Merge pull request 'feat(ai-chat): resumable SSE — клиент + удаление поллинга/латчей (#381 PR 2)' (#387) from feat/381-resumable-sse-pr2 into feat/381-resumable-sse-pr1

Reviewed-on: #387

.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

@@ -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
 
@@ -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/components/ai-chat-window.tsx

@@ -41,17 +41,10 @@ import { extractPageSlugId } from "@/lib";
 import {
   AI_CHATS_RQ_KEY,
   AI_CHAT_MESSAGES_RQ_KEY,
-  AI_CHAT_RUN_RQ_KEY,
   useAiChatMessagesQuery,
-  useAiChatRunQuery,
   useAiChatsQuery,
   useAiRolesQuery,
 } from "@/features/ai-chat/queries/ai-chat-query.ts";
-import {
-  shouldClearLatchOnQueryError,
-  shouldClearStoppingLatch,
-  shouldObserveRun,
-} from "@/features/ai-chat/utils/run-polling.ts";
 import { workspaceAtom } from "@/features/user/atoms/current-user-atom";
 import ConversationList from "@/features/ai-chat/components/conversation-list.tsx";
 import ChatThread from "@/features/ai-chat/components/chat-thread.tsx";
@@ -85,6 +78,12 @@ const MIN_HEIGHT = 400;
 // Margin kept between the window and the viewport edges while dragging.
 const EDGE_MARGIN = 8;
 
+// #184 phase 1.5: hard cap on the degraded-poll fallback. The poll is armed when
+// a resume attempt could not attach to the live run and disarmed by the thread on
+// settle / local stream; this cap is the ONLY backstop against an endless tick
+// (a stuck 'streaming' row before the boot-sweep, or a user-tail 204 with no run).
+const DEGRADED_POLL_MAX_MS = 10 * 60_000;
+
 /** Compact token formatter: 1.2M / 3.4k / 950. */
 function formatTokens(n: number): string {
   if (n >= 1_000_000) return `${(n / 1_000_000).toFixed(1)}M`;
@@ -242,150 +241,62 @@ export default function AiChatWindow() {
     [roles],
   );
 
+  // #184 phase 1.5: degraded-poll fallback (replaces the F4/F5/F7 latches). When
+  // ChatThread could not attach to a still-running run it arms this via
+  // onResumeFallback(true); the thread disarms it on settle / local stream. The
+  // window only OWNS the timer (armedAtRef stamps when it was armed for the cap).
+  const [degradedPoll, setDegradedPoll] = useState(false);
+  const armedAtRef = useRef(0);
+  const onResumeFallback = useCallback((active: boolean): void => {
+    if (active) armedAtRef.current = Date.now();
+    setDegradedPoll(active);
+  }, []);
+  // Reset the degraded poll whenever the open chat changes: it is scoped to the
+  // resume attempt of the previously-open chat (invariant 8).
+  useEffect(() => {
+    setDegradedPoll(false);
+  }, [activeChatId]);
+
   const { data: messageRows, isLoading: messagesLoading } =
-    useAiChatMessagesQuery(activeChatId ?? undefined);
+    useAiChatMessagesQuery(
+      activeChatId ?? undefined,
+      // DELIBERATELY DUMB (invariant 8 / task 2.4): poll every 2.5s while armed
+      // and under the 10-min cap; otherwise off. NO error checks (TanStack v5
+      // resets fetchFailureCount each fetch, so consecutive errors are not
+      // expressible — and the poll must survive a server restart) and NO tail
+      // checks (the settled/local-stream semantics live in ChatThread, which
+      // disarms via onResumeFallback(false)). The time cap is the only backstop.
+      () =>
+        degradedPoll === true &&
+        Date.now() - armedAtRef.current < DEGRADED_POLL_MAX_MS
+          ? 2500
+          : false,
+    );
 
   // #184 reconnect-and-live-follow. Whether detached agent runs are enabled for
-  // this workspace. The reconnect endpoint itself is NOT flag-gated server-side
-  // (it is only owner-gated and returns `{ run: null }` when the chat has no
-  // run); but when the feature is off no runs are ever created, so polling it
-  // would always come back empty — we gate it off here to avoid pointless polls.
+  // this workspace. When the feature is off no runs are ever created, so the
+  // resume attempt would only ever 204; gating ChatThread's resume on it avoids a
+  // pointless attach round-trip.
   const workspace = useAtomValue(workspaceAtom);
   const autonomousRunsEnabled =
     workspace?.settings?.ai?.autonomousRuns === true;
 
-  // Whether THIS tab is the one actively streaming the open chat's run locally
-  // (it started the run here and holds the SSE). Reported up from ChatThread. We
-  // are the STREAMER while true and a passive OBSERVER while false — the basis of
-  // the observer-vs-streamer detection. Reset to false by the fresh ChatThread's
-  // mount effect on every chat switch.
-  const [localStreaming, setLocalStreaming] = useState(false);
-  const onStreamingChange = useCallback((streaming: boolean) => {
-    setLocalStreaming(streaming);
-  }, []);
-
-  // #184 Stop wiring. While a detached run is being stopped we SUPPRESS the
-  // observer merge so the stopping run's still-persisting output does not
-  // re-stream back into view between the moment the user pressed Stop and the run
-  // actually settling as 'aborted' server-side. Polling itself keeps running (so
-  // the terminal transition is still detected) — only the visual merge is gated.
-  // Cleared when the run is observed terminal (below) or the chat is switched.
-  const [stoppingRun, setStoppingRun] = useState(false);
-  // Reset the stopping latch whenever the open chat changes: it is scoped to the
-  // run of the previously-open chat.
-  useEffect(() => {
-    setStoppingRun(false);
-  }, [activeChatId]);
-
   // Authoritative stop of the open chat's detached run (the Stop button in
-  // autonomous mode). Latch "stopping" first (suppresses the re-stream flash),
-  // then request the server stop — the ONLY thing that ends a detached run; a mere
-  // local SSE abort is a client disconnect the server ignores. On failure we
-  // release the latch so the observer resumes (better to show the live run than to
-  // freeze the view) and surface the error.
+  // autonomous mode). Request the server stop — the ONLY thing that ends a
+  // detached run; a mere local SSE abort is a client disconnect the server
+  // ignores. On failure surface the error.
   const handleServerStop = useCallback(
     (chatId: string): void => {
-      setStoppingRun(true);
-      // #234 F4: drop the PREVIOUS turn's run from the cache so `run` becomes null
-      // until the CURRENT turn's run is fetched fresh. Without this, once the local
-      // stream aborts (localStreaming -> false) the run query re-enables and
-      // react-query SYNCHRONOUSLY returns the still-cached prior terminal run; the
-      // terminal effect would then clear the stopping latch against that STALE run
-      // before the current turn's (still-running, detached, growing) run is ever
-      // observed — re-opening the observer merge and flashing the growing output
-      // over the frozen row. With the cache cleared the terminal effect's
-      // `if (!run) return` holds the latch until the current run itself is observed
-      // terminal (see shouldClearStoppingLatch).
-      queryClient.removeQueries({ queryKey: AI_CHAT_RUN_RQ_KEY(chatId) });
       void stopRun(chatId).catch(() => {
-        setStoppingRun(false);
         notifications.show({
           message: t("Failed to stop the run"),
           color: "red",
         });
       });
     },
-    [t, queryClient],
+    [t],
   );
 
-  // Poll the latest run of the open chat ONLY when we are a passive observer:
-  // feature on, a chat is open, and we are NOT the local streamer (the streamer
-  // already has the live SSE — polling/merging too would double-render). The
-  // query's own status-keyed refetchInterval stops once the run is terminal.
-  const { data: runData, isError: runQueryFailed } = useAiChatRunQuery(
-    activeChatId ?? undefined,
-    autonomousRunsEnabled && !localStreaming,
-  );
-  const run = runData?.run ?? null;
-
-  // Safety net (#234 F4 review): after handleServerStop clears the run cache,
-  // `run` is null until the current turn's run is fetched fresh, and the terminal
-  // effect below holds the latch via `if (!run) return`. If that refetch instead
-  // ERRORS PERMANENTLY (the GET-run keeps failing) while we are no longer the
-  // streamer, the run stays null, its status-keyed refetchInterval is off, and
-  // nothing would ever observe a terminal run — freezing the view with the
-  // observer merge suppressed. Release the latch on that error so the live view
-  // resumes rather than stays stuck (the local stopRun may already have succeeded
-  // independently).
-  //
-  // #234 F7: this must NOT fire on a TRANSIENT error while `run` is still an
-  // ACTIVE held run. In TanStack Query v5 (retry:false) the query's `data` is
-  // RETAINED on error, so `runQueryFailed` can be true while `run` is still
-  // pending/running — releasing then would re-open the observer merge and flash
-  // the growing detached run over the frozen row (the very flash F4 prevents). The
-  // decision is the pure, unit-tested `shouldClearLatchOnQueryError`, which gates
-  // on the run NOT being active: it cures only the genuine permanent-null-freeze
-  // (`run === null`) and never releases against an active run.
-  useEffect(() => {
-    if (
-      shouldClearLatchOnQueryError({
-        stoppingRun,
-        isLocalStreaming: localStreaming,
-        runQueryFailed,
-        run,
-      })
-    )
-      setStoppingRun(false);
-  }, [stoppingRun, localStreaming, runQueryFailed, run]);
-  // The run's incrementally-persisted assistant message to merge into the thread,
-  // but only while we are an observer (never when we are the streamer — guards
-  // against a stale poll fighting the live stream). Includes a terminal run so the
-  // final persisted output is shown on reopen.
-  const observedRow =
-    shouldObserveRun(run, localStreaming) && !stoppingRun
-      ? (runData?.message ?? null)
-      : null;
-
-  // When the observed run reaches a terminal status, do a final messages refetch
-  // so the persisted final state (token/context badge, export source) is shown,
-  // then the query's refetchInterval has already stopped polling. Deduped per run
-  // id so it fires exactly once per run, not on every subsequent poll-less render.
-  const finalizedRunIdRef = useRef<string | null>(null);
-  useEffect(() => {
-    if (!run || !activeChatId) return;
-    if (run.status === "pending" || run.status === "running") {
-      // Active again (a new run) — re-arm so its terminal transition fires once.
-      finalizedRunIdRef.current = null;
-      return;
-    }
-    // Terminal: a stop we requested has landed (or the run finished on its own),
-    // so release the stopping latch — the observer merge can now show the final
-    // persisted (aborted/finished) output without any live re-stream. The decision
-    // is the pure, unit-tested `shouldClearStoppingLatch` (run-polling.ts): release
-    // ONLY when we requested a stop, this tab is no longer the streamer, AND the
-    // CURRENT run is terminal. The #234 F4 cache removal in handleServerStop makes
-    // `run` null (this branch's `if (!run) return` above holds) until the current
-    // turn's run is fetched fresh, so the latch can never clear against a stale
-    // cached run.
-    if (shouldClearStoppingLatch({ stoppingRun, run, isLocalStreaming: localStreaming }))
-      setStoppingRun(false);
-    if (finalizedRunIdRef.current === run.id) return;
-    finalizedRunIdRef.current = run.id;
-    queryClient.invalidateQueries({
-      queryKey: AI_CHAT_MESSAGES_RQ_KEY(activeChatId),
-    });
-  }, [run, activeChatId, queryClient, stoppingRun, localStreaming]);
-
   // The page the user is currently viewing. AiChatWindow lives in a pathless
   // parent layout route, so useParams() can't see :pageSlug. Match the full
   // pathname against the authenticated page route instead so "the current page"
@@ -1034,16 +945,13 @@ export default function AiChatWindow() {
               assistantName={currentRole?.name}
               onTurnFinished={onTurnFinished}
               onServerChatId={onServerChatId}
-              // #184: live-follow a still-running run when we reopened the chat as
-              // a passive observer; null when there is nothing to observe or this
-              // tab is the streamer. onStreamingChange lets the window stop polling
-              // while we are the streamer.
-              observedRow={observedRow}
-              onStreamingChange={onStreamingChange}
+              // #184 phase 1.5: arm/disarm the degraded-poll fallback when a
+              // resume attempt could not attach to the live run; the thread
+              // disarms it on settle / local stream.
+              onResumeFallback={onResumeFallback}
               // #184: in autonomous mode the Stop button must hit the authoritative
               // server stop (a local SSE abort is a client disconnect the server
-              // ignores). onServerStop also arms the "stopping" latch above so the
-              // stopped run's output does not re-stream via the observer merge.
+              // ignores).
               autonomousRunsEnabled={autonomousRunsEnabled}
               onServerStop={handleServerStop}
             />

apps/client/src/features/ai-chat/components/chat-thread.test.tsx

@@ -1,6 +1,13 @@
-import { describe, it, expect, beforeEach, vi } from "vitest";
-import { render, screen, fireEvent, act, cleanup } from "@testing-library/react";
+import { describe, it, expect, beforeEach, afterEach, vi } from "vitest";
+import {
+  render,
+  screen,
+  fireEvent,
+  act,
+  cleanup,
+} from "@testing-library/react";
 import { MantineProvider } from "@mantine/core";
+import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
 
 // Shared, hoisted mock state so the @ai-sdk/react and "ai" module mocks (hoisted
 // above the imports) can expose the captured useChat callbacks / transport and
@@ -12,50 +19,61 @@ const h = vi.hoisted(() => ({
     sendMessage: vi.fn(),
     stop: vi.fn(),
     setMessages: vi.fn(),
+    resumeStream: vi.fn(),
+    // The messages array useChat was seeded with (to assert strip/seed behavior).
+    seededMessages: null as null | unknown[],
     transport: null as null | {
-      prepareSendMessagesRequest: (arg: {
+      prepareSendMessagesRequest?: (arg: {
         messages: unknown[];
         body: Record<string, unknown>;
       }) => { body: Record<string, unknown> };
+      prepareReconnectToStreamRequest?: () => { api?: string };
+      fetch?: (input: unknown, init?: { method?: string }) => Promise<unknown>;
     },
   },
 }));
 
-// Mock useChat: capture onFinish, return the spies and the controllable status.
+// Mock useChat: capture onFinish + seeded messages, return the spies and the
+// controllable status.
 vi.mock("@ai-sdk/react", () => ({
-  useChat: (opts: { onFinish?: (arg: Record<string, unknown>) => void }) => {
+  useChat: (opts: {
+    messages?: unknown[];
+    onFinish?: (arg: Record<string, unknown>) => void;
+  }) => {
     h.state.onFinish = opts.onFinish ?? null;
+    h.state.seededMessages = opts.messages ?? null;
     return {
       messages: [],
       sendMessage: h.state.sendMessage,
       status: h.state.status,
       stop: h.state.stop,
       error: null,
-      // #184: ChatThread reads setMessages to merge a polled observer run.
       setMessages: h.state.setMessages,
+      resumeStream: h.state.resumeStream,
     };
   },
 }));
 
 // Mock "ai": deterministic ids + a transport that records its options so the test
-// can invoke prepareSendMessagesRequest and assert the `interrupted` flag.
+// can invoke prepareSendMessagesRequest / prepareReconnectToStreamRequest / fetch.
 vi.mock("ai", () => {
   let counter = 0;
   return {
     generateId: () => `gid-${counter++}`,
     DefaultChatTransport: class {
-      constructor(opts: {
-        prepareSendMessagesRequest: (arg: {
-          messages: unknown[];
-          body: Record<string, unknown>;
-        }) => { body: Record<string, unknown> };
-      }) {
-        h.state.transport = opts;
+      constructor(opts: Record<string, unknown>) {
+        h.state.transport = opts as never;
       }
     },
   };
 });
 
+// Keep the ai-chat-query import light: ChatThread only needs the messages RQ key,
+// so stub the module to avoid pulling axios / i18n transitively.
+vi.mock("@/features/ai-chat/queries/ai-chat-query.ts", () => ({
+  AI_CHAT_MESSAGES_RQ_KEY: (chatId: string) => ["ai-chat-messages", chatId],
+}));
+
 // Stub the heavy children: MessageList (markdown/render) and ChatInput (the
 // composer). The ChatInput stub exposes a button that queues a message, the only
 // interaction this test needs to populate the queue while "streaming".
@@ -63,49 +81,90 @@ vi.mock("@/features/ai-chat/components/message-list.tsx", () => ({
   default: () => <div data-testid="message-list" />,
 }));
 vi.mock("@/features/ai-chat/components/chat-input.tsx", () => ({
-  default: ({ onQueue }: { onQueue: (text: string) => void }) => (
-    <button data-testid="queue-btn" onClick={() => onQueue("queued text")}>
-      queue
-    </button>
+  default: ({
+    onQueue,
+    onStop,
+  }: {
+    onQueue: (text: string) => void;
+    onStop: () => void;
+  }) => (
+    <>
+      <button data-testid="queue-btn" onClick={() => onQueue("queued text")}>
+        queue
+      </button>
+      <button aria-label="Stop" onClick={() => onStop()}>
+        stop
+      </button>
+    </>
   ),
 }));
 
 import ChatThread from "./chat-thread";
+import type { IAiChatMessageRow } from "@/features/ai-chat/types/ai-chat.types.ts";
 
-function renderThread() {
+function row(
+  id: string,
+  role: string,
+  status?: string,
+  text = "",
+): IAiChatMessageRow {
+  return { id, role, content: text, status, createdAt: "2026-01-01T00:00:00Z" };
+}
+
+function renderThread(props?: {
+  chatId?: string | null;
+  initialRows?: IAiChatMessageRow[];
+  autonomousRunsEnabled?: boolean;
+}) {
   const onTurnFinished = vi.fn();
-  render(
-    <MantineProvider>
-      <ChatThread chatId="c1" initialRows={[]} onTurnFinished={onTurnFinished} />
-    </MantineProvider>,
+  const onResumeFallback = vi.fn();
+  const onServerStop = vi.fn();
+  const queryClient = new QueryClient({
+    defaultOptions: { queries: { retry: false } },
+  });
+  const invalidateSpy = vi.spyOn(queryClient, "invalidateQueries");
+  const { unmount } = render(
+    <QueryClientProvider client={queryClient}>
+      <MantineProvider>
+        <ChatThread
+          chatId={props?.chatId === undefined ? "c1" : props.chatId}
+          initialRows={props?.initialRows ?? []}
+          autonomousRunsEnabled={props?.autonomousRunsEnabled}
+          onTurnFinished={onTurnFinished}
+          onResumeFallback={onResumeFallback}
+          onServerStop={onServerStop}
+        />
+      </MantineProvider>
+    </QueryClientProvider>,
   );
-  return { onTurnFinished };
+  return { onTurnFinished, onResumeFallback, onServerStop, invalidateSpy, unmount };
+}
+
+function resetState() {
+  h.state.status = "streaming";
+  h.state.onFinish = null;
+  h.state.seededMessages = null;
+  h.state.transport = null;
+  h.state.sendMessage.mockClear();
+  h.state.stop.mockClear();
+  h.state.setMessages.mockClear();
+  h.state.resumeStream.mockClear();
 }
 
 describe("ChatThread — send now (#198)", () => {
-  beforeEach(() => {
-    h.state.status = "streaming";
-    h.state.onFinish = null;
-    h.state.sendMessage.mockClear();
-    h.state.stop.mockClear();
-    h.state.transport = null;
-  });
+  beforeEach(resetState);
 
   it("aborts the current turn and resends the queued message on the abort", () => {
     renderThread();
 
-    // Queue a message while the turn is streaming.
     fireEvent.click(screen.getByTestId("queue-btn"));
     const sendNowBtn = screen.getByLabelText("Send now");
     expect(sendNowBtn).toBeTruthy();
 
-    // "Send now" interrupts the current turn (stop), but does NOT send yet —
-    // the resend happens once the abort lands in onFinish.
     fireEvent.click(sendNowBtn);
     expect(h.state.stop).toHaveBeenCalledTimes(1);
     expect(h.state.sendMessage).not.toHaveBeenCalled();
 
-    // The abort we triggered reaches onFinish: the promoted head is flushed.
     act(() => {
       h.state.onFinish?.({
         message: { id: "a", role: "assistant", parts: [] },
@@ -122,10 +181,8 @@ describe("ChatThread — send now (#198)", () => {
     fireEvent.click(screen.getByTestId("queue-btn"));
     fireEvent.click(screen.getByLabelText("Send now"));
 
-    const prep = h.state.transport!.prepareSendMessagesRequest;
-    // The send right after "send now" carries interrupted: true...
+    const prep = h.state.transport!.prepareSendMessagesRequest!;
     expect(prep({ messages: [], body: {} }).body.interrupted).toBe(true);
-    // ...and only that one (the flag is read-and-cleared).
     expect(prep({ messages: [], body: {} }).body.interrupted).toBe(false);
   });
 
@@ -136,42 +193,24 @@ describe("ChatThread — send now (#198)", () => {
     fireEvent.click(screen.getByTestId("queue-btn"));
     fireEvent.click(screen.getByLabelText("Send now"));
 
-    // No turn to interrupt: sent straight away, no abort, not flagged.
     expect(h.state.stop).not.toHaveBeenCalled();
     expect(h.state.sendMessage).toHaveBeenCalledWith({ text: "queued text" });
-    const prep = h.state.transport!.prepareSendMessagesRequest;
+    const prep = h.state.transport!.prepareSendMessagesRequest!;
     expect(prep({ messages: [], body: {} }).body.interrupted).toBe(false);
   });
 });
 
-// The turn-end decision lives in the `onFinish` handler: given the terminal
-// outcome of a turn (`isAbort` / `isDisconnect` / `isError`, or none = clean),
-// it decides whether to CONTINUE (flush the next queued message) or END (leave
-// the queue intact for the user), and which stop notice — if any — to show.
-// `sendNow` is exercised above; these tests pin down the plain outcomes.
 describe("ChatThread — turn-end decision (onFinish)", () => {
-  beforeEach(() => {
-    h.state.status = "streaming";
-    h.state.onFinish = null;
-    h.state.sendMessage.mockClear();
-    h.state.stop.mockClear();
-    h.state.transport = null;
-  });
+  beforeEach(resetState);
 
