div?.b)}`,
+          );
+        }
+      }),
+      { numRuns: NUM_RUNS, seed: SEED },
+    );
+  });
+
+  it('P2 — byte fixpoint: pmToMd(mdToPm(pmToMd(d))) === pmToMd(d)', async () => {
+    await fc.assert(
+      fc.asyncProperty(fuzzUnion, async (doc) => {
+        const { md1, md2 } = await roundTrip(doc);
+        expect(md2).toBe(md1);
+      }),
+      { numRuns: NUM_RUNS, seed: SEED },
+    );
+  });
+
+  it('P3 — totality: neither converter throws', async () => {
+    await fc.assert(
+      fc.asyncProperty(fuzzUnion, async (doc) => {
+        // Throwing here fails the property; fast-check shrinks to a minimal doc.
+        await roundTrip(doc);
+      }),
+      { numRuns: NUM_RUNS, seed: SEED },
+    );
+  });
+});

packages/prosemirror-markdown/test/generative/nested-roundtrip.property.test.ts

@@ -0,0 +1,185 @@
+import { describe, expect, it, vi } from 'vitest';
+import fc from 'fast-check';
+// Real converters. Importing markdownToProseMirror (transitively, via index)
+// mutates the global DOM via jsdom at module load — expected, required for
+// @tiptap/html's generateJSON under Node (same as the flat sibling suite).
+import {
+  convertProseMirrorToMarkdown,
+  markdownToProseMirror,
+  docsCanonicallyEqual,
+  canonicalizeContent,
+} from '../../src/lib/index.js';
+import { firstDivergence } from '../roundtrip-helpers.js';
+import { schema, docArb } from './doc-generator.js';
+
+// Each run does a real convert + jsdom parse; give ample headroom so the suite
+// is deterministic under parallel worker load (matching the flat sibling suite).
+vi.setConfig({ testTimeout: 60000 });
+
+// ---------------------------------------------------------------------------
+// #351 PR 2 — GENERATIVE round-trip over NESTED (whole-document) docs produced
+// by the ContentMatch random walk (doc-generator.ts). The invariants mirror the
+// flat suite, plus a parser-fuzz totality property (P4):
+//
+//   P1 — semantic round-trip: docsCanonicallyEqual(mdToPm(pmToMd(d)), d)
+//   P2 — byte fixpoint (2nd pass): pmToMd(mdToPm(pmToMd(d))) === pmToMd(d)
+//        (the FIRST pass may normalize once; the SECOND pass must be a fixpoint)
+//   P3 — totality: neither converter throws; bounded.
+//   P4 — parser fuzz totality: for ANY string, markdownToProseMirror does NOT
+//        throw and returns a SCHEMA-VALID document.
+//
+// GUARDRAIL: a P1/P2/P3/P4 failure means the generator FOUND A REAL CONVERTER
+// BUG. These invariants are kept STRICT — no it.fails / skip / weakening. A
+// failure prints the shrunk minimal counterexample for triage.
+// ---------------------------------------------------------------------------
+
+const SEED = 20250705;
+// The nested walk builds far heavier docs than the flat suite (each P1/P2 run
+// parses the emitted markdown through jsdom), so keep the run count moderate to
+// hold runtime and worker memory in budget while still exercising deep
+// structures. P4 (cheap string parsing) runs at a higher count below.
+const NUM_RUNS = 100;
+
+const pmToMd = (doc: unknown): string => convertProseMirrorToMarkdown(doc);
+const mdToPm = (md: string): Promise<any> => markdownToProseMirror(md);
+
+async function roundTrip(doc: unknown): Promise<{ md1: string; md2: string; doc2: any }> {
+  const md1 = pmToMd(doc);
+  const doc2 = await mdToPm(md1);
+  const md2 = pmToMd(doc2);
+  return { md1, md2, doc2 };
+}
+
+describe('#351 nested generative round-trip — generator validity', () => {
+  it('every generated nested doc passes schema.nodeFromJSON(...).check()', () => {
+    // A nested generator that emits an invalid ProseMirror document is a
+    // GENERATOR bug — the ContentMatch walk must only produce schema-valid docs.
+    fc.assert(
+      fc.property(docArb, (doc) => {
+        schema.nodeFromJSON(doc).check(); // throws on an invalid doc
+        return true;
+      }),
+      { numRuns: NUM_RUNS * 2, seed: SEED },
+    );
+  });
+});
+
+// ── STATUS: P1/P2/P3/P4 all GREEN. The nested generator originally surfaced a
+// batch of real converter bugs; all were fixed in the serializer/parser (see the
+// #351 hand-off). For the record, the classes it found and that are now fixed:
+//   • Loose (multi-block) list items / task items / callouts / details bodies were
+//     joined with a single "\n", so every block after the first merged into the
+//     first paragraph on re-parse (silent content loss) — now blank-line separated.
+//   • Paragraph `textAlign` was dropped inside a TIGHT list item (no <p> host).
+//   • A nested codeBlock lost its trailing newline on the raw-HTML path.
+//   • Media (embed/video/youtube/drawio/excalidraw) inside `columns` churned a
+//     default `data-align` and coerced embed's numeric width/height to strings.
+//   • pageBreak / pageEmbed / subpages / transclusion were dropped when nested in
+//     blockquote / callout / details / list item (standalone-comment position).
+//   • Callouts nested in a list item or a blockquote (`  > [!type]` / `> > [!type]`)
+//     were re-parsed as plain blockquotes (prefix-unaware callout preprocessor).
+//   • Two adjacent sibling lists sharing a marker family (bulletList/taskList →
+//     `<ul>`; orderedList → `<ol>`) merged into one list on re-parse — and for the
+//     cross-type case (taskList beside bulletList) the merged `<ul>` LOST every
+//     taskItem checkbox. The serializer now emits a `<!-- -->` separator between
+//     such adjacent lists (markdown-converter.ts renderBlockChildren), so they stay
+//     distinct and round-trip; the generator therefore emits them freely again.
+describe('#351 nested generative round-trip — properties', () => {
+  it('P1 — semantic round-trip: docsCanonicallyEqual(mdToPm(pmToMd(d)), d)', async () => {
+    await fc.assert(
+      fc.asyncProperty(docArb, async (doc) => {
+        const { doc2 } = await roundTrip(doc);
+        if (!docsCanonicallyEqual(doc2, doc)) {
+          const div = firstDivergence(
+            JSON.parse(JSON.stringify(canonicalizeContent(doc2))),
+            JSON.parse(JSON.stringify(canonicalizeContent(doc))),
+          );
+          throw new Error(
+            `P1 divergence @ ${div?.path}: got=${JSON.stringify(div?.a)} want=${JSON.stringify(div?.b)}`,
+          );
+        }
+      }),
+      { numRuns: NUM_RUNS, seed: SEED },
+    );
+  });
+
+  it('P2 — byte fixpoint: pmToMd(mdToPm(pmToMd(d))) === pmToMd(d)', async () => {
+    await fc.assert(
+      fc.asyncProperty(docArb, async (doc) => {
+        const { md1, md2 } = await roundTrip(doc);
+        expect(md2).toBe(md1);
+      }),
+      { numRuns: NUM_RUNS, seed: SEED },
+    );
+  });
+
+  it('P3 — totality: neither converter throws', async () => {
+    await fc.assert(
+      fc.asyncProperty(docArb, async (doc) => {
+        await roundTrip(doc);
+      }),
+      { numRuns: NUM_RUNS, seed: SEED },
+    );
+  });
+});
+
+// ---------------------------------------------------------------------------
+// P4 — parser fuzz. Independent of the doc generator: for ANY input string the
+// PARSER (markdownToProseMirror) must be TOTAL — never throw — and must always
+// return a schema-valid document. The corpus mixes raw unicode strings with
+// strings assembled from markdown-significant fragments (headings, list bullets,
+// fences, pipes, thematic breaks, HTML-ish snippets) to probe the block/inline
+// parsers on hostile but plausible input.
+// ---------------------------------------------------------------------------
+
+const mdFragmentArb: fc.Arbitrary<string> = fc.constantFrom(
+  '# ', '## ', '### ###', '- ', '* ', '+ ', '1. ', '> ', '>> ',
+  '```', '```js', '~~~', '---', '***', '___', '| a | b |', '|---|---|',
+  '[link](http://x)', '![img](http://x)', '**', '__', '~~', '`code`',
+  '<div>', '</div>', '<b>', '<!-- c -->', '<table>', '<br>', '&amp;',
+  '\t', '\n', '  ', '\\', '^[fn]', '[^1]:', '- [ ] ', '- [x] ',
+  '$$', '$x$', ':::', '{.class}', '\u0000', '\uFEFF', '😀', 'مرحبا',
+);
+
+// Full-unicode strings (fast-check v4 replaced fullUnicodeString with the
+// `unit: 'binary'` string option, which draws over the whole code-point range).
+const fullUnicodeStringArb = (max?: number) =>
+  fc.string({ unit: 'binary', ...(max !== undefined ? { maxLength: max } : {}) });
+
+const assembledMarkdownArb: fc.Arbitrary<string> = fc
+  .array(fc.oneof(mdFragmentArb, fc.string(), fullUnicodeStringArb(8)), {
+    minLength: 1,
+    maxLength: 12,
+  })
+  .map((parts) => parts.join(''));
+
+const parserInputArb: fc.Arbitrary<string> = fc.oneof(
+  { weight: 2, arbitrary: fc.string() },
+  { weight: 2, arbitrary: fullUnicodeStringArb() },
+  { weight: 3, arbitrary: assembledMarkdownArb },
+  { weight: 1, arbitrary: fc.array(mdFragmentArb, { minLength: 1, maxLength: 8 }).map((p) => p.join('\n')) },
+);
+
+describe('#351 parser fuzz — totality on arbitrary input (P4)', () => {
+  it('P4 — markdownToProseMirror never throws and always returns a schema-valid doc', async () => {
+    await fc.assert(
+      fc.asyncProperty(parserInputArb, async (s) => {
+        let result: any;
+        try {
+          result = await mdToPm(s);
+        } catch (e: any) {
+          throw new Error(`P4 parser THREW on input ${JSON.stringify(s)}: ${e?.message ?? e}`);
+        }
+        try {
+          schema.nodeFromJSON(result).check();
+        } catch (e: any) {
+          throw new Error(
+            `P4 parser produced an INVALID doc for input ${JSON.stringify(s)}: ${e?.message ?? e}\n` +
+              `doc=${JSON.stringify(result).slice(0, 600)}`,
+          );
+        }
+      }),
+      { numRuns: NUM_RUNS * 2, seed: SEED },
+    );
+  });
+});

