perf(mcp): кэшировать collab-токен в клиенте — чтобы кэш CollabSession (#400) реально попадал (#435)

CollabSession-реестр (#400/#431) ключуется на `wsUrl + " " + pageId + " " +
token`. Клиент чеканил СВЕЖИЙ collab-токен на КАЖДУЮ мутацию (in-app —
пере-подписывает JWT с новым iat/exp раз в секунду; внешний MCP — POST
/auth/collab-token на каждый вызов), поэтому token-компонент ключа менялся
каждый раз и сессия почти никогда не переиспользовалась (connect-штормы, 25s
таймауты, зомби-сессии).

- новое пер-инстансное поле collabTokenCache {token, mintedAt};
- getCollabTokenWithReauth(forceRefresh=false) сначала смотрит кэш (гейт
  !forceRefresh && ttl>0 && свежесть), оборачивает ОБЕ ветки чеканки
  (provider-путь in-app агента И REST POST /auth/collab-token), обе через
  rememberCollabToken (пустой токен не кэшируется);
- readCollabTokenTtlMs() читает env MCP_COLLAB_TOKEN_TTL_MS свежо; дефолт 5
  мин (сильно ниже 24h жизни токена и <= max-age сессии, окно устаревания
  прав не расширяется сверх #431); ЯВНЫЙ 0/отрицательное отключают кэш
  (rollback-knob = точный fetch-per-call), unset/непарсибельное -> дефолт;
- reauth-ретрай (401/403) в обеих ветках рекурсит forceRefresh=true (обход
  кэша -> свежий токен), гард !forceRefresh ограничивает ровно одним
  повтором;
- кэш сбрасывается на КАЖДОЙ смене идентичности: в login() и в
  response-интерцепторе (где this.token зануляется перед пере-логином) —
  инвариант-4 (изоляция идентичностей) сохранён, кэш пер-клиентный.

Внутренний цикл: 1 проход внутреннего ревью (APPROVE WITH SUGGESTIONS);
правка по ревью — уточнён docstring readCollabTokenTtlMs (расхождение с
поведением: непарсибельное значение даёт дефолт 5мин с ВКЛючённым кэшем, а
не отключает; отключает только явный 0/отрицательное).

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

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

@@ -86,11 +86,19 @@ 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
+// #184 phase 1.5 / #430: backstop for the degraded-poll fallback. The poll is
-// a resume attempt could not attach to the live run and disarmed by the thread on
+// armed when a resume attempt could not attach to the live run and disarmed by the
-// settle / local stream; this cap is the ONLY backstop against an endless tick
+// thread on settle / local stream; this cap is the ONLY backstop against an endless
-// (a stuck 'streaming' row before the boot-sweep, or a user-tail 204 with no run).
+// tick (a stuck 'streaming' row before the boot-sweep, or a user-tail 204 with no
-const DEGRADED_POLL_MAX_MS = 10 * 60_000;
+// run).
+//
+// #430: measured from RUN ACTIVITY, not from arm-time. A real autonomous run takes
+// 11-25 min — longer than a fixed 10-min-from-start cap, which used to cut the poll
+// off mid-run. Instead we cap on INACTIVITY: keep polling as long as the run is
+// still making progress (its persisted rows keep changing), and only give up after
+// this long with NO new activity. A genuinely stuck run produces no row changes, so
+// the idle cap still bounds it; a long-but-progressing run polls to completion.
+const DEGRADED_POLL_IDLE_MAX_MS = 10 * 60_000;
 
 /** Compact token formatter: 1.2M / 3.4k / 950. */
 function formatTokens(n: number): string {
@@ -254,9 +262,12 @@ export default function AiChatWindow() {
   // 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);
+  // #430: timestamp of the LAST run activity while the poll is armed — stamped on
+  // arm and re-stamped whenever the polled rows change (see the effect below). The
+  // idle cap is measured from this, so a long-but-progressing run keeps polling.
+  const lastActivityAtRef = useRef(0);
   const onResumeFallback = useCallback((active: boolean): void => {
-    if (active) armedAtRef.current = Date.now();
+    if (active) lastActivityAtRef.current = Date.now();
     setDegradedPoll(active);
   }, []);
   // Reset the degraded poll whenever the open chat changes: it is scoped to the
@@ -269,18 +280,28 @@ export default function AiChatWindow() {
     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
+      // and while the run is still active (#430: under the INACTIVITY cap, not a
-      // resets fetchFailureCount each fetch, so consecutive errors are not
+      // fixed-from-start cap); otherwise off. NO error checks (TanStack v5 resets
-      // expressible — and the poll must survive a server restart) and NO tail
+      // fetchFailureCount each fetch, so consecutive errors are not expressible —
-      // checks (the settled/local-stream semantics live in ChatThread, which
+      // and the poll must survive a server restart) and NO tail checks (the
-      // disarms via onResumeFallback(false)). The time cap is the only backstop.
+      // settled/local-stream semantics live in ChatThread, which disarms via
+      // onResumeFallback(false)). The idle cap is the only backstop.
       () =>
         degradedPoll === true &&
-        Date.now() - armedAtRef.current < DEGRADED_POLL_MAX_MS
+        Date.now() - lastActivityAtRef.current < DEGRADED_POLL_IDLE_MAX_MS
           ? 2500
           : false,
     );
 
+  // #430: re-stamp the activity clock whenever the polled rows change while the
+  // poll is armed. TanStack keeps the same `messageRows` reference across refetches
+  // that return deep-equal data (structural sharing), so a new reference means the
+  // run genuinely progressed — which extends the inactivity cap above. A stuck run
+  // yields no reference change, so the cap eventually fires and stops the poll.
+  useEffect(() => {
+    if (degradedPoll) lastActivityAtRef.current = Date.now();
+  }, [degradedPoll, messageRows]);
+
   // #184 reconnect-and-live-follow. Whether detached agent runs are enabled for
   // 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

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

@@ -739,3 +739,170 @@ function renderResumable(initialRows: IAiChatMessageRow[]) {
     act(() => view.rerender(<Wrapper rows={rows} />));
   return { rerender, onResumeFallback };
 }
+
+// #430: auto-reconnect to a DETACHED run after a LIVE SSE disconnect. The mount
+// path only resumes on mount/reload; these cover the missing trigger — a live
+// `isDisconnect` on onFinish must (backoff-)re-attach WITHOUT a reload, pin+strip
+// the live row to avoid duplicates, fall back to the degraded poll on a 204, and
+// exhaust to a manual Retry.
+describe("ChatThread — live reconnect after isDisconnect (#430)", () => {
+  // A LIVE local turn that just dropped: the settled tail existed before, and the
+  // partial assistant row lives only in `messages` (not persisted as a tail).
+  const settledTail = () => [
+    row("u1", "user", undefined, "hi"),
+    row("a1", "assistant", "succeeded", "done"),
+  ];
+  // The partial assistant message onFinish hands us for the dropped LIVE turn.
+  const liveMsg = {
+    id: "a2",
+    role: "assistant",
+    parts: [{ type: "text", text: "partial live answer" }],
+  };
+
+  beforeEach(() => {
+    resetState();
+    // status "ready": with a live disconnect the mock is not streaming, so the
+    // status==="streaming" auto-clear effect stays out of the way.
+    h.state.status = "ready";
+    vi.useFakeTimers();
+  });
+  afterEach(() => {
+    vi.useRealTimers();
+    cleanup();
+  });
+
+  // Render a NON-resuming mount (settled tail -> no mount resume) with autonomous
+  // runs on, then simulate a live disconnect via onFinish.
+  function renderLiveThenDisconnect() {
+    const view = renderThread({
+      autonomousRunsEnabled: true,
+      initialRows: settledTail(),
+    });
+    // The settled tail must NOT have triggered a mount resume.
+    expect(h.state.resumeStream).not.toHaveBeenCalled();
+    act(() => {
+      h.state.onFinish?.({
+        message: liveMsg,
+        isAbort: false,
+        isDisconnect: true,
+        isError: false,
+      });
+    });
+    return view;
+  }
+
+  // Fire the pending (scheduled) attempt for `attempt` (backoff = 1s,2s,4s,...).
+  function advanceToAttempt(attempt: number) {
+    act(() => {
+      vi.advanceTimersByTime(1000 * 2 ** (attempt - 1));
+    });
+  }
+
+  // Simulate the reconnect GET returning 204 (nothing live) so the transport's
+  // no-active-stream recovery runs.
+  async function reconnect204() {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({ status: 204, ok: false }),
+    );
+    await act(async () => {
+      await h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+  }
+
+  // Simulate the reconnect GET returning a live 2xx stream.
+  async function reconnect200() {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({ status: 200, ok: true }),
+    );
+    await act(async () => {
+      await h.state.transport!.fetch!("http://x", { method: "GET" });
+    });
+  }
+
+  it("calls resumeStream POST-mount (a live disconnect triggers a backoff reconnect)", () => {
+    renderLiveThenDisconnect();
+    // The banner shows immediately; the attach itself fires after the first backoff.
+    expect(screen.getByText(/reconnecting/i)).toBeTruthy();
+    expect(h.state.resumeStream).not.toHaveBeenCalled();
+    advanceToAttempt(1);
+    // resumeStream is now called AFTER mount — the bug was it only ever fired once
+    // on mount. The reconnect URL pins expect=live&anchor to OUR run.
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(1);
+    expect(h.state.transport!.prepareReconnectToStreamRequest!().api).toBe(
+      "/api/ai-chat/runs/c1/stream?expect=live&anchor=a2",
+    );
+  });
+
+  it("strips the pinned live row before replay so content is NOT duplicated", () => {
+    renderLiveThenDisconnect();
+    advanceToAttempt(1);
+    // The attempt strips the anchor row from the store (the live replay rebuilds
+    // it). Apply the setMessages updater to prove it removes exactly the anchor.
+    const updater = h.state.setMessages.mock.calls.at(-1)![0] as (
+      prev: { id: string }[],
+    ) => { id: string }[];
+    expect(updater([{ id: "u1" }, { id: "a2" }])).toEqual([{ id: "u1" }]);
+  });
+
+  it("a live re-attach (2xx) clears the reconnect banner", async () => {
+    renderLiveThenDisconnect();
+    advanceToAttempt(1);
+    await reconnect200();
+    expect(screen.queryByText(/reconnecting/i)).toBeNull();
+  });
+
+  it("a 204 arms the degraded poll and backs off to the next attempt", async () => {
+    const { onResumeFallback } = renderLiveThenDisconnect();
+    advanceToAttempt(1);
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(1);
+    await reconnect204();
+    // Fallback engaged: the degraded poll is armed (204 -> onNoActiveStream).
+    expect(onResumeFallback).toHaveBeenCalledWith(true);
+    // Still reconnecting — the banner advanced to attempt 2/5.
+    expect(screen.getByText(/reconnecting.*2\/5/i)).toBeTruthy();
+    // The next backoff fires attempt 2 (another resumeStream).
+    advanceToAttempt(2);
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(2);
+  });
+
+  it("exhausts the attempt limit into a manual Retry, which restarts the sequence", async () => {
+    renderLiveThenDisconnect();
+    // Drive all 5 attempts, each failing with a 204.
+    for (let n = 1; n <= 5; n++) {
+      advanceToAttempt(n);
+      expect(h.state.resumeStream).toHaveBeenCalledTimes(n);
+      await reconnect204();
+    }
+    // The 5th 204 exhausted the cap -> the manual Retry replaces the banner.
+    expect(screen.queryByText(/reconnecting/i)).toBeNull();
+    const retry = screen.getByText("Retry");
+    expect(retry).toBeTruthy();
+    // Retry fires attempt 1 immediately (no backoff) — a 6th resumeStream.
+    act(() => {
+      fireEvent.click(retry);
+    });
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(6);
+    expect(screen.getByText(/reconnecting/i)).toBeTruthy();
+  });
+
+  it("does NOT reconnect when autonomous runs are disabled", () => {
+    renderThread({ autonomousRunsEnabled: false, initialRows: settledTail() });
+    act(() => {
+      h.state.onFinish?.({
+        message: liveMsg,
+        isAbort: false,
+        isDisconnect: true,
+        isError: false,
+      });
+    });
+    expect(screen.queryByText(/reconnecting/i)).toBeNull();
+    // The terminal "connection lost" notice is shown instead (unchanged behavior).
+    expect(
+      screen.getByText("Connection lost — the answer was interrupted."),
+    ).toBeTruthy();
+    advanceToAttempt(1);
+    expect(h.state.resumeStream).not.toHaveBeenCalled();
+  });
+});

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

@@ -1,7 +1,17 @@
 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 {
+  ActionIcon,
+  Alert,
+  Box,
+  Button,
+  Group,
+  Loader,
+  Stack,
+  Text,
+  Tooltip,
+} from "@mantine/core";
 import {
   IconClockHour4,
   IconPlayerPlayFilled,
@@ -51,6 +61,15 @@ import classes from "@/features/ai-chat/components/ai-chat.module.css";
 // from the token rate.
 const STREAM_THROTTLE_MS = 50;
 
+// #430: auto-reconnect after a LIVE SSE disconnect of a DETACHED (autonomous) run.
+// The run keeps executing server-side, so instead of a dead "Lost connection"
+// banner we re-attach to the live tail through the SAME resumable machinery the
+// mount path uses. Attempts back off exponentially and are capped; on exhaustion
+// the user gets a manual Retry (the degraded poll keeps catching up underneath).
+const RECONNECT_MAX_ATTEMPTS = 5;
+// Backoff before attempt N (1-based): 1s, 2s, 4s, 8s, 16s.
+const RECONNECT_BASE_DELAY_MS = 1000;
+
 /** The page the user is currently viewing, sent as chat context. */
 export interface OpenPageContext {
   id: string;
@@ -175,6 +194,10 @@ export default function ChatThread({
   const reconcileTailRef = useRef(false);
   const noStreamHandledRef = useRef(false);
   const onNoActiveStreamRef = useRef<(() => void) | null>(null);
+  // #430: called from the transport's reconnect-GET success branch when a live
+  // stream re-attached (2xx, not 204) — clears the reconnect banner. Kept in a ref
+  // because the transport's fetch closure (useMemo([])) reads it live.
+  const onReconnectAttachedRef = 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
@@ -378,6 +401,10 @@ export default function ChatThread({
             // NOT drop the in-progress row or stop tracking the durable run.
             if (response.status === 204 || !response.ok)
               onNoActiveStreamRef.current?.();
+            // #430: a 2xx stream re-attached (live tail or finished-replay). Signal
+            // the reconnect controller to clear its banner. No-op outside an active
+            // reconnect sequence (e.g. the mount attach), so it is safe here.
+            else onReconnectAttachedRef.current?.();
             return response;
           } catch (err) {
             // Network throw: same no-onFinish recovery, then rethrow so the SDK
@@ -481,6 +508,31 @@ export default function ChatThread({
           );
         }
       }
+      // (2b) #430: a LIVE (non-resumed) detached run whose SSE just dropped. The
+      // server run keeps executing, so instead of a dead "Lost connection" banner
+      // start a reconnect sequence: pin the CURRENT streaming assistant row as the
+      // strip/anchor (the live tail is the already-shown partial in `messages`, not
+      // a persistent row) and re-attach to the live tail via the resumable machinery.
+      const startedReconnect =
+        isDisconnect &&
+        !wasResumed &&
+        autonomousRunsEnabled === true &&
+        mountedRef.current &&
+        message?.role === "assistant" &&
+        typeof message.id === "string";
+      if (startedReconnect) {
+        beginReconnect({
+          id: message.id,
+          role: "assistant",
+          content: "",
+          status: "streaming",
+          createdAt: new Date().toISOString(),
+          // Preserve the partial parts so a 204 restore (onNoActiveStream) re-shows
+          // what was on screen while the degraded poll catches the run up to
+          // terminal (rowToUiMessage prefers metadata.parts).
+          metadata: { parts: message.parts },
+        });
+      }
       // (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
@@ -490,9 +542,11 @@ export default function ChatThread({
       onTurnFinished(extractServerChatId(message), threadKey);
       // Show a neutral "stopped" marker for an aborted turn; the red error banner
       // (via `error`) already covers isError, and a clean finish clears any marker.
+      // On a live disconnect that STARTED a reconnect, suppress the terminal
+      // "connection lost" notice — the reconnect banner takes over (#430).
       if (isError) setStopNotice(null);
       else if (isAbort) setStopNotice("manual");
-      else if (isDisconnect) setStopNotice("disconnect");
+      else if (isDisconnect) setStopNotice(startedReconnect ? null : "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
@@ -579,6 +633,106 @@ export default function ChatThread({
 
   const isStreaming = status === "submitted" || status === "streaming";
 
+  // #430: live-disconnect reconnect controller. `null` = idle; `{ trying, attempt }`
+  // = a backoff sequence is running (drives the "reconnecting… (N/max)" banner);
+  // `{ failed }` = attempts exhausted (drives the manual Retry). Mirrored into a ref
+  // so the transport/onNoActiveStream closures branch on the LIVE value.
+  type ReconnectState =
+    | null
+    | { phase: "trying"; attempt: number }
+    | { phase: "failed" };
+  const [reconnectState, setReconnectState] = useState<ReconnectState>(null);
+  const reconnectStateRef = useRef<ReconnectState>(null);
+  const setReconnectStatePair = useCallback((s: ReconnectState) => {
+    reconnectStateRef.current = s;
+    setReconnectState(s);
+  }, []);
+  const reconnectTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
+  const clearReconnectTimer = useCallback(() => {
+    if (reconnectTimerRef.current) {
+      clearTimeout(reconnectTimerRef.current);
+      reconnectTimerRef.current = null;
+    }
+  }, []);
+
+  // One reconnect attempt — MIRRORS the mount strip/anchor path for the LIVE case.
+  // beginReconnect pinned strippedRowRef/stripRef to the run's assistant row, so:
+  //  - remove that row from the store (the mount path strips it from the SEED; here
+  //    it is already shown, so filter it out) — the live replay's `text-start` then
+  //    rebuilds it without DUPLICATING parts (the main dedup risk, #430);
+  //  - reset the one-shot 204 guard so onNoActiveStream can fire for THIS attempt;
+  //  - mark the turn resumed (invariant 7/8) so onFinish runs the recovery block and
+  //    never flushes the queue;
+  //  - resumeStream() -> prepareReconnectToStreamRequest builds
+  //    ?expect=live&anchor=<pinned id>, pinning the replay to OUR run (invariant 6).
+  const attemptReconnectOnce = useCallback(
+    (attempt: number) => {
+      if (!mountedRef.current) return;
+      const anchor = strippedRowRef.current;
+      if (anchor) {
+        setMessages((prev) => prev.filter((m) => m.id !== anchor.id));
+      }
+      noStreamHandledRef.current = false;
+      setResumedTurnPair(true);
+      setReconnectStatePair({ phase: "trying", attempt });
+      void resumeStream();
+    },
+    [setMessages, setResumedTurnPair, setReconnectStatePair, resumeStream],
+  );
+
+  // Schedule attempt `attempt` after an exponential backoff.
+  const scheduleReconnectAttempt = useCallback(
+    (attempt: number) => {
+      clearReconnectTimer();
+      setReconnectStatePair({ phase: "trying", attempt });
+      reconnectTimerRef.current = setTimeout(
+        () => attemptReconnectOnce(attempt),
+        RECONNECT_BASE_DELAY_MS * 2 ** (attempt - 1),
+      );
+    },
+    [clearReconnectTimer, setReconnectStatePair, attemptReconnectOnce],
+  );
+
+  // Start a fresh reconnect sequence, pinning `anchorRow` (the live run's assistant
+  // row) as the strip/anchor reused by every attempt.
+  const beginReconnect = useCallback(
+    (anchorRow: IAiChatMessageRow) => {
+      if (!autonomousRunsEnabled || !mountedRef.current) return;
+      strippedRowRef.current = anchorRow;
+      stripRef.current = true;
+      scheduleReconnectAttempt(1);
+    },
+    [autonomousRunsEnabled, scheduleReconnectAttempt],
+  );
+
+  // Manual Retry (shown once attempts are exhausted): restart at attempt 1 and fire
+  // immediately (the user asked for it now — no backoff).
+  const retryReconnect = useCallback(() => {
+    clearReconnectTimer();
+    attemptReconnectOnce(1);
+  }, [clearReconnectTimer, attemptReconnectOnce]);
+
+  // Live SSE re-attached (the reconnect GET returned a 2xx stream): clear the
+  // banner + any pending backoff. No-op outside a sequence (e.g. the mount attach).
+  const onReconnectAttached = useCallback(() => {
+    if (!mountedRef.current || !reconnectStateRef.current) return;
+    clearReconnectTimer();
+    setReconnectStatePair(null);
+  }, [clearReconnectTimer, setReconnectStatePair]);
+  onReconnectAttachedRef.current = onReconnectAttached;
+
+  // The reconnect GET could not attach (204 / error). onNoActiveStream has already
+  // armed the degraded poll (the robust fallback that drives the row to terminal
+  // from the DB), so this only decides the LIVE-attach retry: back off and try
+  // again up to the cap, else surface the manual Retry.
+  const onReconnectNoStream = useCallback(() => {
+    const s = reconnectStateRef.current;
+    if (s?.phase !== "trying") return;
+    if (s.attempt < RECONNECT_MAX_ATTEMPTS)
+      scheduleReconnectAttempt(s.attempt + 1);
+    else setReconnectStatePair({ phase: "failed" });
+  }, [scheduleReconnectAttempt, setReconnectStatePair]);
+
   // 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
@@ -610,7 +764,17 @@ export default function ChatThread({
     // (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]);
+    // (e) #430: if this 204/error landed during a live-disconnect reconnect
+    // sequence, back off and retry the live attach (or give up to the manual
+    // Retry). The degraded poll armed in (c) is the fallback either way.
+    onReconnectNoStream();
+  }, [
+    setMessages,
+    queryClient,
+    onResumeFallback,
+    setResumedTurnPair,
+    onReconnectNoStream,
+  ]);
   onNoActiveStreamRef.current = onNoActiveStream;
 