-  // Drive a fresh onFinish with the given terminal flags after queueing a
-  // message, and report both what the parent was told and whether the queue was
-  // flushed (a resend to the sendMessage spy).
   function finishWith(flags: {
     isAbort?: boolean;
     isDisconnect?: boolean;
     isError?: boolean;
   }) {
-    // Tear down any prior render so the loop-driven "every outcome" case does
-    // not leave duplicate queue buttons in the DOM.
     cleanup();
     h.state.sendMessage.mockClear();
     const { onTurnFinished } = renderThread();
-    // Populate the queue while the turn is streaming.
     fireEvent.click(screen.getByTestId("queue-btn"));
     act(() => {
       h.state.onFinish?.({
@@ -187,16 +226,12 @@ describe("ChatThread — turn-end decision (onFinish)", () => {
 
   it("CONTINUES — flushes the next queued message on a clean finish", () => {
     finishWith({});
-    // Clean finish (no terminal flag): the queued message is auto-sent.
     expect(h.state.sendMessage).toHaveBeenCalledWith({ text: "queued text" });
-    // A clean finish shows no stop notice.
     expect(screen.queryByText("Response stopped.")).toBeNull();
   });
 
   it("ENDS — keeps the queue intact on a user abort and shows the stopped notice", () => {
     finishWith({ isAbort: true });
-    // A plain Stop (not the sendNow interrupt path) must NOT auto-resend: the
-    // queue is preserved for the user to decide.
     expect(h.state.sendMessage).not.toHaveBeenCalled();
     expect(screen.getByText("Response stopped.")).toBeTruthy();
   });
@@ -211,15 +246,11 @@ describe("ChatThread — turn-end decision (onFinish)", () => {
 
   it("ENDS — keeps the queue intact on a stream error (no auto-retry, no stopped notice)", () => {
     finishWith({ isError: true });
-    // Blindly retrying after a failure would be wrong; the queue is left alone.
     expect(h.state.sendMessage).not.toHaveBeenCalled();
-    // isError clears the neutral notice (the error banner covers this case).
     expect(screen.queryByText("Response stopped.")).toBeNull();
   });
 
   it("notifies the parent on EVERY terminal outcome", () => {
-    // The chat-list refresh / new-chat id adoption must run on success and on
-    // every failure path alike.
     for (const flags of [
       {},
       { isAbort: true },
@@ -232,55 +263,411 @@ describe("ChatThread — turn-end decision (onFinish)", () => {
   });
 });
 
-// #184 passive-observer merge: when reconnecting to a still-running run, the
-// parent feeds the polled run message via `observedRow`; ChatThread merges it via
-// setMessages — but ONLY when this tab is NOT itself streaming (the streamer's
-// SSE owns the view, so a stale observedRow must never overwrite it).
-describe("ChatThread — observer run merge (#184)", () => {
-  beforeEach(() => {
-    h.state.onFinish = null;
-    h.state.setMessages.mockReset();
+// #184 phase 1.5: the resumable-SSE client. A reopened tab resumes the live run
+// via the SDK's reconnect transport (attach: replay + tail) instead of polling.
+describe("ChatThread — resume (attach) machinery (#184)", () => {
+  const streamingTail = () => [
+    row("u1", "user", undefined, "hi"),
+    row("a1", "assistant", "streaming", "partial"),
+  ];
+  const settledTail = () => [
+    row("u1", "user", undefined, "hi"),
+    row("a1", "assistant", "succeeded", "done"),
+  ];
+  const userTail = () => [row("u1", "user", undefined, "hi")];
+
+  const visibleMsg = {
+    id: "a1",
+    role: "assistant",
+    parts: [{ type: "text", text: "streamed answer" }],
+  };
+  const emptyMsg = { id: "a1", role: "assistant", parts: [] };
+
+  beforeEach(resetState);
+  // NOTE: do NOT vi.unstubAllGlobals() here — vitest.setup.ts installs
+  // matchMedia/localStorage via vi.stubGlobal and unstubbing wipes them for the
+  // rest of the file. Fetch is re-stubbed per test that needs it.
+  afterEach(cleanup);
+
+  it("resumes on mount only when the flag is on, chatId is set, and the tail is not a settled assistant", () => {
+    // streaming tail -> resume
+    renderThread({ autonomousRunsEnabled: true, initialRows: streamingTail() });
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(1);
+
+    // user tail -> resume (the assistant row may not be seeded yet)
+    cleanup();
+    h.state.resumeStream.mockClear();
+    renderThread({ autonomousRunsEnabled: true, initialRows: userTail() });
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(1);
+
+    // settled assistant tail -> NO resume
+    cleanup();
+    h.state.resumeStream.mockClear();
+    renderThread({ autonomousRunsEnabled: true, initialRows: settledTail() });
+    expect(h.state.resumeStream).not.toHaveBeenCalled();
+
+    // flag off -> NO resume
+    cleanup();
+    h.state.resumeStream.mockClear();
+    renderThread({ autonomousRunsEnabled: false, initialRows: streamingTail() });
+    expect(h.state.resumeStream).not.toHaveBeenCalled();
+
+    // no chatId -> NO resume
+    cleanup();
+    h.state.resumeStream.mockClear();
+    renderThread({
+      autonomousRunsEnabled: true,
+      chatId: null,
+      initialRows: streamingTail(),
+    });
+    expect(h.state.resumeStream).not.toHaveBeenCalled();
   });
 
-  const observedRow = {
-    id: "a-run",
-    role: "assistant",
-    content: "step 1\nstep 2",
-    metadata: {
-      parts: [{ type: "text", text: "step 1\nstep 2" }],
-    },
-    createdAt: "2026-01-01T00:00:00Z",
-  } as const;
+  it("strips the streaming tail from the seed, but keeps a user tail whole", () => {
+    renderThread({ autonomousRunsEnabled: true, initialRows: streamingTail() });
+    // 2 rows in, streaming tail stripped -> 1 seeded message.
+    expect(h.state.seededMessages).toHaveLength(1);
 
-  function renderObserver(status: string) {
-    h.state.status = status;
-    render(
+    cleanup();
+    renderThread({ autonomousRunsEnabled: true, initialRows: userTail() });
+    // user tail is not stripped.
+    expect(h.state.seededMessages).toHaveLength(1);
+  });
+
+  it("builds the attach URL with expect=live&anchor only when the streaming tail was stripped", () => {
+    renderThread({ autonomousRunsEnabled: true, initialRows: streamingTail() });
+    expect(h.state.transport!.prepareReconnectToStreamRequest!().api).toBe(
+      "/api/ai-chat/runs/c1/stream?expect=live&anchor=a1",
+    );
+
+    cleanup();
+    renderThread({ autonomousRunsEnabled: true, initialRows: userTail() });
+    expect(h.state.transport!.prepareReconnectToStreamRequest!().api).toBe(
+      "/api/ai-chat/runs/c1/stream",
+    );
+  });
+
+  async function fetch204() {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({ status: 204, ok: false }),
+    );
+    await act(async () => {
+      await h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+  }
+
+  it("204 on a user tail: no crash, no restore, reconcile+invalidate, onResumeFallback(true)", async () => {
+    const { onResumeFallback, invalidateSpy } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: userTail(),
+    });
+    await fetch204();
+    // No stripped row -> no restore merge.
+    expect(h.state.setMessages).not.toHaveBeenCalled();
+    expect(invalidateSpy).toHaveBeenCalledWith({
+      queryKey: ["ai-chat-messages", "c1"],
+    });
+    expect(onResumeFallback).toHaveBeenCalledWith(true);
+  });
+
+  it("204 on a streaming tail: restore + invalidate + onResumeFallback(true)", async () => {
+    const { onResumeFallback, invalidateSpy } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    await fetch204();
+    // Stripped row is restored to the store.
+    expect(h.state.setMessages).toHaveBeenCalledTimes(1);
+    expect(invalidateSpy).toHaveBeenCalledWith({
+      queryKey: ["ai-chat-messages", "c1"],
+    });
+    expect(onResumeFallback).toHaveBeenCalledWith(true);
+  });
+
+  it("F7 restart-survival: a 500 attach failure restores the stripped row AND arms the poll (not lost)", async () => {
+    const { onResumeFallback, invalidateSpy } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({ status: 500, ok: false }),
+    );
+    await act(async () => {
+      await h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+    expect(h.state.setMessages).toHaveBeenCalledTimes(1); // stripped row restored
+    expect(invalidateSpy).toHaveBeenCalledWith({
+      queryKey: ["ai-chat-messages", "c1"],
+    });
+    expect(onResumeFallback).toHaveBeenCalledWith(true); // degraded poll armed
+  });
+
+  it("F7 restart-survival: a network throw restores the stripped row AND arms the poll", async () => {
+    const { onResumeFallback, invalidateSpy } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockRejectedValue(new Error("network down")),
+    );
+    await act(async () => {
+      await h.state
+        .transport!.fetch!("http://x", { method: "GET" })
+        .catch(() => undefined); // the wrapper rethrows; swallow here
+    });
+    expect(h.state.setMessages).toHaveBeenCalledTimes(1);
+    expect(invalidateSpy).toHaveBeenCalledWith({
+      queryKey: ["ai-chat-messages", "c1"],
+    });
+    expect(onResumeFallback).toHaveBeenCalledWith(true);
+  });
+
+  it("unmount during a pending attach aborts the controller and gates late callbacks", async () => {
+    const { onResumeFallback, invalidateSpy, unmount } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    let abortSeen = false;
+    let resolveFetch!: (v: unknown) => void;
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockImplementation((_input: unknown, init: RequestInit) => {
+        init.signal?.addEventListener("abort", () => {
+          abortSeen = true;
+        });
+        return new Promise((res) => {
+          resolveFetch = res;
+        });
+      }),
+    );
+    // Kick a reconnect GET (stays pending).
+    let pending!: Promise<unknown>;
+    act(() => {
+      pending = h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+    // Unmount: the cleanup aborts the in-flight attach.
+    unmount();
+    expect(abortSeen).toBe(true);
+    // A late 204 landing after unmount must NOT arm a poll / invalidate the (now
+    // different) chat.
+    onResumeFallback.mockClear();
+    invalidateSpy.mockClear();
+    await act(async () => {
+      resolveFetch({ status: 204, ok: false });
+      await pending;
+    });
+    expect(onResumeFallback).not.toHaveBeenCalledWith(true);
+    expect(invalidateSpy).not.toHaveBeenCalled();
+  });
+
+  it("a resume fetch error clears resumedTurn so the next local turn flushes the queue", async () => {
+    renderThread({ autonomousRunsEnabled: true, initialRows: streamingTail() });
+    h.state.status = "ready";
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({ status: 500, ok: false }),
+    );
+    await act(async () => {
+      await h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+    // Queue then clean-finish: suppression was cleared, so the queue flushes.
+    fireEvent.click(screen.getByTestId("queue-btn"));
+    act(() => {
+      h.state.onFinish?.({
+        message: visibleMsg,
+        isAbort: false,
+        isDisconnect: false,
+        isError: false,
+      });
+    });
+    expect(h.state.sendMessage).toHaveBeenCalledWith({ text: "queued text" });
+  });
+
+  it("a resumed turn's onFinish does NOT flush the queue", () => {
+    renderThread({ autonomousRunsEnabled: true, initialRows: streamingTail() });
+    fireEvent.click(screen.getByTestId("queue-btn"));
+    act(() => {
+      h.state.onFinish?.({
+        message: visibleMsg,
+        isAbort: false,
+        isDisconnect: false,
+        isError: false,
+      });
+    });
+    expect(h.state.sendMessage).not.toHaveBeenCalled();
+  });
+
+  it("a healthy resumed finish (visible content) arms nothing and keeps the store", () => {
+    h.state.status = "ready";
+    const { onResumeFallback } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    h.state.setMessages.mockClear();
+    onResumeFallback.mockClear();
+    act(() => {
+      h.state.onFinish?.({
+        message: visibleMsg,
+        isAbort: false,
+        isDisconnect: false,
+        isError: false,
+      });
+    });
+    // No restore (would clobber the fuller streamed message), no poll arm.
+    expect(h.state.setMessages).not.toHaveBeenCalled();
+    expect(onResumeFallback).not.toHaveBeenCalledWith(true);
+  });
+
+  it("isDisconnect WITH visible content arms the poll but does NOT restore", () => {
+    h.state.status = "ready";
+    const { onResumeFallback, invalidateSpy } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    h.state.setMessages.mockClear();
+    onResumeFallback.mockClear();
+    invalidateSpy.mockClear();
+    act(() => {
+      h.state.onFinish?.({
+        message: visibleMsg,
+        isAbort: false,
+        isDisconnect: true,
+        isError: false,
+      });
+    });
+    expect(onResumeFallback).toHaveBeenCalledWith(true);
+    expect(invalidateSpy).toHaveBeenCalledWith({
+      queryKey: ["ai-chat-messages", "c1"],
+    });
+    // Restore forbidden: the on-screen partial must not roll back.
+    expect(h.state.setMessages).not.toHaveBeenCalled();
+  });
+
+  it("an empty resumed message (starved replay) restores the stripped row AND arms the poll", () => {
+    h.state.status = "ready";
+    const { onResumeFallback } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    h.state.setMessages.mockClear();
+    onResumeFallback.mockClear();
+    act(() => {
+      h.state.onFinish?.({
+        message: emptyMsg,
+        isAbort: false,
+        isDisconnect: false,
+        isError: false,
+      });
+    });
+    expect(h.state.setMessages).toHaveBeenCalledTimes(1); // restore
+    expect(onResumeFallback).toHaveBeenCalledWith(true); // arm
+  });
+
+  it("degraded-merge: merges the tail per initialRows update, and settles disarm the poll", async () => {
+    h.state.status = "ready";
+    const { rerender, onResumeFallback } = renderResumable(streamingTail());
+    // Arm reconcile via a 204.
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({ status: 204, ok: false }),
+    );
+    await act(async () => {
+      await h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+    h.state.setMessages.mockClear();
+    onResumeFallback.mockClear();
+
+    // A streaming-tail update: merge, poll stays armed.
+    rerender([
+      row("u1", "user", undefined, "hi"),
+      row("a1", "assistant", "streaming", "step 1\nstep 2"),
+    ]);
+    expect(h.state.setMessages).toHaveBeenCalledTimes(1);
+    expect(onResumeFallback).not.toHaveBeenCalledWith(false);
+
+    // A settled-tail update: merge + disarm.
+    h.state.setMessages.mockClear();
+    rerender([
+      row("u1", "user", undefined, "hi"),
+      row("a1", "assistant", "succeeded", "final"),
+    ]);
+    expect(h.state.setMessages).toHaveBeenCalledTimes(1);
+    expect(onResumeFallback).toHaveBeenCalledWith(false);
+  });
+
+  it("a local stream disarms both the merge and the poll", () => {
+    h.state.status = "streaming";
+    const { rerender, onResumeFallback } = renderResumable(streamingTail());
+    onResumeFallback.mockClear();
+    // A re-render while streaming: the reconciliation effect disarms.
+    rerender(streamingTail());
+    expect(onResumeFallback).toHaveBeenCalledWith(false);
+  });
+
+  it("Send now is hidden on a resumed turn but visible on a local stream", () => {
+    // Resumed turn: hidden.
+    renderThread({ autonomousRunsEnabled: true, initialRows: streamingTail() });
+    fireEvent.click(screen.getByTestId("queue-btn"));
+    expect(screen.queryByLabelText("Send now")).toBeNull();
+
+    // Local streaming turn (no resume): visible.
+    cleanup();
+    resetState();
+    renderThread({ initialRows: [] });
+    fireEvent.click(screen.getByTestId("queue-btn"));
+    expect(screen.getByLabelText("Send now")).toBeTruthy();
+  });
+
+  it("handleStop aborts the attach controller and calls onServerStop", async () => {
+    const { onServerStop } = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: streamingTail(),
+    });
+    // Establish an attach controller via a (pending) reconnect GET.
+    let abortSeen = false;
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockImplementation((_input: unknown, init: RequestInit) => {
+        init.signal?.addEventListener("abort", () => {
+          abortSeen = true;
+        });
+        return new Promise(() => undefined); // never resolves
+      }),
+    );
+    act(() => {
+      void h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+    fireEvent.click(screen.getByLabelText("Stop"));
+    expect(abortSeen).toBe(true);
+    expect(onServerStop).toHaveBeenCalledWith("c1");
+  });
+});
+
+// Helper: render a resumable thread and expose a rerender that only swaps
+// initialRows (the degraded-merge effect depends on it).
+function renderResumable(initialRows: IAiChatMessageRow[]) {
+  const onResumeFallback = vi.fn();
+  const queryClient = new QueryClient({
+    defaultOptions: { queries: { retry: false } },
+  });
+  const Wrapper = ({ rows }: { rows: IAiChatMessageRow[] }) => (
+    <QueryClientProvider client={queryClient}>
       <MantineProvider>
         <ChatThread
           chatId="c1"
-          initialRows={[]}
+          initialRows={rows}
+          autonomousRunsEnabled
           onTurnFinished={vi.fn()}
-          observedRow={observedRow as never}
+          onResumeFallback={onResumeFallback}
         />
-      </MantineProvider>,
-    );
-  }
-
-  it("merges the polled run message when this tab is a passive observer", () => {
-    renderObserver("ready");
-    expect(h.state.setMessages).toHaveBeenCalledTimes(1);
-    // The updater replaces/append the observed assistant row by id.
-    const updater = h.state.setMessages.mock.calls[0][0] as (
-      prev: { id: string; parts: { text: string }[] }[],
-    ) => { id: string; parts: { text: string }[] }[];
-    const merged = updater([{ id: "u1", parts: [{ text: "hi" }] }]);
-    expect(merged).toHaveLength(2);
-    expect(merged[1].id).toBe("a-run");
-    expect(merged[1].parts[0].text).toBe("step 1\nstep 2");
-  });
-
-  it("does NOT merge while THIS tab is the streamer (no double-render)", () => {
-    renderObserver("streaming");
-    expect(h.state.setMessages).not.toHaveBeenCalled();
-  });
-});
+      </MantineProvider>
+    </QueryClientProvider>
+  );
+  const view = render(<Wrapper rows={initialRows} />);
+  const rerender = (rows: IAiChatMessageRow[]) =>
+    act(() => view.rerender(<Wrapper rows={rows} />));
+  return { rerender, onResumeFallback };
+}

apps/client/src/features/ai-chat/components/chat-thread.tsx

@@ -1,4 +1,5 @@
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
+import { useQueryClient } from "@tanstack/react-query";
 import { generateId } from "ai";
 import { ActionIcon, Box, Group, Stack, Text, Tooltip } from "@mantine/core";
 import {
@@ -24,7 +25,14 @@ import {
 } from "@/features/ai-chat/utils/role-launch.ts";
 import { describeChatError } from "@/features/ai-chat/utils/error-message.ts";
 import { extractServerChatId } from "@/features/ai-chat/utils/adopt-chat-id.ts";
-import { mergeObservedMessage } from "@/features/ai-chat/utils/run-polling.ts";
+import { assistantMessageHasVisibleContent } from "@/features/ai-chat/utils/message-content.ts";
+import {
+  isStreamingTail,
+  isSettledAssistantTail,
+  seedRows,
+  mergeById,
+} from "@/features/ai-chat/utils/resume-helpers.ts";
+import { AI_CHAT_MESSAGES_RQ_KEY } from "@/features/ai-chat/queries/ai-chat-query.ts";
 import {
   dequeue,
   enqueueMessage,
@@ -87,19 +95,13 @@ interface ChatThreadProps {
    *  Copy/export button available mid-stream). Distinct from onTurnFinished,
    *  which fires only at the terminal outcome. */
   onServerChatId?: (serverChatId?: string) => void;
-  /** #184 reconnect-and-live-follow. When THIS tab reopened a chat whose agent
-   *  run is still going (it is a PASSIVE OBSERVER — it did not start the run here),
-   *  the parent polls the reconnect endpoint and feeds the run's incrementally-
-   *  persisted assistant message here; we merge it into the live list so new
-   *  steps/tool-calls appear as they are persisted. Null when there is nothing to
-   *  observe (no run, feature off, or this tab IS the streamer). The merge is
-   *  ADDITIONALLY guarded by our own `isStreaming`, so a stale value can never
-   *  fight the local stream when we are the streamer. */
-  observedRow?: IAiChatMessageRow | null;
-  /** Report this tab's live streaming status up to the parent, so it can stop
-   *  polling the run while WE are the active streamer (the SSE owns the view) and
-   *  resume once we go idle. Called from an effect on every transition. */
-  onStreamingChange?: (streaming: boolean) => void;
+  /** #184 phase 1.5: arm/disarm the parent's degraded-poll fallback for THIS
+   *  chat's window. Called `true` when a resume attempt could not attach to the
+   *  live run (attach 204 / starved-or-torn resumed finish), so the window starts
+   *  a dumb timed poll of the message history to follow the detached run to settle;
+   *  called `false` the moment a local stream starts or the terminal settled row is
+   *  merged (invariant 8). The window owns the timer + its 10-min cap. */
+  onResumeFallback?: (active: boolean) => void;
   /** #184: whether detached/autonomous agent runs are enabled for this workspace.
    *  When true the Stop button must additionally hit the AUTHORITATIVE server stop
    *  (via onServerStop) — aborting only the local SSE is just a client disconnect,
@@ -155,15 +157,58 @@ export default function ChatThread({
   assistantName,
   onTurnFinished,
   onServerChatId,
-  observedRow,
-  onStreamingChange,
+  onResumeFallback,
   autonomousRunsEnabled,
   onServerStop,
 }: ChatThreadProps) {
   const { t } = useTranslation();
+  const queryClient = useQueryClient();
+
+  // resume machinery refs (#184 phase 1.5)
+  const attachAbortRef = useRef<AbortController | null>(null);
+  const reconcileTailRef = useRef(false);
+  const noStreamHandledRef = useRef(false);
+  const onNoActiveStreamRef = useRef<(() => void) | null>(null);
+  // Live mount flag. The attach GET and the resumed `onFinish` are async and can
+  // land AFTER this thread unmounts (the parent remounts per chat via `key`); with
+  // chatIdRef then pointing at the NEW chat, an ungated late callback would arm a
+  // spurious poll + foreign invalidation on the newly-opened chat. Every parent-
+  // facing resume side-effect is gated on this.
+  const mountedRef = useRef(true);
+  const [resumedTurn, setResumedTurn] = useState(false);
+  const resumedTurnRef = useRef(false);
+  // Identity-stable pair setter (bare useState setter + ref write): it is closed
+  // over by the transport useMemo([]), so it MUST NOT capture state.
+  const setResumedTurnPair = useCallback((v: boolean) => {
+    resumedTurnRef.current = v;
+    setResumedTurn(v);
+  }, []);
+
+  // Mount-time resume gating (in refs — computed once for this mount; the parent
+  // remounts per chat via `key`).
+  //
+  // Attempt resume for any non-settled tail: a streaming tail (strip + expect
+  // live replay) or a user tail (the run may exist but its assistant row is not
+  // seeded yet — attach to the pre-opened registry entry and wait for frames).
+  // A settled assistant tail must NEVER resume: replaying a finished run into a
+  // store that already contains its message duplicates parts (SDK text-start
+  // always pushes a new part).
+  const stripRef = useRef(chatId !== null && isStreamingTail(initialRows ?? []));
+  const attemptResumeRef = useRef(
+    autonomousRunsEnabled === true &&
+      chatId !== null &&
+      !isSettledAssistantTail(initialRows ?? []),
+  );
+  const strippedRowRef = useRef<IAiChatMessageRow | null>(
+    stripRef.current ? (initialRows ?? [])[initialRows!.length - 1] : null,
+  );
 
   const initialMessages = useMemo<UIMessage[]>(
-    () => (initialRows ?? []).map(rowToUiMessage),
+    () =>
+      seedRows(
+        initialRows ?? [],
+        attemptResumeRef.current && stripRef.current,
+      ).map(rowToUiMessage),
     [initialRows],
   );
 