packages/prosemirror-markdown/test/generative/node-generators.ts

@@ -0,0 +1,310 @@
+/**
+ * Flat single-node document generators (#351, PR 1).
+ *
+ * For every schema node type that can stand alone, a fast-check arbitrary
+ * producing `{ type:'doc', content:[ <the target node> ] }` with generated attrs
+ * (via nodeAttrsArb) and the minimal REQUIRED immediate children the schema
+ * demands (a heading's inline text, a listItem's one paragraph, a table's
+ * minimal rows, details' summary+content, a callout's one paragraph). Kept
+ * FLAT: a single target node, no deep nesting — nested structural generation is
+ * PR 2.
+ *
+ * The `mode` threads through to the attribute arbitraries:
+ *   - 'p1'   : the round-trip-safe attribute space (P1 semantic round-trip).
+ *   - 'fuzz' : adds degenerate attribute states (P2 byte-fixpoint tolerates the
+ *              one-time normalization; P3 only needs totality).
+ *
+ * A COMPLETENESS CONTRACT (see flat-roundtrip.property.test.ts) enumerates the
+ * whole schema and asserts every node/mark is EITHER produced by a generator
+ * here OR listed in KNOWN_UNCOVERED with a reason — so a new schema type with no
+ * generator turns the suite RED.
+ */
+import fc from 'fast-check';
+import { type AttrMode, nodeAttrsArb } from './attr-arbitraries.js';
+import {
+  inlineContentArb,
+  headingInlineContentArb,
+  plainInlineContentArb,
+  phraseArb,
+  markedTextRunArb,
+} from './text-arbitraries.js';
+
+const doc = (node: any) => ({ type: 'doc', content: [node] });
+const para = (content: any[]) => ({ type: 'paragraph', content });
+
+/** A named flat-document generator. */
+export interface NamedGen {
+  name: string;
+  arb: fc.Arbitrary<any>;
+}
+
+// ---------------------------------------------------------------------------
+// Per-target generators, each a function of mode.
+// ---------------------------------------------------------------------------
+
+const gen = {
+  paragraph: (m: AttrMode) =>
+    fc.tuple(nodeAttrsArb('paragraph', m), inlineContentArb).map(([attrs, content]) =>
+      doc({ type: 'paragraph', attrs, content }),
+    ),
+
+  heading: (m: AttrMode) =>
+    fc.tuple(nodeAttrsArb('heading', m), headingInlineContentArb).map(([attrs, content]) =>
+      doc({ type: 'heading', attrs, content }),
+    ),
+
+  blockquote: (_m: AttrMode) =>
+    inlineContentArb.map((content) => doc({ type: 'blockquote', content: [para(content)] })),
+
+  bulletList: (_m: AttrMode) =>
+    fc
+      .array(inlineContentArb, { minLength: 1, maxLength: 3 })
+      .map((items) =>
+        doc({
+          type: 'bulletList',
+          content: items.map((c) => ({ type: 'listItem', content: [para(c)] })),
+        }),
+      ),
+
+  orderedList: (m: AttrMode) =>
+    fc
+      .tuple(nodeAttrsArb('orderedList', m), fc.array(inlineContentArb, { minLength: 1, maxLength: 3 }))
+      .map(([attrs, items]) =>
+        doc({
+          type: 'orderedList',
+          attrs,
+          content: items.map((c) => ({ type: 'listItem', content: [para(c)] })),
+        }),
+      ),
+
+  taskList: (m: AttrMode) =>
+    fc
+      .array(fc.tuple(nodeAttrsArb('taskItem', m), inlineContentArb), { minLength: 1, maxLength: 3 })
+      .map((items) =>
+        doc({
+          type: 'taskList',
+          content: items.map(([attrs, c]) => ({ type: 'taskItem', attrs, content: [para(c)] })),
+        }),
+      ),
+
+  codeBlock: (m: AttrMode) =>
+    fc
+      .tuple(
+        nodeAttrsArb('codeBlock', m),
+        // A fenced code block always re-imports with a TRAILING NEWLINE in its
+        // text (empirically confirmed). Author the newline so the doc is already
+        // at the round-trip fixpoint (supported-space shaping, not a masked bug).
+        fc.array(phraseArb, { minLength: 1, maxLength: 3 }).map((lines) => lines.join('\n') + '\n'),
+      )
+      .map(([attrs, code]) =>
+        doc({ type: 'codeBlock', attrs, content: [{ type: 'text', text: code }] }),
+      ),
+
+  horizontalRule: (_m: AttrMode) => fc.constant(doc({ type: 'horizontalRule' })),
+
+  pageBreak: (_m: AttrMode) => fc.constant(doc({ type: 'pageBreak' })),
+
+  image: (m: AttrMode) => nodeAttrsArb('image', m).map((attrs) => doc({ type: 'image', attrs })),
+
+  callout: (m: AttrMode) =>
+    fc.tuple(nodeAttrsArb('callout', m), inlineContentArb).map(([attrs, content]) =>
+      doc({ type: 'callout', attrs, content: [para(content)] }),
+    ),
+
+  mathBlock: (m: AttrMode) =>
+    nodeAttrsArb('mathBlock', m).map((attrs) => doc({ type: 'mathBlock', attrs })),
+
+  details: (m: AttrMode) =>
+    fc
+      .tuple(nodeAttrsArb('details', m), plainInlineContentArb, inlineContentArb)
+      .map(([attrs, summary, body]) =>
+        doc({
+          type: 'details',
+          attrs,
+          content: [
+            { type: 'detailsSummary', content: summary },
+            { type: 'detailsContent', content: [para(body)] },
+          ],
+        }),
+      ),
+
+  table: (_m: AttrMode) =>
+    fc.integer({ min: 1, max: 3 }).chain((cols) => {
+      // GFM alignment is column-wide (encoded in the header separator), so a
+      // column's alignment must be identical on the header and every body cell,
+      // else the second export re-aligns and churns. Pick ONE align per column.
+      const alignsArb = fc.array(fc.constantFrom(undefined, 'left', 'center', 'right'), {
+        minLength: cols,
+        maxLength: cols,
+      });
+      const cell = (header: boolean, align?: string) =>
+        phraseArb.map((t) => ({
+          type: header ? 'tableHeader' : 'tableCell',
+          // colspan/rowspan pinned to 1 (GFM cannot express spans); optional
+          // column-consistent align.
+          attrs: { colspan: 1, rowspan: 1, ...(align ? { align } : {}) },
+          content: [para([{ type: 'text', text: t }])],
+        }));
+      return alignsArb.chain((aligns) => {
+        const headerRow = fc
+          .tuple(...aligns.map((a) => cell(true, a)))
+          .map((cells) => ({ type: 'tableRow', content: cells }));
+        const bodyRow = fc
+          .tuple(...aligns.map((a) => cell(false, a)))
+          .map((cells) => ({ type: 'tableRow', content: cells }));
+        return fc
+          .tuple(headerRow, fc.array(bodyRow, { minLength: 1, maxLength: 2 }))
+          .map(([h, body]) => doc({ type: 'table', content: [h, ...body] }));
+      });
+    }),
+
+  columns: (m: AttrMode) =>
+    // Couple the column count to the layout so the two stay consistent
+    // (two_equal/left_sidebar/right_sidebar -> 2, three_equal -> 3).
+    fc
+      .constantFrom('two_equal', 'three_equal', 'left_sidebar', 'right_sidebar')
+      .chain((layout) => {