   // Mount effect: kick off the resume attempt for a non-settled tail. Marking the
@@ -628,6 +792,9 @@ export default function ChatThread({
     return () => {
       mountedRef.current = false;
       attachAbortRef.current?.abort();
+      // #430: drop any pending reconnect backoff so it can't fire against the next
+      // chat this thread's refs are reused for.
+      if (reconnectTimerRef.current) clearTimeout(reconnectTimerRef.current);
     };
     // Mount-only by design; the parent remounts per chat via `key`.
     // eslint-disable-next-line react-hooks/exhaustive-deps
@@ -666,12 +833,27 @@ export default function ChatThread({
     if (tail.status !== "streaming") {
       reconcileTailRef.current = false;
       onResumeFallback?.(false);
+      // #430: the run reached its terminal state via the degraded poll — there is
+      // no live tail left to reconnect to, so drop any reconnect banner / Retry.
+      clearReconnectTimer();
+      setReconnectStatePair(null);
     }
     // onResumeFallback intentionally omitted (parent-stable callback); deps are
     // fixed by the resume design.
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [initialRows, isStreaming, setMessages]);
 
+  // #430: a real stream is live again — the reconnect re-attached to the live tail
+  // (status -> "streaming") OR the user started a new local turn. Either way clear
+  // the reconnect banner + any pending backoff. Gated on "streaming" (not the
+  // broader "submitted") so a still-pending attach GET does not clear prematurely.
+  useEffect(() => {
+    if (status === "streaming") {
+      clearReconnectTimer();
+      setReconnectStatePair(null);
+    }
+  }, [status, clearReconnectTimer, setReconnectStatePair]);
+
   // "Send now" on a queued message: interrupt the current turn and immediately
   // send THIS message, keeping the agent's partial output. Other queued messages
   // stay queued and flush normally after the new turn. Reuses the existing
@@ -719,6 +901,9 @@ export default function ChatThread({
     // observer's Stop would otherwise leave the attach fetch running.
     attachAbortRef.current?.abort();
     stop();
+    // #430: pressing Stop also cancels an in-progress reconnect sequence.
+    clearReconnectTimer();
+    setReconnectStatePair(null);
     if (!autonomousRunsEnabled) return;
     if (chatIdRef.current) {
       onServerStop?.(chatIdRef.current);
@@ -740,7 +925,13 @@ export default function ChatThread({
       // for this fix. Documented so a future change can address the abort-ordering.
       stopPendingRef.current = true;
     }
-  }, [stop, autonomousRunsEnabled, onServerStop]);
+  }, [
+    stop,
+    autonomousRunsEnabled,
+    onServerStop,
+    clearReconnectTimer,
+    setReconnectStatePair,
+  ]);
 
   // Clear the stopped marker as soon as a new turn begins streaming, and drop any
   // stale "Send now" interrupt flags. On the legit interrupt path both refs are
@@ -825,6 +1016,43 @@ export default function ChatThread({
           detail={errorView.detail}
           mb="xs"
         />
+      ) : reconnectState ? (
+        // #430: while auto-reconnecting to a detached run's live tail, show progress
+        // instead of a dead "Lost connection" banner; once attempts are exhausted,
+        // offer a manual Retry (the degraded poll keeps catching up underneath).
+        <Alert
+          variant="light"
+          color="gray"
+          p="xs"
+          mb="xs"
+          style={{ flexShrink: 0 }}
+        >
+          <Group gap={8} wrap="nowrap" align="center">
+            {reconnectState.phase === "trying" ? (
+              <>
+                <Loader size={14} color="gray" style={{ flex: "none" }} />
+                <Text size="sm" lh={1.3} c="dimmed">
+                  {t("Connection lost — reconnecting…")}
+                  {` (${reconnectState.attempt}/${RECONNECT_MAX_ATTEMPTS})`}
+                </Text>
+              </>
+            ) : (
+              <>
+                <Text size="sm" lh={1.3} c="dimmed" style={{ flex: 1 }}>
+                  {t("Couldn't reconnect to the answer.")}
+                </Text>
+                <Button
+                  size="compact-xs"
+                  variant="light"
+                  color="gray"
+                  onClick={retryReconnect}
+                >
+                  {t("Retry")}
+                </Button>
+              </>
+            )}
+          </Group>
+        </Alert>
       ) : stopNotice ? (
         <ChatStoppedNotice
           text={

apps/client/src/features/editor/components/footnote/footnote-definition-view.tsx

@@ -49,19 +49,14 @@ export default function FootnoteDefinitionView(props: NodeViewProps) {
       className={classes.definition}
       style={{ ["--footnote-number" as any]: `"${number}"` }}
     >
-      {/* #146: contentDOM MUST be the first child — a non-editable marker before
+      {/* #146: contentDOM MUST be the first child — non-editable chrome before
           it makes click hit-testing snap the caret above. Content first; the
-          marker + back-link follow in DOM and are placed left/right via CSS
+          back-link follows in DOM and is placed on the right via CSS flex. The
-          flex `order`. The second #146 mitigation lives in
+          decorative "N." number is rendered inline via the .definitionContent
+          ::before rule (from the --footnote-number var), so no marker element
+          precedes the content. The second #146 mitigation lives in
           editor-paste-handler.tsx (reflowAfterPaste). */}
       <NodeViewContent className={classes.definitionContent} />
-      <span
-        className={classes.definitionMarker}
-        contentEditable={false}
-        aria-hidden="true"
-      >
-        {number}.
-      </span>
       {refCount > 1 ? (
         // Multiple references -> ↩ followed by one lettered link per occurrence.
         <span

apps/client/src/features/editor/components/footnote/footnote.module.css

@@ -81,34 +81,34 @@
 .definition {
   display: flex;
   align-items: flex-start;
-  /* Tight number→text spacing (~one space) so it reads like "1. text"
+  /* Tight spacing between the content and the trailing ↩ back-link. */
-     instead of leaving a wide gap after the period. */
+  gap: 0.3em;
-  gap: 0.4em;
   padding: 2px 0;
+  /* Footnotes read smaller than body text (16px). Matches .listHeading. */
+  font-size: var(--mantine-font-size-sm);
 }
 
-.definitionMarker {
+/* The "N." number is decorative (from the --footnote-number CSS var on the
-  order: -1; /* keep the "N." marker on the LEFT though it follows content in DOM (#146) */
+   wrapper, never in the document model) and is rendered inline at the start of
-  flex: 0 0 auto;
+   the first content line via ::before. This keeps text and wrapped lines flush
-  min-width: 1.5em;
+   to the left margin — no hanging indent — while the editable contentDOM stays
-  /* Right-align within the narrow column so the period sits next to the text
+   the FIRST DOM child (#146). */
-     and multi-digit numbers (10, 11, …) stay aligned on their right edge. */
-  text-align: right;
-  font-variant-numeric: tabular-nums;
-  color: var(--mantine-color-dimmed);
-  user-select: none;
-}
-
 .definitionContent {
   flex: 1 1 auto;
   min-width: 0;
 }
 
-/* The inner editable paragraph inherits `.ProseMirror p { margin: 0.5em 0 }`,
+.definitionContent > :first-child::before {
-   which pushes the first text line ~0.5em below the "N." marker (aligned to
+  content: var(--footnote-number, "?") ". ";
-   flex-start), making the number float above the text. Drop the outer margins
+  color: var(--mantine-color-dimmed);
-   so the marker and the first line share the same top edge — same approach
+  font-variant-numeric: tabular-nums;
-   used for callouts in core.css. */
+  user-select: none;
+}
+
+/* The inner editable paragraph inherits `.ProseMirror p { margin: 0.5em 0 }`.
+   Drop the outer margins so the definition sits tight to the heading above and
+   the ::before number aligns with the top of the row — same approach used for
+   callouts in core.css. */
 .definitionContent > :first-child {
   margin-top: 0;
 }

apps/server/src/collaboration/extensions/hocuspocus-unload-race.spec.ts

@@ -0,0 +1,140 @@
+/**
+ * gitmost #401 — regression test for the connect-vs-unload race in
+ * @hocuspocus/server 3.4.4 (patched via patches/@hocuspocus__server@3.4.4.patch).
+ *
+ * The race (unpatched): when the last client disconnects, storeDocumentHooks'
+ * `finally` schedules an async `unloadDocument`. That unload runs its
+ * `beforeUnloadDocument` hooks asynchronously and, meanwhile, records an
+ * in-flight promise in `this.unloadingDocuments`. In the original 3.4.4
+ * `createDocument`, a NEW connection arriving in that window falls straight
+ * through to the `loadingDocuments`/`documents` checks — it never consults
+ * `unloadingDocuments`. So the new connection can start loading (or reuse) a
+ * document while the old instance is still being torn down; the re-check inside
+ * unload (`shouldUnloadDocument`, which sees 0 connections because async auth
+ * hooks have not registered the new connection yet) then deletes/destroys the
+ * doc out from under the freshly-connected client → orphaned Document → later
+ * redis-sync takes the "doc not loaded" path → sync never completes → the
+ * provider hangs until its ~25s timeout.
+ *
+ * The patch: `createDocument` first awaits any in-flight
+ * `unloadingDocuments.get(name)` before proceeding. Once that settles, the
+ * decision is deterministic — either the doc was fully unloaded (gone from
+ * `documents`, so a clean fresh load) or the unload aborted (healthy doc still
+ * in `documents`, reused). The new connection can never hand-shake onto an
+ * about-to-be-destroyed Document.
+ *
+ * These tests exercise the REAL patched `Hocuspocus.createDocument` (the class
+ * is directly constructible) by seeding `unloadingDocuments` with a controllable
+ * in-flight unload and observing that createDocument waits for it.
+ */
+import { Hocuspocus } from '@hocuspocus/server';
+
+// A promise we can resolve on demand, to model an unload that is mid-flight.
+function deferred<T = void>() {
+  let resolve!: (v: T) => void;
+  const promise = new Promise<T>((r) => {
+    resolve = r;
+  });
+  return { promise, resolve };
+}
+
+describe('gitmost #401 — hocuspocus createDocument awaits in-flight unload', () => {
+  it('does NOT start loading a new doc until the in-flight unload settles, then loads fresh', async () => {
+    const hp = new Hocuspocus();
+    const name = 'page.race';
+
+    // Observe loadDocument: on the unpatched code it is invoked synchronously
+    // within createDocument (before the unload settles); on the patched code it
+    // must be deferred until unloadingDocuments resolves.
+    const freshDoc = { name, __fresh: true } as any;
+    const loadSpy = jest
+      .spyOn(hp as any, 'loadDocument')
+      .mockResolvedValue(freshDoc);
+
+    // Model an unload in progress: an entry sits in unloadingDocuments and, when
+    // it completes, it removes the doc from `documents` (a real full unload).
+    const unload = deferred();
+    (hp as any).documents.set(name, { name, __dying: true });
+    (hp as any).unloadingDocuments.set(
+      name,
+      unload.promise.then(() => {
+        (hp as any).documents.delete(name);
+      }),
+    );
+
+    // Kick off a new connection's createDocument but do not await it yet.
+    const createPromise = (hp as any).createDocument(
+      name,
+      {},
+      'socket-1',
+      { isAuthenticated: true, readOnly: false },
+      {},
+    );
+
+    // Let all currently-schedulable microtasks run. The patched createDocument is
+    // now parked on `await unloadingDocuments.get(name)`, so loadDocument must
+    // NOT have been called yet, and it must NOT have returned the dying doc.
+    await Promise.resolve();
+    await Promise.resolve();
+    expect(loadSpy).not.toHaveBeenCalled();
+
+    // The unload completes (doc removed from `documents`).
+    unload.resolve();
+
+    // createDocument now proceeds: sees no existing doc → fresh load.
+    const doc = await createPromise;
+    expect(loadSpy).toHaveBeenCalledTimes(1);
+    expect(doc).toBe(freshDoc);
+    // The freshly-loaded doc is the one registered — never the dying instance.
+    expect((hp as any).documents.get(name)).toBe(freshDoc);
+  });
+
+  it('reuses the live doc when the in-flight unload aborts (doc left in documents)', async () => {
+    const hp = new Hocuspocus();
+    const name = 'page.abort';
+
+    const loadSpy = jest.spyOn(hp as any, 'loadDocument');
+
+    // Model an unload that ABORTS (e.g. a new connection reappeared before the
+    // sync re-check): it settles WITHOUT deleting the doc from `documents`.
+    const unload = deferred();
+    const liveDoc = { name, __live: true } as any;
+    (hp as any).documents.set(name, liveDoc);
+    (hp as any).unloadingDocuments.set(name, unload.promise); // no-op unload
+
+    const createPromise = (hp as any).createDocument(
+      name,
+      {},
+      'socket-2',
+      { isAuthenticated: true, readOnly: false },
+      {},
+    );
+
+    unload.resolve();
+    const doc = await createPromise;
+
+    // The still-present live doc is reused; no fresh load happened.
+    expect(doc).toBe(liveDoc);
+    expect(loadSpy).not.toHaveBeenCalled();
+  });
+
+  it('no in-flight unload → behaves normally (fresh load)', async () => {
+    const hp = new Hocuspocus();
+    const name = 'page.normal';
+    const freshDoc = { name } as any;
+    const loadSpy = jest
+      .spyOn(hp as any, 'loadDocument')
+      .mockResolvedValue(freshDoc);
+
+    const doc = await (hp as any).createDocument(
+      name,
+      {},
+      'socket-3',
+      { isAuthenticated: true, readOnly: false },
+      {},
+    );
+
+    expect(loadSpy).toHaveBeenCalledTimes(1);
+    expect(doc).toBe(freshDoc);
+  });
+});

apps/server/src/collaboration/extensions/persistence-load.spec.ts

@@ -0,0 +1,130 @@
+/**
+ * gitmost #401 fix 2 — onLoadDocument applies the DB state directly into the
+ * hook's target document and returns undefined (instead of building a NEW Y.Doc
+ * and returning it, which made hocuspocus re-encode+apply the whole state a
+ * SECOND time on every cold load).
+ *
+ * These tests assert:
+ *  - the hook mutates `data.document` in place so its content equals the DB doc,
+ *  - onLoadDocument returns undefined (so hocuspocus keeps the mutated doc and
+ *    does NOT run its own applyUpdate(encodeStateAsUpdate(...)) merge),
+ *  - both the raw-ydoc branch and the json→ydoc conversion branch behave so.
+ *
+ * Returning undefined is the observable signal that the double-encode is gone
+ * (the old code returned a new Y.Doc, which made hocuspocus re-encode+apply the
+ * state a second time); we assert that contract rather than counting internal
+ * encode calls, which is brittle given the encodes inside toYdoc and the test's
+ * own `expected` fixtures.
+ */
+import * as Y from 'yjs';
+import { Document } from '@hocuspocus/server';
+import { TiptapTransformer } from '@hocuspocus/transformer';
+import { PersistenceExtension } from './persistence.extension';
+import { tiptapExtensions } from '../collaboration.util';
+
+// A fresh hocuspocus Document (extends Y.Doc, adds isEmpty()) as hocuspocus
+// hands to onLoadDocument on a cold load.
+const freshDoc = () => new Document(`page.${PAGE_ID}`, {});
+
+const PAGE_ID = '550e8400-e29b-41d4-a716-446655440000';
+
+const doc = (text: string) => ({
+  type: 'doc',
+  content: [{ type: 'paragraph', content: [{ type: 'text', text }] }],
+});
+
+const jsonOf = (ydoc: Y.Doc) =>
+  TiptapTransformer.fromYdoc(ydoc, 'default');
+
+describe('PersistenceExtension.onLoadDocument — #401 fix 2 (apply-into-hook-doc)', () => {
+  let ext: PersistenceExtension;
+  let pageRepo: { findById: jest.Mock };
+
+  beforeEach(() => {
+    pageRepo = { findById: jest.fn() };
+    ext = new PersistenceExtension(
+      pageRepo as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+    );
+    jest.spyOn(ext['logger'], 'debug').mockImplementation(() => undefined);
+    jest.spyOn(ext['logger'], 'warn').mockImplementation(() => undefined);
+  });
+
+  const load = (document: Document) =>
+    ext.onLoadDocument({ documentName: `page.${PAGE_ID}`, document } as any);
+
+  it('raw ydoc branch: mutates the hook doc to the DB state and returns undefined', async () => {
+    // Source doc representing the persisted ydoc state.
+    const source = TiptapTransformer.toYdoc(
+      doc('DB CONTENT'),
+      'default',
+      tiptapExtensions,
+    );
+    const dbState = Buffer.from(Y.encodeStateAsUpdate(source));
+    pageRepo.findById.mockResolvedValue({ id: PAGE_ID, ydoc: dbState });
+
+    // The hook target is a fresh empty doc (as hocuspocus supplies on cold load).
+    const target = freshDoc();
+    const result = await load(target);
+
+    // Return undefined so hocuspocus keeps `target` as-is (no second merge).
+    expect(result).toBeUndefined();
+    // The hook document now carries the DB content.
+    expect(jsonOf(target)).toEqual(jsonOf(source));
+  });
+
+  it('json→ydoc branch: converts page.content into the hook doc and returns undefined', async () => {
+    pageRepo.findById.mockResolvedValue({
+      id: PAGE_ID,
+      ydoc: null,
+      content: doc('JSON CONTENT'),
+    });
+
+    const target = freshDoc();
+    const result = await load(target);
+
+    // Returning undefined is what keeps hocuspocus from re-encoding+applying the
+    // state a second time (the old code returned the doc, forcing that extra
+    // encode). We assert the observable contract here — the return value and the
+    // resulting content — rather than counting internal encode calls, which is
+    // brittle: toYdoc and the `expected` build below both encode too.
+    expect(result).toBeUndefined();
+
+    // The converted content landed in the hook document.
+    const expected = TiptapTransformer.toYdoc(
+      doc('JSON CONTENT'),
+      'default',
+      tiptapExtensions,
+    );
+    expect(jsonOf(target)).toEqual(jsonOf(expected));
+  });
+
+  it('live doc already non-empty: early return, no DB read', async () => {
+    // A hocuspocus Document carrying live content (isEmpty('default') === false).
+    const target = freshDoc();
+    const live = TiptapTransformer.toYdoc(
+      doc('LIVE'),
+      'default',
+      tiptapExtensions,
+    );
+    Y.applyUpdate(target, Y.encodeStateAsUpdate(live));
+
+    const result = await load(target);
+    expect(result).toBeUndefined();
+    expect(pageRepo.findById).not.toHaveBeenCalled();
+  });
+
+  it('no persisted state: leaves the fresh empty doc untouched, returns undefined', async () => {
+    pageRepo.findById.mockResolvedValue({ id: PAGE_ID, ydoc: null, content: null });
+    const target = freshDoc();
+    const result = await load(target);
+    expect(result).toBeUndefined();
+    expect(target.isEmpty('default')).toBe(true);
+  });
+});

apps/server/src/collaboration/extensions/persistence.extension.ts

@@ -171,15 +171,21 @@ export class PersistenceExtension implements Extension {
       return;
     }
 
+    // #401 fix 2 — apply the DB state DIRECTLY into the hook's target document
+    // (`document` === `data.document`) and return undefined. When onLoadDocument
+    // returns undefined, hocuspocus keeps the mutated hook document as-is; only
+    // when the hook RETURNS a Y.Doc does hocuspocus re-`applyUpdate(document,
+    // encodeStateAsUpdate(returned))` — a second full encode+apply of the whole
+    // (e.g. 315KB) state on every cold load. Mutating in place performs a single
+    // apply and avoids the throwaway `new Y.Doc()` allocation.
     if (page.ydoc) {
       this.logger.debug(`ydoc loaded from db: ${pageId}`);
 
-      const doc = new Y.Doc();
       const dbState = new Uint8Array(page.ydoc);
 
-      Y.applyUpdate(doc, dbState);
+      Y.applyUpdate(document, dbState);
       observeCollabLoad(dbState.length, (performance.now() - startedAt) / 1000);
-      return doc;
+      return;
     }
 
     // if no ydoc state in db convert json in page.content to Ydoc.
@@ -192,18 +198,23 @@ export class PersistenceExtension implements Extension {
         tiptapExtensions,
       );
 
-      // Reuse this single encode for the size label (do NOT add a second one).
+      // Encode the converted doc ONCE, reuse the bytes for both the size label
+      // and the single apply into the hook document (previously this encode's
+      // result was returned and hocuspocus re-encoded+applied it a second time).
       const encoded = Y.encodeStateAsUpdate(ydoc);
+      Y.applyUpdate(document, encoded);
       observeCollabLoad(
         encoded.byteLength,
         (performance.now() - startedAt) / 1000,
       );
-      return ydoc;
+      return;
     }
 
+    // No persisted state: the hook document is already a fresh empty Y.Doc, so
+    // leave it untouched and return undefined (no re-encode of an empty doc).
     this.logger.debug(`creating fresh ydoc: ${pageId}`);
     observeCollabLoad(0, (performance.now() - startedAt) / 1000);
-    return new Y.Doc();
+    return;
   }
 
   async onStoreDocument(data: onStoreDocumentPayload) {

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

@@ -17,10 +17,24 @@ import { Injectable, Logger, OnModuleDestroy } from '@nestjs/common';
 /** 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;
+ * Per-run replay buffer cap. Past this the buffer is dropped (attach -> 204, and
+ * the client falls back to its restore + degraded-poll path, #430).
+ *
+ * Raised from 4MB to 32MB (#430): marathon autonomous runs (11-25 min observed)
+ * stream far more than 4MB of SSE frames, so a live disconnect mid-run would find
+ * an already-overflowed buffer and could only degrade-poll instead of re-attaching
+ * to the live tail. 32MB comfortably covers those runs while staying bounded.
+ *
+ * Memory cost: this is the WORST-CASE retained size PER ACTIVE run (the buffer is
+ * freed on finish + retention, or dropped immediately on overflow). With the small
+ * number of concurrent autonomous runs a single workspace realistically has, 32MB
+ * each is an acceptable ceiling; the overflow->204->degraded-poll fallback remains
+ * the backstop for anything larger, so correctness never depends on this bound.
+ */
+export const RUN_STREAM_MAX_BUFFER_BYTES = 32 * 1024 * 1024;
 