@@ -261,9 +306,12 @@ export default function ChatThread({
     const { head, rest } = dequeue(queuedRef.current);
     if (!head) return false;
     setQueue(rest);
+    // Local send: clear any resume-suppression flag so this genuine local turn's
+    // onFinish flushes normally (invariant 8).
+    setResumedTurnPair(false);
     sendMessageRef.current?.({ text: head.text });
     return true;
-  }, [setQueue]);
+  }, [setQueue, setResumedTurnPair]);
 
   const enqueue = useCallback(
     (text: string) => {
@@ -283,6 +331,47 @@ export default function ChatThread({
       new DefaultChatTransport<UIMessage>({
         api: "/api/ai-chat/stream",
         credentials: "include",
+        prepareReconnectToStreamRequest: () => ({
+          // SDK default URL uses the useChat STORE id — always build from the real chat id.
+          // ?expect=live&anchor=<row id> ONLY when we stripped a streaming tail: expect=live
+          // is the only case where a finished-retained replay is safe (the row is stripped,
+          // replay rebuilds it), and the anchor pins the replay to OUR run — a mismatching
+          // (newer) run must 204 into the restore+poll path instead of replaying a foreign
+          // transcript into this store.
+          api: `/api/ai-chat/runs/${chatIdRef.current}/stream${
+            stripRef.current
+              ? `?expect=live&anchor=${strippedRowRef.current!.id}`
+              : ""
+          }`,
+        }),
+        fetch: async (input: RequestInfo | URL, init: RequestInit = {}) => {
+          if ((init.method ?? "GET") !== "GET") return fetch(input, init); // send path untouched
+          // Reconnect GET: the SDK passes no AbortSignal, so wire our own controller
+          // for observer Stop / unmount abort.
+          const controller = new AbortController();
+          attachAbortRef.current = controller;
+          try {
+            const response = await fetch(input, {
+              ...init,
+              signal: controller.signal,
+            });
+            // No onFinish will come for a 204 (silent no-op) OR any non-2xx
+            // (5xx/502 — a server restart mid-attach). Both run the same
+            // no-active-stream recovery: restore the stripped row, invalidate, and
+            // arm the degraded poll (idempotent via noStreamHandledRef; its part-d
+            // also clears the resumedTurn flag). This is the restart-survival path
+            // the removed F7 latch used to guard — a transient attach failure must
+            // NOT drop the in-progress row or stop tracking the durable run.
+            if (response.status === 204 || !response.ok)
+              onNoActiveStreamRef.current?.();
+            return response;
+          } catch (err) {
+            // Network throw: same no-onFinish recovery, then rethrow so the SDK
+            // still surfaces the error to its own machinery.
+            onNoActiveStreamRef.current?.();
+            throw err;
+          }
+        },
         // Inject the chat id and the currently-open page alongside the useChat
         // messages so the server can resolve an existing chat (or create one
         // when null) and tell the agent which page "this page" refers to. Both
@@ -312,7 +401,15 @@ export default function ChatThread({
     [],
   );
 
-  const { messages, sendMessage, status, stop, error, setMessages } = useChat({
+  const {
+    messages,
+    sendMessage,
+    status,
+    stop,
+    error,
+    setMessages,
+    resumeStream,
+  } = useChat({
     // Stable per-mount key. Existing chats use their real id; new chats use a
     // generated client id (never `undefined`) so the store is NOT re-created on
     // every render mid-stream (see `chatStoreId` above).
@@ -330,6 +427,38 @@ export default function ChatThread({
     // would be wrong, so on Stop/disconnect/error the queue is left intact for
     // the user to decide.
     onFinish: ({ message, isAbort, isDisconnect, isError }) => {
+      // (1) Capture whether THIS finish belongs to a resumed (attach) turn and
+      // immediately clear the flag so it can never suppress a LATER local turn.
+      const wasResumed = resumedTurnRef.current;
+      setResumedTurnPair(false);
+      // (2) Recovery after a starved/torn resumed finish (invariant 9). The arm
+      // and the stripped-row restore are gated DIFFERENTLY. Skip entirely once
+      // unmounted (an abort-triggered onFinish landing after a chat switch must
+      // not arm a poll / invalidate on the new chat).
+      if (wasResumed && mountedRef.current) {
+        const hasVisibleContent = assistantMessageHasVisibleContent(message);
+        // ARM the reconcile + degraded poll when the resumed message carries no
+        // visible content (starved replay) OR the connection dropped mid-run — in
+        // both cases the poll must drive the row to its real terminal state.
+        if (isDisconnect || !hasVisibleContent) {
+          reconcileTailRef.current = true;
+          queryClient.invalidateQueries({
+            queryKey: AI_CHAT_MESSAGES_RQ_KEY(chatIdRef.current),
+          });
+          onResumeFallback?.(true);
+        }
+        // RESTORE the stripped streaming row ONLY when the resumed message has no
+        // visible content. On isDisconnect WITH visible content restore is
+        // FORBIDDEN: the live stream may have advanced far past the mount-time
+        // snapshot, so restoring would clobber on-screen content (invariant 9) —
+        // the arm above suffices, the poll reaches the true terminal.
+        if (!hasVisibleContent && strippedRowRef.current) {
+          setMessages((prev) =>
+            mergeById(prev, rowToUiMessage(strippedRowRef.current!)),
+          );
+        }
+      }
+      // (3) Standard branches.
       // Forward the authoritative server chatId (streamed on the assistant
       // message metadata) so the parent adopts the REAL created chat id for a new
       // chat — see adopt-chat-id.ts for the full #137 design. `threadKey` lets the
@@ -342,6 +471,10 @@ export default function ChatThread({
       else if (isAbort) setStopNotice("manual");
       else if (isDisconnect) setStopNotice("disconnect");
       else setStopNotice(null);
+      // A resumed turn NEVER flushes the queue (invariant 7): skip BOTH the
+      // flush-on-abort branch and the plain flush. The local streamer is the only
+      // tab that owns the queue.
+      if (wasResumed) return;
       // "Send now": WE triggered this abort to interrupt the current turn and
       // immediately send the promoted head. Flush it even though the turn was
       // aborted (the normal abort path below keeps the queue intact). The
@@ -423,26 +556,98 @@ export default function ChatThread({
 
   const isStreaming = status === "submitted" || status === "streaming";
 
-  // #184: report our live streaming status up so the parent stops polling the run
-  // while WE are the streamer (the SSE owns the view) and resumes once we go idle.
-  // Effect (not render) so it never updates parent state during our own render;
-  // fires on mount with `false`, which also re-syncs the parent after a chat
-  // switch remounts this thread (a fresh mount is idle until the user sends).
-  useEffect(() => {
-    onStreamingChange?.(isStreaming);
-  }, [isStreaming, onStreamingChange]);
+  // 204-handler (`onNoActiveStream`): the attach returned 204 — nothing live to
+  // resume (overflow / begin-failure / after retention / anchor-mismatch). One-
+  // shot via noStreamHandledRef (we do NOT null onNoActiveStreamRef). Exactly four
+  // parts. Kept in a ref (read by the transport's fetch closure) and refreshed
+  // each render below.
+  const onNoActiveStream = useCallback(() => {
+    // A late attach outcome after unmount must not arm a poll / invalidate on the
+    // now-different chat this thread's refs were reused for.
+    if (!mountedRef.current) return;
+    if (noStreamHandledRef.current) return;
+    noStreamHandledRef.current = true;
+    // (a) Restore the stripped streaming row to the store — ONLY when we actually
+    // stripped one (a user-tail 204 does NOT reach here with a stripped row, so do
+    // not dereference null).
+    if (strippedRowRef.current) {
+      setMessages((prev) =>
+        mergeById(prev, rowToUiMessage(strippedRowRef.current!)),
+      );
+    }
+    // (b) Reconcile the tail from the message history + invalidate it so the
+    // degraded poll starts from a fresh fetch.
+    reconcileTailRef.current = true;
+    queryClient.invalidateQueries({
+      queryKey: AI_CHAT_MESSAGES_RQ_KEY(chatIdRef.current),
+    });
+    // (c) Arm the degraded poll (a dumb timer with a 10-min cap in the window);
+    // the thread disarms it via onResumeFallback(false) on settle / local stream.
+    onResumeFallback?.(true);
+    // (d) 204 means onFinish will NOT fire — clear the suppression flag so it
+    // cannot swallow the NEXT local turn's queue flush.
+    setResumedTurnPair(false);
+  }, [setMessages, queryClient, onResumeFallback, setResumedTurnPair]);
+  onNoActiveStreamRef.current = onNoActiveStream;
 
-  // #184 passive-observer merge: when the parent feeds a polled run message (we
-  // reopened a chat whose run is still going and did NOT start it here), merge it
-  // into the live list so new steps/tool-calls appear as they are persisted. Hard-
-  // gated by `!isStreaming`: if THIS tab is actually the streamer, the local SSE
-  // owns the view and a stale observedRow must never overwrite it. `observedRow`
-  // is a stable per-poll object, so this runs once per poll, not per render.
+  // Mount effect: kick off the resume attempt for a non-settled tail. Marking the
+  // turn as resumed BEFORE resumeStream so onFinish (invariant 7/8) sees it.
   useEffect(() => {
-    if (isStreaming || !observedRow) return;
-    const observed = rowToUiMessage(observedRow);
-    setMessages((prev) => mergeObservedMessage(prev, observed));
-  }, [observedRow, isStreaming, setMessages]);
+    // Re-arm on (re)mount — StrictMode dev-mounts twice, and the cleanup below
+    // flips this false between the two.
+    mountedRef.current = true;
+    if (attemptResumeRef.current) {
+      setResumedTurnPair(true);
+      void resumeStream();
+    }
+    // Unmount: mark unmounted (gates late attach/onFinish side-effects) and abort
+    // the in-flight attach GET so its callbacks don't fire against the next chat.
+    return () => {
+      mountedRef.current = false;
+      attachAbortRef.current?.abort();
+    };
+    // Mount-only by design; the parent remounts per chat via `key`.
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, []);
+
+  // Reconciliation + degraded-merge (invariant 8). Deps are EXACTLY
+  // [initialRows, isStreaming, setMessages].
+  useEffect(() => {
+    // A local stream owns the view: disarm BOTH the merge and the window poll.
+    if (isStreaming) {
+      reconcileTailRef.current = false;
+      onResumeFallback?.(false);
+      return;
+    }
+    if (!reconcileTailRef.current) return;
+    const rows = initialRows ?? [];
+    const tail = rows[rows.length - 1];
+    if (!tail || tail.role !== "assistant") return;
+    // Merge the polled assistant tail on EVERY initialRows update — while the
+    // degraded poll is active this IS the live per-step progress.
+    setMessages((prev) => mergeById(prev, rowToUiMessage(tail)));
+    // Anchor-mismatch coherence: when we restored a stripped streaming row A but a
+    // DIFFERENT run's row B is now the tail (A finished, B replaced the registry
+    // entry, so the attach 204'd), A would otherwise linger forever as an orphan
+    // jumping-dots row over the real run. Settle it from fresh history (where A is
+    // now persisted) so no phantom row survives. No-op in the common case where A
+    // IS the tail (id match).
+    const stripped = strippedRowRef.current;
+    if (stripped && stripped.id !== tail.id) {
+      const historical = rows.find((r) => r.id === stripped.id);
+      if (historical)
+        setMessages((prev) => mergeById(prev, rowToUiMessage(historical)));
+    }
+    // Settled: the terminal merge is done — disarm the flag AND the window poll
+    // explicitly (the window only has a time cap, it will not disarm itself).
+    if (tail.status !== "streaming") {
+      reconcileTailRef.current = false;
+      onResumeFallback?.(false);
+    }
+    // onResumeFallback intentionally omitted (parent-stable callback); deps are
+    // fixed by the resume design.
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [initialRows, isStreaming, setMessages]);
 
   // "Send now" on a queued message: interrupt the current turn and immediately
   // send THIS message, keeping the agent's partial output. Other queued messages
@@ -469,10 +674,12 @@ export default function ChatThread({
         const msg = queuedRef.current.find((m) => m.id === id);
         if (!msg) return;
         setQueue(removeQueuedById(queuedRef.current, id));
+        // Local send: clear any resume-suppression flag (invariant 8).
+        setResumedTurnPair(false);
         sendMessageRef.current?.({ text: msg.text });
       }
     },
-    [setQueue, stop],
+    [setQueue, stop, setResumedTurnPair],
   );
 
   // Stop the current turn. ALWAYS abort the local SSE (`stop()`) so the composer
@@ -485,6 +692,9 @@ export default function ChatThread({
   // is not known yet — a brand-new chat in the first moment of its first turn —
   // only the local abort happens (there is no server-side run handle to stop yet).
   const handleStop = useCallback(() => {
+    // Abort the resume/attach GET first: the SDK does not pass it a signal, so an
+    // observer's Stop would otherwise leave the attach fetch running.
+    attachAbortRef.current?.abort();
     stop();
     if (!autonomousRunsEnabled) return;
     if (chatIdRef.current) {
@@ -617,17 +827,23 @@ export default function ChatThread({
                 <Text size="xs" lineClamp={2} className={classes.queuedText}>
                   {m.text}
                 </Text>
-                <Tooltip label={t("Interrupt and send now")} withArrow>
-                  <ActionIcon
-                    size="xs"
-                    variant="subtle"
-                    color="blue"
-                    onClick={() => sendNow(m.id)}
-                    aria-label={t("Send now")}
-                  >
-                    <IconPlayerPlayFilled size={12} />
-                  </ActionIcon>
-                </Tooltip>
+                {/* "Send now" (interrupt) is hidden on a RESUMED turn: a local
+                    stop() does not abort the resumed attach fetch, so the click
+                    would be swallowed while flushOnAbortRef would fire minutes
+                    later on the natural finish. Only the remove affordance stays. */}
+                {!resumedTurn && (
+                  <Tooltip label={t("Interrupt and send now")} withArrow>
+                    <ActionIcon
+                      size="xs"
+                      variant="subtle"
+                      color="blue"
+                      onClick={() => sendNow(m.id)}
+                      aria-label={t("Send now")}
+                    >
+                      <IconPlayerPlayFilled size={12} />
+                    </ActionIcon>
+                  </Tooltip>
+                )}
                 <ActionIcon
                   size="xs"
                   variant="subtle"
@@ -642,7 +858,11 @@ export default function ChatThread({
           </Stack>
         )}
         <ChatInput
-          onSend={(text) => sendMessage({ text })}
+          onSend={(text) => {
+            // Local send: clear any resume-suppression flag (invariant 8).
+            setResumedTurnPair(false);
+            sendMessage({ text });
+          }}
           onQueue={enqueue}
           onStop={handleStop}
           isStreaming={isStreaming}

apps/client/src/features/ai-chat/queries/ai-chat-messages-query.test.tsx

@@ -0,0 +1,90 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import React from "react";
+import { renderHook, waitFor } from "@testing-library/react";
+import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
+
+// react-i18next / notifications are pulled in transitively by ai-chat-query.ts
+// (the mutation hooks use them); stub so the module imports cleanly.
+vi.mock("react-i18next", () => ({
+  useTranslation: () => ({ t: (key: string) => key }),
+}));
+vi.mock("@mantine/notifications", () => ({
+  notifications: { show: vi.fn() },
+}));
+
+// Mock the service module; only getAiChatMessages is exercised, but the other
+// named exports must exist so ai-chat-query.ts imports resolve.
+vi.mock("@/features/ai-chat/services/ai-chat-service.ts", () => ({
+  getAiChatMessages: vi.fn(),
+  getAiChats: vi.fn(),
+  getAiRoleCatalog: vi.fn(),
+  getAiRoleCatalogBundle: vi.fn(),
+  getAiRoles: vi.fn(),
+  importAiRolesFromCatalog: vi.fn(),
+  createAiRole: vi.fn(),
+  deleteAiChat: vi.fn(),
+  deleteAiRole: vi.fn(),
+  renameAiChat: vi.fn(),
+  updateAiRole: vi.fn(),
+  updateAiRoleFromCatalog: vi.fn(),
+}));
+
+import { getAiChatMessages } from "@/features/ai-chat/services/ai-chat-service.ts";
+import { useAiChatMessagesQuery } from "@/features/ai-chat/queries/ai-chat-query.ts";
+
+const emptyPage = { items: [], meta: { hasNextPage: false, nextCursor: null } };
+
+function createWrapper() {
+  const queryClient = new QueryClient({
+    defaultOptions: { queries: { retry: false } },
+  });
+  return function Wrapper({ children }: { children: React.ReactNode }) {
+    return (
+      <QueryClientProvider client={queryClient}>{children}</QueryClientProvider>
+    );
+  };
+}
+
+// The degraded-poll fallback (#184 phase 1.5) is threaded into this query as a
+// `refetchInterval`; AiChatWindow supplies the deliberately-dumb callback. These
+// pin the plumbing the window depends on: the interval polls the message history,
+// and — critically — fetch ERRORS do NOT stop the tick (TanStack v5 resets the
+// failure count each fetch, so the poll must survive a server restart).
+describe("useAiChatMessagesQuery — degraded refetchInterval", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("re-polls at the interval while the callback returns a duration", async () => {
+    vi.mocked(getAiChatMessages).mockResolvedValue(emptyPage as never);
+    renderHook(() => useAiChatMessagesQuery("c1", () => 30), {
+      wrapper: createWrapper(),
+    });
+    await waitFor(() =>
+      expect(vi.mocked(getAiChatMessages).mock.calls.length).toBeGreaterThan(2),
+    );
+  });
+
+  it("does NOT re-poll when the callback returns false", async () => {
+    vi.mocked(getAiChatMessages).mockResolvedValue(emptyPage as never);
+    renderHook(() => useAiChatMessagesQuery("c1", () => false), {
+      wrapper: createWrapper(),
+    });
+    await waitFor(() =>
+      expect(vi.mocked(getAiChatMessages)).toHaveBeenCalledTimes(1),
+    );
+    // Give any errant interval a chance to fire, then assert it did not.
+    await new Promise((r) => setTimeout(r, 60));
+    expect(vi.mocked(getAiChatMessages)).toHaveBeenCalledTimes(1);
+  });
+
+  it("keeps ticking through fetch errors (errors do not gate the poll)", async () => {
+    vi.mocked(getAiChatMessages).mockRejectedValue(new Error("server down"));
+    renderHook(() => useAiChatMessagesQuery("c1", () => 30), {
+      wrapper: createWrapper(),
+    });
+    await waitFor(() =>
+      expect(vi.mocked(getAiChatMessages).mock.calls.length).toBeGreaterThan(2),
+    );
+  });
+});

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";
@@ -12,7 +13,6 @@ import {
   deleteAiChat,
   deleteAiRole,
   getAiChatMessages,
-  getAiChatRun,
   getAiChats,
   getAiRoleCatalog,
   getAiRoleCatalogBundle,
@@ -25,7 +25,6 @@ import {
 import {
   IAiChat,
   IAiChatMessageRow,
-  IAiChatRunResponse,
   IAiRole,
   IAiRoleCatalog,
   IAiRoleCatalogBundle,
@@ -36,7 +35,6 @@ import {
   IAiRoleUpdateFromCatalogResult,
 } from "@/features/ai-chat/types/ai-chat.types.ts";
 import { IPagination } from "@/lib/types.ts";
-import { runPollInterval } from "@/features/ai-chat/utils/run-polling.ts";
 
 export const AI_CHATS_RQ_KEY = ["ai-chats"];
 export const AI_ROLES_RQ_KEY = ["ai-roles"];
@@ -54,7 +52,6 @@ export const AI_CHAT_MESSAGES_RQ_KEY = (chatId: string) => [
   "ai-chat-messages",
   chatId,
 ];
-export const AI_CHAT_RUN_RQ_KEY = (chatId: string) => ["ai-chat-run", chatId];
 
 /** Paginated list of the current user's chats (auto-loads further pages). */
 export function useAiChatsQuery() {
@@ -88,7 +85,15 @@ export function useAiChatsQuery() {
  * Load all persisted messages of a chat (oldest first), flattening the
  * paginated server response. Used to seed `useChat` initial messages.
  */
-export function useAiChatMessagesQuery(chatId: string | undefined) {
+export function useAiChatMessagesQuery(
+  chatId: string | undefined,
+  // #184 phase 1.5: the degraded-poll fallback. When a tab could not attach to a
+  // still-running run (the attach returned 204 / the resumed stream ended with no
+  // terminal row), the window arms a dumb timed poll of the message history to
+  // follow the detached run to settle. The callback form lives in AiChatWindow;
+  // threaded here verbatim so this query owns the polling. Undefined => no poll.
+  refetchInterval?: number | false | (() => number | false),
+) {
   const query = useInfiniteQuery({
     queryKey: AI_CHAT_MESSAGES_RQ_KEY(chatId ?? ""),
     queryFn: ({ pageParam }) =>
@@ -99,6 +104,7 @@ export function useAiChatMessagesQuery(chatId: string | undefined) {
         ? (lastPage.meta.nextCursor ?? undefined)
         : undefined,
     enabled: !!chatId,
+    refetchInterval,
   });
 
   // useInfiniteQuery only fetches the first page on its own. The hook's contract
@@ -138,34 +144,6 @@ export function useAiChatMessagesQuery(chatId: string | undefined) {
   };
 }
 
-/**
- * Reconnect to a chat's latest agent run and LIVE-FOLLOW it (#184). While the run
- * is active the query re-polls every {@link runPollInterval} ms (driven off the
- * fetched `run.status`, the same status-keyed refetchInterval pattern as the
- * embeddings reindex polling); once the run reaches a terminal status — or there
- * is no run — the interval returns `false` and polling stops on its own. Polling
- * is thus naturally bounded by the run terminating; no separate timeout cap.
- *
- * `enabled` gates the whole thing: callers pass `false` when the autonomous-runs
- * feature is off (the endpoint is NOT flag-gated server-side, but with the feature
- * off the chat has no runs, so polling would only ever return `{ run: null }`) OR
- * when THIS tab is the one actively streaming the run (the live SSE owns the view,
- * so we must not also poll/merge). The global `retry: false` means a failed fetch
- * leaves `data` undefined, so refetchInterval(undefined run) returns false — a
- * failed fetch can never spin a tight loop.
- */
-export function useAiChatRunQuery(
-  chatId: string | undefined,
-  enabled: boolean,
-) {
-  return useQuery<IAiChatRunResponse, Error>({
-    queryKey: AI_CHAT_RUN_RQ_KEY(chatId ?? ""),
-    queryFn: () => getAiChatRun(chatId as string),
-    enabled: !!chatId && enabled,
-    refetchInterval: (query) => runPollInterval(query.state.data?.run),
-  });
-}
-
 export function useRenameAiChatMutation() {
   const queryClient = useQueryClient();
   const { t } = useTranslation();
@@ -307,6 +285,29 @@ 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/ai-chat-run-query.test.tsx