+        const count = layout === 'three_equal' ? 3 : 2;
+        return fc
+          .tuple(
+            nodeAttrsArb('columns', m, { layout, widthMode: 'normal' }),
+            fc.array(inlineContentArb, { minLength: count, maxLength: count }),
+          )
+          .map(([attrs, bodies]) =>
+            doc({
+              type: 'columns',
+              attrs,
+              content: bodies.map((c) => ({ type: 'column', content: [para(c)] })),
+            }),
+          );
+      }),
+
+  subpages: (m: AttrMode) =>
+    nodeAttrsArb('subpages', m).map((attrs) => doc({ type: 'subpages', attrs })),
+
+  audio: (m: AttrMode) => nodeAttrsArb('audio', m).map((attrs) => doc({ type: 'audio', attrs })),
+  video: (m: AttrMode) => nodeAttrsArb('video', m).map((attrs) => doc({ type: 'video', attrs })),
+  pdf: (m: AttrMode) => nodeAttrsArb('pdf', m).map((attrs) => doc({ type: 'pdf', attrs })),
+  youtube: (m: AttrMode) => nodeAttrsArb('youtube', m).map((attrs) => doc({ type: 'youtube', attrs })),
+  embed: (m: AttrMode) => nodeAttrsArb('embed', m).map((attrs) => doc({ type: 'embed', attrs })),
+  drawio: (m: AttrMode) => nodeAttrsArb('drawio', m).map((attrs) => doc({ type: 'drawio', attrs })),
+  excalidraw: (m: AttrMode) =>
+    nodeAttrsArb('excalidraw', m).map((attrs) => doc({ type: 'excalidraw', attrs })),
+  attachment: (m: AttrMode) =>
+    nodeAttrsArb('attachment', m).map((attrs) => doc({ type: 'attachment', attrs })),
+  htmlEmbed: (m: AttrMode) =>
+    nodeAttrsArb('htmlEmbed', m).map((attrs) => doc({ type: 'htmlEmbed', attrs })),
+  pageEmbed: (m: AttrMode) =>
+    nodeAttrsArb('pageEmbed', m).map((attrs) => doc({ type: 'pageEmbed', attrs })),
+  transclusionReference: (m: AttrMode) =>
+    nodeAttrsArb('transclusionReference', m).map((attrs) =>
+      doc({ type: 'transclusionReference', attrs }),
+    ),
+
+  transclusionSource: (m: AttrMode) =>
+    fc.tuple(nodeAttrsArb('transclusionSource', m), inlineContentArb).map(([attrs, content]) =>
+      doc({ type: 'transclusionSource', attrs, content: [para(content)] }),
+    ),
+
+  // A footnote reference PLUS its definition (the reference has no standalone
+  // markdown form without its definition — see KNOWN_UNCOVERED note for the
+  // bare reference). Both carry the same id. The definition body uses
+  // headingInlineContentArb (NO hard breaks): a footnote is serialized inline as
+  // `^[...]`, so a hard break inside it collapses to a single space on re-parse
+  // (empirically confirmed) — that is the container's markdown limitation, not
+  // an attribute-level concern. The reference-bearing paragraph is a NORMAL
+  // paragraph and keeps the full inline corpus.
+  footnotes: (m: AttrMode) =>
+    fc.tuple(fc.constantFrom('fn1', 'fn2', 'note'), inlineContentArb, headingInlineContentArb).map(
+      ([id, refText, noteBody]) => ({
+        type: 'doc',
+        content: [
+          para([...refText, { type: 'footnoteReference', attrs: { id } }]),
+          {
+            type: 'footnotesList',
+            content: [{ type: 'footnoteDefinition', attrs: { id }, content: [para(noteBody)] }],
+          },
+        ],
+      }),
+    ),
+
+  // ── inline targets wrapped in a paragraph ────────────────────────────────
+  mention: (m: AttrMode) =>
+    nodeAttrsArb('mention', m).map((attrs) => doc(para([{ type: 'mention', attrs }]))),
+
+  mathInline: (m: AttrMode) =>
+    fc.tuple(phraseArb, nodeAttrsArb('mathInline', m)).map(([t, attrs]) =>
+      doc(para([{ type: 'text', text: t }, { type: 'mathInline', attrs }])),
+    ),
+
+  status: (m: AttrMode) =>
+    nodeAttrsArb('status', m).map((attrs) => doc(para([{ type: 'status', attrs }]))),
+
+  hardBreak: (_m: AttrMode) =>
+    fc.tuple(phraseArb, phraseArb).map(([a, b]) =>
+      doc(para([{ type: 'text', text: a }, { type: 'hardBreak' }, { type: 'text', text: b }])),
+    ),
+
+  // ── marks: a paragraph of marked runs (covers every mark type) ───────────
+  marksOnText: (_m: AttrMode) =>
+    fc.array(markedTextRunArb, { minLength: 1, maxLength: 5 }).map((runs) => {
+      // Merge adjacent same-mark runs (see text-arbitraries.normalizeInline).
+      const out: any[] = [];
+      for (const r of runs) {
+        const prev = out[out.length - 1];
+        if (prev && JSON.stringify(prev.marks ?? []) === JSON.stringify(r.marks ?? [])) {
+          prev.text += r.text;
+        } else out.push({ ...r });
+      }
+      return doc(para(out));
+    }),
+};
+
+/** Build the full list of named generators for a given mode. */
+export function buildGenerators(mode: AttrMode): NamedGen[] {
+  return Object.entries(gen).map(([name, f]) => ({ name, arb: f(mode) }));
+}
+
+// ---------------------------------------------------------------------------
+// Completeness contract support.
+// ---------------------------------------------------------------------------
+
+/**
+ * Schema node/mark types deliberately NOT covered by a P1/P2 generator, each
+ * with a one-line reason. Excluding a type means it is kept OUT of the round-
+ * trip generators — it does NOT weaken any property.
+ *
+ * NOTE (empirical): the candidates the issue flagged for review — pageEmbed,
+ * subpages, transclusionSource/Reference, mention, status — were PROBED against
+ * the live converter and DO round-trip P1/P2 with placeholder ids, so they are
+ * COVERED by real generators rather than allowlisted here. The allowlist below
+ * holds only types with no standalone flat generator by construction.
+ */
+export const KNOWN_UNCOVERED: Record<string, string> = {
+  // The root node; it is the wrapper every generated doc already is, never a
+  // "target" content node, so it has no standalone generator of its own.
+  doc: 'the document root wrapper, not a content node with a standalone generator',
+};
+
+/** Recursively collect every node type and `mark:<type>` under a tree. */
+export function collectTypes(node: any, seen = new Set<string>()): Set<string> {
+  if (!node || typeof node !== 'object') return seen;
+  if (node.type) seen.add(node.type);
+  for (const m of node.marks ?? []) if (m?.type) seen.add(`mark:${m.type}`);
+  for (const c of node.content ?? []) collectTypes(c, seen);
+  return seen;
+}
+
+/**
+ * Sample every generator and return the union of node/mark types they produce.
+ * Deterministic (fixed seed) so the completeness contract is stable.
+ */
+export function coveredTypes(seed = 12345, perGen = 60): Set<string> {
+  const seen = new Set<string>();
+  for (const { arb } of buildGenerators('p1')) {
+    for (const sample of fc.sample(arb, { numRuns: perGen, seed })) {
+      collectTypes(sample, seen);
+    }
+  }
+  return seen;
+}