-// 2x the replay cap: a just-written 4MB replay burst alone can never trip the
+// 2x the replay cap: a just-written full-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;
 

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

@@ -2,6 +2,7 @@ import {
   AiChatStreamRegistryService,
   RUN_STREAM_MAX_BUFFER_BYTES,
   RUN_STREAM_RETAIN_FINISHED_MS,
+  SUBSCRIBER_MAX_BUFFERED_BYTES,
   RunStreamCallbacks,
 } from './ai-chat-stream-registry.service';
 
@@ -210,9 +211,10 @@ describe('AiChatStreamRegistryService', () => {
     const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
     att.start();
 
-    const oneMb = 'x'.repeat(1024 * 1024);
+    // Cap-relative so it survives a buffer-cap change (#430): a quarter-cap frame
-    // 5 x 1MB = 5MB > 4MB cap; the 5th frame is the one that crosses.
+    // means 5 frames comfortably exceed the replay cap; the last one crosses.
-    for (let i = 0; i < 5; i++) src.push(oneMb + i);
+    const chunk = 'x'.repeat(Math.floor(RUN_STREAM_MAX_BUFFER_BYTES / 4));
+    for (let i = 0; i < 5; i++) src.push(chunk + i);
     await flush();
 
     const entry = (registry as any).entries.get(CHAT);
@@ -220,7 +222,7 @@ describe('AiChatStreamRegistryService', () => {
     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);
+    expect(c.frames[4]).toBe(chunk + 4);
 
     // A NEW attach after overflow gets null (replay buffer is gone).
     const c2 = collector();
@@ -240,9 +242,11 @@ describe('AiChatStreamRegistryService', () => {
     const attB = (await registry.attach(CHAT, false, undefined, b.cb))!;
     attB.start();
 
-    const oneMb = 'x'.repeat(1024 * 1024);
+    // Cap-relative so it survives a buffer-cap change (#430): a quarter-of-the-
-    // 9 x 1MB = 9MB > 8MB per-subscriber cap; A's pending overflows, B streams live.
+    // per-subscriber-cap frame means 5 frames exceed A's paused-pending cap while
-    for (let i = 0; i < 9; i++) src.push(oneMb + i);
+    // B streams every frame live.
+    const chunk = 'x'.repeat(Math.floor(SUBSCRIBER_MAX_BUFFERED_BYTES / 4));
+    for (let i = 0; i < 5; i++) src.push(chunk + i);
     await flush();
 
     const entry = (registry as any).entries.get(CHAT);
@@ -250,7 +254,7 @@ describe('AiChatStreamRegistryService', () => {
     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);
+    expect(b.frames).toHaveLength(5);
 
     // A's start() (arriving late) degrades to an immediate end, not a partial replay.
     attA.start();

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

@@ -148,6 +148,53 @@ describe('assistantParts', () => {
     expect(toolPart).not.toHaveProperty('output');
   });
 
+  it('replays the REAL error text for a THROWN tool (tool-error part)', () => {
+    const steps = [
+      {
+        text: '',
+        toolCalls: [
+          { toolCallId: 'c7', toolName: 'editPageText', input: { id: 'p1' } },
+        ],
+        // A thrown tool is a `tool-error` content part; toolResults holds only
+        // successes and stays empty for this call.
+        toolResults: [],
+        content: [
+          {
+            type: 'tool-error',
+            toolCallId: 'c7',
+            toolName: 'editPageText',
+            input: { id: 'p1' },
+            error: new Error('page is locked'),
+          },
+        ],
+      },
+    ];
+    const parts = assistantParts(steps, '') as AnyPart[];
+    const toolPart = parts.find((p) => p.type === 'tool-editPageText');
+    expect(toolPart).toBeDefined();
+    expect(toolPart!.state).toBe('output-error');
+    // The REAL error is replayed, NOT the 'Tool call did not complete.' placeholder.
+    expect(toolPart!.errorText).toBe('page is locked');
+    expect(toolPart).not.toHaveProperty('output');
+  });
+
+  it('keeps the placeholder ONLY for a call with neither result nor tool-error', () => {
+    const steps = [
+      {
+        text: '',
+        toolCalls: [
+          { toolCallId: 'c8', toolName: 'insertNode', input: { node: {} } },
+        ],
+        toolResults: [],
+        content: [], // aborted mid-step: no result AND no tool-error
+      },
+    ];
+    const parts = assistantParts(steps, '') as AnyPart[];
+    const toolPart = parts.find((p) => p.type === 'tool-insertNode');
+    expect(toolPart!.state).toBe('output-error');
+    expect(toolPart!.errorText).toBe('Tool call did not complete.');
+  });
+
   it('skips malformed tool-calls (missing toolName or toolCallId)', () => {
     const steps = [
       {
@@ -195,6 +242,45 @@ describe('serializeSteps', () => {
     expect(trace[0]).toEqual({ toolName: 'getPage', input: { id: 'p1' } });
     expect(trace[1]).toEqual({ toolName: 'getPage', output: { title: 'T' } });
   });
+
+  it('records a THROWN tool failure (tool-error part) with its error message', () => {
+    const trace = serializeSteps([
+      {
+        toolCalls: [{ toolName: 'editPageText', input: { id: 'p1' } }],
+        toolResults: [],
+        content: [
+          {
+            type: 'tool-error',
+            toolName: 'editPageText',
+            error: new Error('page is locked'),
+          },
+        ],
+      },
+    ]) as Array<Record<string, unknown>>;
+    // The call element is followed by a paired error element (mirroring how a
+    // successful result is appended), so the failure survives in the trace.
+    expect(trace).toHaveLength(2);
+    expect(trace[0]).toEqual({ toolName: 'editPageText', input: { id: 'p1' } });
+    expect(trace[1]).toEqual({
+      toolName: 'editPageText',
+      error: 'page is locked',
+    });
+  });
+
+  it('truncates a very long tool-error message to the tool-output limit', () => {
+    const long = 'x'.repeat(5000);
+    const trace = serializeSteps([
+      {
+        toolCalls: [{ toolName: 'editPageText', input: {} }],
+        toolResults: [],
+        content: [{ type: 'tool-error', toolName: 'editPageText', error: long }],
+      },
+    ]) as Array<Record<string, unknown>>;
+    const errorText = trace[1].error as string;
+    // Truncated (not the full 5000 chars) and carries the omission marker.
+    expect(errorText.length).toBeLessThan(long.length);
+    expect(errorText).toContain('chars omitted');
+  });
 });
 
 describe('rowToUiMessage', () => {

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

@@ -1637,6 +1637,17 @@ type StepLike = {
     toolName?: string;
     output?: unknown;
   }>;
+  // ai@6.0.134: a tool that THREW surfaces as a `tool-error` content part
+  // ({ type:'tool-error', toolCallId, toolName, input, error }), NOT as a
+  // `toolResults` entry (which holds only successes). Read from here so failed
+  // calls are persisted with their real error instead of being dropped.
+  content?: ReadonlyArray<{
+    type?: string;
+    toolCallId?: string;
+    toolName?: string;
+    input?: unknown;
+    error?: unknown;
+  }>;
 };
 
 /**
@@ -1739,6 +1750,26 @@ function compactValue(value: unknown, depth: number): unknown {
   return value;
 }
 
+/**
+ * Extract a bounded string message from a `tool-error` part's `error` field for
+ * persistence and history replay. The field may be an `Error`, a string, or an
+ * arbitrary object, so pull a message robustly. The result is passed through
+ * `compactValue` so a very long error honors the SAME truncation limits the file
+ * already applies to tool outputs (no new limit is introduced here).
+ */
+function normalizeToolError(error: unknown): string {
+  const message =
+    error instanceof Error
+      ? error.message
+      : typeof error === 'string'
+        ? error
+        : error != null &&
+            typeof (error as { message?: unknown }).message === 'string'
+          ? (error as { message: string }).message
+          : String(error);
+  return compactValue(message, 0) as string;
+}
+
 /**
  * Rebuild the FULL UIMessage `parts` for an assistant turn from the SDK steps,
  * so multi-turn history replays prior tool-calls/results to the model (not just
@@ -1771,6 +1802,14 @@ export function assistantParts(
     for (const r of step.toolResults ?? []) {
       if (r.toolCallId) resultsById.set(r.toolCallId, r.output);
     }
+    // Index this step's THROWN tool failures (ai@6 `tool-error` content parts)
+    // by tool call id, so a call that failed replays with its real error text.
+    const errorsById = new Map<string, unknown>();
+    for (const part of step.content ?? []) {
+      if (part.type === 'tool-error' && part.toolCallId) {
+        errorsById.set(part.toolCallId, part.error);
+      }
+    }
     for (const call of step.toolCalls ?? []) {
       if (!call.toolName || !call.toolCallId) continue;
       const hasResult = resultsById.has(call.toolCallId);
@@ -1783,9 +1822,21 @@ export function assistantParts(
           input: call.input,
           output: compactToolOutput(resultsById.get(call.toolCallId)),
         });
+      } else if (errorsById.has(call.toolCallId)) {
+        // The tool THREW: replay the REAL error so the model on the next turn
+        // knows WHY the call failed (and does not blindly repeat it). An
+        // output-error round-trips through convertToModelMessages as a balanced
+        // tool-call + tool-result, keeping the rebuilt history valid.
+        parts.push({
+          type: `tool-${call.toolName}`,
+          toolCallId: call.toolCallId,
+          state: 'output-error',
+          input: call.input,
+          errorText: normalizeToolError(errorsById.get(call.toolCallId)),
+        });
       } else {
-        // No paired result (e.g. aborted mid-step). Persisting a bare
+        // No paired result AND no tool-error (e.g. aborted mid-step). Persisting
-        // tool-call (input-available) would replay as an unpaired call and
+        // a bare tool-call (input-available) would replay as an unpaired call and
         // throw MissingToolResultsError on the next turn (convertToModelMessages
         // emits no tool-result for it). Emit a SYNTHETIC paired result instead:
         // an output-error round-trips through convertToModelMessages as a
@@ -2021,10 +2072,19 @@ export function serializeSteps(
   steps: ReadonlyArray<{
     toolCalls?: ReadonlyArray<{ toolName?: string; input?: unknown }>;
     toolResults?: ReadonlyArray<{ toolName?: string; output?: unknown }>;
+    content?: ReadonlyArray<{
+      type?: string;
+      toolName?: string;
+      error?: unknown;
+    }>;
   }>,
 ): unknown {
-  const calls: Array<{ toolName?: string; input?: unknown; output?: unknown }> =
+  const calls: Array<{
-    [];
+    toolName?: string;
+    input?: unknown;
+    output?: unknown;
+    error?: string;
+  }> = [];
   for (const step of steps ?? []) {
     for (const call of step.toolCalls ?? []) {
       calls.push({ toolName: call.toolName, input: call.input });
@@ -2032,6 +2092,18 @@ export function serializeSteps(
     for (const r of step.toolResults ?? []) {
       calls.push({ toolName: r.toolName, output: compactToolOutput(r.output) });
     }
+    // ai@6 surfaces a THROWN tool failure as a `tool-error` content part, NOT as
+    // a `toolResults` entry. Record it as its own paired element (mirroring how a
+    // successful result is appended) so the failure and its reason survive in the
+    // trace instead of leaving an orphaned call with no result.
+    for (const part of step.content ?? []) {
+      if (part.type === 'tool-error') {
+        calls.push({
+          toolName: part.toolName,
+          error: normalizeToolError(part.error),
+        });
+      }
+    }
   }
   return calls.length > 0 ? calls : null;
 }

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

@@ -729,6 +729,35 @@ export class AiChatToolsService {
           }),
       ),
 
+      // Schema + description live in @docmost/mcp's SHARED_TOOL_SPECS (#423).
+      // meta.hash in the result is the baseHash drawioUpdate requires.
+      drawioGet: sharedTool(
+        sharedToolSpecs.drawioGet,
+        async ({ pageId, node, format }) =>
+          await client.drawioGet(pageId, node, format ?? 'xml'),
+      ),
+
+      // Schema + description live in @docmost/mcp's SHARED_TOOL_SPECS (#423).
+      // The flat schema fields are regrouped into the client's `where` object.
+      drawioCreate: sharedTool(
+        sharedToolSpecs.drawioCreate,
+        async ({ pageId, xml, position, anchorNodeId, anchorText, title }) =>
+          await client.drawioCreate(
+            pageId,
+            { position, anchorNodeId, anchorText },
+            xml,
+            title,
+          ),
+      ),
+
+      // Schema + description live in @docmost/mcp's SHARED_TOOL_SPECS (#423).
+      // baseHash is the optimistic lock: mismatch => structured conflict error.
+      drawioUpdate: sharedTool(
+        sharedToolSpecs.drawioUpdate,
+        async ({ pageId, node, xml, baseHash }) =>
+          await client.drawioUpdate(pageId, node, xml, baseHash),
+      ),
+
       // Schema + description now live in @docmost/mcp's SHARED_TOOL_SPECS (#294).
       // The table reference parameter was unified to `table` (was `tableRef`).
       tableInsertRow: sharedTool(

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

@@ -168,6 +168,32 @@ export interface DocmostClientLike {
     url: string,
     opts?: { align?: 'left' | 'center' | 'right'; alt?: string },
   ): Promise<Record<string, unknown>>;
+  // --- draw.io diagrams (#423, stage 1) ---
+  // Read a diagram as decoded mxGraph XML (default) or the raw .drawio.svg.
+  // meta.hash is the optimistic-lock key drawioUpdate expects as baseHash.
+  drawioGet(
+    pageId: string,
+    node: string,
+    format?: 'xml' | 'svg',
+  ): Promise<Record<string, unknown>>;
+  // Lint mxGraph XML, build the .drawio.svg attachment and insert a drawio node.
+  drawioCreate(
+    pageId: string,
+    where: {
+      position: 'before' | 'after' | 'append';
+      anchorNodeId?: string;
+      anchorText?: string;
+    },
+    xml: string,
+    title?: string,
+  ): Promise<Record<string, unknown>>;
+  // Optimistic-locked full replacement of a diagram (baseHash from drawioGet).
+  drawioUpdate(
+    pageId: string,
+    node: string,
+    xml: string,
+    baseHash: string,
+  ): Promise<Record<string, unknown>>;
   tableInsertRow(
     pageId: string,
     tableRef: string,

docs/reading-ai-logs.md

@@ -13,12 +13,14 @@ Read the **Gotchas** section before you trust any error count.
 - Agent chats live in Postgres, DB `docmost`, tables `ai_chat_*`.
 - Each tool invocation is stored as **two** array elements (a `tool-call` part and
   a `tool-result` part), so naive counting double-counts.
-- **A tool that *throws* writes no result part at all.** Its error text is nowhere
+- **A tool that *throws* writes no result part.** Since the #407 fix its error is
-  in the DB — not in `tool_calls`, `content`, or `metadata`. It is shown live in
+  persisted as a dedicated `{toolName, error}` element in `tool_calls` (queryable +
-  the UI only. So `isError` / `success=false` scans under-report by design.
+  replayed to the model). **Rows written before #407 still drop it** — the error is
-- To find where agents fail you need **three** sources: (1) soft-failure markers in
+  nowhere in the DB and shows only in the live UI. So `isError` / `success=false`
-  `tool_calls`, (2) the orphan-gap proxy for thrown errors, (3) server logs / the
+  scans under-report by design, and pre-#407 thrown errors are invisible.
-  live UI for the actual error text.
+- To find where agents fail: (1) soft-failure markers in `tool_calls`, (2) the new
+  `error` field for thrown errors (new rows) / the orphan-gap proxy (old rows),
+  (3) server logs / the live UI for full stack traces beyond the truncated message.
 
 ## Where the data lives
 
@@ -61,13 +63,17 @@ index 0: { "toolName": "getPage",  "input":  { "pageId": "…" } }   ← tool-ca
 index 1: { "toolName": "getPage",  "output": { … } }                ← tool-result (has output, NO input)
 ```
 
-The **only** keys that ever appear on an element are `toolName`, `input`, `output`.
+The keys that appear on an element are `toolName`, `input`, `output`, and — for a
-There is no `state`, no `errorText`, no `type`. Consequences:
+**thrown** failure on rows written after the #407 fix — `error` (the tool's error
+message; see the "Hard failures" section below). There is no `state`, no `errorText`,
+no `type`. On pre-#407 rows a thrown failure has NO paired result element at all
+(silent orphan). Consequences:
 
-1. **Real invocation count = elements that have `output`.** Counting every element
+1. **Real invocation count = elements that have `output` or `error`.** Counting every
-   double-counts (you get ~2× and a spurious "~50% of every tool has no output").
+   element double-counts (you get ~2× and a spurious "~50% of every tool has no output").
-2. **Pairing:** a successful call = a `tool-call` part followed by its `tool-result`
+2. **Pairing:** a call = a `tool-call` part followed by its result part. A success
-   part. Both carry `toolName`, so you can group by tool on either.
+   carries `output`; a thrown failure (post-#407) carries `error` instead. Both carry
+   `toolName`, so you can group by tool on either.
 
 ## The two classes of failure (and which the DB can see)
 
@@ -85,26 +91,38 @@ These are visible in the `tool-result` `output`. The marker differs per tool:
 Note `editPageText` returns `failed: []` on success — filtering on the *presence*
 of the key gives false positives; filter on **non-empty**.
 
-### 2. Hard failures — tool THREW → NOT PERSISTED ❌ (the trap)
+### 2. Hard failures — tool THREW → NOW PERSISTED ✅ (since the #407 fix)
 
 When a tool throws (the classic one is `patchNode` / `insertNode` / `tableUpdateCell`
 → `Failed to encode document to Yjs (fromJSON): Unknown node type: undefined`), the
-runtime writes **no `tool-result` part**. The orphaned `tool-call` part stays, but
+runtime still writes **no `tool-result` part** — the failure is an ai@6 `tool-error`
-the error text is **nowhere in the DB**. It is streamed to the UI live and (until
+content part instead. **Since the #407 fix, that error is persisted**: `serializeSteps`
-rotation) to server logs — that is it.
+appends a dedicated element `{toolName, error: "<message>"}` right after the failed
+call, mirroring how a successful `{toolName, output}` element is appended. So a thrown
+error now leaves a queryable `error` field carrying its (truncated) reason, and the
+same real text is replayed to the model on the next turn (an `output-error` part with
+the real `errorText`, no longer the `'Tool call did not complete.'` placeholder).
 
-So any query like `count(*) FILTER (WHERE output.success = false)` will happily
+**Cutover caveat — old rows keep the old blind shape.** Rows written **before** this
-return **0** for `patchNode` even when the chat is visibly full of red failures.
+change have the two-part shape (`call` + `output` only) and simply **drop** thrown
-That is survivorship bias, not reliability.
+errors, leaving a silent **orphan** (a `call` with no `output` *and* no `error`). Rows
+written **after** the fix additionally carry the `error` element. So:
 
-The only DB-side proxy for a thrown error is an **orphan**: a `tool-call` part with
+- **New rows:** query the `error` field directly (see the hard-error query below) — no
-no matching `tool-result`. Caveat: orphans also appear when a run is **aborted**
+  orphan heuristic needed for thrown failures.
-mid-flight (server restart), so a high-volume tool (`createComment`, `searchInPage`,
+- **Old rows (pre-#407):** the only DB-side proxy is still an **orphan**: a `tool-call`
-`Search_web_search`) shows orphans from aborts, not from real errors. Treat the
+  part with no matching `tool-result` *and* no `error`. Orphans also appear when a run
-orphan gap as an *upper bound* on hard errors, and cross-check the tool: a gap on a
+  is **aborted** mid-flight (server restart), so a high-volume tool (`createComment`,
-structural editor (`patchNode`, `insertNode`, `updatePageJson`, `transformPage`) is
+  `searchInPage`, `Search_web_search`) shows orphans from aborts, not real errors on
+  old rows. Treat the orphan gap as an *upper bound*, and cross-check the tool: a gap on
+  a structural editor (`patchNode`, `insertNode`, `updatePageJson`, `transformPage`) is
   almost certainly a thrown Yjs-encode error; a gap on `createComment` is mostly aborts.
 
+A note on the aborted-call fallback: a call with **neither** a result **nor** a
+`tool-error` (genuinely interrupted mid-step) still replays with the
+`'Tool call did not complete.'` placeholder and persists as an orphan — that path is
+unchanged, and is distinct from a real thrown error, which now carries `error`.
+
 ### 3. Run-level failures → `ai_chat_runs`
 
 `status` ∈ `succeeded | aborted | failed | running`; `error` holds the text. Seen in
@@ -164,14 +182,28 @@ WHERE jsonb_typeof(o->'failed') = 'array'
 GROUP BY 1 ORDER BY 2 DESC;
 ```
 
-**Hard-error proxy — orphan gap per tool, WITH a spread column** (call parts minus
+**Hard errors — persisted `error` field per tool (NEW rows, since #407)** — thrown
-result parts, plus how many distinct chats the gap is spread across):
+tool failures now carry their real reason, so query them directly:
+
+```sql
+SELECT elem->>'toolName' AS tool, count(*) AS thrown_errors,
+       min(elem->>'error') AS sample_error
+FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
+WHERE jsonb_typeof(m.tool_calls) = 'array' AND elem ? 'error'
+GROUP BY 1 ORDER BY 2 DESC;
+```
+
+**Hard-error proxy for OLD rows (pre-#407) — orphan gap per tool, WITH a spread column**
+(call parts minus result parts, plus how many distinct chats the gap is spread across).
+This covers rows written before thrown errors were persisted; on new rows a thrown
+failure now has its own `error` element (use the query above) and an orphan means only
+a genuinely aborted mid-step call:
 
 ```sql
 WITH parts AS (
   SELECT m.chat_id, elem->>'toolName' AS tool,
          (elem ? 'input' AND NOT (elem ? 'output')) AS is_call,
-         (elem ? 'output')                          AS is_result
+         (elem ? 'output' OR elem ? 'error')        AS is_result
   FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
   WHERE jsonb_typeof(m.tool_calls) = 'array' AND m.role = 'assistant'
 ),
@@ -188,8 +220,12 @@ HAVING sum(gap) FILTER (WHERE gap > 0) > 0
 ORDER BY missing_results DESC;
 ```
 