@@ -1,92 +0,0 @@
-import { describe, it, expect, vi, beforeEach } from "vitest";
-import React from "react";
-import { renderHook, waitFor } from "@testing-library/react";
-import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
-import type { IAiChatRunResponse } from "@/features/ai-chat/types/ai-chat.types.ts";
-
-// react-i18next is pulled in transitively by ai-chat-query.ts (the mutation hooks
-// use it); stub it so the module imports cleanly in this hook test.
-vi.mock("react-i18next", () => ({
-  useTranslation: () => ({ t: (key: string) => key }),
-}));
-
-vi.mock("@mantine/notifications", () => ({
-  notifications: { show: vi.fn() },
-}));
-
-// Mock the whole service module; only getAiChatRun is exercised here, but the
-// other named exports must exist so ai-chat-query.ts imports resolve.
-vi.mock("@/features/ai-chat/services/ai-chat-service.ts", () => ({
-  getAiChatRun: vi.fn(),
-  getAiChatMessages: vi.fn(),
-  getAiChats: vi.fn(),
-  getAiRoleCatalog: vi.fn(),
-  getAiRoleCatalogBundle: vi.fn(),
-  getAiRoles: vi.fn(),
-  importAiRolesFromCatalog: vi.fn(),
-  createAiRole: vi.fn(),
-  deleteAiChat: vi.fn(),
-  deleteAiRole: vi.fn(),
-  renameAiChat: vi.fn(),
-  updateAiRole: vi.fn(),
-  updateAiRoleFromCatalog: vi.fn(),
-}));
-
-import { getAiChatRun } from "@/features/ai-chat/services/ai-chat-service.ts";
-import { useAiChatRunQuery } from "@/features/ai-chat/queries/ai-chat-query.ts";
-
-function createWrapper() {
-  const queryClient = new QueryClient({
-    defaultOptions: { queries: { retry: false } },
-  });
-  return function Wrapper({ children }: { children: React.ReactNode }) {
-    return (
-      <QueryClientProvider client={queryClient}>{children}</QueryClientProvider>
-    );
-  };
-}
-
-const runningResponse: IAiChatRunResponse = {
-  run: { id: "run-1", chatId: "c1", status: "running" },
-  message: {
-    id: "a1",
-    role: "assistant",
-    content: "working...",
-    createdAt: "2026-01-01T00:00:00Z",
-  },
-};
-
-describe("useAiChatRunQuery — enable gating", () => {
-  beforeEach(() => {
-    vi.clearAllMocks();
-  });
-
-  it("fetches the run when enabled (passive observer, feature on)", async () => {
-    vi.mocked(getAiChatRun).mockResolvedValue(runningResponse);
-    const { result } = renderHook(() => useAiChatRunQuery("c1", true), {
-      wrapper: createWrapper(),
-    });
-    await waitFor(() => expect(result.current.isSuccess).toBe(true));
-    expect(getAiChatRun).toHaveBeenCalledWith("c1");
-    expect(result.current.data?.run?.status).toBe("running");
-  });
-
-  it("does NOT fetch when disabled (this tab is the streamer / feature off)", async () => {
-    vi.mocked(getAiChatRun).mockResolvedValue(runningResponse);
-    renderHook(() => useAiChatRunQuery("c1", false), {
-      wrapper: createWrapper(),
-    });
-    // Give any errant fetch a chance to fire, then assert none did.
-    await new Promise((r) => setTimeout(r, 20));
-    expect(getAiChatRun).not.toHaveBeenCalled();
-  });
-
-  it("does NOT fetch when there is no chat id", async () => {
-    vi.mocked(getAiChatRun).mockResolvedValue(runningResponse);
-    renderHook(() => useAiChatRunQuery(undefined, true), {
-      wrapper: createWrapper(),
-    });
-    await new Promise((r) => setTimeout(r, 20));
-    expect(getAiChatRun).not.toHaveBeenCalled();
-  });
-});

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/services/ai-chat-service.ts

@@ -5,7 +5,6 @@ import {
   IAiChatListParams,
   IAiChatMessageRow,
   IAiChatMessagesParams,
-  IAiChatRunResponse,
   IAiRole,
   IAiRoleCatalog,
   IAiRoleCatalogBundle,
@@ -43,23 +42,6 @@ export async function getAiChatMessages(
   return req.data;
 }
 
-/**
- * Reconnect to the latest agent run of a chat (#184). Returns the run's
- * persisted lifecycle state and the assistant message it materializes (the
- * partial output while the run is in-flight, the final output once it finished).
- * The DB is the source of truth, so this works for an in-flight run (the browser
- * dropped, the run kept going) and a finished one alike; `{ run: null }` when the
- * chat has never had a run. Owner-gated server-side (the requesting user must own
- * the chat); it is NOT flag-gated — when the feature is off the chat simply has no
- * runs, so the endpoint returns `{ run: null }`.
- */
-export async function getAiChatRun(
-  chatId: string,
-): Promise<IAiChatRunResponse> {
-  const req = await api.post<IAiChatRunResponse>("/ai-chat/run", { chatId });
-  return req.data;
-}
-
 /**
  * Explicitly STOP the active agent run of a chat (#184). This is the ONLY thing
  * that ends a DETACHED run — a mere browser disconnect (aborting the local SSE)

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";
+  }[];
 }
 
 /**
@@ -197,41 +210,14 @@ export interface IAiChatMessageRow {
     // renders a "stopped" marker on interrupted turns.
     finishReason?: string;
   } | null;
+  // Persisted lifecycle status of the row's turn, carried on the wire by
+  // `baseFields`. 'streaming' marks a still-in-progress assistant row (used by
+  // the resume machinery to decide whether a tail is a live stream to attach to
+  // or a settled row that must not be replayed).
+  status?: string;
   createdAt: string;
 }
 
-/**
- * A persisted agent-run row (#184), mirroring the `ai_chat_runs` fields the
- * client reads from `POST /ai-chat/run`. Only `status` is load-bearing for the
- * reconnect-and-live-update UX (it drives the poll cadence); the rest are carried
- * for display/diagnostics. The DB is the source of truth, so this resolves for an
- * in-flight run (the browser dropped, the run kept going) and a finished one.
- */
-export interface IAiChatRun {
-  id: string;
-  chatId: string;
-  // 'pending' | 'running' | 'succeeded' | 'failed' | 'aborted'. The first two are
-  // ACTIVE (keep polling); the rest are TERMINAL (stop polling).
-  status: "pending" | "running" | "succeeded" | "failed" | "aborted" | string;
-  error?: string | null;
-  stepCount?: number;
-  assistantMessageId?: string | null;
-  startedAt?: string | null;
-  finishedAt?: string | null;
-  createdAt?: string;
-  updatedAt?: string;
-}
-
-/**
- * Response of `POST /ai-chat/run` (#184): the latest run of a chat and the
- * assistant message it materializes (the partial/final output, projected from the
- * persisted rows). Both are `null` when the chat has never had a run.
- */
-export interface IAiChatRunResponse {
-  run: IAiChatRun | null;
-  message: IAiChatMessageRow | null;
-}
-
 export interface IAiChatListParams extends QueryParams {}
 
 export interface IAiChatMessagesParams {

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/ai-chat/utils/resume-helpers.test.ts

@@ -0,0 +1,112 @@
+import { describe, it, expect } from "vitest";
+import type { UIMessage } from "@ai-sdk/react";
+import type { IAiChatMessageRow } from "@/features/ai-chat/types/ai-chat.types.ts";
+import {
+  isStreamingTail,
+  isSettledAssistantTail,
+  seedRows,
+  mergeById,
+} from "./resume-helpers.ts";
+
+function row(
+  id: string,
+  role: string,
+  status?: string,
+): IAiChatMessageRow {
+  return { id, role, content: "", status, createdAt: "2026-01-01T00:00:00Z" };
+}
+
+function makeMsg(id: string, text: string): UIMessage {
+  return {
+    id,
+    role: "assistant",
+    parts: [{ type: "text", text }],
+  } as UIMessage;
+}
+
+describe("isStreamingTail", () => {
+  it("is true when the last row is a streaming assistant row", () => {
+    expect(
+      isStreamingTail([row("u1", "user"), row("a1", "assistant", "streaming")]),
+    ).toBe(true);
+  });
+
+  it("is false for a settled assistant tail", () => {
+    expect(isStreamingTail([row("a1", "assistant", "succeeded")])).toBe(false);
+    expect(isStreamingTail([row("a1", "assistant")])).toBe(false);
+  });
+
+  it("is false when the tail is a user row or the list is empty", () => {
+    expect(isStreamingTail([row("u1", "user")])).toBe(false);
+    expect(isStreamingTail([])).toBe(false);
+  });
+});
+
+describe("isSettledAssistantTail", () => {
+  it("is true for an assistant tail whose status is not streaming", () => {
+    expect(isSettledAssistantTail([row("a1", "assistant", "succeeded")])).toBe(
+      true,
+    );
+    expect(isSettledAssistantTail([row("a1", "assistant")])).toBe(true);
+    expect(isSettledAssistantTail([row("a1", "assistant", "aborted")])).toBe(
+      true,
+    );
+  });
+
+  it("is false for a streaming assistant tail", () => {
+    expect(isSettledAssistantTail([row("a1", "assistant", "streaming")])).toBe(
+      false,
+    );
+  });
+
+  it("is false when the tail is a user row or the list is empty", () => {
+    expect(isSettledAssistantTail([row("u1", "user")])).toBe(false);
+    expect(isSettledAssistantTail([])).toBe(false);
+  });
+});
+
+describe("seedRows", () => {
+  const rows = [row("u1", "user"), row("a1", "assistant", "streaming")];
+
+  it("returns the rows unchanged when not stripping", () => {
+    expect(seedRows(rows, false)).toBe(rows);
+  });
+
+  it("drops the last row when stripping", () => {
+    const seeded = seedRows(rows, true);
+    expect(seeded).toHaveLength(1);
+    expect(seeded[0].id).toBe("u1");
+  });
+
+  it("returns an empty list when stripping a single-row list", () => {
+    expect(seedRows([row("a1", "assistant", "streaming")], true)).toHaveLength(
+      0,
+    );
+  });
+});
+
+describe("mergeById", () => {
+  it("replaces the message with the same id in place (per-step growth)", () => {
+    const prev = [makeMsg("u1", "hi"), makeMsg("a1", "step 1")];
+    const incoming = makeMsg("a1", "step 1\nstep 2");
+    const next = mergeById(prev, incoming);
+    expect(next).toHaveLength(2);
+    expect(next[1]).toBe(incoming);
+    expect(next[0]).toBe(prev[0]); // untouched
+    expect(next).not.toBe(prev); // new array (never mutates input)
+  });
+
+  it("appends when the incoming message is not yet present", () => {
+    const prev = [makeMsg("u1", "hi")];
+    const incoming = makeMsg("a1", "first token");
+    const next = mergeById(prev, incoming);
+    expect(next).toHaveLength(2);
+    expect(next[1]).toBe(incoming);
+  });
+
+  it("returns the original list unchanged when there is nothing to merge", () => {
+    const prev = [makeMsg("u1", "hi")];
+    expect(mergeById(prev, null)).toBe(prev);
+    expect(mergeById(prev, undefined)).toBe(prev);
+  });
+});

apps/client/src/features/ai-chat/utils/resume-helpers.ts

@@ -0,0 +1,62 @@
+import type { UIMessage } from "@ai-sdk/react";
+import type { IAiChatMessageRow } from "@/features/ai-chat/types/ai-chat.types.ts";
+
+/**
+ * Pure decisions for the resumable-SSE resume machinery (#184 phase 1.5). A tab
+ * that reopens a chat whose agent run is still going attaches to the server's
+ * run-stream registry (replay + live tail) instead of polling snapshots; these
+ * small predicates decide WHICH tail is safe to resume and how to seed the store,
+ * extracted so they can be unit-tested in isolation.
+ */
+
+/**
+ * A STREAMING tail: the last persisted row is an assistant row still marked
+ * `status === 'streaming'`. Such a tail is stripped from the seed and rebuilt by
+ * the replay (`expect=live`), since the SDK's `text-start` always pushes a new
+ * part and replaying over a seeded in-progress row would duplicate its text.
+ */
+export function isStreamingTail(rows: IAiChatMessageRow[]): boolean {
+  const tail = rows[rows.length - 1];
+  return !!tail && tail.role === "assistant" && tail.status === "streaming";
+}
+
+/**
+ * A SETTLED assistant tail: the last row is an assistant row whose status is
+ * anything OTHER than 'streaming'. A settled assistant tail must NEVER resume —
+ * replaying a finished run into a store that already holds its message duplicates
+ * parts (`text-start` always pushes a new part).
+ */
+export function isSettledAssistantTail(rows: IAiChatMessageRow[]): boolean {
+  const tail = rows[rows.length - 1];
+  return !!tail && tail.role === "assistant" && tail.status !== "streaming";
+}
+
+/**
+ * Seed rows for `useChat`: return the rows unchanged, or without the last row when
+ * `strip` is set (the streaming tail is stripped so the live replay rebuilds it
+ * without duplicating parts).
+ */
+export function seedRows(
+  rows: IAiChatMessageRow[],
+  strip: boolean,
+): IAiChatMessageRow[] {
+  return strip ? rows.slice(0, -1) : rows;
+}
+
+/**
+ * Merge an assistant message into the rendered list by id: replace the message
+ * with the same id in place (the in-progress assistant row is already seeded from
+ * history, so per-step growth replaces it), or append it when absent. Returns a
+ * new array; the input is never mutated.
+ */
+export function mergeById(
+  messages: UIMessage[],
+  incoming: UIMessage | null | undefined,
+): UIMessage[] {
+  if (!incoming) return messages;
+  const idx = messages.findIndex((m) => m.id === incoming.id);
+  if (idx === -1) return [...messages, incoming];
+  const next = messages.slice();
+  next[idx] = incoming;
+  return next;
+}

apps/client/src/features/ai-chat/utils/run-polling.test.ts

@@ -1,303 +0,0 @@
-import { describe, it, expect } from "vitest";
-import type { UIMessage } from "@ai-sdk/react";
-import type { IAiChatRun } from "@/features/ai-chat/types/ai-chat.types.ts";
-import {
-  RUN_POLL_INTERVAL_MS,
-  isRunActive,
-  runPollInterval,
-  shouldObserveRun,
-  shouldClearStoppingLatch,
-  shouldClearLatchOnQueryError,
-  mergeObservedMessage,
-} from "./run-polling.ts";
-
-function makeRun(status: string): IAiChatRun {
-  return { id: "run-1", chatId: "c1", status };
-}
-
-function makeMsg(id: string, text: string): UIMessage {
-  return {
-    id,
-    role: "assistant",
-    parts: [{ type: "text", text }],
-  } as UIMessage;
-}
-
-describe("isRunActive", () => {
-  it("treats pending and running as active", () => {
-    expect(isRunActive(makeRun("pending"))).toBe(true);
-    expect(isRunActive(makeRun("running"))).toBe(true);
-  });
-
-  it("treats terminal / unknown / nullish as not active", () => {
-    expect(isRunActive(makeRun("succeeded"))).toBe(false);
-    expect(isRunActive(makeRun("failed"))).toBe(false);
-    expect(isRunActive(makeRun("aborted"))).toBe(false);
-    expect(isRunActive(makeRun("weird-future-status"))).toBe(false);
-    expect(isRunActive(null)).toBe(false);
-    expect(isRunActive(undefined)).toBe(false);
-  });
-});
-
-describe("runPollInterval (the refetchInterval helper)", () => {
-  it("returns 2000ms while the run is pending/running", () => {
-    expect(runPollInterval(makeRun("pending"))).toBe(RUN_POLL_INTERVAL_MS);
-    expect(runPollInterval(makeRun("running"))).toBe(RUN_POLL_INTERVAL_MS);
-    expect(RUN_POLL_INTERVAL_MS).toBe(2000);
-  });
-
-  it("returns false (stop polling) once the run is terminal", () => {
-    expect(runPollInterval(makeRun("succeeded"))).toBe(false);
-    expect(runPollInterval(makeRun("failed"))).toBe(false);
-    expect(runPollInterval(makeRun("aborted"))).toBe(false);
-  });
-
-  it("returns false (no polling) when there is no run", () => {
-    expect(runPollInterval(null)).toBe(false);
-    expect(runPollInterval(undefined)).toBe(false);
-  });
-});
-
-describe("shouldObserveRun (observer-vs-streamer decision)", () => {
-  it("observes an active run when this tab is NOT the local streamer", () => {
-    expect(shouldObserveRun(makeRun("running"), false)).toBe(true);
-    expect(shouldObserveRun(makeRun("pending"), false)).toBe(true);
-  });
-
-  it("observes a terminal run too (so the final output shows on reopen)", () => {
-    expect(shouldObserveRun(makeRun("succeeded"), false)).toBe(true);
-  });
-
-  it("does NOT observe when this tab IS the streamer (no double-render)", () => {
-    expect(shouldObserveRun(makeRun("running"), true)).toBe(false);
-    expect(shouldObserveRun(makeRun("succeeded"), true)).toBe(false);
-  });
-
-  it("does NOT observe when there is no run", () => {
-    expect(shouldObserveRun(null, false)).toBe(false);
-    expect(shouldObserveRun(undefined, false)).toBe(false);
-  });
-});
-
-describe("shouldClearStoppingLatch (#234 latch-release decision)", () => {
-  // The one case the latch SHOULD clear: we requested a stop, we are the passive
-  // observer (not streaming), and the CURRENT run is terminal.
-  it("clears only when stopping, observing, and the run is terminal", () => {
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: makeRun("aborted"),
-        isLocalStreaming: false,
-      }),
-    ).toBe(true);
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: makeRun("succeeded"),
-        isLocalStreaming: false,
-      }),
-    ).toBe(true);
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: makeRun("failed"),
-        isLocalStreaming: false,
-      }),
-    ).toBe(true);
-  });
-
-  // Round-3 regression: clearing while THIS tab is still the local streamer would
-  // re-open the flash for the current turn the moment we switch to observer role.
-  // A predicate lacking the streaming gate would (wrongly) return true here.
-  it("does NOT clear while this tab is the local streamer", () => {
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: makeRun("aborted"),
-        isLocalStreaming: true,
-      }),
-    ).toBe(false);
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: makeRun("succeeded"),
-        isLocalStreaming: true,
-      }),
-    ).toBe(false);
-  });
-
-  // The detached run keeps growing after a local abort — while it is still
-  // active the latch MUST hold so the observer merge stays suppressed.
-  it("does NOT clear while the run is still active", () => {
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: makeRun("running"),
-        isLocalStreaming: false,
-      }),
-    ).toBe(false);
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: makeRun("pending"),
-        isLocalStreaming: false,
-      }),
-    ).toBe(false);
-  });
-
-  // #234 F4: on Stop the stale PREVIOUS-turn run is removed from the cache, so the
-  // observed `run` is null until the current turn's run is fetched fresh. A null
-  // run HOLDS the latch — it can never clear against the just-removed stale run,
-  // only against the current turn's own terminal run once observed.
-  it("does NOT clear against a removed/absent run (F4 stale-run guard)", () => {
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: null,
-        isLocalStreaming: false,
-      }),
-    ).toBe(false);
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: true,
-        run: undefined,
-        isLocalStreaming: false,
-      }),
-    ).toBe(false);
-  });
-
-  it("does NOT clear when no stop was requested", () => {
-    expect(
-      shouldClearStoppingLatch({
-        stoppingRun: false,
-        run: makeRun("aborted"),
-        isLocalStreaming: false,
-      }),
-    ).toBe(false);
-  });
-});
-
-describe("shouldClearLatchOnQueryError (#234 F7 error-safety-net decision)", () => {
-  // This guards the REAL anti-flash decision the component's run-query-error
-  // safety-net effect uses (ai-chat-window.tsx wires the effect to THIS helper,
-  // not a copy — so the test is non-vacuous vs the live code).
-
-  // (b) The F7 hole: a TRANSIENT run-query error while `run` is STILL ACTIVE must
-  // NOT clear the latch. TanStack Query v5 retains `data` on error, so
-  // runQueryFailed can be true while the held run is still pending/running.
-  // Against the PRE-F7 condition (without `!isRunActive(run)`) this would return
-  // true — so this assertion fails on the buggy code (non-vacuous).
-  it("does NOT clear on a transient error while the run is still ACTIVE (F7)", () => {
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: true,
-        isLocalStreaming: false,
-        runQueryFailed: true,
-        run: makeRun("running"),
-      }),
-    ).toBe(false);
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: true,
-        isLocalStreaming: false,
-        runQueryFailed: true,
-        run: makeRun("pending"),
-      }),
-    ).toBe(false);
-  });
-
-  // (a) The genuine permanent-null-freeze: run cache cleared by removeQueries +
-  // the refetch keeps ERRORING, so `run === null`. This is the ONLY case the
-  // safety-net exists to cure — it MUST clear so the frozen view resumes.
-  it("clears on a permanent error when the run is null (permanent-null-freeze)", () => {
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: true,
-        isLocalStreaming: false,
-        runQueryFailed: true,
-        run: null,
-      }),
-    ).toBe(true);
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: true,
-        isLocalStreaming: false,
-        runQueryFailed: true,
-        run: undefined,
-      }),
-    ).toBe(true);
-  });
-
-  // A TERMINAL run also satisfies `!isRunActive`; clearing then is harmless — the
-  // terminal effect (shouldClearStoppingLatch) already clears for a terminal run,
-  // so this only ever agrees with it. Asserted so the (c) reasoning is pinned.
-  it("clears on an error when the run is terminal (harmless, agrees with terminal effect)", () => {
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: true,
-        isLocalStreaming: false,
-        runQueryFailed: true,
-        run: makeRun("aborted"),
-      }),
-    ).toBe(true);
-  });
-
-  it("does NOT clear without an actual query error", () => {
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: true,
-        isLocalStreaming: false,
-        runQueryFailed: false,
-        run: null,
-      }),
-    ).toBe(false);
-  });
-
-  it("does NOT clear while this tab is the local streamer", () => {
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: true,
-        isLocalStreaming: true,
-        runQueryFailed: true,
-        run: null,
-      }),
-    ).toBe(false);
-  });
-
-  it("does NOT clear when no stop was requested", () => {
-    expect(
-      shouldClearLatchOnQueryError({
-        stoppingRun: false,
-        isLocalStreaming: false,
-        runQueryFailed: true,
-        run: null,
-      }),
-    ).toBe(false);
-  });
-});
-
-describe("mergeObservedMessage", () => {
-  it("replaces the message with the same id in place (per-step growth)", () => {
-    const prev = [makeMsg("u1", "hi"), makeMsg("a1", "step 1")];
-    const observed = makeMsg("a1", "step 1\nstep 2");
-    const next = mergeObservedMessage(prev, observed);
-    expect(next).toHaveLength(2);
-    expect(next[1]).toBe(observed);
-    expect(next[0]).toBe(prev[0]); // untouched
-    expect(next).not.toBe(prev); // new array (never mutates input)
-  });
-
-  it("appends when the observed message is not yet present", () => {
-    const prev = [makeMsg("u1", "hi")];
-    const observed = makeMsg("a1", "first token");
-    const next = mergeObservedMessage(prev, observed);
-    expect(next).toHaveLength(2);
-    expect(next[1]).toBe(observed);
-  });
-
-  it("returns the original list unchanged when there is nothing to merge", () => {
-    const prev = [makeMsg("u1", "hi")];
-    expect(mergeObservedMessage(prev, null)).toBe(prev);
-    expect(mergeObservedMessage(prev, undefined)).toBe(prev);
-  });
-});