packages/prosemirror-markdown/test/generative/text-arbitraries.ts

@@ -0,0 +1,258 @@
+/**
+ * Hostile inline-text corpus for the generative flat-document round-trip suite
+ * (#351, PR 1).
+ *
+ * These arbitraries are a DIRECT PORT of the "supported space" guardrails that
+ * `test/markdown-roundtrip.property.test.ts` proved empirically against the live
+ * converter. That file's long header documents WHY each guardrail exists; rather
+ * than re-derive them, we reuse the exact same shapes here so the attribute-level
+ * generative suite inherits the same byte-stable text space. Each guardrail is
+ * cited back to that file below.
+ *
+ * The corpus deliberately spans the CommonMark / canon hostile alphabet
+ * (`* _ [ ] ( ) { } | < > & # ! ~ = + -`), unicode / emoji / RTL, and the legal
+ * mark combinations on runs (including the `code` mark, which the schema's
+ * `excludes: "_"` makes suppress every co-occurring mark — so it is never
+ * combined with another mark in the byte-stable space).
+ */
+import fc from 'fast-check';
+
+// ---------------------------------------------------------------------------
+// Words and the hostile special-character alphabet.
+// (Ported from markdown-roundtrip.property.test.ts, "Inline text arbitraries".)
+// ---------------------------------------------------------------------------
+
+/** Alphanumeric "word" (no markdown-significant characters). Length 1..6. */
+export const wordArb = fc
+  .stringMatching(/^[A-Za-z0-9]{1,6}$/)
+  .filter((w) => w.length > 0);
+
+/**
+ * A SINGLE markdown-significant character, emitted only as an isolated,
+ * space-flanked token. Every char the task calls out plus a few more; each was
+ * verified byte-stable in this position by the sibling property test.
+ *
+ * NOTE: the backtick (`) is DELIBERATELY excluded from free-floating plain text
+ * (it is a code-span delimiter that re-pairs globally). It is exercised only via
+ * the `code` mark and code blocks — see markdown-roundtrip.property.test.ts.
+ */
+export const specialCharArb = fc.constantFrom(
+  '*', '_', '[', ']', '(', ')', '{', '}', '|', '<', '>', '&', '#', '!', '~', '=', '+', '-',
+);
+
+// A pinch of unicode / emoji / RTL, always word-like (no markdown specials) so
+// it stays inside the space-flanked corpus. Kept letter/emoji-bearing so it is
+// never coerced to a number (see letterPhraseArb rationale).
+export const unicodeWordArb = fc.constantFrom(
+  'café', 'naïve', 'Zürich', 'Москва', 'こんにちは', '你好', '😀', '🚀x', 'مرحبا', 'שלום',
+);
+
+/**
+ * A "safe special" text string: a space-joined sequence of tokens that always
+ * BEGINS and ENDS with an alphanumeric word, with any isolated special chars (or
+ * unicode words) confined to the MIDDLE, each space-flanked by words.
+ *
+ * Both boundary guarantees matter (verbatim from the sibling test):
+ *   * Leading word: the line never opens with a block/inline trigger
+ *     (">", "*", "-", "#", "1." ...).
+ *   * Trailing word: adjacent text runs CONCATENATE with no separator, so a run
+ *     ending in a bare "<" beside a run starting with a letter would form a fake
+ *     HTML tag. Ending every run with a word keeps every special internal and
+ *     space-flanked even after concatenation.
+ */
+export const safeTextArb: fc.Arbitrary<string> = fc
+  .tuple(
+    wordArb,
+    fc.array(fc.oneof(wordArb, specialCharArb, unicodeWordArb), {
+      minLength: 0,
+      maxLength: 3,
+    }),
+    wordArb,
+  )
+  .map(([first, middle, last]) => [first, ...middle, last].join(' '));
+
+/**
+ * A plain alphanumeric phrase (1..3 words) for places where even isolated
+ * specials are not wanted (e.g. code-block language, mention labels, status
+ * text, table cells rendered on the plain-markdown path).
+ */
+export const phraseArb: fc.Arbitrary<string> = fc
+  .array(wordArb, { minLength: 1, maxLength: 3 })
+  .map((ws) => ws.join(' '));
+
+/**
+ * A phrase guaranteed to contain at least one letter. Used for image/media alt
+ * text and link titles: a PURELY numeric alt/title (e.g. "0") is parsed back as
+ * a NUMBER and then dropped by the converter's `value || ""` coercion — not
+ * byte-stable. A letter anywhere keeps it a string. (Ported verbatim.)
+ */
+export const letterPhraseArb: fc.Arbitrary<string> = fc
+  .tuple(
+    fc.stringMatching(/^[A-Za-z]{1,4}$/),
+    fc.array(wordArb, { minLength: 0, maxLength: 2 }),
+  )
+  .map(([head, rest]) => [head, ...rest].join(' '));
+
+/** A paren/space-free URL — safe inside markdown link/image `(...)` syntax. */
+export const urlArb: fc.Arbitrary<string> = fc
+  .webUrl()
+  .filter((u) => !/[()\s]/.test(u));
+
+// ---------------------------------------------------------------------------
+// Marked inline runs.
+// (Ported from markdown-roundtrip.property.test.ts "markedTextRunArb".)
+// ---------------------------------------------------------------------------
+
+/**
+ * A text run with an OPTIONAL single non-code formatting mark (bold/italic/
+ * strike/underline/superscript/subscript/spoiler), or a SOLE `code` mark, or a
+ * link, or an inline comment anchor. `code` is NEVER combined with another mark
+ * in the byte-stable space (that combination is a documented converter
+ * limitation — the schema's `code` mark declares `excludes: "_"`). Marks wrap
+ * `safeTextArb`, which stays stable even when it contains isolated specials.
+ *
+ * The mark set here is broadened past the sibling test's {bold,italic,strike}
+ * to also cover underline / superscript / subscript / spoiler / textStyle /
+ * highlight (all single, non-code marks), so the marks-on-text generator
+ * exercises every mark the schema declares except the deliberately-excluded
+ * `code`+other combination.