-**`missing_results` mixes thrown errors AND aborted/interrupted runs — you cannot
+The `is_result` predicate counts an `error` element as a paired result too, so on new
-split them from `output` alone** (a positional "what follows the orphan" heuristic
+rows a persisted thrown error no longer inflates the orphan gap; a remaining gap is an
+aborted/interrupted call.
+
+**On OLD rows, `missing_results` mixes thrown errors AND aborted/interrupted runs — you
+cannot split them from `output` alone** (a positional "what follows the orphan" heuristic
 breaks on parallel tool batches, which persist as `call,call,…,result,result`). Use
 `chats_spread` to disambiguate:
 
@@ -244,18 +280,20 @@ docker compose -p gitmost logs -f --tail=100 # whole stack
 ```
 
 Logging is `json-file`, `max-size=10m max-file=5` → ~50 MB retained, then rotated,
-and **wiped on container recreate**. So thrown-tool error text is only reliably
+and **wiped on container recreate**. Since the #407 fix, thrown-tool error text is
-caught **in real time** (or in the live chat UI, which renders the failed part with
+**persisted in the `error` field** of `tool_calls` (see the hard-error query above), so
-its message). There is no durable, queryable store of hard tool errors today — if you
+you no longer depend on live logs for it. Logs/live UI remain useful for **pre-#407
-need one, that is a feature to add (persist `output-error` parts, or emit a
+rows** (whose thrown errors were dropped) and for full stack traces beyond the
-`tool_calls_total{tool,status}` metric to VictoriaMetrics).
+truncated stored message. A per-tool `tool_calls_total{tool,status}` metric to
+VictoriaMetrics is still a possible future add for aggregate dashboards.
 
 ## Gotchas checklist
 
-- [ ] Counting every `tool_calls` element → **2× overcount**. Count elements with `output`.
+- [ ] Counting every `tool_calls` element → **overcount**. Count `output` elements; add `error` elements for thrown failures (new rows), but don't count both as invocations.
-- [ ] `isError` / `success=false` ≈ 0 does **not** mean "no errors" — thrown errors aren't persisted.
+- [ ] `isError` / `success=false` ≈ 0 does **not** mean "no errors" — thrown errors are a separate `error` element (new rows) or dropped entirely (pre-#407 rows).
+- [ ] Thrown errors persist only on rows written **after the #407 fix** — pre-#407 rows still drop them (orphan only). Mind the cutover when trending over time.
 - [ ] `editPageText.failed` is `[]` on success — test for **non-empty**, not presence.
-- [ ] Orphan gap mixes thrown errors **and** aborted runs — split by tool before concluding.
+- [ ] Orphan gap on OLD rows mixes thrown errors **and** aborted runs — split by tool. On NEW rows a thrown error is its own `error` element, so a gap ≈ aborted call.
 - [ ] `aborted` runs = server restarts, `failed` runs = provider overload — not agent mistakes.
 - [ ] Never dump a raw `tool_calls` cell — it can be hundreds of KB.
 - [ ] Logs are ephemeral (≤50 MB, wiped on recreate) — grab hard-error text live.

package.json

@@ -97,7 +97,8 @@
     "patchedDependencies": {
       "scimmy@1.3.5": "patches/scimmy@1.3.5.patch",
       "yjs@13.6.30": "patches/yjs@13.6.30.patch",
-      "ai@6.0.134": "patches/ai@6.0.134.patch"
+      "ai@6.0.134": "patches/ai@6.0.134.patch",
+      "@hocuspocus/server@3.4.4": "patches/@hocuspocus__server@3.4.4.patch"
     },
     "overrides": {
       "prosemirror-changeset": "2.4.0",

packages/mcp/package.json

@@ -52,6 +52,7 @@
     "form-data": "^4.0.0",
     "jsdom": "^27.4.0",
     "marked": "^17.0.1",
+    "pako": "^2.0.3",
     "re2": "^1.21.0",
     "ws": "^8.19.0",
     "y-prosemirror": "1.3.7",

packages/mcp/src/index.ts

@@ -13,6 +13,11 @@ import { SHARED_TOOL_SPECS, SharedToolSpec } from "./tool-specs.js";
 export { DocmostClient } from "./client.js";
 export type { DocmostMcpConfig } from "./client.js";
 
+// Teardown for the live per-page CollabSession cache (issue #400). An embedding
+// HTTP host (the gitmost NestJS server) should call this from its own shutdown
+// hook so no cached collab provider outlives the process.
+export { destroyAllSessions } from "./lib/collab-session.js";
+
 // Re-export the zod-agnostic shared tool-spec registry so the in-app AI-SDK
 // service can read it off the loaded module (it cannot import the ESM package's
 // internals directly; it goes through loadDocmostMcp()).
@@ -47,7 +52,7 @@ const VERSION = packageJson.version;
 export const SERVER_INSTRUCTIONS =
   "Docmost editing guide — choose the tool by intent.\n" +
   "READ: find a page -> search (workspace-wide full-text); list -> list_pages / list_spaces. Locate blocks and their ids CHEAPLY -> get_outline (compact top-level map; start here, not get_page_json). One block's subtree -> get_node (by attrs.id, or \"#<index>\" for tables, which carry no id). Find every occurrence of a string/regex ON a page (and where each is) -> search_in_page, NOT block-by-block get_node — it returns each hit's node ref + block index + context for a targeted comment. Whole page -> get_page (Markdown, lossy; inline <span data-comment-id> tags are comment anchors — markup, not text) or get_page_json (lossless ProseMirror with block ids). Hand a huge page (with images) to an external consumer without pulling it through the model context -> stash_page (returns a short-lived anonymous URL).\n" +
-  "EDIT: fix wording/typos/numbers -> edit_page_text (find/replace inside blocks, no node id needed). Change ONE block (paragraph/heading/callout/etc.) structurally -> patch_node (by attrs.id from get_outline). Add a block -> insert_node (before/after a block by attrs.id or by anchor text, or append). Remove a block -> delete_node (by attrs.id). Tables -> table_get / table_update_cell / table_insert_row / table_delete_row (address by \"#<index>\" from get_outline; table nodes have no attrs.id). Images -> insert_image (add from a web URL) / replace_image (swap an existing image). Footnotes -> insert_footnote. Bulk/structural rewrite -> update_page_json (full ProseMirror replace; prefer the granular tools above to avoid resending the whole ~100KB+ document). Complex/scripted rewrite (multiple coordinated edits, renumbering) -> docmost_transform: write a JS `(doc, ctx) => doc` transform, preview the diff with dryRun (default), then apply with dryRun:false; ctx.helpers includes commentsToFootnotes for turning inline comments into numbered footnotes.\n" +
+  "EDIT: fix wording/typos/numbers -> edit_page_text (find/replace inside blocks, no node id needed). Change ONE block (paragraph/heading/callout/etc.) structurally -> patch_node (by attrs.id from get_outline). Add a block -> insert_node (before/after a block by attrs.id or by anchor text, or append). Remove a block -> delete_node (by attrs.id). Tables -> table_get / table_update_cell / table_insert_row / table_delete_row (address by \"#<index>\" from get_outline; table nodes have no attrs.id). Images -> insert_image (add from a web URL) / replace_image (swap an existing image). Draw.io diagrams -> drawio_create (create from mxGraph XML and insert), drawio_get (read a diagram as mxGraph XML + a hash), drawio_update (replace a diagram; pass the hash from drawio_get as baseHash for optimistic locking). Footnotes -> insert_footnote. Bulk/structural rewrite -> update_page_json (full ProseMirror replace; prefer the granular tools above to avoid resending the whole ~100KB+ document). Complex/scripted rewrite (multiple coordinated edits, renumbering) -> docmost_transform: write a JS `(doc, ctx) => doc` transform, preview the diff with dryRun (default), then apply with dryRun:false; ctx.helpers includes commentsToFootnotes for turning inline comments into numbered footnotes.\n" +
   "PAGES: new -> create_page (Markdown). Rename (title only) -> rename_page. Move -> move_page. Delete -> delete_page (SOFT delete — the page goes to trash and is restorable; nothing is permanent). Copy/replace a page's whole content from another page (server-side, no document through the model) -> copy_page_content. Sharing -> share_page / unshare_page / list_shares; share_page makes the page PUBLICLY accessible — do it only when explicitly asked.\n" +
   "COMMENTS: create_comment is always inline and requires an EXACT selection — contiguous text from a single block, <=250 chars (fails rather than leaving an unanchored comment); reply to a thread via parentCommentId. Propose a concrete text fix for one-click human approval -> create_comment with suggestedText (the exact plain-text replacement for the selection; the selection must then be UNIQUE in the page — extend it with context if needed); prefer this over editing directly when the change is subjective or needs the author's sign-off. Manage -> list_comments, update_comment, resolve_comment (resolve/reopen, reversible — prefer over delete to close), delete_comment, check_new_comments.\n" +
   "HISTORY: review what changed -> diff_page_versions (a historyId vs current, or two versions). List saved versions -> list_page_history. Undo a bad edit -> restore_page_version (writes a past version back as current; itself revertible). Lossless markdown round-trip (download, edit, re-upload, incl. comment anchors) -> export_page_markdown / import_page_markdown.";
@@ -460,6 +465,38 @@ registerShared(
   },
 );
 
+// Tool: drawio_get — read a draw.io diagram as mxGraph XML (or the raw SVG).
+registerShared(
+  SHARED_TOOL_SPECS.drawioGet,
+  async ({ pageId, node, format }) => {
+    const result = await docmostClient.drawioGet(pageId, node, format ?? "xml");
+    return jsonContent(result);
+  },
+);
+
+// Tool: drawio_create — lint mxGraph XML, build the .drawio.svg, insert a node.
+registerShared(
+  SHARED_TOOL_SPECS.drawioCreate,
+  async ({ pageId, xml, position, anchorNodeId, anchorText, title }) => {
+    const result = await docmostClient.drawioCreate(
+      pageId,
+      { position, anchorNodeId, anchorText },
+      xml,
+      title,
+    );
+    return jsonContent(result);
+  },
+);
+
+// Tool: drawio_update — optimistic-locked full replacement of a diagram.
+registerShared(
+  SHARED_TOOL_SPECS.drawioUpdate,
+  async ({ pageId, node, xml, baseHash }) => {
+    const result = await docmostClient.drawioUpdate(pageId, node, xml, baseHash);
+    return jsonContent(result);
+  },
+);
+
 // Tool: share_page
 // Schema + description now live in the shared registry (#294). The execute body
 // keeps this transport's own `searchIndexing ?? true` default.

packages/mcp/src/lib/collab-session.ts

@@ -0,0 +1,668 @@
+import { HocuspocusProvider } from "@hocuspocus/provider";
+import { TiptapTransformer } from "@hocuspocus/transformer";
+import * as Y from "yjs";
+import WebSocket from "ws";
+import {
+  buildCollabWsUrl,
+  applyDocToFragment,
+  MutationResult,
+} from "./collaboration.js";
+import { summarizeChange } from "./diff.js";
+
+/**
+ * Live per-page collaboration session cache (issue #400).
+ *
+ * The one-shot write path (collaboration.mutatePageContent /
+ * client.mutateLiveContentUnlocked) used to open a NEW HocuspocusProvider, run
+ * the full connect -> auth -> onLoadDocument -> initial-sync handshake, apply a
+ * single edit, wait for persistence, and then `provider.destroy()` — for EVERY
+ * content mutation. Disconnecting after every edit means that once the pause
+ * between calls exceeds the server's write debounce, the server does a full
+ * store -> unload -> reload per cell, causing 25s connect timeouts and
+ * event-loop lag under a burst of edits on one page.
+ *
+ * This module keeps ONE live provider + ydoc per (wsUrl, pageId, token) alive
+ * across a SERIES of edits. While the provider stays connected the server never
+ * enters store -> unload -> reload, its debounce coalesces N writes into 1-2
+ * stores, and the repeated auth/load/initial-sync disappears.
+ *
+ * The synchronous read -> transform -> write section and the per-edit
+ * persistence-ack logic are preserved VERBATIM from the one-shot machine — the
+ * only change is that they run on a persistent provider instead of a throwaway
+ * one. See CollabSession.mutate.
+ */
+
+/** Time we wait for the initial handshake/sync before giving up. */
+const CONNECT_TIMEOUT_MS = 25000;
+/** Time we wait for the server to acknowledge our write before giving up. */
+const PERSIST_TIMEOUT_MS = 20000;
+
+/**
+ * Tunables, read fresh from the environment on every acquire so tests (and a
+ * live rollback) can change them without reloading the module. Mirrors how
+ * http.ts parses MCP_SESSION_IDLE_MS.
+ *   - MCP_COLLAB_SESSION_IDLE_MS: idle TTL, reset after every op. Default 60s.
+ *     0 (or negative) DISABLES the cache — every op opens its own provider and
+ *     destroys it after the op, i.e. the exact legacy per-op-provider behavior
+ *     (the rollback path).
+ *   - MCP_COLLAB_SESSION_MAX_AGE_MS: hard lifetime checked at acquire; bounds
+ *     the permission-staleness window. Default 10 min.
+ *   - MCP_COLLAB_SESSION_MAX_ENTRIES: registry cap; the least-recently-used
+ *     session is destroy-evicted when the cap is reached. Default 32.
+ */
+interface SessionConfig {
+  idleMs: number;
+  maxAgeMs: number;
+  maxEntries: number;
+}
+
+function parseEnvInt(value: string | undefined, fallback: number): number {
+  const parsed = parseInt(value ?? "", 10);
+  return Number.isFinite(parsed) ? parsed : fallback;
+}
+
+function readConfig(): SessionConfig {
+  // idleMs: allow 0 (disable). A malformed value falls back to the default.
+  const idleRaw = parseInt(process.env.MCP_COLLAB_SESSION_IDLE_MS ?? "", 10);
+  const idleMs = Number.isFinite(idleRaw) ? Math.max(0, idleRaw) : 60 * 1000;
+  const maxAgeMs = Math.max(
+    0,
+    parseEnvInt(process.env.MCP_COLLAB_SESSION_MAX_AGE_MS, 10 * 60 * 1000),
+  );
+  const maxEntriesRaw = parseEnvInt(
+    process.env.MCP_COLLAB_SESSION_MAX_ENTRIES,
+    32,
+  );
+  const maxEntries = maxEntriesRaw > 0 ? maxEntriesRaw : 32;
+  return { idleMs, maxAgeMs, maxEntries };
+}
+
+/**
+ * The subset of HocuspocusProvider this module depends on, so the provider can
+ * be replaced with a fake in unit tests (there is no server in the test env).
+ */
+export interface CollabProviderLike {
+  synced: boolean;
+  unsyncedChanges: number;
+  destroy(): void;
+  on(event: "unsyncedChanges", handler: (data: { number: number }) => void): void;
+  off(event: "unsyncedChanges", handler: (data: { number: number }) => void): void;
+}
+
+/** The configuration object passed to the provider factory. */
+export interface CollabProviderConfig {
+  url: string;
+  name: string;
+  document: Y.Doc;
+  token: string;
+  WebSocketPolyfill: unknown;
+  onConnect: () => void;
+  onSynced: () => void;
+  onDisconnect: () => void;
+  onClose: () => void;
+  onAuthenticationFailed: () => void;
+}
+
+export type CollabProviderFactory = (
+  config: CollabProviderConfig,
+) => CollabProviderLike;
+
+const defaultProviderFactory: CollabProviderFactory = (config) =>
+  // @ts-ignore - WebSocketPolyfill is required for the Node.js environment.
+  new HocuspocusProvider(config) as unknown as CollabProviderLike;
+
+let providerFactory: CollabProviderFactory = defaultProviderFactory;
+
+/**
+ * TEST SEAM: swap the provider factory (pass null to restore the real one).
+ * Not part of the public API — used only by the unit tests, which cannot reach
+ * a real collaboration server.
+ */
+export function __setCollabProviderFactory(
+  factory: CollabProviderFactory | null,
+): void {
+  providerFactory = factory ?? defaultProviderFactory;
+}
+
+/** Optional per-acquire hooks (metrics), passed through from the call site. */
+export interface AcquireOptions {
+  /** Invoked when the initial connect handshake times out (CONNECT_TIMEOUT_MS). */
+  onConnectTimeout?: () => void;
+}
+
+type SessionState = "connecting" | "ready" | "dead";
+
+/**
+ * One live provider + ydoc for a single (wsUrl, pageId, token) triple.
+ *
+ * Lifecycle: connecting -> ready -> dead. A session becomes `dead` on the first
+ * disconnect/close/auth-failure at ANY time, on an idle/eviction/max-age
+ * teardown, or on an explicit destroy(); death is terminal and removes the
+ * session from the registry so the next acquire opens a fresh one. We never use
+ * the provider's auto-reconnect — destroying on the first disconnect closes the
+ * "reconnect drove unsyncedChanges to 0 without retransmitting our write" class
+ * of false success.
+ */
+export class CollabSession {
+  readonly key: string;
+  readonly pageId: string;
+  readonly wsUrl: string;
+  readonly token: string;
+  readonly createdAt: number;
+  state: SessionState = "connecting";
+  /**
+   * Set true on disconnect/close/auth-failure so a reconnect-driven
+   * unsyncedChanges->0 cannot be mistaken for a successful persist of our
+   * write (preserved verbatim from the one-shot machine).
+   */
+  connectionLost = false;
+
+  provider: CollabProviderLike | undefined;
+  private readonly ydoc: Y.Doc;
+  private readonly cfg: SessionConfig;
+  /**
+   * Ephemeral sessions (cache disabled, MCP_COLLAB_SESSION_IDLE_MS<=0) are never
+   * registered and self-destroy after their single op — the legacy
+   * provider-per-op behavior.
+   */
+  private readonly ephemeral: boolean;
+  private readonly opts: AcquireOptions | undefined;
+
+  private dead = false;
+  private connectTimer: ReturnType<typeof setTimeout> | undefined;
+  private idleTimer: ReturnType<typeof setTimeout> | undefined;
+  private openPromise: Promise<void> | undefined;
+  private openResolve: (() => void) | undefined;
+  private openReject: ((err: Error) => void) | undefined;
+  private openSettled = false;
+  /**
+   * The rejector of the CURRENT in-flight mutate, if any. A disconnect/close/
+   * auth-failure or timeout at ANY time rejects the in-flight op through this
+   * with the SAME error text the one-shot machine emitted.
+   */
+  private inflightReject: ((err: Error) => void) | undefined;
+
+  constructor(
+    key: string,
+    pageId: string,
+    wsUrl: string,
+    token: string,
+    cfg: SessionConfig,
+    ephemeral: boolean,
+    opts: AcquireOptions | undefined,
+  ) {
+    this.key = key;
+    this.pageId = pageId;
+    this.wsUrl = wsUrl;
+    this.token = token;
+    this.cfg = cfg;
+    this.ephemeral = ephemeral;
+    this.opts = opts;
+    this.createdAt = Date.now();
+    this.ydoc = new Y.Doc();
+  }
+
+  /**
+   * A cached session may be reused only when it is fully ready, still synced,
+   * has not lost its connection, and has not exceeded its max age (invariant 5
+   * "validate on reuse" + the max-age acquire check).
+   */
+  isReusable(): boolean {
+    return (
+      !this.dead &&
+      this.state === "ready" &&
+      !this.connectionLost &&
+      !!this.provider &&
+      this.provider.synced === true &&
+      Date.now() - this.createdAt < this.cfg.maxAgeMs
+    );
+  }
+
+  /**
+   * Connect and wait for the initial sync (onSynced) within CONNECT_TIMEOUT_MS.
+   * Idempotent: repeated calls return the same in-flight/settled promise.
+   */
+  open(): Promise<void> {
+    if (this.openPromise) return this.openPromise;
+    this.openPromise = new Promise<void>((resolve, reject) => {
+      this.openResolve = resolve;
+      this.openReject = reject;
+
+      this.connectTimer = setTimeout(() => {
+        // The 25s connect timeout: the collab connection never became ready.
+        this.opts?.onConnectTimeout?.();
+        this.teardown(
+          new Error("Connection timeout to collaboration server"),
+          false,
+        );
+      }, CONNECT_TIMEOUT_MS);
+
+      if (process.env.DEBUG)
+        console.error(`Connecting to WebSocket: ${this.wsUrl}`);
+
+      this.provider = providerFactory({
+        url: this.wsUrl,
+        name: `page.${this.pageId}`,
+        document: this.ydoc,
+        token: this.token,
+        WebSocketPolyfill: WebSocket,
+        onConnect: () => {
+          if (process.env.DEBUG) console.error("WS Connect");
+        },
+        // An unexpected disconnect/close at ANY time (during the connect-wait,
+        // between edits, or during a persistence wait) makes the session dead:
+        // surface it now instead of hanging, reject any in-flight op with the
+        // same error text as the one-shot machine, and remove ourselves from
+        // the registry so the next acquire opens fresh. `teardown` is idempotent
+        // so the onClose our own destroy() triggers is a harmless no-op.
+        onDisconnect: () => {
+          if (process.env.DEBUG) console.error("WS Disconnect");
+          this.teardown(
+            new Error(
+              "Collaboration connection closed before the update was persisted/synced",
+            ),
+            true,
+          );
+        },
+        onClose: () => {
+          if (process.env.DEBUG) console.error("WS Close");
+          this.teardown(
+            new Error(
+              "Collaboration connection closed before the update was persisted/synced",
+            ),
+            true,
+          );
+        },
+        onSynced: () => {
+          if (this.dead || this.openSettled) return;
+          if (process.env.DEBUG) console.error("Connected and synced!");
+          if (this.connectTimer) {
+            clearTimeout(this.connectTimer);
+            this.connectTimer = undefined;
+          }
+          this.state = "ready";
+          this.openSettled = true;
+          this.openResolve?.();
+        },
+        onAuthenticationFailed: () => {
+          this.teardown(