apps/client/src/features/ai-chat/utils/run-polling.ts

@@ -1,151 +0,0 @@
-import type { UIMessage } from "@ai-sdk/react";
-import type { IAiChatRun } from "@/features/ai-chat/types/ai-chat.types.ts";
-
-/**
- * Reconnect-and-live-follow helpers (#184). When a chat is reopened while its
- * agent run is STILL going, this tab is a PASSIVE OBSERVER: it did not start the
- * run here (no local SSE stream), so it catches up by POLLING the reconnect
- * endpoint (`POST /ai-chat/run`) and merging the run's incrementally-persisted
- * assistant message into the rendered thread. These are the small pure decisions
- * that machinery hangs off, extracted so they can be unit-tested in isolation
- * (mirrors how reindex polling / editor-sync-state are tested).
- */
-
-/** How often to re-poll the reconnect endpoint while a run is ACTIVE. */
-export const RUN_POLL_INTERVAL_MS = 2000;
-
-// 'pending' and 'running' are the two ACTIVE statuses; 'succeeded' | 'failed' |
-// 'aborted' are TERMINAL (and any unknown future status is treated as terminal,
-// so a stale/odd value never polls forever).
-const ACTIVE_STATUSES = new Set(["pending", "running"]);
-
-/** Whether a run is still going (worth polling / merging live updates from). */
-export function isRunActive(run: IAiChatRun | null | undefined): boolean {
-  return !!run && ACTIVE_STATUSES.has(run.status);
-}
-
-/**
- * The TanStack Query `refetchInterval` value for the run query: poll every
- * {@link RUN_POLL_INTERVAL_MS} while the run is active, and `false` (stop) once
- * it is terminal or there is no run. Polling is thus naturally bounded by the run
- * reaching a terminal status — no separate timeout cap is needed.
- */
-export function runPollInterval(
-  run: IAiChatRun | null | undefined,
-): number | false {
-  return isRunActive(run) ? RUN_POLL_INTERVAL_MS : false;
-}
-
-/**
- * Observer-vs-streamer decision. We render the polled run message (catch up +
- * keep advancing) ONLY when this tab is a passive observer: there IS a run AND
- * this tab is NOT the one locally streaming it (we reconnected, we didn't start
- * it here). When this tab is the streamer, the live SSE stream owns the view, so
- * we neither poll nor merge — avoiding a double-render fight. Terminal runs still
- * merge (so the final persisted output is shown on reopen); the poll itself is
- * stopped separately by {@link runPollInterval}.
- */
-export function shouldObserveRun(
-  run: IAiChatRun | null | undefined,
-  localStreaming: boolean,
-): boolean {
-  return !!run && !localStreaming;
-}
-
-/**
- * Should the "stopping" latch — which suppresses the observer re-stream flash
- * after the user pressed Stop — be RELEASED now? All three must hold:
- *  - `stoppingRun`: we actually requested a stop (otherwise nothing to release);
- *  - `!isLocalStreaming`: this tab is NOT the local streamer. While we are the
- *    streamer the run query is disabled, so the observed `run` is not the run we
- *    are following — releasing the latch then would re-open the flash for the
- *    current turn the instant we switch to observer role;
- *  - the observed `run` EXISTS and has reached a TERMINAL status.
- *
- * The null / still-active `run` case is the #234 F4 invariant. On Stop the stale
- * PREVIOUS-turn run is removed from the query cache (`removeQueries`), so `run`
- * is null until the CURRENT turn's run is re-fetched fresh; a null or active run
- * therefore HOLDS the latch, so it can only ever clear against the current turn's
- * OWN terminal run — never a stale cached one. (The cache removal itself is
- * integration-level in AiChatWindow; this predicate encodes the decision given
- * whatever run is currently observed, and a stale terminal run is
- * indistinguishable from a current terminal run at the predicate level — hence
- * the cache removal is what guarantees only the current run is ever passed here.)
- */
-export function shouldClearStoppingLatch(args: {
-  stoppingRun: boolean;
-  run: IAiChatRun | null | undefined;
-  isLocalStreaming: boolean;
-}): boolean {
-  const { stoppingRun, run, isLocalStreaming } = args;
-  if (!stoppingRun || isLocalStreaming) return false;
-  return !!run && !isRunActive(run);
-}
-
-/**
- * Should the "stopping" latch be RELEASED by the run-query ERROR safety-net?
- * (#234 F7 — a NEW path of the same re-stream flash the F4 latch exists to
- * prevent.) After Stop, `handleServerStop` clears the run cache; the terminal
- * effect then holds the latch via `if (!run) return` until the CURRENT turn's run
- * is fetched fresh. If that refetch instead ERRORS permanently, `run` stays null,
- * its status-keyed refetchInterval is off, and nothing would ever observe a
- * terminal run — freezing the view with the observer merge suppressed. This
- * safety-net cures ONLY that genuine permanent-null-freeze.
- *
- * All four must hold:
- *  - `stoppingRun`: we actually requested a stop (otherwise nothing to release);
- *  - `!isLocalStreaming`: this tab is NOT the local streamer (same reason as
- *    {@link shouldClearStoppingLatch});
- *  - `runQueryFailed`: the run query is in its error state (TanStack Query v5 with
- *    retry:false — isError);
- *  - `!isRunActive(run)`: the observed `run` is NOT an active (pending/running)
- *    held run. This is the F7 gate. In TanStack Query v5 the query's `data` is
- *    RETAINED on error, so `runQueryFailed` can be true while `run` is STILL an
- *    ACTIVE run (a single transient GET-run failure in the window between Stop and
- *    settle). Without this gate a transient error would release the latch early —
- *    re-opening the observer merge and flashing the growing detached run over the
- *    frozen row (exactly the F4 flash). Gating on the run NOT being active means we
- *    only ever cure the permanent-null-freeze (`run === null`, so
- *    `isRunActive(null)` is false), never release against an active run.
- *
- * (A terminal `run` also satisfies `!isRunActive(run)`; clearing then is harmless
- * — the terminal effect's {@link shouldClearStoppingLatch} already clears the
- * latch for a terminal run, so this only ever agrees with it, never conflicts.)
- *
- * INVARIANT (do not break): clearing the latch on the `run === null` branch is safe
- * ONLY because the run query's `refetchInterval` (see {@link runPollInterval}) stops
- * polling when the data is empty — so after we clear on null+error there is no
- * subsequent auto-poll that could return a still-active detached run and re-open the
- * merge. If `refetchInterval` is ever changed to keep polling on `run === null`/on
- * error, this null-branch clear would re-open the F7 flash through the null path.
- * Do not change the run query's refetchInterval without re-checking this path.
- */
-export function shouldClearLatchOnQueryError(args: {
-  stoppingRun: boolean;
-  isLocalStreaming: boolean;
-  runQueryFailed: boolean;
-  run: IAiChatRun | null | undefined;
-}): boolean {
-  const { stoppingRun, isLocalStreaming, runQueryFailed, run } = args;
-  return (
-    stoppingRun && !isLocalStreaming && runQueryFailed && !isRunActive(run)
-  );
-}
-
-/**
- * Merge an observed assistant message into the rendered list: replace the message
- * with the same id in place (the in-progress assistant row is already seeded from
- * history, so per-step growth replaces it), or append it when absent. Returns a
- * new array; the input is never mutated.
- */
-export function mergeObservedMessage(
-  messages: UIMessage[],
-  observed: UIMessage | null | undefined,
-): UIMessage[] {
-  if (!observed) return messages;
-  const idx = messages.findIndex((m) => m.id === observed.id);
-  if (idx === -1) return [...messages, observed];
-  const next = messages.slice();
-  next[idx] = observed;
-  return next;
-}

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/src/core/ai-chat/ai-chat-stream-registry.service.ts

@@ -0,0 +1,329 @@
+import { Injectable, Logger, OnModuleDestroy } from '@nestjs/common';
+
+/**
+ * In-memory run-stream registry (#184 phase 1.5). A durable agent run tees its
+ * SSE frames here (via `pipeUIMessageStreamToResponse({ consumeSseStream })`)
+ * so a LATE tab — one that reloaded, or opened after the starter dropped — can
+ * attach through `GET /ai-chat/runs/:chatId/stream`, replay the frames buffered
+ * so far, and then follow the live tail as a normal streamer.
+ *
+ * This is deliberately single-process and best-effort: it holds nothing the DB
+ * does not (the run + assistant row are the source of truth), so a process
+ * restart simply drops in-flight entries and the client falls back to its
+ * restore + degraded-poll path. The async `attach` return type is the seam for a
+ * future phase-2 cross-process backend (Redis) — the interface does not change.
+ */
+
+/** How long a finished entry is retained for late attach (replay + immediate end). */
+export const RUN_STREAM_RETAIN_FINISHED_MS = 30_000;
+
+/** Per-run replay buffer cap. Past this the buffer is dropped (attach -> 204). */
+export const RUN_STREAM_MAX_BUFFER_BYTES = 4 * 1024 * 1024;
+
+// 2x the replay cap: a just-written 4MB replay burst alone can never trip the
+// per-subscriber cap (see controller); only a genuinely stalled socket can.
+export const SUBSCRIBER_MAX_BUFFERED_BYTES = 2 * RUN_STREAM_MAX_BUFFER_BYTES;
+
+export interface RunStreamCallbacks {
+  onFrame: (frame: string) => void;
+  onEnd: () => void;
+}
+
+export interface RunStreamAttachment {
+  replay: string[];
+  finished: boolean;
+  start(): void; // drain pending frames (order preserved) and go live
+  unsubscribe(): void; // safe to call at any point, idempotent
+}
+
+interface Subscriber extends RunStreamCallbacks {
+  started: boolean;
+  pending: string[];
+  // Byte size of `pending`, capped at SUBSCRIBER_MAX_BUFFERED_BYTES. `start()` is
+  // called in the SAME tick as `attach()` today (see attach), so `pending` never
+  // holds more than one microtask of frames — but the async `attach` signature is
+  // a phase-2 seam: an await between attach and start would let a stalled paused
+  // subscriber buffer the WHOLE run here. The cap is the structural backstop.
+  pendingBytes: number;
+  overflowed: boolean;
+  pendingEnd: boolean;
+}
+
+interface Entry {
+  runId: string;
+  // The persisted assistant row id of this run (set at bind; undefined if the
+  // seed failed). Used by the attach anchor check (invariant 6).
+  assistantMessageId?: string;
+  frames: string[];
+  bytes: number;
+  overflowed: boolean;
+  finished: boolean;
+  subscribers: Set<Subscriber>;
+  retainTimer?: NodeJS.Timeout;
+}
+
+@Injectable()
+export class AiChatStreamRegistryService implements OnModuleDestroy {
+  private readonly logger = new Logger(AiChatStreamRegistryService.name);
+  private readonly entries = new Map<string, Entry>(); // key: chatId
+
+  /**
+   * Register a fresh entry at the START of a run (before any frame), so a tab
+   * that attaches in the begin->seed window finds an entry to wait on. If an
+   * entry already exists for this chat (a previous, possibly still-live run whose
+   * tee loop is draining), it is terminated MIRRORING the done-path (invariant 3)
+   * so its subscribers are released and its retention timer is cleared; a late
+   * `done` from that old tee then fires against the closed-over old reference and,
+   * thanks to identity checks, never touches this new entry.
+   */
+  open(chatId: string, runId: string): void {
+    const existing = this.entries.get(chatId);
+    if (existing) {
+      if (existing.retainTimer) {
+        clearTimeout(existing.retainTimer);
+        existing.retainTimer = undefined;
+      }
+      // Started subscribers get exactly one onEnd() and are removed; paused ones
+      // are marked pendingEnd (their start() will end them). finished=true guards
+      // any later done from the old tee loop from double-notifying.
+      this.terminateSubscribers(existing);
+    }
+    this.entries.set(chatId, {
+      runId,
+      frames: [],
+      bytes: 0,
+      overflowed: false,
+      finished: false,
+      subscribers: new Set<Subscriber>(),
+    });
+  }
+
+  /**
+   * Tee a run's SSE frame stream into its entry (called from consumeSseStream).
+   * No-op with a warning when there is no entry or the entry belongs to a
+   * different run (invariant 1). The reader loop is fire-and-forget: the tee
+   * branch outlives the client socket by design.
+   */
+  bind(
+    chatId: string,
+    runId: string,
+    assistantMessageId: string | undefined,
+    stream: ReadableStream<string>,
+  ): void {
+    const entry = this.entries.get(chatId);
+    if (!entry || entry.runId !== runId) {
+      // Invariant 1: only the matching run may mutate the entry.
+      this.logger.warn(
+        `bind: no matching run-stream entry for chat=${chatId} run=${runId}`,
+      );
+      return;
+    }
+    entry.assistantMessageId = assistantMessageId;
+    const reader = stream.getReader();
+    const pump = async (): Promise<void> => {
+      try {
+        for (;;) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          this.ingestFrame(entry, value);
+        }
+        this.finalizeEntry(chatId, entry);
+      } catch {
+        // A read error is a terminal event too — release subscribers.
+        this.finalizeEntry(chatId, entry);
+      }
+    };
+    void pump();
+  }
+
+  /**
+   * Terminate a run's entry from the OUTER catch of the stream method (a failure
+   * before/while wiring the pipe, so `done` will never arrive). Identity-checked
+   * on runId (invariant 1); the shared terminal path is idempotent.
+   */
+  abortEntry(chatId: string, runId: string): void {
+    const entry = this.entries.get(chatId);
+    if (!entry || entry.runId !== runId) return;
+    this.finalizeEntry(chatId, entry);
+  }
+
+  /**
+   * Attach to a run's stream. Async only for the phase-2 Redis seam — the body
+   * runs synchronously so the replay snapshot and the subscriber registration
+   * happen in ONE tick with no await between them (invariant 4): a frame ingested
+   * concurrently cannot slip into the gap and be lost or duplicated.
+   *
+   * Returns null (-> the caller answers 204) when:
+   *  - there is no entry, or it overflowed (replay is gone);
+   *  - expect=live with an anchor that does not match this run's assistant id
+   *    (invariant 6: a stripped tab must never replay a FOREIGN run's transcript);
+   *  - the run finished and the caller did not expect a live tail.
+   * A finished run with expect=live yields a replay-only attachment (no
+   * subscriber registered). Otherwise a paused subscriber is registered and the
+   * caller replays `replay`, then calls start() to drain and go live.
+   */
+  async attach(
+    chatId: string,
+    expectLive: boolean,
+    anchor: string | undefined,
+    cb: RunStreamCallbacks,
+  ): Promise<RunStreamAttachment | null> {
+    const entry = this.entries.get(chatId);
+    if (!entry || entry.overflowed) return null;
+    // Invariant 6: cross-run replay is forbidden. Before bind, assistantMessageId
+    // is undefined and mismatches any anchor -> 204 -> client restore+poll path.
+    if (expectLive && anchor && entry.assistantMessageId !== anchor) return null;
+    if (entry.finished && !expectLive) return null;
+    if (entry.finished && expectLive) {
+      // Replay-only: the run is done, no subscriber is registered.
+      return {
+        replay: entry.frames.slice(),
+        finished: true,
+        start: () => undefined,
+        unsubscribe: () => undefined,
+      };
+    }
+
+    const sub: Subscriber = {
+      onFrame: cb.onFrame,
+      onEnd: cb.onEnd,
+      started: false,
+      pending: [],
+      pendingBytes: 0,
+      overflowed: false,
+      pendingEnd: false,
+    };
+    entry.subscribers.add(sub);
+    // Snapshot in the SAME synchronous block as the registration (invariant 4).
+    const replay = entry.frames.slice();
+    // CONTRACT: the caller MUST call start() in the SAME tick as this attach()
+    // returns — no await between them. While a subscriber is paused, every frame
+    // is buffered in sub.pending; a delayed start() lets a whole run accumulate
+    // there. The pendingBytes cap (see ingestFrame) is the structural backstop if
+    // that contract is ever broken (e.g. the phase-2 Redis await seam).
+    return {
+      replay,
+      finished: false,
+      start: () => {
+        if (sub.overflowed) {
+          // The pending buffer overflowed while paused: end the stream instead of
+          // replaying a partial (a 204-equivalent post-attach degrade).
+          try {
+            sub.onEnd();
+          } catch {
+            // The socket is gone; nothing to end.
+          }
+          entry.subscribers.delete(sub);
+          return;
+        }
+        // Deliver frames buffered while paused, in order, then go live.
+        for (const frame of sub.pending) {
+          try {
+            sub.onFrame(frame);
+          } catch {
+            entry.subscribers.delete(sub);
+            return;
+          }
+        }
+        sub.pending = [];
+        sub.started = true;
+        if (sub.pendingEnd) {
+          try {
+            sub.onEnd();
+          } catch {
+            // The socket is gone; nothing to end.
+          }
+          entry.subscribers.delete(sub);
+        }
+      },
+      unsubscribe: () => {
+        entry.subscribers.delete(sub);
+      },
+    };
+  }
+
+  onModuleDestroy(): void {
+    for (const entry of this.entries.values()) {
+      if (entry.retainTimer) clearTimeout(entry.retainTimer);
+    }
+    this.entries.clear();
+  }
+
+  /** Buffer + fan-out a single frame. See invariant/overflow semantics inline. */
+  private ingestFrame(entry: Entry, frame: string): void {
+    entry.bytes += Buffer.byteLength(frame);
+    if (!entry.overflowed) {
+      entry.frames.push(frame);
+      if (entry.bytes > RUN_STREAM_MAX_BUFFER_BYTES) {
+        // The crossing frame was already counted AND (below) fanned out; only the
+        // replay buffer is dropped. After overflow no more frames are buffered,
+        // but live fan-out continues.
+        entry.overflowed = true;
+        entry.frames = [];
+        this.logger.warn(
+          `run-stream buffer overflow for run=${entry.runId}; ` +
+            `late attach will 204 until the run ends`,
+        );
+      }
+    }
+    for (const sub of entry.subscribers) {
+      if (sub.started) {
+        try {
+          sub.onFrame(frame);
+        } catch {
+          entry.subscribers.delete(sub);
+        }
+      } else {
+        sub.pending.push(frame);
+        sub.pendingBytes += Buffer.byteLength(frame);
+        if (sub.pendingBytes > SUBSCRIBER_MAX_BUFFERED_BYTES) {
+          // The paused subscriber's buffer overflowed — only possible if start()
+          // was delayed past the same-tick contract (the phase-2 await seam).
+          // Drop it rather than buffer the whole run; on start() it degrades to an
+          // immediate end (a 204-equivalent) instead of replaying a partial.
+          sub.overflowed = true;
+          sub.pending = [];
+          entry.subscribers.delete(sub);
+        }
+      }
+    }
+  }
+
+  /**
+   * Shared terminal path for done / read-error / external-abort. Idempotent: a
+   * second call (already finished) is a no-op, so an open()-replaced or
+   * abort-then-done entry is never double-armed or double-ended.
+   */
+  private finalizeEntry(chatId: string, entry: Entry): void {
+    if (entry.finished) return;
+    this.terminateSubscribers(entry);
+    const timer = setTimeout(() => {
+      // Invariant 2: only delete OUR entry (a replacement may already own the key).
+      if (this.entries.get(chatId) === entry) this.entries.delete(chatId);
+    }, RUN_STREAM_RETAIN_FINISHED_MS);
+    timer.unref?.();
+    entry.retainTimer = timer;
+  }
+
+  /**
+   * Mark the entry finished and release its subscribers, mirroring the done-path:
+   * started subscribers get exactly one onEnd() and are removed; paused ones are
+   * flagged pendingEnd so their start() ends them. Deleting the current element
+   * during Set iteration is safe.
+   */
+  private terminateSubscribers(entry: Entry): void {
+    entry.finished = true;
+    for (const sub of entry.subscribers) {
+      if (sub.started) {
+        try {
+          sub.onEnd();
+        } catch {
+          // The socket is gone; nothing to end.
+        }
+        entry.subscribers.delete(sub);
+      } else {
+        sub.pendingEnd = true;
+      }
+    }
+  }
+}