+ */
+export const markedTextRunArb: fc.Arbitrary<any> = fc.oneof(
+  // Plain text.
+  safeTextArb.map((t) => ({ type: 'text', text: t })),
+  // Single formatting mark (attribute-free marks).
+  fc
+    .tuple(
+      safeTextArb,
+      fc.constantFrom('bold', 'italic', 'strike', 'underline', 'superscript', 'subscript', 'spoiler'),
+    )
+    .map(([t, m]) => ({ type: 'text', text: t, marks: [{ type: m }] })),
+  // highlight with a color attr.
+  fc
+    .tuple(safeTextArb, fc.constantFrom('#ffcc00', '#a0e0ff', 'yellow'))
+    .map(([t, color]) => ({ type: 'text', text: t, marks: [{ type: 'highlight', attrs: { color } }] })),
+  // textStyle with a color attr.
+  fc
+    .tuple(safeTextArb, fc.constantFrom('#123456', '#ff0000', '#00aa88'))
+    .map(([t, color]) => ({ type: 'text', text: t, marks: [{ type: 'textStyle', attrs: { color } }] })),
+  // Sole code mark (backtick span). safeTextArb is backtick-free, so the span
+  // content cannot contain an inner backtick.
+  safeTextArb.map((t) => ({ type: 'text', text: t, marks: [{ type: 'code' }] })),
+  // Link with safe text, a paren/space-free href, optionally a letter-bearing
+  // title (a purely numeric title is coerced to a number and dropped).
+  fc
+    .tuple(phraseArb, urlArb, fc.option(letterPhraseArb, { nil: undefined }))
+    .map(([t, href, title]) => ({
+      type: 'text',
+      text: t,
+      marks: [{ type: 'link', attrs: title ? { href, title } : { href } }],
+    })),
+  // Inline comment anchor: a span[data-comment-id] that must survive byte-for-
+  // byte. commentId is an alphanumeric token; `resolved` rides only when true.
+  fc
+    .tuple(safeTextArb, fc.stringMatching(/^[A-Za-z0-9]{4,10}$/), fc.boolean())
+    .map(([t, commentId, resolved]) => ({
+      type: 'text',
+      text: t,
+      marks: [
+        { type: 'comment', attrs: resolved ? { commentId, resolved: true } : { commentId } },
+      ],
+    })),
+);
+
+// ---------------------------------------------------------------------------
+// Inline atoms and inline-content assembly.
+// (Ported from markdown-roundtrip.property.test.ts.)
+// ---------------------------------------------------------------------------
+
+/** Inline math node carrying LaTeX that includes the `a < b` the task asks for. */
+export const mathInlineArb: fc.Arbitrary<any> = fc
+  .constantFrom('a < b', 'x^2 + y^2', 'a < b < c', '\\frac{1}{2}', 'E = mc^2')
+  .map((text) => ({ type: 'mathInline', attrs: { text } }));
+
+/** Mention node; label/id/entity are plain phrases / uuids. */
+export const mentionArb: fc.Arbitrary<any> = fc
+  .tuple(phraseArb, fc.uuid(), fc.uuid())
+  .map(([label, id, entityId]) => ({
+    type: 'mention',
+    attrs: { id, label, entityType: 'user', entityId },
+  }));
+
+export const hardBreakArb: fc.Arbitrary<any> = fc.constant({ type: 'hardBreak' });
+
+const sameMarks = (a: any[] | undefined, b: any[] | undefined): boolean =>
+  JSON.stringify(a ?? []) === JSON.stringify(b ?? []);
+
+/**
+ * Canonicalize a generated inline-content array the way ProseMirror stores it,
+ * then trim the markdown-fragile edges. (Ported verbatim from
+ * markdown-roundtrip.property.test.ts "normalizeInline":)
+ *   1) MERGE adjacent text runs with IDENTICAL marks (the editor coalesces
+ *      them; split same-mark runs export to ambiguous "**a****b**").
+ *   2) Collapse CONSECUTIVE hard breaks (two render a blank line marked eats).
+ *   3) Drop a TRAILING hard break (removed by the converter's .trim()).
+ */
+export function normalizeInline(nodes: any[]): any[] {
+  const out: any[] = [];
+  for (const node of nodes) {
+    const prev = out[out.length - 1];
+    if (node.type === 'hardBreak' && prev && prev.type === 'hardBreak') continue;
+    if (
+      node.type === 'text' &&
+      prev &&
+      prev.type === 'text' &&
+      sameMarks(prev.marks, node.marks)
+    ) {
+      prev.text += node.text;
+      continue;
+    }
+    out.push(node.type === 'text' ? { ...node } : node);
+  }
+  while (out.length > 1 && out[out.length - 1].type === 'hardBreak') out.pop();
+  return out;
+}
+
+/**
+ * Inline content for a paragraph: at least one marked text run, optionally with
+ * inline atoms (math/mention) and hard breaks interspersed. Always starts with a
+ * text run so the paragraph never opens with a block trigger. (Ported.)
+ */
+export const inlineContentArb: fc.Arbitrary<any[]> = fc
+  .tuple(
+    markedTextRunArb,
+    fc.array(
+      fc.oneof(
+        { weight: 5, arbitrary: markedTextRunArb },
+        { weight: 1, arbitrary: mathInlineArb },
+        { weight: 1, arbitrary: mentionArb },
+        { weight: 1, arbitrary: hardBreakArb },
+      ),
+      { minLength: 0, maxLength: 4 },
+    ),
+  )
+  .map(([first, rest]) => normalizeInline([first, ...rest]));
+
+/**
+ * Inline content for a HEADING — identical to a paragraph's, but WITHOUT hard
+ * breaks. A hard break inside an ATX heading is not byte-stable (marked splits
+ * the heading). (Ported.)
+ */
+export const headingInlineContentArb: fc.Arbitrary<any[]> = fc
+  .tuple(
+    markedTextRunArb,
+    fc.array(
+      fc.oneof(
+        { weight: 5, arbitrary: markedTextRunArb },
+        { weight: 1, arbitrary: mathInlineArb },
+        { weight: 1, arbitrary: mentionArb },
+      ),
+      { minLength: 0, maxLength: 4 },
+    ),
+  )
+  .map(([first, rest]) => normalizeInline([first, ...rest]));
+
+/** Simple plain-text inline content (single run) for containers rendered on the
+ * raw-HTML path (table cells / column bodies) where fancy inline is undesirable. */
+export const plainInlineContentArb: fc.Arbitrary<any[]> = phraseArb.map((t) => [
+  { type: 'text', text: t },
+]);