+            new Error("Authentication failed for collaboration connection"),
+            true,
+          );
+        },
+      });
+    });
+    return this.openPromise;
+  }
+
+  /**
+   * Run one atomic read -> transform -> write against the LIVE doc and wait for
+   * the server to acknowledge the write.
+   *
+   * INVARIANT 1 (read->write atomicity): between `TiptapTransformer.fromYdoc`
+   * and `applyDocToFragment` there is NO `await`. Yjs applies remote updates
+   * only when the event loop yields, so this synchronous block sees a consistent
+   * live doc and no concurrent human edit can interleave and be clobbered —
+   * exactly as in the one-shot onSynced code, just on a persistent provider.
+   *
+   * INVARIANT 2 (per-edit ack): after the write, resolve immediately if
+   * unsyncedChanges is already 0, else wait for the unsyncedChanges->0 event
+   * (PERSIST_TIMEOUT_MS), guarded by connectionLost so a reconnect handshake
+   * cannot report a false success.
+   *
+   * CONCURRENCY: not safe to invoke concurrently on ONE session — the caller
+   * MUST serialize (hold the per-page lock), mirroring acquireCollabSession.
+   * The in-flight op is tracked in a single `inflightReject` field, so an
+   * overlapping second call would clobber the first's rejector and leave it
+   * hanging on disconnect. A fail-fast guard below rejects the overlap instead.
+   * Sequential (awaited) mutates are fine: localFinish clears inflightReject
+   * before the promise settles, so the guard is clear by the time the next runs.
+   */
+  mutate(
+    transform: (liveDoc: any) => any | null,
+  ): Promise<MutationResult> {
+    // Belt-and-suspenders (acquire already validated): refuse to write on a
+    // session that is not in a live, synced, ready state.
+    if (
+      this.dead ||
+      this.state !== "ready" ||
+      this.connectionLost ||
+      !this.provider ||
+      this.provider.synced !== true
+    ) {
+      return Promise.reject(
+        new Error("Collaboration session is not in a ready state"),
+      );
+    }
+
+    // Fail-fast on concurrent use: a second overlapping mutate would overwrite
+    // the first's inflightReject, so a disconnect would only reject the second
+    // and hang the first until PERSIST_TIMEOUT_MS. Reject the overlap WITHOUT
+    // touching the in-flight op's state (no localFinish/teardown here).
+    if (this.inflightReject) {
+      return Promise.reject(
+        new Error(
+          "mutate already in-flight; caller must serialize (hold the page lock)",
+        ),
+      );
+    }
+
+    return new Promise<MutationResult>((resolve, reject) => {
+      let settled = false;
+      let persistTimer: ReturnType<typeof setTimeout> | undefined;
+      let unsyncedHandler:
+        | ((data: { number: number }) => void)
+        | undefined;
+      // The verifiable result resolved on every success/abort path. Set on
+      // abort (no-op report) and after a real write (computed change report).
+      let mutationResult: MutationResult;
+
+      const localFinish = (err: Error | null, value?: MutationResult) => {
+        if (settled) return;
+        settled = true;
+        if (persistTimer) clearTimeout(persistTimer);
+        if (unsyncedHandler && this.provider) {
+          try {
+            this.provider.off("unsyncedChanges", unsyncedHandler);
+          } catch (e) {}
+        }
+        this.inflightReject = undefined;
+        if (err) reject(err);
+        else resolve(value as MutationResult);
+        // Post-settle lifecycle: an ephemeral (cache-disabled) session dies with
+        // its single op; a cached session that is still alive re-arms its idle
+        // TTL so the clock starts from the LAST op.
+        if (this.ephemeral) {
+          this.destroy("ephemeral op complete");
+        } else if (!this.dead) {
+          this.armIdle();
+        }
+      };
+
+      // Register so a disconnect/close/auth-failure/teardown rejects THIS op
+      // with the connection-loss error text. localFinish's `settled` guard makes
+      // a racing teardown + normal resolve safe (first one wins).
+      this.inflightReject = (e: Error) => localFinish(e);
+
+      // Resolve once the server acknowledges our update: the provider increments
+      // unsyncedChanges when the local update is sent and decrements it on the
+      // server's SyncStatus(applied=true); reaching 0 means the authoritative
+      // in-memory ydoc on the server now contains our write.
+      const waitForPersistence = () => {
+        if (settled) return;
+        // A missing provider is a failure, not a success: without it the write
+        // can never have been acknowledged.
+        if (!this.provider) {
+          localFinish(new Error("collab provider gone before persistence"));
+          return;
+        }
+        if (this.provider.unsyncedChanges === 0) {
+          localFinish(null, mutationResult);
+          return;
+        }
+        persistTimer = setTimeout(() => {
+          localFinish(
+            new Error(
+              "Timeout waiting for collaboration server to persist the update",
+            ),
+          );
+        }, PERSIST_TIMEOUT_MS);
+        unsyncedHandler = (data: { number: number }) => {
+          // Only treat unsyncedChanges->0 as success when the connection is
+          // still up. A transient disconnect + reconnect handshake can drive the
+          // counter back to 0 without our write being re-transmitted; in that
+          // case let the disconnect/close error win instead.
+          if (data.number === 0 && !this.connectionLost) {
+            localFinish(null, mutationResult);
+          }
+        };
+        this.provider.on("unsyncedChanges", unsyncedHandler);
+      };
+
+      // CRITICAL: everything between reading the live doc and writing it back
+      // must stay synchronous (no await). While the JS event loop is not
+      // yielded, no incoming remote update can interleave, so any already-synced
+      // concurrent edits are preserved in liveDoc.
+      let newDoc: any;
+      let beforeDoc: any;
+      try {
+        let liveDoc = TiptapTransformer.fromYdoc(this.ydoc, "default");
+        if (
+          !liveDoc ||
+          typeof liveDoc !== "object" ||
+          !Array.isArray(liveDoc.content)
+        ) {
+          liveDoc = { type: "doc", content: [] };
+        }
+
+        // Snapshot the before-doc for the change report. Docs are
+        // JSON-serializable, so this is a safe deep clone.
+        beforeDoc = JSON.parse(JSON.stringify(liveDoc));
+
+        newDoc = transform(liveDoc);
+
+        if (newDoc == null) {
+          // Transform aborted — write nothing, return the live doc with a no-op
+          // change report.
+          mutationResult = {
+            doc: liveDoc,
+            verify: {
+              changed: false,
+              textInserted: 0,
+              textDeleted: 0,
+              blocksChanged: 0,
+              marks: {},
+              summary: "no changes (transform aborted)",
+            },
+          };
+          localFinish(null, mutationResult);
+          return;
+        }
+
+        // Structural diff into the live fragment (issue #152): preserves the Yjs
+        // ids of unchanged nodes, so an open editor's cursor is not yanked to the
+        // end of the document on every agent write.
+        applyDocToFragment(this.ydoc, newDoc);
+      } catch (e) {
+        // Includes errors thrown by transform (e.g. "afterText not found",
+        // "text not found"): propagate them verbatim to the caller.
+        localFinish(e instanceof Error ? e : new Error(String(e)));
+        return;
+      }
+
+      // Compute the verifiable change report AFTER the transact write: it only
+      // needs the JSON before/after, so it cannot affect the atomic read->write
+      // window, and summarizeChange never throws.
+      mutationResult = {
+        doc: newDoc,
+        verify: summarizeChange(beforeDoc, newDoc),
+      };
+      if (process.env.DEBUG)
+        console.error("Content written, waiting for server to persist...");
+      waitForPersistence();
+    });
+  }
+
+  /** (Re)arm the idle TTL so the clock starts from the most recent activity. */
+  armIdle(): void {
+    if (this.dead || this.ephemeral) return;
+    if (this.idleTimer) clearTimeout(this.idleTimer);
+    if (this.cfg.idleMs > 0) {
+      this.idleTimer = setTimeout(() => {
+        this.destroy("idle timeout");
+      }, this.cfg.idleMs);
+      // Never let the idle timer keep the process alive.
+      (this.idleTimer as any).unref?.();
+    }
+  }
+
+  /**
+   * Idempotent teardown: mark dead, clear timers, remove from the registry, fail
+   * any pending open/in-flight op, and destroy the provider. `inflightError` is
+   * the error a pending open or in-flight op is rejected with; `connectionLoss`
+   * marks the session as connection-lost so the ack guard cannot report a false
+   * success on a racing unsyncedChanges->0.
+   */
+  private teardown(inflightError: Error | null, connectionLoss: boolean): void {
+    if (this.dead) return;
+    this.dead = true;
+    this.state = "dead";
+    if (connectionLoss) this.connectionLost = true;
+
+    if (this.connectTimer) {
+      clearTimeout(this.connectTimer);
+      this.connectTimer = undefined;
+    }
+    if (this.idleTimer) {
+      clearTimeout(this.idleTimer);
+      this.idleTimer = undefined;
+    }
+
+    // Remove ourselves from the registry (only if we are still the live entry —
+    // a re-open under the same key must not be evicted by our teardown).
+    if (sessions.get(this.key) === this) {
+      sessions.delete(this.key);
+    }
+
+    // Fail a pending open() and any in-flight mutate with the terminal error.
+    if (!this.openSettled) {
+      this.openSettled = true;
+      this.openReject?.(
+        inflightError ?? new Error("Collaboration session destroyed"),
+      );
+    }
+    if (this.inflightReject) {
+      const rej = this.inflightReject;
+      this.inflightReject = undefined;
+      rej(inflightError ?? new Error("Collaboration session destroyed"));
+    }
+
+    if (this.provider) {
+      try {
+        this.provider.destroy();
+      } catch (e) {}
+      this.provider = undefined;
+    }
+  }
+
+  /**
+   * Public idempotent teardown used by the acquire/eviction paths and by a
+   * caller that wants the session dropped after a failed op ("next call
+   * reconnects fresh").
+   */
+  destroy(reason: string): void {
+    if (this.dead) return;
+    if (process.env.DEBUG)
+      console.error(`Destroying collab session ${this.pageId}: ${reason}`);
+    this.teardown(new Error(`Collaboration session destroyed: ${reason}`), false);
+  }
+}
+
+/** key = wsUrl + pageId + collabToken (identity isolation: invariant 4). */
+const sessions = new Map<string, CollabSession>();
+
+function sessionKey(wsUrl: string, pageId: string, token: string): string {
+  // The token is part of the key so sessions are NEVER shared between different
+  // users' MCP sessions (HTTP mode), and a token rotation makes a new entry
+  // while the old one idles out.
+  return `${wsUrl}${pageId}${token}`;
+}
+
+/**
+ * Get a live, synced CollabSession for a page, reusing a cached one when it is
+ * still valid or opening a fresh one otherwise. Does NOT take the per-page lock
+ * — the caller MUST already hold it (both call sites run inside withPageLock,
+ * which is not reentrant, so acquiring the lock here would deadlock
+ * mutateLiveContentUnlocked).
+ */
+export async function acquireCollabSession(
+  pageId: string,
+  collabToken: string,
+  baseUrl: string,
+  opts?: AcquireOptions,
+): Promise<CollabSession> {
+  const cfg = readConfig();
+  const wsUrl = buildCollabWsUrl(baseUrl);
+
+  // Cache disabled (rollback path): open an unregistered ephemeral session that
+  // self-destroys after its single op — the exact legacy per-op-provider flow.
+  if (cfg.idleMs <= 0) {
+    const session = new CollabSession(
+      sessionKey(wsUrl, pageId, collabToken),
+      pageId,
+      wsUrl,
+      collabToken,
+      cfg,
+      true,
+      opts,
+    );
+    await session.open();
+    return session;
+  }
+
+  const key = sessionKey(wsUrl, pageId, collabToken);
+  const existing = sessions.get(key);
+  if (existing) {
+    if (existing.isReusable()) {
+      // Reuse. Refresh LRU order (re-insert = most recently used) and re-arm the
+      // idle TTL so the reuse counts as activity.
+      sessions.delete(key);
+      sessions.set(key, existing);
+      existing.armIdle();
+      if (process.env.DEBUG)
+        console.error(`Reusing collab session for page ${pageId}`);
+      return existing;
+    }
+    // Stale (not synced / past max age / lost): drop it and open fresh.
+    existing.destroy("stale on reuse");
+  }
+
+  // Enforce the registry cap before inserting: destroy-evict the least recently
+  // used (the first entry in insertion order) until there is room.
+  while (sessions.size >= cfg.maxEntries) {
+    const oldestKey: string | undefined = sessions.keys().next().value;
+    if (oldestKey === undefined) break;
+    const victim = sessions.get(oldestKey);
+    if (victim) victim.destroy("evicted (LRU cap)");
+    // destroy() removes it from the map; guard against a no-op destroy.
+    if (sessions.has(oldestKey)) sessions.delete(oldestKey);
+  }
+
+  const session = new CollabSession(
+    key,
+    pageId,
+    wsUrl,
+    collabToken,
+    cfg,
+    false,
+    opts,
+  );
+  sessions.set(key, session);
+  try {
+    await session.open();
+  } catch (e) {
+    // Failed connect/sync: make sure it is not left cached.
+    session.destroy("open failed");
+    throw e;
+  }
+  session.armIdle();
+  if (process.env.DEBUG)
+    console.error(`Opened new collab session for page ${pageId}`);
+  return session;
+}
+
+/**
+ * Destroy every cached session. Wired into the process shutdown so a hanging
+ * session does not keep a doc loaded on the server past exit.
+ */
+export function destroyAllSessions(): void {
+  for (const session of [...sessions.values()]) {
+    session.destroy("process shutdown");
+  }
+  sessions.clear();
+}
+
+/** TEST-ONLY: number of currently cached sessions. */
+export function __sessionCountForTests(): number {
+  return sessions.size;
+}

packages/mcp/src/lib/collaboration.ts

@@ -1,4 +1,3 @@
-import { HocuspocusProvider } from "@hocuspocus/provider";
 import { TiptapTransformer } from "@hocuspocus/transformer";
 import * as Y from "yjs";
 import WebSocket from "ws";
@@ -16,7 +15,8 @@ import { docmostExtensions, docmostSchema } from "./docmost-schema.js";
 import { withPageLock } from "./page-lock.js";
 import { sanitizeForYjs, findUnstorableAttr } from "./node-ops.js";
 import { canonicalizeFootnotes } from "./footnote-canonicalize.js";
-import { summarizeChange, VerifyReport } from "./diff.js";
+import { VerifyReport } from "./diff.js";
+import { acquireCollabSession } from "./collab-session.js";
 
 export { markdownToProseMirror };
 
@@ -194,26 +194,27 @@ export function assertYjsEncodable(doc: any): void {
   }
 }
 
-/** Time we wait for the initial handshake/sync before giving up. */
-const CONNECT_TIMEOUT_MS = 25000;
-/** Time we wait for the server to acknowledge our write before giving up. */
-const PERSIST_TIMEOUT_MS = 20000;
-
 /**
  * Safely mutate the live content of a page over the collaboration websocket.
  *
  * This is the single safe write path for every MCP content mutation. It:
  *   1. serializes per-page writes through withPageLock (no two MCP writes on
  *      the same page overlap);
- *   2. connects to Hocuspocus and waits for the initial sync so the local ydoc
+ *   2. acquires a LIVE, synced CollabSession for the page (issue #400) — a
- *      mirrors the authoritative server doc — INCLUDING edits/comments/images
+ *      cached provider whose local ydoc mirrors the authoritative server doc
- *      that are not yet in the debounced REST snapshot;
+ *      (INCLUDING edits/comments/images not yet in the debounced REST snapshot),
- *   3. inside onSynced, SYNCHRONOUSLY reads the live doc, runs `transform`, and
+ *      reused across a series of edits instead of a fresh connect/auth/sync per
- *      writes the result back — with no `await` between read and write so no
+ *      call;
- *      remote update can interleave and clobber concurrent human edits;
+ *   3. SYNCHRONOUSLY reads the live doc, runs `transform`, and writes the result
+ *      back — with no `await` between read and write so no remote update can
+ *      interleave and clobber concurrent human edits (CollabSession.mutate);
  *   4. waits for the server to acknowledge the write (unsyncedChanges -> 0)
  *      before resolving, so the next operation observes our change.
  *
+ * On any mutate failure the session is destroyed so the next call reconnects
+ * fresh; the page lock is held for the whole acquire+mutate so the session's
+ * synchronous read->write window never overlaps another MCP write on the page.
+ *
  * `transform` receives the live ProseMirror doc and returns the NEW full
  * ProseMirror doc to write, or `null` to abort with no write (a no-op). If
  * `transform` throws, the error is propagated to the caller (not swallowed).
@@ -230,7 +231,7 @@ export async function mutatePageContent(
   baseUrl: string,
   transform: (liveDoc: any) => any | null,
 ): Promise<MutationResult> {
-  return withPageLock(pageId, () => {
+  return withPageLock(pageId, async () => {
     if (process.env.DEBUG) {
       console.error(`Starting realtime content mutate for page ${pageId}`);
       // Token prefix is sensitive; only log it under DEBUG.
@@ -239,202 +240,15 @@ export async function mutatePageContent(
       );
     }
 
-    const ydoc = new Y.Doc();
+    const session = await acquireCollabSession(pageId, collabToken, baseUrl);
-    const wsUrl = buildCollabWsUrl(baseUrl);
-    if (process.env.DEBUG) console.error(`Connecting to WebSocket: ${wsUrl}`);
-
-    return new Promise<MutationResult>((resolve, reject) => {
-      let provider: HocuspocusProvider | undefined;
-      let applied = false; // onSynced may fire again on reconnect — apply once.
-      let settled = false;
-      // Set true on disconnect/close so a reconnect-driven unsyncedChanges->0
-      // cannot be mistaken for a successful persist of our write.
-      let connectionLost = false;
-      let connectTimer: ReturnType<typeof setTimeout> | undefined;
-      let persistTimer: ReturnType<typeof setTimeout> | undefined;
-      let unsyncedHandler: ((data: { number: number }) => void) | undefined;
-
-      const cleanup = () => {
-        if (connectTimer) clearTimeout(connectTimer);
-        if (persistTimer) clearTimeout(persistTimer);
-        if (provider) {
-          if (unsyncedHandler) {
     try {
-              provider.off("unsyncedChanges", unsyncedHandler);
+      return await session.mutate(transform);
-            } catch (err) {}
-          }
-          try {
-            provider.destroy();
-          } catch (err) {}
-        }
-      };
-
-      const finish = (err: Error | null, value?: MutationResult) => {
-        if (settled) return;
-        settled = true;
-        cleanup();
-        if (err) reject(err);
-        else resolve(value as MutationResult);
-      };
-
-      connectTimer = setTimeout(() => {
-        finish(new Error("Connection timeout to collaboration server"));
-      }, CONNECT_TIMEOUT_MS);
-
-      // Resolve once the server has acknowledged our update. The provider
-      // increments unsyncedChanges when our local update is sent and
-      // decrements it when the server replies with a SyncStatus(applied=true);
-      // reaching 0 means the authoritative in-memory ydoc on the server now
-      // contains our write.
-      const waitForPersistence = () => {
-        if (settled) return;
-        // A missing provider is a failure, not a success: without it the write
-        // can never have been acknowledged. Only an actual unsyncedChanges===0
-        // on a live provider counts as persisted.
-        if (!provider) {
-          finish(new Error("collab provider gone before persistence"));
-          return;
-        }
-        if (provider.unsyncedChanges === 0) {
-          finish(null, mutationResult);
-          return;
-        }
-        persistTimer = setTimeout(() => {
-          finish(
-            new Error(
-              "Timeout waiting for collaboration server to persist the update",
-            ),
-          );
-        }, PERSIST_TIMEOUT_MS);
-        unsyncedHandler = (data: { number: number }) => {
-          // Only treat unsyncedChanges->0 as success when the connection is
-          // still up. A transient disconnect + reconnect handshake can drive
-          // the counter back to 0 without our write being re-transmitted; in
-          // that case let the disconnect/close error win instead.
-          if (data.number === 0 && !connectionLost) {
-            finish(null, mutationResult);
-          }
-        };
-        provider.on("unsyncedChanges", unsyncedHandler);
-      };
-
-      // The verifiable result resolved on every success/abort path. Set on
-      // abort (no-op report) and after a real write (computed change report).
-      let mutationResult: MutationResult;
-
-      provider = new HocuspocusProvider({
-        url: wsUrl,
-        name: `page.${pageId}`,
-        document: ydoc,
-        token: collabToken,
-        // @ts-ignore - Required for Node.js environment