apps/server/src/core/ai-chat/ai-chat-stream-registry.spec.ts

@@ -0,0 +1,388 @@
+import {
+  AiChatStreamRegistryService,
+  RUN_STREAM_MAX_BUFFER_BYTES,
+  RUN_STREAM_RETAIN_FINISHED_MS,
+  RunStreamCallbacks,
+} from './ai-chat-stream-registry.service';
+
+/**
+ * Unit tests for the in-memory run-stream registry (#184 phase 1.5). The registry
+ * is the whole of the resumable-transport contract: replay ordering, paused ->
+ * live hand-off, overflow, retention, the anchor check (invariant 6), and the
+ * mirror-the-done-path replace semantics (invariant 3). Every enumerated case in
+ * the issue's task 1.5 has a test here.
+ */
+
+// A ReadableStream whose frames the test pushes explicitly, plus close/error.
+function makePushStream(): {
+  stream: ReadableStream<string>;
+  push: (f: string) => void;
+  close: () => void;
+  error: (e?: unknown) => void;
+} {
+  let controller!: ReadableStreamDefaultController<string>;
+  const stream = new ReadableStream<string>({
+    start(c) {
+      controller = c;
+    },
+  });
+  return {
+    stream,
+    push: (f) => controller.enqueue(f),
+    close: () => controller.close(),
+    error: (e) => controller.error(e ?? new Error('read error')),
+  };
+}
+
+// Let the fire-and-forget pump drain queued frames (reader.read() resolves on a
+// macrotask boundary for an already-enqueued value).
+const flush = (): Promise<void> => new Promise((r) => setTimeout(r, 0));
+
+function collector(): {
+  cb: RunStreamCallbacks;
+  frames: string[];
+  ended: () => number;
+} {
+  const frames: string[] = [];
+  let ends = 0;
+  return {
+    frames,
+    ended: () => ends,
+    cb: {
+      onFrame: (f) => frames.push(f),
+      onEnd: () => {
+        ends += 1;
+      },
+    },
+  };
+}
+
+describe('AiChatStreamRegistryService', () => {
+  const CHAT = 'chat-1';
+  let registry: AiChatStreamRegistryService;
+
+  beforeEach(() => {
+    registry = new AiChatStreamRegistryService();
+    jest.spyOn((registry as any).logger, 'warn').mockImplementation(() => {});
+  });
+
+  afterEach(() => {
+    registry.onModuleDestroy();
+  });
+
+  it('replays frames in arrival order (live attach)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    src.push('b');
+    src.push('c');
+    await flush();
+
+    const c = collector();
+    const att = await registry.attach(CHAT, false, undefined, c.cb);
+    expect(att).not.toBeNull();
+    expect(att!.replay).toEqual(['a', 'b', 'c']);
+    expect(att!.finished).toBe(false);
+  });
+
+  it('late attach gets the full prefix as replay plus the live tail', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    src.push('b');
+    await flush();
+
+    const c = collector();
+    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
+    expect(att.replay).toEqual(['a', 'b']);
+    att.start();
+    // Live tail arrives after start().
+    src.push('c');
+    src.push('d');
+    await flush();
+    expect(c.frames).toEqual(['c', 'd']);
+  });
+
+  it('a paused subscriber receives frames buffered during pause in order, then live (no loss/reorder)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    await flush();
+
+    const c = collector();
+    // Attach (paused). Frames that arrive BEFORE start() must queue, not drop.
+    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
+    expect(att.replay).toEqual(['a']);
+    src.push('b'); // arrives while paused -> pending
+    src.push('c');
+    await flush();
+    expect(c.frames).toEqual([]); // nothing delivered yet (paused)
+    att.start(); // drains pending in order
+    expect(c.frames).toEqual(['b', 'c']);
+    src.push('d'); // now live
+    await flush();
+    expect(c.frames).toEqual(['b', 'c', 'd']);
+  });
+
+  it('a run that finishes while a subscriber is paused ends it on start()', async () => {
+    registry.open(CHAT, 'run-1');
+    const c = collector();
+    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
+    // Terminate the run while the subscriber is still paused.
+    registry.abortEntry(CHAT, 'run-1');
+    expect(c.ended()).toBe(0); // paused: not ended yet
+    att.start();
+    expect(c.ended()).toBe(1); // start() drains + ends
+  });
+
+  it('finished + expect=live returns a replay WITHOUT registering a subscriber', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    src.push('b');
+    src.close();
+    await flush();
+
+    const c = collector();
+    const att = (await registry.attach(CHAT, true, undefined, c.cb))!;
+    expect(att.finished).toBe(true);
+    expect(att.replay).toEqual(['a', 'b']);
+    // No subscriber registered: start()/unsubscribe are no-ops and the entry has
+    // zero subscribers.
+    const entry = (registry as any).entries.get(CHAT);
+    expect(entry.subscribers.size).toBe(0);
+    att.start();
+    expect(c.frames).toEqual([]);
+  });
+
+  it('finished WITHOUT expect=live returns null', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    src.close();
+    await flush();
+
+    const c = collector();
+    expect(await registry.attach(CHAT, false, undefined, c.cb)).toBeNull();
+  });
+
+  it('anchor mismatch with expect=live returns null (and null before bind sets assistantMessageId)', async () => {
+    registry.open(CHAT, 'run-1');
+    const c = collector();
+    // Before bind: assistantMessageId is undefined -> mismatches any anchor.
+    expect(
+      await registry.attach(CHAT, true, 'assist-1', c.cb),
+    ).toBeNull();
+
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    await flush();
+    // Wrong anchor -> null (cross-run replay forbidden, invariant 6).
+    expect(await registry.attach(CHAT, true, 'other-id', c.cb)).toBeNull();
+  });
+
+  it('matching anchor with expect=live attaches', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    await flush();
+
+    const c = collector();
+    const att = await registry.attach(CHAT, true, 'assist-1', c.cb);
+    expect(att).not.toBeNull();
+    expect(att!.replay).toEqual(['a']);
+  });
+
+  it('overflow: attach returns null, but the LIVE subscriber keeps receiving (incl. the crossing frame)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+
+    // A live (started) subscriber attached before the flood.
+    const c = collector();
+    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
+    att.start();
+
+    const oneMb = 'x'.repeat(1024 * 1024);
+    // 5 x 1MB = 5MB > 4MB cap; the 5th frame is the one that crosses.
+    for (let i = 0; i < 5; i++) src.push(oneMb + i);
+    await flush();
+
+    const entry = (registry as any).entries.get(CHAT);
+    expect(entry.overflowed).toBe(true);
+    expect(entry.bytes).toBeGreaterThan(RUN_STREAM_MAX_BUFFER_BYTES);
+    // The live subscriber received ALL 5 frames, including the crossing one.
+    expect(c.frames).toHaveLength(5);
+    expect(c.frames[4]).toBe(oneMb + 4);
+
+    // A NEW attach after overflow gets null (replay buffer is gone).
+    const c2 = collector();
+    expect(await registry.attach(CHAT, false, undefined, c2.cb)).toBeNull();
+  });
+
+  it('a paused subscriber whose pending buffer overflows is dropped and ends on start(); other subscribers keep receiving', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+
+    // A: paused (start() deliberately delayed to simulate the phase-2 await seam).
+    const a = collector();
+    const attA = (await registry.attach(CHAT, false, undefined, a.cb))!;
+    // B: live (started) — its delivery must be unaffected by A's overflow.
+    const b = collector();
+    const attB = (await registry.attach(CHAT, false, undefined, b.cb))!;
+    attB.start();
+
+    const oneMb = 'x'.repeat(1024 * 1024);
+    // 9 x 1MB = 9MB > 8MB per-subscriber cap; A's pending overflows, B streams live.
+    for (let i = 0; i < 9; i++) src.push(oneMb + i);
+    await flush();
+
+    const entry = (registry as any).entries.get(CHAT);
+    // A was dropped from the subscriber set on overflow; B (started) remains.
+    expect(entry.subscribers.size).toBe(1);
+    expect(a.frames).toEqual([]); // paused + overflowed: nothing was delivered
+    // B received every frame live (delivery unaffected by A's overflow).
+    expect(b.frames).toHaveLength(9);
+
+    // A's start() (arriving late) degrades to an immediate end, not a partial replay.
+    attA.start();
+    expect(a.frames).toEqual([]);
+    expect(a.ended()).toBe(1);
+  });
+
+  it('open() over a LIVE entry ends started subscribers exactly once and a late done does not touch the new entry (invariant 3)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    await flush();
+
+    const c = collector();
+    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
+    att.start(); // started subscriber on run-1
+
+    // run-2 starts on the same chat while run-1's tee is still reading.
+    registry.open(CHAT, 'run-2');
+    expect(c.ended()).toBe(1); // exactly one onEnd from the replace
+
+    const newEntry = (registry as any).entries.get(CHAT);
+    expect(newEntry.runId).toBe('run-2');
+    expect(newEntry.finished).toBe(false);
+
+    // The old tee now completes: its late done must NOT double-end nor delete the
+    // new entry.
+    src.push('b');
+    src.close();
+    await flush();
+    expect(c.ended()).toBe(1); // still exactly one
+    const still = (registry as any).entries.get(CHAT);
+    expect(still).toBe(newEntry);
+    expect(still.runId).toBe('run-2');
+  });
+
+  it('bind with a foreign runId is a no-op (invariant 1)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'WRONG-run', 'assist-x', src.stream);
+    src.push('a');
+    await flush();
+    const entry = (registry as any).entries.get(CHAT);
+    // Frames were NOT ingested (bind bailed), assistantMessageId untouched.
+    expect(entry.frames).toEqual([]);
+    expect(entry.assistantMessageId).toBeUndefined();
+  });
+
+  it('abortEntry with a foreign runId is a no-op (invariant 1)', async () => {
+    registry.open(CHAT, 'run-1');
+    registry.abortEntry(CHAT, 'WRONG-run');
+    const entry = (registry as any).entries.get(CHAT);
+    expect(entry.finished).toBe(false);
+  });
+
+  it('a throwing onFrame ejects only that subscriber; the ingest loop stays alive', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+
+    const bad = collector();
+    const badAtt = (await registry.attach(CHAT, false, undefined, {
+      onFrame: () => {
+        throw new Error('boom');
+      },
+      onEnd: bad.cb.onEnd,
+    }))!;
+    badAtt.start();
+
+    const good = collector();
+    const goodAtt = (await registry.attach(CHAT, false, undefined, good.cb))!;
+    goodAtt.start();
+
+    src.push('a'); // bad throws on this frame -> ejected
+    src.push('b'); // good still receives both
+    await flush();
+
+    const entry = (registry as any).entries.get(CHAT);
+    expect(entry.subscribers.size).toBe(1); // bad ejected, good remains
+    expect(good.frames).toEqual(['a', 'b']);
+  });
+});
+
+/**
+ * Retention + replace timer behavior. Fake timers, and entries are finalized via
+ * the synchronous abortEntry() path so no stream pump / microtask juggling is
+ * needed.
+ */
+describe('AiChatStreamRegistryService retention timers', () => {
+  const CHAT = 'chat-r';
+  let registry: AiChatStreamRegistryService;
+
+  beforeEach(() => {
+    jest.useFakeTimers();
+    registry = new AiChatStreamRegistryService();
+    jest.spyOn((registry as any).logger, 'warn').mockImplementation(() => {});
+  });
+
+  afterEach(() => {
+    registry.onModuleDestroy();
+    jest.useRealTimers();
+  });
+
+  it('a finished entry is removed after the retention window', () => {
+    registry.open(CHAT, 'run-1');
+    registry.abortEntry(CHAT, 'run-1'); // finalize -> retention armed
+    expect((registry as any).entries.get(CHAT)).toBeDefined();
+    jest.advanceTimersByTime(RUN_STREAM_RETAIN_FINISHED_MS + 1);
+    expect((registry as any).entries.get(CHAT)).toBeUndefined();
+  });
+
+  it('retention deletes ONLY its own entry (invariant 2)', () => {
+    registry.open(CHAT, 'run-1');
+    registry.abortEntry(CHAT, 'run-1'); // arm retention for entry A
+    // Simulate the race where the key was replaced without clearing A's timer.
+    const sentinel = { marker: true };
+    (registry as any).entries.set(CHAT, sentinel);
+    jest.advanceTimersByTime(RUN_STREAM_RETAIN_FINISHED_MS + 1);
+    // A's timer saw entries.get(CHAT) !== A, so it did NOT delete the successor.
+    expect((registry as any).entries.get(CHAT)).toBe(sentinel);
+  });
+
+  it('open() over a retained entry clears its timer and the successor survives', () => {
+    registry.open(CHAT, 'run-1');
+    registry.abortEntry(CHAT, 'run-1'); // retained, timer armed
+    const clearSpy = jest.spyOn(global, 'clearTimeout');
+    registry.open(CHAT, 'run-2'); // must clear run-1's retain timer
+    expect(clearSpy).toHaveBeenCalled();
+    jest.advanceTimersByTime(RUN_STREAM_RETAIN_FINISHED_MS + 1);
+    const entry = (registry as any).entries.get(CHAT);
+    expect(entry).toBeDefined();
+    expect(entry.runId).toBe('run-2');
+  });
+});

apps/server/src/core/ai-chat/ai-chat.controller.attach.spec.ts

@@ -0,0 +1,423 @@
+import { ForbiddenException } from '@nestjs/common';
+import { AiChatController } from './ai-chat.controller';
+import type {
+  RunStreamAttachment,
+  RunStreamCallbacks,
+} from './ai-chat-stream-registry.service';
+import { SUBSCRIBER_MAX_BUFFERED_BYTES } from './ai-chat-stream-registry.service';
+import type { User, Workspace } from '@docmost/db/types/entity.types';
+
+/**
+ * Wiring spec for the #184 phase 1.5 attach endpoint
+ * (`GET /ai-chat/runs/:chatId/stream`). Owner-gated via assertOwnedChat; the
+ * registry is mocked so this exercises ONLY the controller's replay/live/204/
+ * cleanup wiring against a fake raw socket. Constructor order is (aiChatService,
+ * aiChatRunService, aiChatRepo, aiChatMessageRepo, aiTranscription, pageRepo,
+ * streamRegistry, environment).
+ */
+describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
+  const user = { id: 'u1' } as User;
+  const workspace = { id: 'ws1' } as Workspace;
+
+  function makeRawRes() {
+    const raw: any = {
+      writableEnded: false,
+      writableLength: 0,
+      destroyed: false,
+      written: [] as string[],
+      head: null as any,
+      write: jest.fn((f: string) => {
+        raw.written.push(f);
+        return true;
+      }),
+      writeHead: jest.fn((code: number, headers: any) => {
+        raw.head = { code, headers };
+        return raw;
+      }),
+      flushHeaders: jest.fn(),
+      end: jest.fn(() => {
+        raw.writableEnded = true;
+      }),
+      destroy: jest.fn(() => {
+        raw.destroyed = true;
+      }),
+      on: jest.fn(),
+      once: jest.fn(),
+    };
+    const res: any = {
+      raw,
+      status: jest.fn(() => res),
+      send: jest.fn(),
+      hijack: jest.fn(),
+    };
+    return { res, raw };
+  }
+
+  function makeReq(destroyed = false) {
+    const handlers: Record<string, () => void> = {};
+    const raw: any = {
+      destroyed,
+      once: jest.fn((ev: string, fn: () => void) => {
+        handlers[ev] = fn;
+      }),
+    };
+    return { req: { raw } as any, raw, fireClose: () => handlers['close']?.() };
+  }
+
+  function makeAttachment(
+    over: Partial<RunStreamAttachment> = {},
+  ): RunStreamAttachment {
+    return {
+      replay: [],
+      finished: false,
+      start: jest.fn(),
+      unsubscribe: jest.fn(),
+      ...over,
+    };
+  }
+
+  function makeController(opts: {
+    chat?: unknown;
+    attachment?: RunStreamAttachment | null;
+  }) {
+    const aiChatRepo = { findById: jest.fn().mockResolvedValue(opts.chat) };
+    let capturedCb: RunStreamCallbacks | undefined;
+    const streamRegistry = {
+      attach: jest.fn(
+        (
+          _chatId: string,
+          _live: boolean,
+          _anchor: string | undefined,
+          cb: RunStreamCallbacks,
+        ) => {
+          capturedCb = cb;
+          return Promise.resolve(
+            opts.attachment === undefined ? makeAttachment() : opts.attachment,
+          );
+        },
+      ),
+    };
+    const environment = { isAiChatResumableStreamEnabled: () => true };
+    const controller = new AiChatController(
+      {} as never, // aiChatService
+      {} as never, // aiChatRunService
+      aiChatRepo as never,
+      {} as never, // aiChatMessageRepo
+      {} as never, // aiTranscription
+      {} as never, // pageRepo
+      streamRegistry as never,
+      environment as never,
+    );
+    return {
+      controller,
+      aiChatRepo,
+      streamRegistry,
+      getCb: () => capturedCb!,
+    };
+  }
+
+  const owned = { id: 'c1', creatorId: 'u1' };
+
+  it('owner-gates: a foreign chat throws ForbiddenException and never attaches', async () => {
+    const { controller, streamRegistry } = makeController({
+      chat: { id: 'c1', creatorId: 'someone-else' },
+    });
+    const { res } = makeRawRes();
+    const { req } = makeReq();
+    await expect(
+      controller.attachRunStream(
+        'c1',
+        undefined,
+        undefined,
+        req,
+        res,
+        user,
+        workspace,
+      ),
+    ).rejects.toBeInstanceOf(ForbiddenException);
+    expect(streamRegistry.attach).not.toHaveBeenCalled();
+  });
+
+  it('answers 204 when the registry has nothing to resume (no entry / finished / anchor-mismatch)', async () => {
+    const { controller } = makeController({ chat: owned, attachment: null });
+    const { res } = makeRawRes();
+    const { req } = makeReq();
+    await controller.attachRunStream(
+      'c1',
+      undefined,
+      undefined,
+      req,
+      res,
+      user,
+      workspace,
+    );
+    expect(res.status).toHaveBeenCalledWith(204);
+    expect(res.send).toHaveBeenCalled();
+    expect(res.hijack).not.toHaveBeenCalled();
+  });
+
+  it('threads expect=live and anchor through to the registry', async () => {
+    const { controller, streamRegistry } = makeController({
+      chat: owned,
+      attachment: null,
+    });
+    const { res } = makeRawRes();
+    const { req } = makeReq();
+    await controller.attachRunStream(
+      'c1',
+      'live',
+      'anchor-1',
+      req,
+      res,
+      user,
+      workspace,
+    );
+    expect(streamRegistry.attach).toHaveBeenCalledWith(
+      'c1',
+      true,
+      'anchor-1',
+      expect.anything(),
+    );
+  });
+
+  it('passes expect=false when the query is absent', async () => {
+    const { controller, streamRegistry } = makeController({
+      chat: owned,
+      attachment: null,
+    });
+    const { res } = makeRawRes();
+    const { req } = makeReq();
+    await controller.attachRunStream(
+      'c1',
+      undefined,
+      undefined,
+      req,
+      res,
+      user,
+      workspace,
+    );
+    expect(streamRegistry.attach).toHaveBeenCalledWith(
+      'c1',
+      false,
+      undefined,
+      expect.anything(),
+    );
+  });
+
+  it('hijacks, writes headers + replay, registers a close cleanup, then goes live', async () => {
+    const start = jest.fn();
+    const attachment = makeAttachment({
+      replay: ['f1', 'f2'],
+      finished: false,
+      start,
+    });
+    const { controller } = makeController({ chat: owned, attachment });
+    const { res, raw } = makeRawRes();
+    const { req } = makeReq();
+    await controller.attachRunStream(
+      'c1',
+      undefined,
+      undefined,
+      req,
+      res,
+      user,
+      workspace,
+    );
+    expect(res.hijack).toHaveBeenCalled();
+    expect(raw.writeHead).toHaveBeenCalledWith(
+      200,
+      expect.objectContaining({ 'content-type': 'text/event-stream' }),
+    );
+    expect(raw.written).toEqual(['f1', 'f2']); // replay
+    expect(start).toHaveBeenCalled(); // go live after replay
+    expect(req.raw.once).toHaveBeenCalledWith('close', expect.any(Function));
+  });
+
+  it('finished replay ends the response immediately without going live', async () => {
+    const start = jest.fn();
+    const attachment = makeAttachment({
+      replay: ['f1'],
+      finished: true,
+      start,
+    });
+    const { controller } = makeController({ chat: owned, attachment });
+    const { res, raw } = makeRawRes();
+    const { req } = makeReq();
+    await controller.attachRunStream(
+      'c1',
+      'live',
+      'a1',
+      req,
+      res,
+      user,
+      workspace,
+    );
+    expect(raw.written).toEqual(['f1']);
+    expect(raw.end).toHaveBeenCalled();
+    expect(start).not.toHaveBeenCalled(); // finished -> returns before start()
+  });
+
+  it('a close during the awaits (req.raw.destroyed) unsubscribes and writes nothing', async () => {
+    const attachment = makeAttachment({ replay: ['f1'] });
+    const { controller } = makeController({ chat: owned, attachment });
+    const { res, raw } = makeRawRes();
+    const { req } = makeReq(true); // destroyed already at registration time
+    await controller.attachRunStream(
+      'c1',
+      undefined,
+      undefined,
+      req,
+      res,
+      user,
+      workspace,
+    );
+    expect(attachment.unsubscribe).toHaveBeenCalled();
+    expect(raw.writeHead).not.toHaveBeenCalled();
+    expect(raw.written).toEqual([]);
+  });
+
+  it('the registered close handler unsubscribes the attachment', async () => {
+    const attachment = makeAttachment({ replay: [] });
+    const { controller } = makeController({ chat: owned, attachment });
+    const { res } = makeRawRes();
+    const { req, fireClose } = makeReq();
+    await controller.attachRunStream(
+      'c1',
+      undefined,
+      undefined,
+      req,
+      res,
+      user,
+      workspace,
+    );
+    expect(attachment.unsubscribe).not.toHaveBeenCalled();
+    fireClose(); // socket closed
+    expect(attachment.unsubscribe).toHaveBeenCalled();
+  });
+
+  it('onFrame destroys the socket when the buffered length exceeds the cap', async () => {
+    const attachment = makeAttachment({ replay: [] });
+    const { controller, getCb } = makeController({ chat: owned, attachment });
+    const { res, raw } = makeRawRes();
+    const { req } = makeReq();
+    await controller.attachRunStream(
+      'c1',
+      undefined,
+      undefined,
+      req,
+      res,
+      user,
+      workspace,
+    );
+    const cb = getCb();
+    // Normal live frame writes.
+    raw.writableLength = 0;
+    cb.onFrame('live-1');
+    expect(raw.written).toContain('live-1');
+    // A stalled socket over the cap is destroyed instead of buffering.
+    raw.writableLength = SUBSCRIBER_MAX_BUFFERED_BYTES + 1;
+    raw.write.mockClear();
+    cb.onFrame('too-much');
+    expect(raw.destroy).toHaveBeenCalled();
+    expect(raw.write).not.toHaveBeenCalled();
+  });
+});
+
+/**
+ * The begin-hook `open()` flag gate (#184 phase 1.5). `open()` lives ONLY in the
+ * stream() begin-hook, gated on the resumable flag. If it regressed, a flag-off
+ * turn would create an EMPTY registry entry (never bound, never finished) and a
+ * later attach would find a non-null paused attachment -> a hung SSE that never
+ * gets a frame and never ends, instead of a clean 204. These drive stream() only
+ * far enough to capture the runHooks it hands to the service, then invoke the
+ * begin-hook and assert whether the registry was opened.
+ */
+describe('AiChatController begin-hook open() flag gate (#184 phase 1.5)', () => {
+  const user = { id: 'u1' } as User;
+  const workspace = {
+    id: 'ws1',
+    settings: { ai: { chat: true, autonomousRuns: true } },
+  } as unknown as Workspace;
+
+  function makeController(opts: { resumable: boolean }) {
+    let capturedArgs: any;
+    const aiChatService = {
+      resolveRoleForRequest: jest.fn(async () => null),
+      getChatModel: jest.fn(async () => ({})),
+      stream: jest.fn(async (args: any) => {
+        capturedArgs = args;
+      }),
+    };
+    const aiChatRunService = {
+      getActiveForChat: jest.fn(async () => undefined),
+      beginRun: jest.fn(async () => ({
+        runId: 'run-1',
+        signal: new AbortController().signal,
+      })),
+    };
+    const streamRegistry = { open: jest.fn(), attach: jest.fn() };
+    const environment = {
+      isAiChatResumableStreamEnabled: () => opts.resumable,
+    };
+    const controller = new AiChatController(
+      aiChatService as never,
+      aiChatRunService as never,
+      {} as never, // aiChatRepo
+      {} as never, // aiChatMessageRepo
+      {} as never, // aiTranscription
+      {} as never, // pageRepo
+      streamRegistry as never,
+      environment as never,
+    );
+    return {
+      controller,
+      streamRegistry,
+      aiChatRunService,
+      getRunHooks: () => capturedArgs?.runHooks,
+    };
+  }
+
+  function makeReqRes() {
+    const req: any = {
+      raw: { sessionId: 'sess-1', once: jest.fn() },
+      body: {
+        messages: [
+          { id: 'm1', role: 'user', parts: [{ type: 'text', text: 'hi' }] },
+        ],
+      },
+    };
+    const res: any = {
+      raw: { once: jest.fn(), on: jest.fn(), headersSent: false, writableEnded: false },
+      hijack: jest.fn(),
+    };
+    return { req, res };
+  }
+
+  it('flag OFF: the begin-hook does NOT open a registry entry (no hung empty entry)', async () => {
+    const { controller, streamRegistry, aiChatRunService, getRunHooks } =
+      makeController({ resumable: false });
+    const { req, res } = makeReqRes();
+    await controller.stream(req, res, user, workspace);
+
+    const runHooks = getRunHooks();
+    expect(runHooks).toBeDefined();
+    const handle = await runHooks.begin('chat-1');
+    // The run still begins (the durable-run feature is independent of resume)...
+    expect(aiChatRunService.beginRun).toHaveBeenCalled();
+    expect(handle).toEqual({ runId: 'run-1', signal: expect.anything() });
+    // ...but with the flag off the registry entry is NEVER opened.
+    expect(streamRegistry.open).not.toHaveBeenCalled();
+  });
+
+  it('flag ON: the begin-hook opens the registry entry with (chatId, runId)', async () => {
+    const { controller, streamRegistry, getRunHooks } = makeController({
+      resumable: true,
+    });
+    const { req, res } = makeReqRes();
+    await controller.stream(req, res, user, workspace);
+
+    const runHooks = getRunHooks();
+    await runHooks.begin('chat-1');
+    expect(streamRegistry.open).toHaveBeenCalledWith('chat-1', 'run-1');
+  });
+});