packages/prosemirror-markdown/test/gitmost-transcript-neutralization.test.ts

@@ -0,0 +1,124 @@
+import { describe, expect, it } from "vitest";
+// Import DIRECTLY from src (NOT the docmost-client barrel, which pulls in
+// collaboration.ts and mutates global DOM at import time).
+import { convertProseMirrorToMarkdown } from "../src/lib/markdown-converter.js";
+import { markdownToProseMirror } from "../src/lib/markdown-to-prosemirror.js";
+
+/**
+ * gitmost #377 (round-1 review, finding #1) — proof, against the REAL
+ * converter, that the transcript-insert boundary defense survives git-sync.
+ *
+ * The web bridge (apps/client .../gitmost/gitmost-recording.ts,
+ * `gitmostInsertTranscriptIntoEditor`) appends each transcript line as a
+ * PARAGRAPH text node. The paragraph serializer here (`case "paragraph"`) emits
+ * that text VERBATIM with no block-escape, so a line whose text begins with a
+ * col-0 markdown block trigger would, on the doc -> markdown -> doc git-sync
+ * cycle, silently re-parse into a heading / list / quote / callout / code block.
+ * That missing block-escape is the pre-existing root cause; the bridge's
+ * boundary defense prepends an invisible zero-width space (U+200B) to a line
+ * that begins with such a trigger, shifting it off column 0.
+ *
+ * This test keeps a COPY of the bridge's trigger regex (the bridge is in a
+ * different package and can't be imported here) and asserts:
+ *   1. bare trigger lines DO corrupt (documents the root cause), and
+ *   2. the ZWSP-neutralized form round-trips as a single PARAGRAPH with the
+ *      text byte-preserved.
+ */
+
+const ZWSP = "​"; // U+200B
+
+// MUST stay in sync with GITMOST_MD_BLOCK_TRIGGER_RE in the client bridge.
+const MD_BLOCK_TRIGGER_RE =
+  /^(?:#{1,6}(?:\s|$)|[-*+](?:\s|$)|>|\d+[.)](?:\s|$)|```|~~~|\||([-*_])(?:\s*\1){2,}\s*$)/;
+
+const doc = (...nodes: any[]) => ({ type: "doc", content: nodes });
+const para = (t: string) => ({
+  type: "paragraph",
+  content: [{ type: "text", text: t }],
+});
+
+const roundtrip = async (text: string) => {
+  const md = convertProseMirrorToMarkdown(doc(para(text)));
+  const back = await markdownToProseMirror(md);
+  return back.content as any[];
+};
+
+describe("gitmost transcript neutralization (git-sync round-trip)", () => {
+  // Lines that, at column 0, the serializer's missing block-escape would let
+  // git-sync re-parse into a non-paragraph block.
+  const triggerLines = [
+    "- dash",
+    "* star",
+    "+ plus",
+    "> quote",
+    "# hash",
+    "1. one",
+    "1) one",
+    "> [!info] note",
+    "```js",
+    "~~~",
+    // Solid + spaced thematic breaks — these re-parse into a `horizontalRule`,
+    // which carries NO text, so a bare separator line LOSES its text entirely
+    // (round-2 finding). `_` also only forms a block via this construct.
+    "---",
+    "***",
+    "___",
+    "- - -", // spaced dash break (solid form is caught by [-*+]\s too, but this is the break)
+    "_ _ _",
+  ];
+
+  it("BARE trigger lines corrupt into non-paragraph blocks (root cause)", async () => {
+    for (const line of triggerLines) {
+      const blocks = await roundtrip(line);
+      // At least one produced block is NOT a paragraph — i.e. corruption.
+      const allParagraphs = blocks.every((b) => b.type === "paragraph");
+      expect(
+        allParagraphs,
+        `expected "${line}" to corrupt when inserted bare`,
+      ).toBe(false);
+    }
+  });
+
+  it("BARE solid thematic breaks corrupt into a text-LOSING horizontalRule", async () => {
+    // The severe case: no text node survives. Documents why neutralization
+    // matters more here than for list/quote (where the text survived).
+    for (const line of ["---", "***", "___"]) {
+      const blocks = await roundtrip(line);
+      expect(blocks.map((b) => b.type)).toContain("horizontalRule");
+      // No block carries the original text anywhere.
+      const flat = JSON.stringify(blocks);
+      expect(flat).not.toContain(line);
+    }
+  });
+
+  it("ZWSP-neutralized trigger lines round-trip as a single paragraph, text preserved", async () => {
+    for (const line of triggerLines) {
+      // The regex must actually classify each as a trigger.
+      expect(MD_BLOCK_TRIGGER_RE.test(line), `regex missed "${line}"`).toBe(
+        true,
+      );
+      const neutralized = ZWSP + line;
+      const blocks = await roundtrip(neutralized);
+
+      expect(blocks).toHaveLength(1);
+      expect(blocks[0].type).toBe("paragraph");
+      // Text is byte-preserved (ZWSP + original line), so the display is the
+      // original line with only an invisible leading character.
+      expect(blocks[0].content[0].text).toBe(neutralized);
+    }
+  });
+
+  it("normal host-prefixed lines never match the trigger regex and round-trip byte-exact", async () => {
+    for (const line of [
+      "You: hello there",
+      "Speaker 1: - and then a dash mid-line",
+      "Speaker 2: 1. not a list",
+    ]) {
+      expect(MD_BLOCK_TRIGGER_RE.test(line)).toBe(false);
+      const blocks = await roundtrip(line);
+      expect(blocks).toHaveLength(1);
+      expect(blocks[0].type).toBe("paragraph");
+      expect(blocks[0].content[0].text).toBe(line);
+    }
+  });
+});