-        WebSocketPolyfill: WebSocket,
-        onConnect: () => {
-          if (process.env.DEBUG) console.error("WS Connect");
-        },
-        // An unexpected disconnect/close while we are still waiting (during the
-        // connect-wait before onSynced, or during the persistence wait after the
-        // write) means the update will never be acknowledged — surface it now
-        // instead of hanging until the connect/persist timeout fires. `finish`
-        // is idempotent via the `settled` flag, so the onClose that our own
-        // cleanup()->provider.destroy() triggers (after settled=true is set) is
-        // a harmless no-op and cannot cause a double-resolve.
-        onDisconnect: () => {
-          if (process.env.DEBUG) console.error("WS Disconnect");
-          // Mark BEFORE finish so the unsyncedChanges handler (if it races)
-          // sees the connection as lost and won't report a false success.
-          connectionLost = true;
-          finish(
-            new Error(
-              "Collaboration connection closed before the update was persisted/synced",
-            ),
-          );
-        },
-        onClose: () => {
-          if (process.env.DEBUG) console.error("WS Close");
-          // Mark BEFORE finish so the unsyncedChanges handler (if it races)
-          // sees the connection as lost and won't report a false success.
-          connectionLost = true;
-          finish(
-            new Error(
-              "Collaboration connection closed before the update was persisted/synced",
-            ),
-          );
-        },
-        onSynced: () => {
-          if (applied || settled) return;
-          applied = true;
-          if (process.env.DEBUG) console.error("Connected and synced!");
-
-          // CRITICAL: everything between reading the live doc and writing it
-          // back must stay synchronous (no await). While the JS event loop is
-          // not yielded, no incoming remote update can interleave, so any
-          // already-synced concurrent edits are preserved in liveDoc.
-          let newDoc: any;
-          let beforeDoc: any;
-          try {
-            let liveDoc = TiptapTransformer.fromYdoc(ydoc, "default");
-            if (
-              !liveDoc ||
-              typeof liveDoc !== "object" ||
-              !Array.isArray(liveDoc.content)
-            ) {
-              liveDoc = { type: "doc", content: [] };
-            }
-
-            // Snapshot the before-doc for the change report. Docs are
-            // JSON-serializable, so this is a safe deep clone.
-            beforeDoc = JSON.parse(JSON.stringify(liveDoc));
-
-            newDoc = transform(liveDoc);
-
-            if (newDoc == null) {
-              // Transform aborted — write nothing, return the live doc with a
-              // no-op change report.
-              mutationResult = {
-                doc: liveDoc,
-                verify: {
-                  changed: false,
-                  textInserted: 0,
-                  textDeleted: 0,
-                  blocksChanged: 0,
-                  marks: {},
-                  summary: "no changes (transform aborted)",
-                },
-              };
-              finish(null, mutationResult);
-              return;
-            }
-
-            // Structural diff into the live fragment (issue #152): preserves
-            // the Yjs ids of unchanged nodes, so an open editor's cursor is not
-            // yanked to the end of the document on every agent write.
-            applyDocToFragment(ydoc, newDoc);
     } catch (e) {
-            // Includes errors thrown by transform (e.g. "afterText not found",
+      // Drop the session on any failure so the next call reconnects fresh (this
-            // "text not found"): propagate them verbatim to the caller.
+      // also closes the "reconnect drove the counter to 0" false-success class).
-            finish(e instanceof Error ? e : new Error(String(e)));
+      session.destroy("mutate failed");
-            return;
+      throw e;
     }
-
-          // Compute the verifiable change report AFTER the transact write: it
-          // only needs the JSON before/after, so it cannot affect the atomic
-          // read->write window, and summarizeChange never throws.
-          mutationResult = {
-            doc: newDoc,
-            verify: summarizeChange(beforeDoc, newDoc),
-          };
-          if (process.env.DEBUG)
-            console.error("Content written, waiting for server to persist...");
-          waitForPersistence();
-        },
-        onAuthenticationFailed: () => {
-          finish(
-            new Error("Authentication failed for collaboration connection"),
-          );
-        },
-      });
-    });
   });
 }
 

packages/mcp/src/lib/comment-anchor.ts

@@ -17,8 +17,23 @@
  * comparing and match across maximal runs of consecutive text nodes within a
  * single block, while mapping every normalized character back to its raw index
  * so the mark lands on the exact original characters.
+ *
+ * MARKDOWN-STRIP FALLBACK: when the agent copies a selection that still carries
+ * inline markdown (`**bold**`, `` `code` ``, `[t](u)`), the raw locator will not
+ * match the document's plain text. Exactly like edit_page_text's json-edit
+ * fallback, we first try the verbatim selection and, ONLY if it anchors nowhere
+ * in the whole document, retry with `stripInlineMarkdown` applied. `canAnchorInDoc`,
+ * `getAnchoredText` and `applyAnchorInDoc` share this decision via
+ * `resolveAnchorSelection`. `countAnchorMatches` keeps its OWN parallel exact-wins
+ * implementation (it needs a raw match COUNT, not a single resolved locator), kept
+ * deliberately in sync with `resolveAnchorSelection`: raw match ⇒ use raw, else fall
+ * back to the stripped count. All four therefore agree on which locator matched —
+ * the suggestion-uniqueness gate depends on count and can/get never disagreeing, so
+ * these two exact-wins implementations MUST stay in sync if either is changed.
  */
 
+import { stripInlineMarkdown } from "./text-normalize.js";
+
 /** Typographic double-quote variants mapped to ASCII `"`. */
 const DOUBLE_QUOTES = "«»„“”‟〝〞＂";
 /** Typographic single-quote/apostrophe variants mapped to ASCII `'`. */
@@ -214,15 +229,17 @@ function reconstructRawText(blockContent: any[], match: AnchorMatch): string {
  * un-appliable (spurious 409).
  */
 export function getAnchoredText(doc: any, selection: string): string | null {
+  const { selection: effective, found } = resolveAnchorSelection(doc, selection);
+  if (!found) return null;
   const visit = (node: any, depth: number): string | null => {
     if (depth > MAX_DEPTH || !node || typeof node !== "object") return null;
     if (!Array.isArray(node.content)) return null;
-    const match = findAnchorInBlock(node.content, selection);
+    const match = findAnchorInBlock(node.content, effective);
     if (match) return reconstructRawText(node.content, match);
     for (const child of node.content) {
       if (child && typeof child === "object" && Array.isArray(child.content)) {
-        const found = visit(child, depth + 1);
+        const foundText = visit(child, depth + 1);
-        if (found !== null) return found;
+        if (foundText !== null) return foundText;
       }
     }
     return null;
@@ -231,12 +248,11 @@ export function getAnchoredText(doc: any, selection: string): string | null {
 }
 
 /**
- * Depth-first, document-order check for whether `selection` can be anchored
+ * RAW (no markdown-strip fallback) depth-first check that `selection` anchors
- * anywhere in `doc`. At each node with an array `content`, first try to match
+ * somewhere in `doc`. This is the primitive `resolveAnchorSelection` builds on;
- * within that node's own content, then recurse into children that themselves
+ * public callers should use `canAnchorInDoc`, which adds the strip fallback.
- * have a `content` array.
  */
-export function canAnchorInDoc(doc: any, selection: string): boolean {
+function rawCanAnchorInDoc(doc: any, selection: string): boolean {
   const visit = (node: any, depth: number): boolean => {
     if (depth > MAX_DEPTH || !node || typeof node !== "object") return false;
     if (!Array.isArray(node.content)) return false;
@@ -251,6 +267,43 @@ export function canAnchorInDoc(doc: any, selection: string): boolean {
   return visit(doc, 0);
 }
 
+/**
+ * Decide the locator that ACTUALLY anchors `selection` in `doc`, applying the
+ * markdown-strip fallback once (so every public entry point agrees):
+ *  - EXACT WINS: if the verbatim selection anchors anywhere, use it as-is.
+ *  - FALLBACK: only if the verbatim selection anchors nowhere, and the
+ *    markdown-stripped form differs and DOES anchor, use the stripped form and
+ *    flag `normalized` so callers can surface a soft warning.
+ *  - otherwise `found` is false and `selection` is returned unchanged.
+ *
+ * The stripped form is used ONLY to LOCATE the anchor; getAnchoredText still
+ * reconstructs and stores the RAW document substring, so the strip never leaks
+ * into what gets persisted.
+ */
+export function resolveAnchorSelection(
+  doc: any,
+  selection: string,
+): { selection: string; found: boolean; normalized: boolean } {
+  if (rawCanAnchorInDoc(doc, selection)) {
+    return { selection, found: true, normalized: false };
+  }
+  const stripped = stripInlineMarkdown(selection);
+  if (stripped !== selection && rawCanAnchorInDoc(doc, stripped)) {
+    return { selection: stripped, found: true, normalized: true };
+  }
+  return { selection, found: false, normalized: false };
+}
+
+/**
+ * Depth-first, document-order check for whether `selection` can be anchored
+ * anywhere in `doc` (with the markdown-strip fallback). At each node with an
+ * array `content`, first try to match within that node's own content, then
+ * recurse into children that themselves have a `content` array.
+ */
+export function canAnchorInDoc(doc: any, selection: string): boolean {
+  return resolveAnchorSelection(doc, selection).found;
+}
+
 /**
  * Split the matched text nodes and splice the comment mark across the range.
  * `blockContent` is mutated IN PLACE. `match.startChild..endChild` are all text
@@ -315,7 +368,7 @@ function spliceCommentMark(
  * not use this. (Note: counts OCCURRENCES, not just matching blocks, so two
  * occurrences inside one block are correctly reported as 2.)
  */
-export function countAnchorMatches(doc: any, selection: string): number {
+function rawCountAnchorMatches(doc: any, selection: string): number {
   const normSel = normalizeForMatch(selection).norm.trim();
   if (normSel.length === 0) return 0;
 
@@ -369,6 +422,25 @@ export function countAnchorMatches(doc: any, selection: string): number {
   return total;
 }
 
+/**
+ * Uniqueness gate for suggestions, with the SAME markdown-strip fallback as the
+ * other entry points so count never disagrees with can/get/apply. EXACT WINS: if
+ * the verbatim selection occurs at all, return its raw occurrence count (so a
+ * selection that is unique raw stays unique — the fallback never runs and cannot
+ * introduce a spurious second match). Only when the verbatim selection is absent
+ * do we count occurrences of the markdown-stripped form.
+ */
+export function countAnchorMatches(doc: any, selection: string): number {
+  const raw = rawCountAnchorMatches(doc, selection);
+  if (raw > 0) return raw;
+  const stripped = stripInlineMarkdown(selection);
+  if (stripped !== selection) {
+    const strippedCount = rawCountAnchorMatches(doc, stripped);
+    if (strippedCount > 0) return strippedCount;
+  }
+  return 0;
+}
+
 /**
  * Depth-first (same order as canAnchorInDoc) over `doc`; on the FIRST block
  * whose content matches `selection`, splice the comment mark across the matched
@@ -380,10 +452,12 @@ export function applyAnchorInDoc(
   selection: string,
   commentId: string,
 ): boolean {
+  const { selection: effective, found } = resolveAnchorSelection(doc, selection);
+  if (!found) return false;
   const visit = (node: any, depth: number): boolean => {
     if (depth > MAX_DEPTH || !node || typeof node !== "object") return false;
     if (!Array.isArray(node.content)) return false;
-    const match = findAnchorInBlock(node.content, selection);
+    const match = findAnchorInBlock(node.content, effective);
     if (match) {
       spliceCommentMark(node.content, match, commentId);
       return true;

packages/mcp/src/lib/drawio-preview.ts

@@ -0,0 +1,193 @@
+// Pure-TS schematic SVG preview for draw.io diagrams (issue #423, stage 1).
+//
+// HARD CONSTRAINT: no backend rendering. This is a dependency-free string
+// builder — given the parsed mxGraph cells it draws a rough schematic (rects,
+// ellipses, diamonds, edges + labels) that stands in as the diagram's visible
+// image UNTIL a human first opens it in the draw.io editor and saves, at which
+// point the client replaces this with the pixel-perfect export SVG. It is
+// deliberately approximate: it exists so a freshly-agent-created diagram is not
+// an empty box in the page.
+
+import type { DrawioCell, DrawioBBox } from "./drawio-xml.js";
+import { absolutePos } from "./drawio-xml.js";
+
+function esc(s: string): string {
+  return s
+    .replace(/&/g, "&")
+    .replace(/</g, "&lt;")
+    .replace(/>/g, "&gt;")
+    .replace(/"/g, "&quot;");
+}
+
+/**
+ * Strip HTML markup from a cell value (draw.io labels are HTML when html=1),
+ * decode the handful of entities we care about, and collapse whitespace so the
+ * label fits on the schematic. `<br>` becomes a space (this is a one-line
+ * preview label, not a faithful multi-line render).
+ */
+function labelText(value: string): string {
+  return value
+    .replace(/<br\s*\/?>/gi, " ")
+    .replace(/<[^>]+>/g, "")
+    .replace(/&lt;/g, "<")
+    .replace(/&gt;/g, ">")
+    .replace(/&quot;/g, '"')
+    .replace(/&#39;|&apos;/g, "'")
+    .replace(/&#xa;|&#10;/gi, " ")
+    .replace(/&amp;/g, "&")
+    .replace(/\s+/g, " ")
+    .trim();
+}
+
+function centeredLabel(cx: number, cy: number, value: string, color = "#000000"): string {
+  const text = labelText(value);
+  if (!text) return "";
+  return (
+    `<text x="${round(cx)}" y="${round(cy)}" ` +
+    `font-family="Helvetica, Arial, sans-serif" font-size="12" ` +
+    `text-anchor="middle" dominant-baseline="middle" fill="${esc(color)}">` +
+    `${esc(text)}</text>`
+  );
+}
+
+function round(n: number): number {
+  return Math.round(n * 100) / 100;
+}
+
+interface ShapeKind {
+  kind: "ellipse" | "rhombus" | "triangle" | "rect";
+}
+
+/**
+ * Decide which schematic primitive to draw for a vertex. A shape can be named
+ * either as the style's base token (e.g. "ellipse;…") or as a key (e.g.
+ * "shape=rhombus" / "ellipse=1"), so both the base style and the map are
+ * checked.
+ */
+function shapeKind(
+  styleMap: Record<string, string>,
+  baseStyle?: string,
+): ShapeKind {
+  const shape = styleMap.shape ?? baseStyle;
+  const has = (name: string) => shape === name || styleMap[name] != null;
+  if (has("ellipse")) return { kind: "ellipse" };
+  if (has("rhombus")) return { kind: "rhombus" };
+  if (has("triangle")) return { kind: "triangle" };
+  // Everything else — including unknown stencils (shape=mxgraph.*), swimlanes,
+  // and plain boxes — is drawn as a (rounded) rectangle.
+  return { kind: "rect" };
+}
+
+function fill(styleMap: Record<string, string>): string {
+  const c = styleMap.fillColor;
+  if (!c || c.toLowerCase() === "none") return "#ffffff";
+  return c;
+}
+
+function stroke(styleMap: Record<string, string>): string {
+  const c = styleMap.strokeColor;
+  if (!c || c.toLowerCase() === "none") return "#000000";
+  return c;
+}
+
+/**
+ * Render the schematic shapes as the INNER content of the `.drawio.svg` (the
+ * outer <svg> wrapper is added by drawio-xml.buildDrawioSvg). Coordinates are
+ * absolute (container children are resolved via the parent chain).
+ */
+export function renderDiagramShapes(cells: DrawioCell[], _bbox: DrawioBBox): string {
+  const byId = new Map(cells.map((c) => [c.id, c]));
+  const parts: string[] = [];
+
+  // Edges first so vertices sit on top of their connectors.
+  for (const c of cells) {
+    if (!c.edge) continue;
+    parts.push(renderEdge(c, byId));
+  }
+
+  for (const c of cells) {
+    if (!c.vertex || !c.geometry.hasGeometry) continue;
+    const g = c.geometry;
+    if (g.width == null || g.height == null) continue;
+    const { x, y } = absolutePos(c, byId);
+    parts.push(renderVertex(c, x, y, g.width, g.height));
+  }
+
+  return `<g>${parts.filter(Boolean).join("")}</g>`;
+}
+
+function renderVertex(
+  c: DrawioCell,
+  x: number,
+  y: number,
+  w: number,
+  h: number,
+): string {
+  const f = esc(fill(c.styleMap));
+  const s = esc(stroke(c.styleMap));
+  const { kind } = shapeKind(c.styleMap, c.baseStyle);
+  const cx = x + w / 2;
+  const cy = y + h / 2;
+  let shape = "";
+  switch (kind) {
+    case "ellipse":
+      shape =
+        `<ellipse cx="${round(cx)}" cy="${round(cy)}" rx="${round(w / 2)}" ` +
+        `ry="${round(h / 2)}" fill="${f}" stroke="${s}"/>`;
+      break;
+    case "rhombus": {
+      const pts = [
+        `${round(cx)},${round(y)}`,
+        `${round(x + w)},${round(cy)}`,
+        `${round(cx)},${round(y + h)}`,
+        `${round(x)},${round(cy)}`,
+      ].join(" ");
+      shape = `<polygon points="${pts}" fill="${f}" stroke="${s}"/>`;
+      break;
+    }
+    case "triangle": {
+      const pts = [
+        `${round(x)},${round(y)}`,
+        `${round(x + w)},${round(cy)}`,
+        `${round(x)},${round(y + h)}`,
+      ].join(" ");
+      shape = `<polygon points="${pts}" fill="${f}" stroke="${s}"/>`;
+      break;
+    }
+    default: {
+      const rounded = c.styleMap.rounded === "1";
+      const rx = rounded ? Math.min(12, w / 2, h / 2) : 0;
+      shape =
+        `<rect x="${round(x)}" y="${round(y)}" width="${round(w)}" ` +
+        `height="${round(h)}" rx="${round(rx)}" ry="${round(rx)}" ` +
+        `fill="${f}" stroke="${s}"/>`;
+    }
+  }
+  return shape + centeredLabel(cx, cy, c.value, c.styleMap.fontColor || "#000000");
+}
+
+function renderEdge(c: DrawioCell, byId: Map<string, DrawioCell>): string {
+  const src = c.source != null ? byId.get(c.source) : undefined;
+  const tgt = c.target != null ? byId.get(c.target) : undefined;
+  const p1 = anchorPoint(src, byId);
+  const p2 = anchorPoint(tgt, byId);
+  if (!p1 || !p2) return ""; // a floating endpoint with no fixed point: skip
+  const line =
+    `<line x1="${round(p1.x)}" y1="${round(p1.y)}" ` +
+    `x2="${round(p2.x)}" y2="${round(p2.y)}" ` +
+    `stroke="#000000" stroke-width="1"/>`;
+  const mid = { x: (p1.x + p2.x) / 2, y: (p1.y + p2.y) / 2 };
+  return line + centeredLabel(mid.x, mid.y, c.value);
+}
+
+/** Center point of a vertex used as an edge anchor (approximate). */
+function anchorPoint(
+  cell: DrawioCell | undefined,
+  byId: Map<string, DrawioCell>,
+): { x: number; y: number } | null {
+  if (!cell || !cell.geometry.hasGeometry) return null;
+  const g = cell.geometry;
+  if (g.width == null || g.height == null) return null;
+  const { x, y } = absolutePos(cell, byId);
+  return { x: x + g.width / 2, y: y + g.height / 2 };
+}

packages/mcp/src/lib/drawio-xml.ts

@@ -0,0 +1,771 @@
+// draw.io (mxGraph) XML support for the MCP drawio tools (issue #423, stage 1).
+//
+// This module owns everything that is pure data-plumbing for draw.io diagrams:
+//   - the DECODE CHAIN that turns a stored `diagram.drawio.svg` attachment back
+//     into mxGraph XML (handles both the plain nested-XML form Docmost writes
+//     and draw.io's own COMPRESSED `<diagram>` payload — base64 + raw-deflate);
+//   - the ENCODE side that wraps mxGraph XML into the `.drawio.svg` attachment
+//     using the exact same contract as the import service's createDrawioSvg;
+//   - a deterministic LINTER that rejects the structural mistakes generators
+//     make before anything is written (each violation carries the offending
+//     cellId + position so the model can auto-retry);
+//   - a stable HASH over the normalized XML, used as the optimistic-lock key.
+//
+// HARD CONSTRAINT: no backend rendering. Nothing here shells out or renders a
+// bitmap; the only runtime dependencies are jsdom (already used across this
+// package for XML parsing) and pako (raw-inflate for the compressed format).
+
+import { createHash } from "node:crypto";
+import { JSDOM } from "jsdom";
+import pako from "pako";
+
+// --- shared XML parser -----------------------------------------------------
+
+// A single reusable JSDOM window; constructing one per parse is wasteful and
+// these tools are low-frequency. Only the DOMParser is used.
+let _window: any = null;
+function xmlWindow(): any {
+  if (!_window) _window = new JSDOM("").window;
+  return _window;
+}
+
+/** Default mxGraphModel attributes used when the server wraps a cell list. */
+const DEFAULT_MODEL_ATTRS =
+  'dx="0" dy="0" grid="1" gridSize="10" page="1" pageWidth="850" pageHeight="1100"';
+
+// --- structured lint errors ------------------------------------------------
+
+export interface DrawioLintIssue {
+  /** Machine-readable rule id, e.g. "edge-geometry". */
+  rule: string;
+  /** Human-readable explanation the model can act on. */
+  message: string;
+  /** The offending cell's id, when the rule is cell-scoped. */
+  cellId?: string;
+  /** Extra location info: cell index in <root>, or a parser line:col. */
+  position?: string;