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

@@ -4,11 +4,15 @@ import {
   ConflictException,
   Controller,
   ForbiddenException,
+  Get,
   HttpCode,
   HttpException,
   HttpStatus,
   Logger,
+  Param,
+  ParseUUIDPipe,
   Post,
+  Query,
   Req,
   Res,
   ServiceUnavailableException,
@@ -54,6 +58,12 @@ import {
 } from './dto/ai-chat.dto';
 import { describeProviderError } from '../../integrations/ai/ai-error.util';
 import { buildChatMarkdown } from './chat-markdown.util';
+import {
+  AiChatStreamRegistryService,
+  SUBSCRIBER_MAX_BUFFERED_BYTES,
+} from './ai-chat-stream-registry.service';
+import { startSseHeartbeat } from './sse-resilience';
+import { EnvironmentService } from '../../integrations/environment/environment.service';
 
 /**
  * Per-user AI chat API (§6.1). Routes are POST to match this codebase's
@@ -72,6 +82,11 @@ export class AiChatController {
     private readonly aiChatMessageRepo: AiChatMessageRepo,
     private readonly aiTranscription: AiTranscriptionService,
     private readonly pageRepo: PageRepo,
+    // #184 phase 1.5. OPTIONAL so existing positional constructions (controller
+    // specs) compile unchanged; Nest always injects the real providers in
+    // production. Only touched on the resumable-stream (flag-on) path.
+    private readonly streamRegistry?: AiChatStreamRegistryService,
+    private readonly environment?: EnvironmentService,
   ) {}
 
   /** List the requesting user's chats in this workspace (paginated). */
@@ -233,6 +248,102 @@ export class AiChatController {
     return { stopped };
   }
 
+  /**
+   * Attach to a chat's live run stream (#184 phase 1.5). A late/reloaded tab
+   * replays the frames buffered so far and then follows the live tail as a normal
+   * streamer. Owner-gated via assertOwnedChat (same gate as getRun). When there is
+   * nothing to resume — no entry, a finished run without expect=live, an
+   * overflowed buffer, or an anchor that pins a DIFFERENT run — the endpoint
+   * answers 204, the ONLY "nothing to resume" signal the AI SDK's reconnect
+   * accepts (it maps 204 to a silent no-op). With AI_CHAT_RESUMABLE_STREAM off the
+   * registry is never populated, so attach always 204s.
+   *
+   * `expect=live` opts into replaying a finished-but-retained run (safe only when
+   * the client stripped the streaming tail); `anchor` is the client's assistant
+   * row id, which must match this run's (invariant 6) or a foreign run's
+   * transcript would be replayed into the store.
+   */
+  @SkipTransform()
+  @UseGuards(JwtAuthGuard, UserThrottlerGuard)
+  @Throttle({ [AI_CHAT_THROTTLER]: { limit: 60, ttl: 60000 } })
+  @Get('runs/:chatId/stream')
+  async attachRunStream(
+    @Param('chatId', new ParseUUIDPipe()) chatId: string,
+    @Query('expect') expect: string | undefined,
+    @Query('anchor') anchor: string | undefined,
+    @Req() req: FastifyRequest,
+    @Res() res: FastifyReply,
+    @AuthUser() user: User,
+    @AuthWorkspace() workspace: Workspace,
+  ): Promise<void> {
+    await this.assertOwnedChat(chatId, user, workspace); // same gate as getRun
+    let stopHeartbeat: () => void = () => undefined;
+    const attachment = await this.streamRegistry?.attach(
+      chatId,
+      expect === 'live',
+      anchor,
+      {
+        onFrame: (frame) => {
+          // Backpressure guard: 2x the replay cap, so the initial replay burst
+          // alone can never trip it; only a genuinely stalled socket can.
+          try {
+            if (res.raw.writableLength > SUBSCRIBER_MAX_BUFFERED_BYTES) {
+              res.raw.destroy(); // 'close' fires -> unsubscribe below
+              return;
+            }
+            if (!res.raw.writableEnded) res.raw.write(frame);
+          } catch {
+            res.raw.destroy();
+          }
+        },
+        onEnd: () => {
+          stopHeartbeat();
+          if (!res.raw.writableEnded) res.raw.end();
+        },
+      },
+    );
+    if (!attachment) {
+      res.status(204).send(); // the ONLY "nothing to resume" signal the SDK accepts
+      return;
+    }
+    res.hijack();
+    // Cleanup BEFORE any write (invariant 5): a torn-down socket must not orphan
+    // a paused subscriber whose pending queue would buffer the whole run.
+    req.raw.once('close', () => {
+      attachment.unsubscribe();
+      stopHeartbeat();
+    });
+    // A close emitted DURING the awaits above was missed by the listener — check.
+    // (Healthy pending GETs have req.raw.destroyed === false, so no false
+    // positives; returning without end() is fine — the socket is gone.)
+    if (req.raw.destroyed) {
+      attachment.unsubscribe();
+      return;
+    }
+    res.raw.on('error', () => undefined);
+    try {
+      res.raw.writeHead(200, {
+        'content-type': 'text/event-stream',
+        'cache-control': 'no-cache',
+        'x-vercel-ai-ui-message-stream': 'v1',
+        'x-accel-buffering': 'no',
+        // deliberately NO Connection/Keep-Alive (hop-by-hop; Safari/HTTP2)
+      });
+      res.raw.flushHeaders?.();
+      for (const frame of attachment.replay) res.raw.write(frame);
+      if (attachment.finished) {
+        res.raw.end();
+        return;
+      }
+      stopHeartbeat = startSseHeartbeat(res.raw, 15_000);
+      attachment.start(); // drain pending accumulated during replay, go live
+    } catch {
+      attachment.unsubscribe();
+      stopHeartbeat();
+      res.raw.destroy();
+    }
+  }
+
   /** Rename a chat. */
   @HttpCode(HttpStatus.OK)
   @Post('rename')
@@ -344,13 +455,25 @@ export class AiChatController {
     // its progress, and settle its terminal status — see AiChatRunService.
     const runHooks: AiChatRunHooks | undefined = autonomousRuns
       ? {
-          begin: (chatId) =>
-            this.aiChatRunService.beginRun({
+          begin: async (chatId) => {
+            const handle = await this.aiChatRunService.beginRun({
               chatId,
               workspaceId: workspace.id,
               userId: user.id,
               trigger: 'user',
-            }),
+            });
+            // #184 phase 1.5: register the run-stream entry at BEGIN (before any
+            // frame) so a tab that attaches in the begin->seed window finds an
+            // entry to wait on. Gated on AI_CHAT_RESUMABLE_STREAM: with the flag
+            // off nothing is registered and attach always 204s.
+            if (
+              handle?.runId &&
+              this.environment?.isAiChatResumableStreamEnabled?.()
+            ) {
+              this.streamRegistry?.open(chatId, handle.runId);
+            }
+            return handle;
+          },
           onAssistantSeeded: (runId, messageId) =>
             this.aiChatRunService.linkAssistantMessage(
               runId,

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

@@ -4,6 +4,7 @@ import { TokenModule } from '../auth/token.module';
 import { AiChatController } from './ai-chat.controller';
 import { AiChatService } from './ai-chat.service';
 import { AiChatRunService } from './ai-chat-run.service';
+import { AiChatStreamRegistryService } from './ai-chat-stream-registry.service';
 import { AiTranscriptionService } from './ai-transcription.service';
 import { AiChatToolsService } from './tools/ai-chat-tools.service';
 import { EmbeddingModule } from './embedding/embedding.module';
@@ -44,6 +45,7 @@ import { PublicShareChatToolsService } from './tools/public-share-chat-tools.ser
   providers: [
     AiChatService,
     AiChatRunService,
+    AiChatStreamRegistryService,
     AiTranscriptionService,
     AiChatToolsService,
     PublicShareChatService,

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

@@ -1,4 +1,13 @@
-import { ForbiddenException } from '@nestjs/common';
+import { ForbiddenException, Logger } from '@nestjs/common';
+// Mock ONLY streamText so a driven stream() call can capture the pipe-options
+// object (consumeSseStream / generateMessageId). Everything else in the AI SDK
+// stays REAL (requireActual), so the pure-helper suites in this file are
+// unaffected — none of them call stream()/streamText.
+jest.mock('ai', () => ({
+  ...jest.requireActual('ai'),
+  streamText: jest.fn(),
+}));
+import { streamText } from 'ai';
 import {
   AiChatService,
   compactToolOutput,
@@ -1059,3 +1068,181 @@ describe('isInterruptResume', () => {
     expect(isInterruptResume(withPrev(null), true)).toBe(false);
   });
 });
+
+/**
+ * #184 phase 1.5 — the run-wrapped pipe options (unit). Drives stream() to the
+ * pipe call with streamText mocked, capturing the options object, and asserts:
+ *  - flag OFF while a runId IS present -> the LEGACY option shape (no
+ *    consumeSseStream, no generateMessageId), and the registry is never touched.
+ *    This is the exact dormancy guarantee this PR rests on.
+ *  - flag ON + runId -> consumeSseStream tees into the registry and
+ *    generateMessageId returns the seeded assistant DB row id.
+ *  - flag ON but no runHooks (runId undefined) -> legacy (the runId gate).
+ *  - flag ON + runId -> the outer catch releases the entry via abortEntry.
+ */
+describe('AiChatService.stream — resumable pipe options (#184 phase 1.5)', () => {
+  const streamTextMock = streamText as unknown as jest.Mock;
+  let pipeMock: jest.Mock;
+
+  beforeEach(() => {
+    streamTextMock.mockReset();
+    pipeMock = jest.fn();
+    streamTextMock.mockReturnValue({
+      consumeStream: jest.fn(),
+      pipeUIMessageStreamToResponse: pipeMock,
+    });
+    // Silence the service's diagnostic logging for a clean test run.
+    jest.spyOn(Logger.prototype, 'log').mockImplementation(() => undefined as never);
+    jest
+      .spyOn(Logger.prototype, 'error')
+      .mockImplementation(() => undefined as never);
+    jest
+      .spyOn(Logger.prototype, 'warn')
+      .mockImplementation(() => undefined as never);
+  });
+
+  afterEach(() => jest.restoreAllMocks());
+
+  // A raw-response stub sufficient for the post-streamText wiring.
+  function makeRes() {
+    return {
+      raw: {
+        writeHead: jest.fn(),
+        write: jest.fn(),
+        once: jest.fn(),
+        on: jest.fn(),
+        flushHeaders: jest.fn(),
+        writableEnded: false,
+        destroyed: false,
+      },
+    };
+  }
+
+  // Wire only the deps reached on the way to the pipe call, plus a spy registry.
+  function makeService(opts: { resumable: boolean }) {
+    const aiChatRepo = {
+      findById: jest.fn(async () => ({ id: 'chat-1', workspaceId: 'ws-1' })),
+      insert: jest.fn(),
+    };
+    const aiChatMessageRepo = {
+      // Both the user insert and the assistant seed return the same row id.
+      insert: jest.fn(async () => ({ id: 'msg-1' })),
+      findAllByChat: jest.fn(async () => []),
+      update: jest.fn(async () => ({ id: 'msg-1' })),
+    };
+    const aiSettings = { resolve: jest.fn(async () => ({})) };
+    const tools = { forUser: jest.fn(async () => ({})) };
+    const mcpClients = {
+      toolsFor: jest.fn(async () => ({
+        tools: {},
+        clients: [],
+        outcomes: [],
+        instructions: [],
+      })),
+    };
+    const streamRegistry = {
+      open: jest.fn(),
+      bind: jest.fn(),
+      abortEntry: jest.fn(),
+    };
+    const svc = new AiChatService(
+      {} as never, // ai (model is injected)
+      aiChatRepo as never,
+      aiChatMessageRepo as never,
+      {} as never, // aiChatPageSnapshotRepo
+      aiSettings as never,
+      tools as never,
+      mcpClients as never,
+      {} as never, // aiAgentRoleRepo
+      {} as never, // pageRepo (openPage undefined -> never touched)
+      {} as never, // pageAccess
+      {
+        isAiChatDeferredToolsEnabled: () => false,
+        isAiChatResumableStreamEnabled: () => opts.resumable,
+      } as never,
+      streamRegistry as never,
+    );
+    return { svc, streamRegistry };
+  }
+
+  const body = {
+    chatId: 'chat-1',
+    messages: [
+      { id: 'm1', role: 'user', parts: [{ type: 'text', text: 'hi' }] },
+    ],
+  };
+
+  const makeRunHooks = () => ({
+    begin: jest.fn(async () => ({
+      runId: 'run-1',
+      signal: new AbortController().signal,
+    })),
+    onAssistantSeeded: jest.fn(),
+    onStep: jest.fn(),
+    onSettled: jest.fn(),
+  });
+
+  async function drive(svc: AiChatService, hooks: unknown): Promise<void> {
+    await svc.stream({
+      user: { id: 'u1' } as never,
+      workspace: { id: 'ws-1' } as never,
+      sessionId: 's1',
+      body: body as never,
+      res: makeRes() as never,
+      signal: new AbortController().signal,
+      model: {} as never,
+      role: null,
+      runHooks: hooks as never,
+    });
+  }
+
+  it('flag OFF + runId present: LEGACY option shape (no consumeSseStream / generateMessageId); registry untouched', async () => {
+    const { svc, streamRegistry } = makeService({ resumable: false });
+    await drive(svc, makeRunHooks());
+    expect(pipeMock).toHaveBeenCalledTimes(1);
+    const options = pipeMock.mock.calls[0][1];
+    // The dormancy guarantee: a live run with the flag off tees NOTHING and does
+    // not stamp a message id — byte-for-byte the pre-1.5 wire.
+    expect(options.consumeSseStream).toBeUndefined();
+    expect(options.generateMessageId).toBeUndefined();
+    expect(streamRegistry.bind).not.toHaveBeenCalled();
+    expect(streamRegistry.abortEntry).not.toHaveBeenCalled();
+  });
+
+  it('flag ON + runId: consumeSseStream tees into the registry; generateMessageId returns the seeded row id', async () => {
+    const { svc, streamRegistry } = makeService({ resumable: true });
+    await drive(svc, makeRunHooks());
+    const options = pipeMock.mock.calls[0][1];
+    expect(typeof options.consumeSseStream).toBe('function');
+    expect(typeof options.generateMessageId).toBe('function');
+    // generateMessageId stamps the seeded assistant DB row id.
+    expect(options.generateMessageId()).toBe('msg-1');
+    // consumeSseStream binds the tee: (chatId, runId, assistantId, stream).
+    const fakeStream = {} as ReadableStream<string>;
+    options.consumeSseStream({ stream: fakeStream });
+    expect(streamRegistry.bind).toHaveBeenCalledWith(
+      'chat-1',
+      'run-1',
+      'msg-1',
+      fakeStream,
+    );
+  });
+
+  it('flag ON but NO runHooks (runId undefined): pipe options stay legacy (the runId gate)', async () => {
+    const { svc, streamRegistry } = makeService({ resumable: true });
+    await drive(svc, undefined);
+    const options = pipeMock.mock.calls[0][1];
+    expect(options.consumeSseStream).toBeUndefined();
+    expect(options.generateMessageId).toBeUndefined();
+    expect(streamRegistry.bind).not.toHaveBeenCalled();
+  });
+
+  it('flag ON + runId: the outer catch calls abortEntry when the stream throws', async () => {
+    const { svc, streamRegistry } = makeService({ resumable: true });
+    streamTextMock.mockImplementation(() => {
+      throw new Error('boom');
+    });
+    await expect(drive(svc, makeRunHooks())).rejects.toThrow('boom');
+    expect(streamRegistry.abortEntry).toHaveBeenCalledWith('chat-1', 'run-1');
+  });
+});

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

@@ -32,6 +32,7 @@ import {
 import { AiChatToolsService } from './tools/ai-chat-tools.service';
 import { McpClientsService } from './external-mcp/mcp-clients.service';
 import { EnvironmentService } from '../../integrations/environment/environment.service';
+import { AiChatStreamRegistryService } from './ai-chat-stream-registry.service';
 import { buildSystemPrompt } from './ai-chat.prompt';
 import {
   CORE_TOOL_KEYS,
@@ -276,6 +277,10 @@ export class AiChatService implements OnModuleInit {
     // Reads the AI_CHAT_DEFERRED_TOOLS toggle (#332). Injected last so existing
     // positional constructor callers (tests) only append one stub.
     private readonly environment: EnvironmentService,
+    // #184 phase 1.5 run-stream registry. OPTIONAL so existing positional
+    // constructions (int-specs) compile unchanged; Nest always injects the real
+    // provider in production. Only ever touched on the run-wrapped + flag-on path.
+    private readonly streamRegistry?: AiChatStreamRegistryService,
   ) {}
 
   /**
@@ -1174,6 +1179,35 @@ export class AiChatService implements OnModuleInit {
         // as the cumulative authoritative usage so the client never jumps DOWN.
         let cumulativeStepUsage: ChatStreamUsage | undefined;
         result.pipeUIMessageStreamToResponse(res.raw, {
+          // #184 phase 1.5: run-wrapped mode only — the legacy path (flag off) stays
+          // byte-for-byte identical, including the absence of start.messageId. Both
+          // fields are gated on `runId` (present only for a durable run) AND the
+          // AI_CHAT_RESUMABLE_STREAM flag; the seed `assistantId` is unconditional,
+          // so gating on `assistantId` alone would change the legacy wire.
+          ...(runId && this.environment?.isAiChatResumableStreamEnabled?.()
+            ? {
+                // Tee the SSE frames into the run-stream registry so late tabs can
+                // attach (replay + live tail).
+                consumeSseStream: ({
+                  stream,
+                }: {
+                  stream: ReadableStream<string>;
+                }) =>
+                  this.streamRegistry?.bind(
+                    chatId,
+                    runId!,
+                    assistantId,
+                    stream,
+                  ),
+                // Stamp the persisted assistant row's DB id onto the streamed
+                // message so every tab renders the SAME id as the DB row (id-based
+                // reconciliation). Seeding is best-effort: when it failed, let the
+                // client generate the id.
+                ...(assistantId
+                  ? { generateMessageId: () => assistantId }
+                  : {}),
+              }
+            : {}),
           headers: { 'X-Accel-Buffering': 'no' },
           // Surface the authoritative chatId on the streamed assistant UI message so
           // the client adopts the REAL id of the row we created, instead of guessing
@@ -1239,6 +1273,12 @@ export class AiChatService implements OnModuleInit {
       // finalizeRun (onSettled) is idempotent — a settle here and a settle from a
       // streamText callback collapse to a single terminal write.
       if (runId) {
+        // #184 phase 1.5: a failure here means the tee `done` will never arrive,
+        // so release the registry entry's subscribers explicitly — otherwise an
+        // attached tab hangs forever. Same flag gate as the tee wiring above.
+        if (this.environment?.isAiChatResumableStreamEnabled?.()) {
+          this.streamRegistry?.abortEntry(chatId, runId);
+        }
         await runHooks?.onSettled?.(
           runId,
           'error',

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/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/environment/environment.service.ts

@@ -292,6 +292,23 @@ export class EnvironmentService {
     return enabled === 'true';
   }
 
+  /**
+   * Resumable SSE transport for durable agent runs (#184 phase 1.5). When
+   * enabled, a run tees its SSE frames into the in-memory run-stream registry so
+   * a late/reloaded tab can attach (replay + live tail) via
+   * `GET /ai-chat/runs/:chatId/stream`. Defaults to DISABLED: PR 1 ships the
+   * server code dormant — with the flag off, `open`/`bind`/`generateMessageId`
+   * are never called and attach always answers 204, so the legacy and #184
+   * phase-1 wire paths stay byte-for-byte identical. Set
+   * AI_CHAT_RESUMABLE_STREAM=true to activate it (paired with the PR 2 client).
+   */
+  isAiChatResumableStreamEnabled(): boolean {
+    const enabled = this.configService
+      .get<string>('AI_CHAT_RESUMABLE_STREAM', 'false')
+      .toLowerCase();
+    return enabled === 'true';
+  }
+
   getPostHogHost(): string {
     return this.configService.get<string>('POSTHOG_HOST');
   }

apps/server/src/integrations/metrics/http-metrics.hook.ts

@@ -2,15 +2,36 @@ import { FastifyReply, FastifyRequest } from 'fastify';
 import { isStreamingResponse } from './metrics.constants';
 import { observeHttp } from './metrics.registry';
 