packages/prosemirror-markdown/test/markdown-converter-gaps.test.ts

@@ -374,7 +374,9 @@ describe('converter gap coverage — emission branches (specs 1–11)', () => {
         ],
       }),
     );
-    expect(out).toBe('- [ ] top\n  - child');
+    // Block children of a task item are blank-line separated (loose list) per the
+    // #351 fix; the sublist stays at the fixed 2-column continuation indent.
+    expect(out).toBe('- [ ] top\n\n  - child');
   });
 
   // 10. A bulletList inside a blockquote: each list line independently prefixed.

packages/prosemirror-markdown/test/markdown-converter.test.ts

@@ -198,7 +198,11 @@ describe('convertProseMirrorToMarkdown', () => {
         }),
       );
       // First line carries the marker; the nested list is indented 2 columns.
-      expect(out).toBe('- parent\n  - child');
+      // Block children of a list item are separated by a BLANK line (loose list):
+      // this is the #351 fix — a single "\n" let a following block merge into the
+      // first paragraph on re-parse (silent content loss). The blank line stays
+      // inside the item, so the sublist remains nested at the 2-col marker column.
+      expect(out).toBe('- parent\n\n  - child');
     });
 
     it('nested ordered list indents by the wider 3-col marker width', () => {
@@ -219,8 +223,9 @@ describe('convertProseMirrorToMarkdown', () => {
           ],
         }),
       );
-      // "1. " is 3 columns wide, so the continuation indent is 3 spaces.
-      expect(out).toBe('1. parent\n   1. child');
+      // "1. " is 3 columns wide, so the continuation indent is 3 spaces. Block
+      // children are blank-line separated (loose list) per the #351 fix.
+      expect(out).toBe('1. parent\n\n   1. child');
     });
   });
 
@@ -539,11 +544,12 @@ describe('convertProseMirrorToMarkdown', () => {
       );
 
       // The 10th marker is the 4-column "10. "; the nested sublist line must be
-      // indented exactly 4 spaces (prefix.length 3 + 1), NOT 3.
-      expect(out).toContain('10. j\n    1. x');
+      // indented exactly 4 spaces (prefix.length 3 + 1), NOT 3. Block children are
+      // blank-line separated (loose list) per the #351 fix.
+      expect(out).toContain('10. j\n\n    1. x');
       // Guard against the off-by-one (3-space) regression that would re-parse
       // the sublist as loose/sibling content on import.
-      expect(out).not.toContain('10. j\n   1. x');
+      expect(out).not.toContain('10. j\n\n   1. x');
       // And the single-digit items keep the narrower 3-column marker (no body
       // continuation here, but the marker itself must stay "1. ".."9. ").
       expect(out.startsWith('1. a\n2. b\n')).toBe(true);
@@ -580,17 +586,18 @@ describe('convertProseMirrorToMarkdown', () => {
           content: [para(text('line1')), para(text('line2'))],
         }),
       );
-      // NOTE(review): the spec predicted ':::warning\nline1\n\nline2\n:::' (a
-      // The converter joins the callout's rendered children with a single '\n'
-      // and emits an Obsidian-native callout: a `> [!type]` opener plus one
-      // `>`-prefixed body line per content line. We pin the lowercasing
-      // (WARNING -> warning) and the multi-child join.
-      expect(out).toBe('> [!warning]\n> line1\n> line2');
+      // The converter emits an Obsidian-native callout: a `> [!type]` opener plus
+      // one `>`-prefixed body line per content line. Block children are separated
+      // by a blank `>` line (#351 fix): a single '\n' let the two paragraphs merge
+      // into one on re-parse. We pin the lowercasing (WARNING -> warning) and the
+      // blank-line-separated multi-child join.
+      expect(out).toBe('> [!warning]\n> line1\n>\n> line2');
       // The type is lowercased (an uppercase `[!WARNING]` would not re-import).
       expect(out.startsWith('> [!warning]\n')).toBe(true);
       expect(out).not.toContain('[!WARNING]');
-      // Both paragraph children are present, each blockquote-prefixed.
-      expect(out).toContain('> line1\n> line2');
+      // Both paragraph children are present, each blockquote-prefixed, blank-`>`
+      // separated so they stay distinct paragraphs on re-parse.
+      expect(out).toContain('> line1\n>\n> line2');
     });
 