+}
+
+/**
+ * Thrown by the linter and by decode/prepare when the input is unusable. Carries
+ * the full list of issues so the caller can surface a structured tool-error the
+ * model auto-retries against.
+ */
+export class DrawioLintError extends Error {
+  issues: DrawioLintIssue[];
+  constructor(issues: DrawioLintIssue[]) {
+    const summary = issues
+      .map((i) => {
+        const where = [
+          i.cellId != null ? `cellId=${i.cellId}` : null,
+          i.position != null ? `at ${i.position}` : null,
+        ]
+          .filter(Boolean)
+          .join(", ");
+        return `[${i.rule}] ${i.message}${where ? ` (${where})` : ""}`;
+      })
+      .join("; ");
+    super(`drawio lint failed: ${summary}`);
+    this.name = "DrawioLintError";
+    this.issues = issues;
+  }
+}
+
+// --- parsed-cell model -----------------------------------------------------
+
+export interface DrawioGeometry {
+  x?: number;
+  y?: number;
+  width?: number;
+  height?: number;
+  relative: boolean;
+  hasGeometry: boolean;
+}
+
+export interface DrawioCell {
+  id: string;
+  parent?: string;
+  source?: string;
+  target?: string;
+  vertex: boolean;
+  edge: boolean;
+  value: string;
+  style: string;
+  styleMap: Record<string, string>;
+  /** Non-key/value leading token of the style (a base stylename), if any. */
+  baseStyle?: string;
+  geometry: DrawioGeometry;
+}
+
+export interface DrawioBBox {
+  width: number;
+  height: number;
+}
+
+// --- style parsing ---------------------------------------------------------
+
+/**
+ * Parse a draw.io style string into { baseStyle, map }. Grammar:
+ *   [stylename;]key=value;key=value;...
+ * A single leading token without '=' is the base stylename (e.g. "text" or
+ * "ellipse"). Every other non-empty segment must be exactly one key=value pair.
+ * Returns `null` (the segment index) on the first malformed segment so the
+ * linter can report a precise error.
+ */
+export function parseStyle(
+  style: string,
+): { baseStyle?: string; map: Record<string, string>; badSegment?: string } {
+  const map: Record<string, string> = {};
+  let baseStyle: string | undefined;
+  const segments = style.split(";");
+  for (let i = 0; i < segments.length; i++) {
+    const seg = segments[i].trim();
+    if (seg === "") continue; // trailing/empty segments are fine
+    const eq = seg.indexOf("=");
+    if (eq === -1) {
+      // A bare token is only valid as the FIRST meaningful segment (base style).
+      if (baseStyle === undefined && Object.keys(map).length === 0) {
+        baseStyle = seg;
+        continue;
+      }
+      return { baseStyle, map, badSegment: seg };
+    }
+    // A second '=' inside the same segment is malformed.
+    if (seg.indexOf("=", eq + 1) !== -1) {
+      return { baseStyle, map, badSegment: seg };
+    }
+    const key = seg.slice(0, eq).trim();
+    const val = seg.slice(eq + 1).trim();
+    if (key === "") return { baseStyle, map, badSegment: seg };
+    map[key] = val;
+  }
+  return { baseStyle, map };
+}
+
+// --- low-level XML helpers -------------------------------------------------
+
+function parseXml(xml: string): { doc: any; error: string | null } {
+  const parser = new (xmlWindow().DOMParser)();
+  const doc = parser.parseFromString(xml, "application/xml");
+  const err = doc.getElementsByTagName("parsererror");
+  if (err.length > 0) {
+    // jsdom prefixes the message with "line:col:" — keep it as the position.
+    return { doc, error: (err[0].textContent || "malformed XML").trim() };
+  }
+  return { doc, error: null };
+}
+
+function num(v: string | null): number | undefined {
+  if (v == null || v === "") return undefined;
+  const n = Number(v);
+  return Number.isFinite(n) ? n : undefined;
+}
+
+/** Extract the raw `<mxGraphModel …>…</mxGraphModel>` substring, or null. */
+function sliceModel(xml: string): string | null {
+  const open = xml.indexOf("<mxGraphModel");
+  if (open === -1) return null;
+  const close = xml.indexOf("</mxGraphModel>", open);
+  if (close === -1) {
+    // Self-closed empty model, e.g. `<mxGraphModel .../>`.
+    const selfClose = xml.indexOf("/>", open);
+    if (selfClose !== -1) return xml.slice(open, selfClose + 2);
+    return null;
+  }
+  return xml.slice(open, close + "</mxGraphModel>".length);
+}
+
+// --- decode chain ----------------------------------------------------------
+
+/**
+ * Read the `content=` attribute out of a `.drawio.svg` string. Docmost stores a
+ * base64 payload there (createDrawioSvg); draw.io's own SVG export may store the
+ * XML entity-encoded instead. The DOM decodes entities for us, so the caller
+ * only has to distinguish "starts with '<'" (raw XML) from base64.
+ */
+export function extractContentAttr(svg: string): string {
+  const { doc, error } = parseXml(svg);
+  if (!error) {
+    const root = doc.documentElement;
+    if (root && root.hasAttribute && root.hasAttribute("content")) {
+      return root.getAttribute("content") || "";
+    }
+  }
+  // Fallback for a malformed wrapper: pull the attribute directly. The content
+  // value itself never contains a double-quote (base64 / entity-encoded XML).
+  const m = /content="([^"]*)"/.exec(svg);
+  if (m) {
+    // Decode the handful of XML entities a raw regex would leave encoded.
+    return m[1]
+      .replace(/&lt;/g, "<")
+      .replace(/&gt;/g, ">")
+      .replace(/&quot;/g, '"')
+      .replace(/&#39;/g, "'")
+      .replace(/&amp;/g, "&");
+  }
+  throw new Error("drawio: SVG has no content= attribute to decode");
+}
+
+/**
+ * Turn a decoded draw.io file (`<mxfile>` or a bare `<mxGraphModel>`, possibly
+ * with a COMPRESSED `<diagram>` payload) into the mxGraphModel XML. For the
+ * plain form the raw substring is returned verbatim so a round-trip stays
+ * byte-stable; the compressed form is inflated (base64 → raw-deflate →
+ * decodeURIComponent), which is how draw.io stores diagrams by default.
+ */
+export function decodeDrawioFileToModel(fileXml: string): string {
+  // Plain, nested XML: return the model substring untouched (byte-stable).
+  const sliced = sliceModel(fileXml);
+  if (sliced) return sliced;
+
+  // Otherwise it must be the compressed `<diagram>…</diagram>` text payload.
+  const open = fileXml.indexOf("<diagram");
+  if (open !== -1) {
+    const gt = fileXml.indexOf(">", open);
+    const close = fileXml.indexOf("</diagram>", gt);
+    if (gt !== -1 && close !== -1) {
+      const payload = fileXml.slice(gt + 1, close).trim();
+      if (payload) {
+        const inflated = inflateDiagramPayload(payload);
+        const model = sliceModel(inflated);
+        if (model) return model;
+        return inflated;
+      }
+    }
+  }
+  throw new Error(
+    "drawio: could not decode file — no <mxGraphModel> and no compressed <diagram> payload",
+  );
+}
+
+/**
+ * Upper bound on the inflated size of a compressed `<diagram>` payload
+ * (decompression-bomb guard). `fetchInternalFile` caps the DOWNLOAD at 64 MiB,
+ * but a tiny crafted compressed payload can inflate to gigabytes and OOM the
+ * process. A real diagram's mxGraphModel XML is small (KBs to low MBs even for
+ * large diagrams), so 16 MiB is far above any legitimate payload while keeping
+ * memory bounded. Chars ~= bytes for the (mostly ASCII) URI-encoded XML.
+ */
+export const MAX_INFLATED_DIAGRAM_BYTES = 16 * 1024 * 1024;
+
+/**
+ * Inflate draw.io's compressed diagram payload:
+ *   base64-decode → raw-inflate (raw deflate, windowBits -15) →
+ *   decodeURIComponent.
+ *
+ * Uses pako's streaming Inflate so we can abort as soon as the decompressed
+ * output exceeds MAX_INFLATED_DIAGRAM_BYTES — the full bomb is never
+ * materialised in memory.
+ */
+export function inflateDiagramPayload(base64: string): string {
+  const bytes = Buffer.from(base64, "base64");
+  const inflator = new pako.Inflate({ raw: true, to: "string" });
+  let total = 0;
+  const passthrough = inflator.onData.bind(inflator);
+  inflator.onData = (chunk: string | Uint8Array) => {
+    total += chunk.length;
+    if (total > MAX_INFLATED_DIAGRAM_BYTES) {
+      // Throwing here propagates out of push(), aborting inflation immediately.
+      throw new Error(
+        `drawio: refusing to decode diagram — decompressed size exceeds ` +
+          `${MAX_INFLATED_DIAGRAM_BYTES} bytes (possible decompression bomb)`,
+      );
+    }
+    passthrough(chunk);
+  };
+  inflator.push(bytes, true);
+  if (inflator.err) {
+    throw new Error(
+      `drawio: failed to inflate compressed <diagram> payload (${inflator.msg || inflator.err})`,
+    );
+  }
+  const uriEncoded = inflator.result as string;
+  return decodeURIComponent(uriEncoded);
+}
+
+/** Full decode chain: `.drawio.svg` string → mxGraphModel XML. */
+export function decodeDrawioSvg(svg: string): string {
+  const content = extractContentAttr(svg).trim();
+  const fileXml = content.startsWith("<")
+    ? content
+    : Buffer.from(content, "base64").toString("utf-8");
+  return decodeDrawioFileToModel(fileXml);
+}
+
+// --- encode side -----------------------------------------------------------
+
+/**
+ * Wrap an mxGraphModel in the plain (uncompressed) `<mxfile><diagram>` envelope.
+ * draw.io opens uncompressed XML fine, and staying uncompressed keeps the
+ * write path deterministic and the round-trip byte-stable.
+ */
+export function encodeDrawioFile(modelXml: string, title = "Page-1"): string {
+  const safeTitle = xmlEscape(title);
+  return `<mxfile host="drawio"><diagram id="page-1" name="${safeTitle}">${modelXml}</diagram></mxfile>`;
+}
+
+/**
+ * Build the `diagram.drawio.svg` attachment. Mirrors the import service's
+ * createDrawioSvg contract exactly:
+ *   <svg xmlns=… xmlns:xlink=… content="${base64(drawioFile)}">${inner}</svg>
+ * plus width/height/viewBox from the diagram bounding box and the schematic
+ * preview as the visible children (`inner`).
+ */
+export function buildDrawioSvg(
+  modelXml: string,
+  inner: string,
+  bbox: DrawioBBox,
+  title = "Page-1",
+): string {
+  const file = encodeDrawioFile(modelXml, title);
+  const base64 = Buffer.from(file, "utf-8").toString("base64");
+  const w = Math.max(1, Math.round(bbox.width));
+  const h = Math.max(1, Math.round(bbox.height));
+  return (
+    `<svg xmlns="http://www.w3.org/2000/svg" ` +
+    `xmlns:xlink="http://www.w3.org/1999/xlink" ` +
+    `width="${w}" height="${h}" viewBox="0 0 ${w} ${h}" ` +
+    `content="${base64}">${inner}</svg>`
+  );
+}
+
+function xmlEscape(s: string): string {
+  return s
+    .replace(/&/g, "&amp;")
+    .replace(/</g, "&lt;")
+    .replace(/>/g, "&gt;")
+    .replace(/"/g, "&quot;");
+}
+
+// --- normalization + hash --------------------------------------------------
+
+/**
+ * Normalize mxGraph XML for hashing / stable comparison: drop the whitespace
+ * between tags and trim. This is intentionally conservative — it never reorders
+ * attributes or cells (that would be lossy) — so two documents hash equal iff
+ * they differ only in inter-tag formatting.
+ */
+export function normalizeXml(xml: string): string {
+  return xml.replace(/>\s+</g, "><").trim();
+}
+
+/** Stable optimistic-lock hash over the normalized model XML (sha256, hex). */
+export function mxHash(modelXml: string): string {
+  return createHash("sha256").update(normalizeXml(modelXml), "utf-8").digest("hex");
+}
+
+// --- cell parsing ----------------------------------------------------------
+
+/** Parse every `<mxCell>` in a model into a structured DrawioCell list. */
+export function parseCells(modelXml: string): DrawioCell[] {
+  const { doc, error } = parseXml(modelXml);
+  if (error) {
+    throw new DrawioLintError([
+      { rule: "well-formed-xml", message: error, position: firstLineCol(error) },
+    ]);
+  }
+  const cells: DrawioCell[] = [];
+  const els = doc.getElementsByTagName("mxCell");
+  for (let i = 0; i < els.length; i++) {
+    cells.push(readCell(els[i]));
+  }
+  return cells;
+}
+
+function readCell(el: any): DrawioCell {
+  const style = el.getAttribute("style") || "";
+  const parsed = parseStyle(style);
+  const geoEl = firstChildByTag(el, "mxGeometry");
+  const geometry: DrawioGeometry = geoEl
+    ? {
+        x: num(geoEl.getAttribute("x")),
+        y: num(geoEl.getAttribute("y")),
+        width: num(geoEl.getAttribute("width")),
+        height: num(geoEl.getAttribute("height")),
+        relative: geoEl.getAttribute("relative") === "1",
+        hasGeometry: true,
+      }
+    : { relative: false, hasGeometry: false };
+  return {
+    id: el.getAttribute("id") ?? "",
+    parent: el.getAttribute("parent") ?? undefined,
+    source: el.getAttribute("source") ?? undefined,
+    target: el.getAttribute("target") ?? undefined,
+    vertex: el.getAttribute("vertex") === "1",
+    edge: el.getAttribute("edge") === "1",
+    value: el.getAttribute("value") ?? "",
+    style,
+    styleMap: parsed.map,
+    baseStyle: parsed.baseStyle,
+    geometry,
+  };
+}
+
+function firstChildByTag(el: any, tag: string): any {
+  for (let i = 0; i < el.childNodes.length; i++) {
+    const c = el.childNodes[i];
+    if (c.nodeType === 1 && c.tagName === tag) return c;
+  }
+  return null;
+}
+
+function firstLineCol(msg: string): string | undefined {
+  const m = /^(\d+:\d+)/.exec(msg);
+  return m ? m[1] : undefined;
+}
+
+// --- bounding box ----------------------------------------------------------
+
+/**
+ * Absolute bounding box of the diagram from its vertex geometries. Container
+ * children are relative, so absolute positions are resolved along the parent
+ * chain before taking the extent. Falls back to a default canvas when empty.
+ */
+export function computeBBox(cells: DrawioCell[]): DrawioBBox {
+  const byId = new Map(cells.map((c) => [c.id, c]));
+  let maxX = 0;
+  let maxY = 0;
+  let any = false;
+  for (const c of cells) {
+    if (!c.vertex || !c.geometry.hasGeometry) continue;
+    const g = c.geometry;
+    if (g.width == null || g.height == null) continue;
+    const { x, y } = absolutePos(c, byId);
+    maxX = Math.max(maxX, x + g.width);
+    maxY = Math.max(maxY, y + g.height);
+    any = true;
+  }
+  if (!any) return { width: 300, height: 200 };
+  // A small margin so borders/labels are not clipped at the edge.
+  return { width: Math.ceil(maxX) + 20, height: Math.ceil(maxY) + 20 };
+}
+
+/** Absolute (x,y) of a vertex, following its parent chain (containers). */
+export function absolutePos(
+  cell: DrawioCell,
+  byId: Map<string, DrawioCell>,
+): { x: number; y: number } {
+  let x = cell.geometry.x ?? 0;
+  let y = cell.geometry.y ?? 0;
+  const seen = new Set<string>([cell.id]);
+  let parentId = cell.parent;
+  while (parentId && !seen.has(parentId)) {
+    seen.add(parentId);
+    const p = byId.get(parentId);
+    // Sentinels (0/1) carry no geometry; stop there.
+    if (!p || !p.vertex || !p.geometry.hasGeometry) break;
+    x += p.geometry.x ?? 0;
+    y += p.geometry.y ?? 0;
+    parentId = p.parent;
+  }
+  return { x, y };
+}
+
+// --- linter ----------------------------------------------------------------
+
+/**
+ * Run every deterministic pre-write rule over a full mxGraphModel string. On any
+ * violation it throws a DrawioLintError carrying one issue per violation, each
+ * with the offending cellId + position. Returns the parsed cells on success.
+ */
+export function lintModel(modelXml: string): {
+  cells: DrawioCell[];
+  warnings: string[];
+} {
+  const issues: DrawioLintIssue[] = [];
+  const warnings: string[] = [];
+
+  // Rule: no XML comments. Checked on the raw string (a comment survives DOM
+  // parsing as a comment node, but the intent is to reject them outright — they
+  // routinely wrap "TODO" cruft that breaks downstream tooling).
+  if (modelXml.includes("<!--")) {
+    issues.push({
+      rule: "no-comments",
+      message: "XML comments (<!-- -->) are not allowed in diagram XML",
+    });
+  }
+
+  // Rule: value escaping + literal newline. Scan raw <mxCell> tags so the error
+  // can name the cell id even when the whole document is otherwise malformed.
+  scanRawValues(modelXml, issues);
+
+  // Well-formedness — everything below needs a parsed DOM.
+  const { doc, error } = parseXml(modelXml);
+  if (error) {
+    issues.push({
+      rule: "well-formed-xml",
+      message: error,
+      position: firstLineCol(error),
+    });
+    throw new DrawioLintError(issues);
+  }
+
+  const root = doc.documentElement;
+  if (!root || root.tagName !== "mxGraphModel") {
+    issues.push({
+      rule: "structure",
+      message: `root element must be <mxGraphModel>, got <${root ? root.tagName : "?"}>`,
+    });
+    throw new DrawioLintError(issues);
+  }
+  if (!firstChildByTag(root, "root")) {
+    issues.push({
+      rule: "structure",
+      message: "<mxGraphModel> must contain a <root> element",
+    });
+    throw new DrawioLintError(issues);
+  }
+
+  const cells = parseCells(modelXml);
+  const ids = new Set<string>();
+
+  // Rule: sentinel cells id="0" and id="1"(parent="0").
+  const cell0 = cells.find((c) => c.id === "0");
+  const cell1 = cells.find((c) => c.id === "1");
+  if (!cell0) {
+    issues.push({
+      rule: "sentinel-cells",
+      message: 'missing the root sentinel cell <mxCell id="0"/>',
+      cellId: "0",
+    });
+  }
+  if (!cell1) {
+    issues.push({
+      rule: "sentinel-cells",
+      message: 'missing the layer sentinel cell <mxCell id="1" parent="0"/>',
+      cellId: "1",
+    });
+  } else if (cell1.parent !== "0") {
+    issues.push({
+      rule: "sentinel-cells",
+      message: 'the layer sentinel <mxCell id="1"> must have parent="0"',
+      cellId: "1",
+    });
+  }
+
+  cells.forEach((c, index) => {
+    const pos = `cell #${index}`;
+    const isSentinel = c.id === "0" || c.id === "1";
+
+    // Rule: unique, non-empty ids; user cells must not reuse 0/1.
+    if (c.id === "") {
+      issues.push({ rule: "cell-id", message: "cell has an empty id", position: pos });
+    } else if (ids.has(c.id)) {
+      issues.push({
+        rule: "duplicate-id",
+        message: `duplicate cell id "${c.id}"`,
+        cellId: c.id,
+        position: pos,
+      });
+    }
+    ids.add(c.id);
+
+    if (isSentinel) return; // sentinels are exempt from the shape rules below
+
+    // Rule: vertex XOR edge (a cell may be neither: groups/containers).
+    if (c.vertex && c.edge) {
+      issues.push({
+        rule: "vertex-edge-exclusive",
+        message: 'a cell cannot be both vertex="1" and edge="1"',
+        cellId: c.id,
+        position: pos,
+      });
+    }
+
+    // Rule: every edge has a child <mxGeometry as="geometry"/>.
+    if (c.edge && !c.geometry.hasGeometry) {
+      issues.push({
+        rule: "edge-geometry",
+        message:
+          'edge is missing its child <mxGeometry relative="1" as="geometry"/> — it will not render',
+        cellId: c.id,
+        position: pos,
+      });
+    }
+
+    // Rule: edge endpoints resolve to existing ids.
+    if (c.edge) {
+      for (const end of ["source", "target"] as const) {
+        const ref = c[end];
+        if (ref != null && ref !== "" && !cellExists(cells, ref)) {
+          issues.push({
+            rule: "edge-endpoint",
+            message: `edge ${end} "${ref}" does not resolve to any cell`,
+            cellId: c.id,
+            position: pos,
+          });
+        }
+      }
+    }
+
+    // Rule: parent must exist.
+    if (c.parent != null && c.parent !== "" && !cellExists(cells, c.parent)) {
+      issues.push({
+        rule: "parent-exists",
+        message: `parent "${c.parent}" does not resolve to any cell`,
+        cellId: c.id,
+        position: pos,
+      });
+    }
+
+    // Rule: style parses as key=value; pairs.
+    if (c.style !== "") {
+      const parsed = parseStyle(c.style);
+      if (parsed.badSegment !== undefined) {
+        issues.push({
+          rule: "style-format",
+          message: `malformed style segment "${parsed.badSegment}" (expected key=value)`,
+          cellId: c.id,
+          position: pos,
+        });
+      }
+    }
+  });
+
+  if (issues.length > 0) throw new DrawioLintError(issues);
+  return { cells, warnings };
+}
+
+function cellExists(cells: DrawioCell[], id: string): boolean {
+  return cells.some((c) => c.id === id);
+}
+
+/**
+ * Raw-string scan of every `value="…"`/`value='…'` on an mxCell tag. Catches an
+ * unescaped `&`/`<`/`>` and a literal newline character inside a value, keyed to
+ * the cell's id. Runs before DOM parsing so a value bug is reported with its
+ * cellId even when the document is otherwise malformed.
+ */
+function scanRawValues(xml: string, issues: DrawioLintIssue[]): void {
+  const tagRe = /<mxCell\b([^>]*?)\/?>/g;
+  let m: RegExpExecArray | null;
+  while ((m = tagRe.exec(xml)) !== null) {
+    const attrs = m[1];
+    const idM = /\bid\s*=\s*"([^"]*)"/.exec(attrs);
+    const cellId = idM ? idM[1] : undefined;
+    const valM = /\bvalue\s*=\s*"([^"]*)"/.exec(attrs) || /\bvalue\s*=\s*'([^']*)'/.exec(attrs);
+    if (!valM) continue;
+    const raw = valM[1];
+    // Literal newline (0x0A / 0x0D) inside the attribute value.
+    if (/[\n\r]/.test(raw)) {
+      issues.push({
+        rule: "value-newline",
+        message:
+          "value contains a literal newline; use &#xa; (or <br> with html=1) instead",
+        cellId,
+      });
+    }
+    // Unescaped '<' or '>' inside a value.
+    if (raw.includes("<") || raw.includes(">")) {
+      issues.push({
+        rule: "value-escaping",
+        message: "value contains an unescaped '<' or '>'; use &lt; / &gt;",