+// URL path prefixes served by @fastify/static (client build output under
+// client/dist). `/assets/` holds the content-hashed bundle (index-*.js,
+// chunk-*.js) — a NEW set of names every deploy, i.e. an UNBOUNDED label set
+// (#362); the others (/vad/, /brand/, /locales/, /icons/ — copied verbatim from
+// public/) have stable names, so they are merely repetitive per-file labels
+// rather than unbounded. Either way none of these belong in the API-route
+// histogram: collapse them all to one bounded `static` label. (Edge latency for
+// static is already measured by Traefik's traefik_router_request_duration_*.)
+const STATIC_PATH_PREFIXES = [
+  '/assets/',
+  '/vad/',
+  '/brand/',
+  '/locales/',
+  '/icons/',
+];
+
 /**
  * Resolve the BOUNDED route label for an HTTP response.
  *
- * HARD REQUIREMENT (#355): use the ROUTE TEMPLATE (`/pages/:id`), NEVER the raw
- * URL (`/pages/abc-123`), so label cardinality stays finite. Fastify exposes the
- * matched template on `req.routeOptions.url`. On 404s (no route matched) that is
- * missing → collapse to the literal `unknown`.
+ * HARD REQUIREMENT (#355): use the ROUTE TEMPLATE (`/pages/:id`), NEVER a raw
+ * URL (`/pages/abc-123` or `/assets/index-CAbxDtto.js`), so label cardinality
+ * stays finite. Fastify exposes the matched template on `req.routeOptions.url`,
+ * BUT @fastify/static serves each file through a route whose matched url is the
+ * raw (hashed) file path — so for static assets that value is itself unbounded.
+ * Detect static requests by their path prefix FIRST and collapse to `static`;
+ * otherwise use the route template; on a 404 (no route matched) → `unknown`.
  */
 export function resolveRouteLabel(req: FastifyRequest): string {
+  const path = (req.url ?? '').split('?', 1)[0];
+  if (STATIC_PATH_PREFIXES.some((p) => path.startsWith(p))) return 'static';
   const url = req.routeOptions?.url;
   return typeof url === 'string' && url.length > 0 ? url : 'unknown';
 }

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/metrics/metrics.spec.ts

@@ -25,6 +25,61 @@ describe('resolveRouteLabel (histogram route label)', () => {
     const req = { url: '/x' } as unknown as FastifyRequest;
     expect(resolveRouteLabel(req)).toBe('unknown');
   });
+
+  it.each([
+    '/assets/index-CAbxDtto.js',
+    '/assets/chunk-3OPIFGDE-CJOt9nr5.js',
+    '/assets/excalidraw-menu-DpsI0kFW.js',
+    '/vad/silero_vad_v5.onnx',
+    '/brand/logo.svg',
+    '/locales/en.json',
+    '/icons/app-icon-192x192.png',
+  ])('collapses hashed/static asset %p to "static" (#362 cardinality)', (url) => {
+    // @fastify/static serves each file through a route whose matched url is the
+    // raw (hashed) file path, so routeOptions.url is itself unbounded here.
+    const req = {
+      url,
+      routeOptions: { url },
+    } as unknown as FastifyRequest;
+    const label = resolveRouteLabel(req);
+    expect(label).toBe('static');
+    expect(label).not.toContain('.js');
+    expect(label).not.toContain('index-');
+  });
+
+  it('strips the query string before the static-prefix check', () => {
+    const req = {
+      url: '/assets/index-CAbxDtto.js?v=2',
+      routeOptions: { url: '/assets/index-CAbxDtto.js' },
+    } as unknown as FastifyRequest;
+    expect(resolveRouteLabel(req)).toBe('static');
+  });
+
+  it('does NOT collapse a real API route that merely mentions assets', () => {
+    // A templated API route is kept as-is; only the static path PREFIXES collapse.
+    const req = {
+      url: '/api/pages/assets-guide',
+      routeOptions: { url: '/api/pages/:id' },
+    } as unknown as FastifyRequest;
+    expect(resolveRouteLabel(req)).toBe('/api/pages/:id');
+  });
+
+  it.each([
+    // The TRAILING SLASH on the prefix is the anti-false-collapse guard: a path
+    // that is the prefix WITHOUT its slash, or merely shares the prefix as a
+    // substring of a longer segment, must NOT collapse. These would collapse
+    // under a buggy `includes('/assets/')` / slashless-prefix impl.
+    '/assets',
+    '/assetsx/foo.js',
+    '/iconset/x.png',
+  ])('does NOT collapse the prefix-boundary case %p', (url) => {
+    const req = {
+      url,
+      routeOptions: { url: '/some/:route' },
+    } as unknown as FastifyRequest;
+    expect(resolveRouteLabel(req)).not.toBe('static');
+    expect(resolveRouteLabel(req)).toBe('/some/:route');
+  });
 });
 
 describe('isStreamingResponse (SSE exclusion)', () => {

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/ai-chat-attach.int-spec.ts

@@ -0,0 +1,564 @@
+import * as http from 'node:http';
+import { Kysely } from 'kysely';
+import {
+  MockLanguageModelV3,
+  convertArrayToReadableStream,
+} from 'ai/test';
+import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
+import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunRepo } from '@docmost/db/repos/ai-chat/ai-chat-run.repo';
+import { AiChatService } from 'src/core/ai-chat/ai-chat.service';
+import { AiChatRunService } from 'src/core/ai-chat/ai-chat-run.service';
+import {
+  AiChatStreamRegistryService,
+  RunStreamCallbacks,
+} from 'src/core/ai-chat/ai-chat-stream-registry.service';
+import {
+  getTestDb,
+  destroyTestDb,
+  createWorkspace,
+  createUser,
+  createChat,
+} from './db';
+
+/**
+ * #184 phase 1.5 — the resumable transport end to end against REAL Postgres,
+ * the REAL `streamText` (seeded via MockLanguageModelV3) and a REAL Node
+ * ServerResponse, driving the REAL `AiChatService.stream` run-wrapped path with a
+ * REAL `AiChatStreamRegistryService`. The run-hooks mirror the controller: they
+ * begin a durable run and `open()` the registry entry at begin, and the service
+ * tees the SSE frames into it via `consumeSseStream` while stamping the DB row id
+ * via `generateMessageId` (both gated on runId + the resumable flag).
+ *
+ * Proven here: a finished run's replay is the full frame sequence incl `[DONE]`
+ * with `start.messageId` == the seeded DB row id; the anchor check (invariant 6);
+ * an attach opened BEFORE the first frame follows the live stream from frame 0; an
+ * explicit stop surfaces `{"type":"abort"}` + `[DONE]` + end to the subscriber;
+ * and the legacy (non-run) path tees nothing.
+ */
+
+const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
+
+async function waitFor(
+  cond: () => Promise<boolean> | boolean,
+  { timeoutMs = 15_000, stepMs = 25 } = {},
+): Promise<void> {
+  const start = Date.now();
+  while (Date.now() - start < timeoutMs) {
+    if (await cond()) return;
+    await sleep(stepMs);
+  }
+  throw new Error('waitFor: condition not met within timeout');
+}
+
+// A real Node ServerResponse wired to a live socket (as in the stream int-spec).
+function makeRealResponse(): Promise<{
+  res: http.ServerResponse;
+  cleanup: () => Promise<void>;
+}> {
+  return new Promise((resolve) => {
+    const server = http.createServer((_req, res) => {
+      resolve({
+        res,
+        cleanup: () =>
+          new Promise<void>((done) => {
+            try {
+              if (!res.writableEnded) res.end();
+            } catch {
+              /* socket already gone */
+            }
+            server.close(() => done());
+          }),
+      });
+    });
+    server.listen(0, () => {
+      const port = (server.address() as any).port;
+      const creq = http.request({ port, method: 'GET' }, (cres) => {
+        cres.resume();
+      });
+      creq.on('error', () => undefined);
+      creq.end();
+    });
+  });
+}
+
+// A full, successful single-step turn.
+function successStream() {
+  return convertArrayToReadableStream([
+    { type: 'stream-start', warnings: [] },
+    { type: 'text-start', id: 't1' },
+    { type: 'text-delta', id: 't1', delta: 'Hello' },
+    { type: 'text-delta', id: 't1', delta: ' there' },
+    { type: 'text-end', id: 't1' },
+    {
+      type: 'finish',
+      finishReason: 'stop',
+      usage: { inputTokens: 10, outputTokens: 5, totalTokens: 15 },
+    },
+  ] as any);
+}
+
+// A stream the test feeds chunk by chunk (to attach mid-flight / drive a stop).
+function makeControlledStream() {
+  let controller!: ReadableStreamDefaultController<any>;
+  const stream = new ReadableStream<any>({
+    start(c) {
+      controller = c;
+    },
+  });
+  return {
+    stream,
+    emit: (chunk: any) => controller.enqueue(chunk),
+    close: () => controller.close(),
+  };
+}
+
+// Collect replay + live frames from an attachment.
+function liveSink(): {
+  cb: RunStreamCallbacks;
+  frames: string[];
+  ended: () => boolean;
+} {
+  const frames: string[] = [];
+  let ended = false;
+  return {
+    frames,
+    ended: () => ended,
+    cb: {
+      onFrame: (f) => frames.push(f),
+      onEnd: () => {
+        ended = true;
+      },
+    },
+  };
+}
+
+// The SSE `start` frame carries the message id; pull it out of a `data: {...}`.
+function parseStartMessageId(frames: string[]): string | undefined {
+  for (const f of frames) {
+    const m = /^data: (\{.*\})\s*$/m.exec(f.trim());
+    if (!m) continue;
+    try {
+      const json = JSON.parse(m[1]);
+      if (json.type === 'start') return json.messageId;
+    } catch {
+      /* not this frame */
+    }
+  }
+  return undefined;
+}
+
+describe('AiChatService run-stream attach [integration]', () => {
+  let db: Kysely<any>;
+  let aiChatRepo: AiChatRepo;
+  let msgRepo: AiChatMessageRepo;
+  let runRepo: AiChatRunRepo;
+  let workspaceId: string;
+  let userId: string;
+
+  const mcpClients = {
+    toolsFor: async () => ({
+      tools: {},
+      clients: [],
+      outcomes: [],
+      instructions: [],
+    }),
+  };
+
+  // Build the service with the run-stream registry wired and the resumable flag
+  // ON (the property under test). Deferred tools OFF (irrelevant here).
+  function buildService(registry: AiChatStreamRegistryService): AiChatService {
+    return new AiChatService(
+      { getChatModel: async () => null } as any,
+      aiChatRepo,
+      msgRepo,
+      {} as any,
+      { resolve: async () => null } as any,
+      { forUser: async () => ({}) } as any,
+      mcpClients as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {
+        isAiChatDeferredToolsEnabled: () => false,
+        isAiChatResumableStreamEnabled: () => true,
+      } as any,
+      registry,
+    );
+  }
+
+  // Run-hooks mirroring the controller: begin the durable run AND open() the
+  // registry entry at begin. Captures the runId so a test can stop it.
+  function makeRunHooks(
+    runService: AiChatRunService,
+    registry: AiChatStreamRegistryService,
+    box: { runId?: string },
+  ) {
+    return {
+      begin: async (chatId: string) => {
+        const handle = await runService.beginRun({
+          chatId,
+          workspaceId,
+          userId,
+          trigger: 'user',
+        });
+        box.runId = handle.runId;
+        registry.open(chatId, handle.runId);
+        return handle;
+      },
+      onAssistantSeeded: (runId: string, messageId: string) =>
+        runService.linkAssistantMessage(runId, workspaceId, messageId),
+      onStep: (runId: string, n: number) =>
+        void runService.recordStep(runId, workspaceId, n),
+      onSettled: (runId: string, status: any, error?: string) =>
+        runService.finalizeRun(runId, workspaceId, status, error),
+    };
+  }
+
+  function userUiMessage(text: string) {
+    return {
+      id: `u-${Math.random()}`,
+      role: 'user',
+      parts: [{ type: 'text', text }],
+    };
+  }
+
+  async function startRun(opts: {
+    registry: AiChatStreamRegistryService;
+    runService?: AiChatRunService;
+    model: MockLanguageModelV3;
+    chatId: string;
+    body: any;
+    box?: { runId?: string };
+  }): Promise<{ res: http.ServerResponse; cleanup: () => Promise<void> }> {
+    const service = buildService(opts.registry);
+    const { res, cleanup } = await makeRealResponse();
+    const runHooks = opts.runService
+      ? makeRunHooks(opts.runService, opts.registry, opts.box ?? {})
+      : undefined;
+    await service.stream({
+      user: { id: userId, workspaceId } as any,
+      workspace: { id: workspaceId, name: 'WS' } as any,
+      sessionId: 'sess-1',
+      body: opts.body,
+      res: { raw: res } as any,
+      signal: new AbortController().signal,
+      model: opts.model as any,
+      role: null,
+      runHooks,
+    } as any);
+    return { res, cleanup };
+  }
+
+  async function assistantRowId(chatId: string): Promise<string> {
+    const rows = await msgRepo.findAllByChat(chatId, workspaceId);
+    const row = rows.find((r: any) => r.role === 'assistant');
+    return row!.id as string;
+  }
+
+  beforeAll(async () => {
+    db = getTestDb();
+    aiChatRepo = new AiChatRepo(db as any);
+    msgRepo = new AiChatMessageRepo(db as any);
+    runRepo = new AiChatRunRepo(db as any);
+    workspaceId = (await createWorkspace(db)).id;
+    userId = (await createUser(db, workspaceId)).id;
+  });
+
+  afterAll(async () => {
+    await destroyTestDb();
+  });
+
+  it('run-wrapped: replay is the full frame sequence incl [DONE], start.messageId == the seeded DB row id', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const registry = new AiChatStreamRegistryService();
+    const runService = new AiChatRunService(runRepo, {
+      isCloud: () => false,
+    } as never);
+    const model = new MockLanguageModelV3({
+      doStream: async () => ({ stream: successStream() }),
+    } as any);
+
+    const box: { runId?: string } = {};
+    const { cleanup } = await startRun({
+      registry,
+      runService,
+      model,
+      chatId,
+      body: { chatId, messages: [userUiMessage('Hi')] },
+      box,
+    });
+    try {
+      // Wait for the assistant row to settle (terminal callbacks run async).
+      await waitFor(async () => {
+        const rows = await msgRepo.findAllByChat(chatId, workspaceId);
+        return rows.some(
+          (r: any) =>
+            r.role === 'assistant' &&
+            ['completed', 'error', 'aborted'].includes(r.status),
+        );
+      });
+      const rowId = await assistantRowId(chatId);
+
+      // Finished-run replay with expect=live + the correct anchor.
+      const sink = liveSink();
+      const att = await registry.attach(chatId, true, rowId, sink.cb);
+      expect(att).not.toBeNull();
+      expect(att!.finished).toBe(true);
+      // The tee captured frames (consumeSseStream was wired).
+      expect(att!.replay.length).toBeGreaterThan(0);
+      // generateMessageId stamped the DB row id onto the streamed start frame.
+      expect(parseStartMessageId(att!.replay)).toBe(rowId);
+      // The full sequence includes the streamed text and the terminal marker.
+      const joined = att!.replay.join('');
+      expect(joined).toContain('Hello');
+      expect(att!.replay.some((f) => f.includes('[DONE]'))).toBe(true);
+    } finally {
+      registry.onModuleDestroy();
+      await cleanup();
+    }
+  });
+
+  it('anchor mismatch with expect=live returns null (invariant 6)', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const registry = new AiChatStreamRegistryService();
+    const runService = new AiChatRunService(runRepo, {
+      isCloud: () => false,
+    } as never);
+    const model = new MockLanguageModelV3({
+      doStream: async () => ({ stream: successStream() }),
+    } as any);
+
+    const { cleanup } = await startRun({
+      registry,
+      runService,
+      model,
+      chatId,
+      body: { chatId, messages: [userUiMessage('Hi')] },
+    });
+    try {
+      await waitFor(async () => {
+        const rows = await msgRepo.findAllByChat(chatId, workspaceId);
+        return rows.some(
+          (r: any) => r.role === 'assistant' && r.status === 'completed',
+        );
+      });
+      const sink = liveSink();
+      // A foreign anchor must NOT replay this run's transcript.
+      expect(
+        await registry.attach(chatId, true, 'a-different-run-row', sink.cb),
+      ).toBeNull();
+    } finally {
+      registry.onModuleDestroy();
+      await cleanup();
+    }
+  });
+
+  it('an attach opened BEFORE the first frame follows the live stream from frame 0', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const registry = new AiChatStreamRegistryService();
+    const runService = new AiChatRunService(runRepo, {
+      isCloud: () => false,
+    } as never);
+    const controlled = makeControlledStream();
+    const model = new MockLanguageModelV3({
+      doStream: async () => ({ stream: controlled.stream }),
+    } as any);
+
+    const { cleanup } = await startRun({
+      registry,
+      runService,
+      model,
+      chatId,
+      body: { chatId, messages: [userUiMessage('Slow please')] },
+    });
+    try {
+      // Attach while the entry exists (opened at begin) but before any frame.
+      const sink = liveSink();
+      const att = (await registry.attach(chatId, false, undefined, sink.cb))!;
+      expect(att.replay).toEqual([]); // nothing streamed yet -> replay from 0
+      att.start(); // go live (drains nothing, then follows)
+
+      // Now emit the whole turn.
+      controlled.emit({ type: 'stream-start', warnings: [] });
+      controlled.emit({ type: 'text-start', id: 't1' });
+      controlled.emit({ type: 'text-delta', id: 't1', delta: 'Zero' });
+      controlled.emit({ type: 'text-end', id: 't1' });
+      controlled.emit({
+        type: 'finish',
+        finishReason: 'stop',
+        usage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
+      });
+      controlled.close();
+
+      await waitFor(() => sink.frames.some((f) => f.includes('[DONE]')));
+      // The subscriber saw the stream from the very first frame (`start`) through
+      // the terminal marker, with the streamed text present.
+      expect(sink.frames.some((f) => f.includes('"type":"start"'))).toBe(true);
+      expect(sink.frames.join('')).toContain('Zero');
+      expect(sink.frames[sink.frames.length - 1]).toContain('[DONE]');
+      expect(sink.ended()).toBe(true);
+    } finally {
+      registry.onModuleDestroy();
+      await cleanup();
+    }
+  });
+
+  it('requestStop surfaces {"type":"abort"} + [DONE] + end to the attached subscriber', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const registry = new AiChatStreamRegistryService();
+    const runService = new AiChatRunService(runRepo, {
+      isCloud: () => false,
+    } as never);
+    // An abort-AWARE model: it streams some partial output, then errors the model
+    // stream with an AbortError when the run signal aborts — exactly as a real
+    // provider network stream is torn down on abort (a plain in-memory stream
+    // would just stall, so streamText would never observe the stop).
+    const model = new MockLanguageModelV3({
+      doStream: async ({ abortSignal }: any) => {
+        const stream = new ReadableStream<any>({
+          start(controller) {
+            controller.enqueue({ type: 'stream-start', warnings: [] });
+            controller.enqueue({ type: 'text-start', id: 't1' });
+            controller.enqueue({ type: 'text-delta', id: 't1', delta: 'partial' });
+            abortSignal?.addEventListener('abort', () => {
+              try {
+                controller.error(
+                  new DOMException('Aborted', 'AbortError'),
+                );
+              } catch {
+                /* already errored/closed */
+              }
+            });
+          },
+        });
+        return { stream };
+      },
+    } as any);
+
+    const box: { runId?: string } = {};
+    const { cleanup } = await startRun({
+      registry,
+      runService,
+      model,
+      chatId,
+      body: { chatId, messages: [userUiMessage('Start then stop')] },
+      box,
+    });
+    try {
+      const sink = liveSink();
+      const att = (await registry.attach(chatId, false, undefined, sink.cb))!;
+      att.start();
+
+      // Give streamText a beat to begin consuming the partial output.
+      await sleep(250);
+      // User presses Stop -> the run signal aborts -> the SDK emits an abort chunk.
+      await runService.requestStop(box.runId!, workspaceId);
+
+      await waitFor(() => sink.ended());
+      expect(sink.frames.some((f) => f.includes('"type":"abort"'))).toBe(true);
+      expect(sink.frames.some((f) => f.includes('[DONE]'))).toBe(true);
+      expect(sink.ended()).toBe(true);
+    } finally {
+      registry.onModuleDestroy();
+      await cleanup();
+    }
+  });
+
+  it('the outer catch calls abortEntry so an open entry is released (finished)', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const registry = new AiChatStreamRegistryService();
+    const runService = new AiChatRunService(runRepo, {
+      isCloud: () => false,
+    } as never);
+    const model = new MockLanguageModelV3({
+      doStream: async () => ({ stream: successStream() }),
+    } as any);
+
+    // A msgRepo whose user-row insert throws: the turn fails AFTER begin (the
+    // entry is already open) but BEFORE the pipe, exercising the outer catch.
+    const throwingMsgRepo = {
+      insert: async () => {
+        throw new Error('db boom');
+      },
+      findAllByChat: (...a: any[]) => (msgRepo as any).findAllByChat(...a),
+      update: (...a: any[]) => (msgRepo as any).update(...a),
+      findById: (...a: any[]) => (msgRepo as any).findById(...a),
+    };
+    const service = new AiChatService(
+      { getChatModel: async () => null } as any,
+      aiChatRepo,
+      throwingMsgRepo as any,
+      {} as any,
+      { resolve: async () => null } as any,
+      { forUser: async () => ({}) } as any,
+      mcpClients as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {
+        isAiChatDeferredToolsEnabled: () => false,
+        isAiChatResumableStreamEnabled: () => true,
+      } as any,
+      registry,
+    );
+    const { res, cleanup } = await makeRealResponse();
+    const box: { runId?: string } = {};
+    try {
+      await expect(
+        service.stream({
+          user: { id: userId, workspaceId } as any,
+          workspace: { id: workspaceId, name: 'WS' } as any,
+          sessionId: 'sess-1',
+          body: { chatId, messages: [userUiMessage('will throw')] },
+          res: { raw: res } as any,
+          signal: new AbortController().signal,
+          model: model as any,
+          role: null,
+          runHooks: makeRunHooks(runService, registry, box),
+        } as any),
+      ).rejects.toThrow();
+      // The entry opened at begin was terminated by abortEntry (from the catch),
+      // so it is finished and a plain attach returns null instead of hanging.
+      const entry = (registry as any).entries.get(chatId);
+      expect(entry).toBeDefined();
+      expect(entry.finished).toBe(true);
+      const sink = liveSink();
+      expect(await registry.attach(chatId, false, undefined, sink.cb)).toBeNull();
+    } finally {
+      registry.onModuleDestroy();
+      await cleanup();
+    }
+  });
+
+  it('legacy (no run-hooks): the registry is never populated', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const registry = new AiChatStreamRegistryService();
+    const model = new MockLanguageModelV3({
+      doStream: async () => ({ stream: successStream() }),
+    } as any);
+
+    // No runHooks -> runId undefined -> the run-wrapped tee is never wired.
+    const { cleanup } = await startRun({
+      registry,
+      model,
+      chatId,
+      body: { chatId, messages: [userUiMessage('Legacy hi')] },
+    });
+    try {
+      await waitFor(async () => {
+        const rows = await msgRepo.findAllByChat(chatId, workspaceId);
+        return rows.some(
+          (r: any) => r.role === 'assistant' && r.status === 'completed',
+        );
+      });
+      const sink = liveSink();
+      // No entry was ever opened; attach always yields null.
+      expect(await registry.attach(chatId, false, undefined, sink.cb)).toBeNull();
+      expect(await registry.attach(chatId, true, 'anything', sink.cb)).toBeNull();
+    } finally {
+      registry.onModuleDestroy();
+      await cleanup();
+    }
+  });
+});

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": {

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/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/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/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/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);
+    }
+  });
+});

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
@@ -626,10 +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
@@ -16358,17 +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