     // Spec 4 — blockquote per-line prefixer over a multi-line nested callout.
@@ -607,13 +614,11 @@ describe('convertProseMirrorToMarkdown', () => {
           ],
         }),
       );
-      // NOTE(review): the spec predicted '> :::info\n> a\n>\n> b\n> :::',
-      // assuming the nested callout body contains a blank line between 'a' and
-      // The nested callout renders as an Obsidian callout '> [!info]\n> a\n> b'
-      // (single-'\n' join, no blank line). The outer blockquote prefixer then
-      // prefixes each of those lines with '> ' again, yielding a doubly-nested
-      // blockquote — the realistic per-line-prefix loop over a multi-line child.
-      expect(out).toBe('> > [!info]\n> > a\n> > b');
+      // The nested callout renders as an Obsidian callout '> [!info]\n> a\n>\n> b'
+      // (blank-`>` separated children per the #351 fix). The outer blockquote
+      // prefixer then prefixes each of those lines with '> ' again, yielding a
+      // doubly-nested blockquote — the per-line-prefix loop over a multi-line child.
+      expect(out).toBe('> > [!info]\n> > a\n> >\n> > b');
       // Every produced line carries the '> ' prefix (no line escapes to col 0).
       for (const line of out.split('\n')) {
         expect(line.startsWith('>')).toBe(true);

patches/ai@6.0.134.patch

@@ -0,0 +1,68 @@
+diff --git a/dist/index.js b/dist/index.js
+index ae447a12f7823ec0a00837ee9f0eb809a610d5f8..a3402b2c2d021ef432cfa76e35d370073d525135 100644
+--- a/dist/index.js
++++ b/dist/index.js
+@@ -6578,9 +6578,19 @@ function createOutputTransformStream(output) {
+         controller.enqueue({ part: chunk, partialOutput: void 0 });
+         return;
+       }
+-      text2 += chunk.text;
+       textChunk += chunk.text;
+       textProviderMetadata = (_a21 = chunk.providerMetadata) != null ? _a21 : textProviderMetadata;
++      if (output == null) {
++        // PATCH(docmost #OOM): no output strategy requested -> publish each
++        // text-delta immediately and do NOT build cumulative partialOutput
++        // snapshots. Unpatched, the default text() output snapshots the ENTIRE
++        // accumulated turn text on every delta (O(n^2) memory) and those
++        // snapshots pile up in the never-consumed leftover tee branch of
++        // DefaultStreamTextResult.baseStream -> heap OOM on long agent turns.
++        publishTextChunk({ controller });
++        return;
++      }
++      text2 += chunk.text;
+       const result = await output.parsePartialOutput({ text: text2 });
+       if (result !== void 0) {
+         const currentJson = JSON.stringify(result.partial);
+@@ -6959,7 +6969,7 @@ var DefaultStreamTextResult = class {
+         })
+       );
+     }
+-    this.baseStream = stream.pipeThrough(createOutputTransformStream(output != null ? output : text())).pipeThrough(eventProcessor);
++    this.baseStream = stream.pipeThrough(createOutputTransformStream(output)).pipeThrough(eventProcessor);
+     const { maxRetries, retry } = prepareRetries({
+       maxRetries: maxRetriesArg,
+       abortSignal
+diff --git a/dist/index.mjs b/dist/index.mjs
+index 663875332e3f9a9bd167c25583c515876f42951b..b840b0502c9894df983e0154805abb80e70e6331 100644
+--- a/dist/index.mjs
++++ b/dist/index.mjs
+@@ -6501,9 +6501,19 @@ function createOutputTransformStream(output) {
+         controller.enqueue({ part: chunk, partialOutput: void 0 });
+         return;
+       }
+-      text2 += chunk.text;
+       textChunk += chunk.text;
+       textProviderMetadata = (_a21 = chunk.providerMetadata) != null ? _a21 : textProviderMetadata;
++      if (output == null) {
++        // PATCH(docmost #OOM): no output strategy requested -> publish each
++        // text-delta immediately and do NOT build cumulative partialOutput
++        // snapshots. Unpatched, the default text() output snapshots the ENTIRE
++        // accumulated turn text on every delta (O(n^2) memory) and those
++        // snapshots pile up in the never-consumed leftover tee branch of
++        // DefaultStreamTextResult.baseStream -> heap OOM on long agent turns.
++        publishTextChunk({ controller });
++        return;
++      }
++      text2 += chunk.text;
+       const result = await output.parsePartialOutput({ text: text2 });
+       if (result !== void 0) {
+         const currentJson = JSON.stringify(result.partial);
+@@ -6882,7 +6892,7 @@ var DefaultStreamTextResult = class {
+         })
+       );
+     }
+-    this.baseStream = stream.pipeThrough(createOutputTransformStream(output != null ? output : text())).pipeThrough(eventProcessor);
++    this.baseStream = stream.pipeThrough(createOutputTransformStream(output)).pipeThrough(eventProcessor);
+     const { maxRetries, retry } = prepareRetries({
+       maxRetries: maxRetriesArg,
+       abortSignal

pnpm-lock.yaml

@@ -44,6 +44,9 @@ overrides:
   ip-address: 10.1.1
 
 patchedDependencies:
+  ai@6.0.134:
+    hash: f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9
+    path: patches/ai@6.0.134.patch
   scimmy@1.3.5:
     hash: 775d80f86830b2c5dd1a250c9802c10f8fc3da3c7898373de5aa0c23993d1673
     path: patches/scimmy@1.3.5.patch
@@ -543,6 +546,9 @@ importers:
       '@docmost/pdf-inspector':
         specifier: 1.9.6
         version: 1.9.6
+      '@docmost/prosemirror-markdown':
+        specifier: workspace:*
+        version: link:../../packages/prosemirror-markdown
       '@fastify/cookie':
         specifier: ^11.0.2
         version: 11.0.2
@@ -623,10 +629,10 @@ importers:
         version: 8.3.0(socket.io-adapter@2.5.4)
       ai:
         specifier: ^6.0.134
-        version: 6.0.134(zod@4.3.6)
+        version: 6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6)
       ai-sdk-ollama:
         specifier: ^3.8.1
-        version: 3.8.1(ai@6.0.134(zod@4.3.6))(zod@4.3.6)
+        version: 3.8.1(ai@6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6))(zod@4.3.6)
       bcrypt:
         specifier: ^6.0.0
         version: 6.0.0
@@ -16355,17 +16361,17 @@ snapshots:
 
   agent-base@7.1.4: {}
 
-  ai-sdk-ollama@3.8.1(ai@6.0.134(zod@4.3.6))(zod@4.3.6):
+  ai-sdk-ollama@3.8.1(ai@6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6))(zod@4.3.6):
     dependencies:
       '@ai-sdk/provider': 3.0.8
       '@ai-sdk/provider-utils': 4.0.21(zod@4.3.6)
-      ai: 6.0.134(zod@4.3.6)
+      ai: 6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6)
       jsonrepair: 3.13.3
       ollama: 0.6.3
     transitivePeerDependencies:
       - zod
 
-  ai@6.0.134(zod@4.3.6):
+  ai@6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6):
     dependencies:
       '@ai-sdk/gateway': 3.0.77(zod@4.3.6)
       '@ai-sdk/provider': 3.0.8