+        cellId,
+      });
+    }
+    // '&' that does not begin a valid entity.
+    const badAmp = /&(?!(amp|lt|gt|quot|apos|#[0-9]+|#x[0-9a-fA-F]+);)/.test(raw);
+    if (badAmp) {
+      issues.push({
+        rule: "value-escaping",
+        message: "value contains an unescaped '&'; use &amp;",
+        cellId,
+      });
+    }
+  }
+}
+
+// --- input normalization + prepare -----------------------------------------
+
+/**
+ * Normalize an accepted tool input into a full mxGraphModel string:
+ *   - a bare `<mxGraphModel>` is used as-is;
+ *   - an `<mxfile>` is decoded to its first page's model;
+ *   - a list of `<mxCell>` is wrapped with the mxGraphModel/root envelope and
+ *     the sentinel cells (id=0, id=1 parent=0) are added when absent.
+ */
+export function normalizeInput(inputXml: string): string {
+  let xml = inputXml.trim();
+  // Strip an optional XML prolog.
+  if (xml.startsWith("<?xml")) {
+    const end = xml.indexOf("?>");
+    if (end !== -1) xml = xml.slice(end + 2).trim();
+  }
+  if (xml.startsWith("<mxfile")) {
+    return decodeDrawioFileToModel(xml);
+  }
+  if (xml.startsWith("<mxGraphModel")) {
+    return xml;
+  }
+  if (xml.includes("<mxCell")) {
+    return wrapCellFragment(xml);
+  }
+  throw new DrawioLintError([
+    {
+      rule: "unrecognized-input",
+      message:
+        "input must be a <mxGraphModel>, an <mxfile>, or a list of <mxCell> elements",
+    },
+  ]);
+}
+
+function wrapCellFragment(fragment: string): string {
+  // Validate the fragment is well-formed (wrapped so a bare list parses) and
+  // discover which sentinels are already present.
+  const { doc, error } = parseXml(`<root>${fragment}</root>`);
+  if (error) {
+    throw new DrawioLintError([
+      {
+        rule: "well-formed-xml",
+        message: error,
+        position: firstLineCol(error),
+      },
+    ]);
+  }
+  const existing = new Set<string>();
+  const els = doc.getElementsByTagName("mxCell");
+  for (let i = 0; i < els.length; i++) {
+    existing.add(els[i].getAttribute("id") ?? "");
+  }
+  let prefix = "";
+  if (!existing.has("0")) prefix += '<mxCell id="0"/>';
+  if (!existing.has("1")) prefix += '<mxCell id="1" parent="0"/>';
+  return `<mxGraphModel ${DEFAULT_MODEL_ATTRS}><root>${prefix}${fragment}</root></mxGraphModel>`;
+}
+
+export interface PreparedModel {
+  /** Canonical (normalized) mxGraphModel XML that gets written. */
+  modelXml: string;
+  cells: DrawioCell[];
+  bbox: DrawioBBox;
+  /** Number of user cells (excludes the id=0/id=1 sentinels). */
+  cellCount: number;
+  warnings: string[];
+  hash: string;
+}
+
+/**
+ * Full pre-write pipeline for create/update: normalize the input into a model,
+ * lint it (throws DrawioLintError on any violation), then compute the canonical
+ * form, bounding box, cell count and hash. Never touches the network.
+ */
+export function prepareModel(inputXml: string): PreparedModel {
+  const rawModel = normalizeInput(inputXml);
+  const { cells, warnings } = lintModel(rawModel);
+  const modelXml = normalizeXml(rawModel);
+  const bbox = computeBBox(cells);
+  const cellCount = cells.filter((c) => c.id !== "0" && c.id !== "1").length;
+  return {
+    modelXml,
+    cells,
+    bbox,
+    cellCount,
+    warnings,
+    hash: mxHash(modelXml),
+  };
+}
+
+/** Cell count of a decoded model (user cells only) — used by drawio_get meta. */
+export function countUserCells(modelXml: string): number {
+  return parseCells(modelXml).filter((c) => c.id !== "0" && c.id !== "1").length;
+}

packages/mcp/src/lib/json-edit.ts

@@ -12,7 +12,11 @@
  * re-import for small wording fixes.
  */
 
-import { stripInlineMarkdown, stripBalancedWrappers } from "./text-normalize.js";
+import {
+  stripInlineMarkdown,
+  stripBalancedWrappers,
+  closestBlockHint,
+} from "./text-normalize.js";
 
 export interface TextEdit {
   find: string;
@@ -381,29 +385,9 @@ export function applyTextEdits(
       } else {
         // Append a bounded "closest text" hint: find the FIRST block that
         // contains the longest whitespace-delimited token (>= 3 chars) of the
-        // (stripped, then raw) locator, and quote that block's plain text.
+        // (stripped, then raw) locator, and quote that block's plain text. Shared
-        reason = "text not found in the document.";
+        // with create_comment via closestBlockHint so both give the same hint.
-        const tokenSource = stripped.length > 0 ? stripped : edit.find;
+        reason = "text not found in the document." + closestBlockHint(blockPlain, edit.find);
-        const longestToken = tokenSource
-          .split(/\s+/)
-          .filter((t) => t.length >= 3)
-          .sort((a, b) => b.length - a.length)[0];
-        if (longestToken) {
-          const hitBlock = blockPlain.find((plain) =>
-            plain.includes(longestToken),
-          );
-          if (hitBlock) {
-            // Truncate by code point (spread iterates by code point) so a
-            // surrogate pair is never split; append the ellipsis only when the
-            // text was actually longer than the limit.
-            const points = [...hitBlock];
-            const snippet =
-              points.length > 120
-                ? points.slice(0, 120).join("") + "…"
-                : hitBlock;
-            reason += ` Closest block text: "${snippet}".`;
-          }
-        }
       }
       failed.push({ find: edit.find, reason });
       continue;

packages/mcp/src/lib/pako.d.ts

@@ -0,0 +1,61 @@
+// Minimal ambient type declaration for `pako` (no @types/pako is installed and
+// pako 2.x ships no bundled .d.ts). We only use the raw-deflate codec to read
+// draw.io's compressed `<diagram>` payload, so declare just that surface.
+declare module "pako" {
+  interface RawOptions {
+    /** When "string", the result is returned as a (binary/UTF-8) string. */
+    to?: "string";
+    /** Raw-deflate window bits; draw.io uses raw deflate (no zlib header). */
+    windowBits?: number;
+    level?: number;
+  }
+
+  /** Raw-inflate (windowBits: -15). `to:"string"` yields a string. */
+  export function inflateRaw(
+    data: Uint8Array | ArrayBuffer | number[],
+    options: RawOptions & { to: "string" },
+  ): string;
+  export function inflateRaw(
+    data: Uint8Array | ArrayBuffer | number[],
+    options?: RawOptions,
+  ): Uint8Array;
+
+  /** Raw-deflate (windowBits: -15). Used only by tests to build fixtures. */
+  export function deflateRaw(
+    data: Uint8Array | string,
+    options?: RawOptions,
+  ): Uint8Array;
+
+  interface InflateStreamOptions {
+    to?: "string";
+    windowBits?: number;
+    /** Raw deflate (no zlib header) — equivalent to windowBits: -15. */
+    raw?: boolean;
+    chunkSize?: number;
+  }
+
+  /**
+   * Streaming inflate. We use it to bound the decompressed size: `onData` is
+   * invoked per output chunk, letting us abort a decompression bomb before the
+   * full output is materialised.
+   */
+  export class Inflate {
+    constructor(options?: InflateStreamOptions);
+    onData: (chunk: string | Uint8Array) => void;
+    onEnd: (status: number) => void;
+    push(
+      data: Uint8Array | ArrayBuffer | number[] | string,
+      flushMode?: boolean | number,
+    ): boolean;
+    result: string | Uint8Array;
+    err: number;
+    msg: string;
+  }
+
+  const _default: {
+    inflateRaw: typeof inflateRaw;
+    deflateRaw: typeof deflateRaw;
+    Inflate: typeof Inflate;
+  };
+  export default _default;
+}

packages/mcp/src/lib/text-normalize.ts

@@ -114,3 +114,37 @@ export function stripInlineMarkdown(s: string): string {
 
   return out;
 }
+
+/**
+ * Build a bounded "closest text" hint for an anchor/find MISS, shared by
+ * edit_page_text (json-edit) and create_comment (client) so both surface the
+ * same self-correction affordance.
+ *
+ * Take the longest whitespace-delimited token (>= 3 chars) of the locator
+ * (markdown-stripped first, so `**bold**` contributes `bold`), find the FIRST
+ * of `blockTexts` that contains it, and return ` Closest block text: "…".` with
+ * the block quoted (truncated to 120 code points + ellipsis). Returns "" when
+ * no token qualifies or no block contains it, so the caller can append it
+ * unconditionally.
+ */
+export function closestBlockHint(
+  blockTexts: string[],
+  locator: string,
+): string {
+  if (typeof locator !== "string" || locator.length === 0) return "";
+  const stripped = stripInlineMarkdown(locator);
+  const tokenSource = stripped.length > 0 ? stripped : locator;
+  const longestToken = tokenSource
+    .split(/\s+/)
+    .filter((t) => t.length >= 3)
+    .sort((a, b) => b.length - a.length)[0];
+  if (!longestToken) return "";
+  const hitBlock = blockTexts.find((plain) => plain.includes(longestToken));
+  if (!hitBlock) return "";
+  // Truncate by code point (spread iterates by code point) so a surrogate pair
+  // is never split; append the ellipsis only when the text was actually longer.
+  const points = [...hitBlock];
+  const snippet =
+    points.length > 120 ? points.slice(0, 120).join("") + "…" : hitBlock;
+  return ` Closest block text: "${snippet}".`;
+}

packages/mcp/src/stdio.ts

@@ -1,6 +1,7 @@
 #!/usr/bin/env node
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 import { createDocmostMcpServer } from "./index.js";
+import { destroyAllSessions } from "./lib/collab-session.js";
 
 // Standalone stdio entrypoint. This restores the original behavior of the
 // package when run as a CLI (`docmost-mcp`): it reads credentials from the
@@ -33,6 +34,20 @@ async function run() {
     console.error("Uncaught exception:", error);
   });
 
+  // Teardown hook (issue #400): destroy every cached live CollabSession on exit
+  // so a hanging session does not keep a doc loaded on the server (which would
+  // also defer the server's afterUnloadDocument cleanup). `exit` runs the
+  // synchronous idempotent teardown; SIGINT/SIGTERM also run it, then exit.
+  process.on("exit", () => {
+    destroyAllSessions();
+  });
+  for (const signal of ["SIGINT", "SIGTERM"] as const) {
+    process.on(signal, () => {
+      destroyAllSessions();
+      process.exit(0);
+    });
+  }
+
   const server = createDocmostMcpServer({
     apiUrl: API_URL!,
     email: EMAIL!,

packages/mcp/src/tool-specs.ts

@@ -771,9 +771,13 @@ export const SHARED_TOOL_SPECS = {
       '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 " +
+      "parent's anchor and take no selection. Always COPY the `selection` " +
-      '"selection not found" error, retry with a corrected EXACT selection ' +
+      'VERBATIM from get_page / search_in_page output — do NOT quote it from ' +
-      'copied verbatim from a single paragraph/block. You may also attach a ' +
+      'memory (stale-memory quoting is the top cause of anchor misses). If the ' +
+      'call fails with a "selection not found" error, the error quotes the ' +
+      "closest block text (or says the selection spans multiple blocks); 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.',
@@ -1115,4 +1119,112 @@ export const SHARED_TOOL_SPECS = {
       alt: z.string().optional(),
     }),
   },
+
+  // --- draw.io diagrams (issue #423, stage 1) ---
+
+  drawioGet: {
+    mcpName: 'drawio_get',
+    inAppKey: 'drawioGet',
+    description:
+      'Read a draw.io diagram on a page as mxGraph XML (default) or as its raw ' +
+      '`.drawio.svg`. `node` is the drawio node\'s attrs.id (from get_outline / ' +
+      'get_page_json) or "#<index>" for a top-level block. Returns the decoded ' +
+      'mxGraphModel XML plus meta { attachmentId, title, width, height, ' +
+      'cellCount, hash }. `hash` is the optimistic-lock key you MUST pass back ' +
+      'as baseHash to drawio_update. Diagrams a human saved from the editor ' +
+      '(including draw.io\'s compressed format) decode losslessly.',
+    tier: 'deferred',
+    catalogLine:
+      'drawioGet — read a draw.io diagram as mxGraph XML (+ hash for updates).',
+    buildShape: (z) => ({
+      pageId: z.string().min(1),
+      node: z
+        .string()
+        .min(1)
+        .describe('The drawio node attrs.id, or "#<index>" for a top-level block.'),
+      format: z
+        .enum(['xml', 'svg'])
+        .optional()
+        .describe('"xml" (default) for mxGraph XML, or "svg" for the raw .drawio.svg.'),
+    }),
+  },
+
+  drawioCreate: {
+    mcpName: 'drawio_create',
+    inAppKey: 'drawioCreate',
+    description:
+      'Create a draw.io diagram from mxGraph XML and insert it as a diagram ' +
+      'block. `xml` is a bare `<mxGraphModel>` OR a list of `<mxCell>` elements ' +
+      '(the server wraps it and adds the id=0 / id=1 sentinel cells). The XML is ' +
+      'LINTED first (well-formedness, sentinel cells, unique ids, vertex XOR ' +
+      'edge, every edge has a child <mxGeometry as="geometry"/>, edge ' +
+      'source/target and every parent resolve, style parses, no XML comments, ' +
+      'value escaping) — a violation returns a structured error naming the rule ' +
+      'and cellId so you can fix and retry. `where` positions the block like ' +
+      'insert_node: position before/after (with exactly one of anchorNodeId or ' +
+      'anchorText) or append. Returns { nodeId, attachmentId, warnings }. The ' +
+      'returned `nodeId` is an index-based "#<index>" handle (drawio nodes carry ' +
+      'no attrs.id): it addresses the new top-level block and can be fed straight ' +
+      'back into drawio_get / drawio_update for THIS document. It is positional, ' +
+      'so if you add or remove blocks before it, re-resolve via get_outline. The ' +
+      'diagram is editable in the draw.io editor and can be re-read with ' +
+      'drawio_get.',
+    tier: 'deferred',
+    catalogLine:
+      'drawioCreate — create a draw.io diagram from mxGraph XML and insert it.',
+    buildShape: (z) => ({
+      pageId: z.string().min(1),
+      xml: z
+        .string()
+        .min(1)
+        .describe(
+          'mxGraph XML: a bare <mxGraphModel> or a list of <mxCell> elements.',
+        ),
+      position: z
+        .enum(['before', 'after', 'append'])
+        .describe('Where to insert relative to the anchor.'),
+      anchorNodeId: z
+        .string()
+        .optional()
+        .describe('Anchor block id (for before/after).'),
+      anchorText: z
+        .string()
+        .optional()
+        .describe('Anchor text fragment (for before/after).'),
+      title: z.string().optional().describe('Optional diagram title.'),
+    }),
+  },
+
+  drawioUpdate: {
+    mcpName: 'drawio_update',
+    inAppKey: 'drawioUpdate',
+    description:
+      'Replace a draw.io diagram\'s content with new mxGraph XML (same lint ' +
+      'pipeline as drawio_create). `baseHash` is MANDATORY: pass the hash from ' +
+      'the drawio_get you based the edit on. If the diagram changed since ' +
+      '(a human or another agent edited it) the hash mismatches and the update ' +
+      'is refused with a conflict error — re-read with drawio_get and retry. On ' +
+      'success it overwrites the diagram attachment and updates the node ' +
+      'width/height. `node` is the drawio node attrs.id or "#<index>".',
+    tier: 'deferred',
+    catalogLine:
+      'drawioUpdate — replace a draw.io diagram (optimistic-locked by baseHash).',
+    buildShape: (z) => ({
+      pageId: z.string().min(1),
+      node: z
+        .string()
+        .min(1)
+        .describe('The drawio node attrs.id, or "#<index>" for a top-level block.'),
+      xml: z
+        .string()
+        .min(1)
+        .describe(
+          'New mxGraph XML: a bare <mxGraphModel> or a list of <mxCell> elements.',
+        ),
+      baseHash: z
+        .string()
+        .min(1)
+        .describe('The meta.hash from the drawio_get this edit is based on.'),
+    }),
+  },
 } satisfies Record<string, SharedToolSpec>;

packages/mcp/test/mock/collab-token-cache.test.mjs

@@ -0,0 +1,282 @@
+// Unit tests for the collab-token cache (issue #435). The live CollabSession
+// registry (#400/#431) keys sessions on (wsUrl, pageId, collabToken), so a token
+// string that changes every op defeats reuse. This cache holds the last minted
+// token per DocmostClient for MCP_COLLAB_TOKEN_TTL_MS so a burst of mutations
+// reuses ONE token -> ONE session. These tests exercise both mint sources:
+//   - the getCollabToken PROVIDER path (in-app agent), via a counting provider fn;
+//   - the REST /auth/collab-token path (external MCP), via a mock http server.
+// getCollabTokenWithReauth is private in TS but a plain method on the compiled
+// build, so the tests call it directly (same convention as reauth.test.mjs).
+import { test, afterEach, after } from "node:test";
+import assert from "node:assert/strict";
+import http from "node:http";
+import { DocmostClient } from "../../build/client.js";
+
+// Restore the env knob after each test so cases do not leak into one another.
+const ENV_KEY = "MCP_COLLAB_TOKEN_TTL_MS";
+afterEach(() => {
+  delete process.env[ENV_KEY];
+});
+
+// ---------------------------------------------------------------------------
+// Small mock server for the REST /auth/collab-token path. Counts collab-token
+// mints and can be told to 401 the first N of them (to drive the reauth retry).
+// ---------------------------------------------------------------------------
+function readBody(req) {
+  return new Promise((resolve) => {
+    let raw = "";
+    req.on("data", (c) => (raw += c));
+    req.on("end", () => resolve(raw));
+  });
+}
+function sendJson(res, status, obj, extra = {}) {
+  res.writeHead(status, { "Content-Type": "application/json", ...extra });
+  res.end(JSON.stringify(obj));
+}
+
+const openServers = [];
+after(async () => {
+  await Promise.all(
+    openServers.map((s) => new Promise((r) => s.close(r))),
+  );
+});
+
+// state: { collabCalls, loginCalls, unauthorizedCollabHits }
+function spawnCollabServer(state, { collabAuthFailsFor = 0 } = {}) {
+  return new Promise((resolve) => {
+    const server = http.createServer(async (req, res) => {
+      await readBody(req);
+      if (req.url === "/api/auth/login") {
+        state.loginCalls++;
+        // A fresh authToken per login so an identity change is observable.
+        sendJson(res, 200, { success: true }, {
+          "Set-Cookie": `authToken=login-${state.loginCalls}; Path=/; HttpOnly`,
+        });
+        return;
+      }
+      if (req.url === "/api/auth/collab-token") {
+        state.collabCalls++;
+        if (state.collabCalls <= collabAuthFailsFor) {
+          sendJson(res, 401, { message: "Unauthorized" });
+          return;
+        }
+        // Unique token per mint so a stale cached value is distinguishable.
+        sendJson(res, 200, { data: { token: `collab-${state.collabCalls}` } });
+        return;
+      }
+      sendJson(res, 404, { message: "not found" });
+    });
+    server.listen(0, "127.0.0.1", () => {
+      openServers.push(server);
+      resolve(`http://127.0.0.1:${server.address().port}/api`);
+    });
+  });
+}
+
+// ===========================================================================
+// PROVIDER path (in-app agent getCollabToken fn)
+// ===========================================================================
+
+// A counting provider that returns a distinct token each call so a cached
+// (reused) token is visibly the SAME string while a fresh mint is different.
+function countingProvider() {
+  let n = 0;
+  const fn = async () => {
+    n++;
+    return `provider-token-${n}`;
+  };
+  return {
+    fn,
+    get calls() {
+      return n;
+    },
+  };
+}
+
+test("within TTL, repeated calls return the SAME token and mint ONCE (provider path)", async () => {
+  process.env[ENV_KEY] = "300000"; // 5 min
+  const p = countingProvider();
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+    getCollabToken: p.fn,
+  });
+
+  const a = await client.getCollabTokenWithReauth();
+  const b = await client.getCollabTokenWithReauth();
+  const c = await client.getCollabTokenWithReauth();
+
+  assert.equal(a, "provider-token-1");
+  assert.equal(b, a, "second call reuses the cached token");
+  assert.equal(c, a, "third call reuses the cached token");
+  assert.equal(p.calls, 1, "the provider is invoked exactly once within the TTL");
+});
+
+test("after TTL expiry a new token is minted (provider path)", async () => {
+  process.env[ENV_KEY] = "20"; // 20ms TTL
+  const p = countingProvider();
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+    getCollabToken: p.fn,
+  });
+
+  const a = await client.getCollabTokenWithReauth();
+  await new Promise((r) => setTimeout(r, 40)); // let the TTL lapse
+  const b = await client.getCollabTokenWithReauth();
+
+  assert.equal(a, "provider-token-1");
+  assert.equal(b, "provider-token-2", "a fresh token is minted after expiry");
+  assert.equal(p.calls, 2);
+});
+
+test("MCP_COLLAB_TOKEN_TTL_MS=0 disables the cache: mint on EVERY call (provider path)", async () => {
+  process.env[ENV_KEY] = "0";
+  const p = countingProvider();
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+    getCollabToken: p.fn,
+  });
+
+  await client.getCollabTokenWithReauth();
+  await client.getCollabTokenWithReauth();
+  await client.getCollabTokenWithReauth();
+
+  assert.equal(p.calls, 3, "cache disabled -> exact fetch-per-call legacy path");
+});
+
+test("a 401 triggers the internal reauth retry, which bypasses the cache and mints fresh (provider path)", async () => {
+  process.env[ENV_KEY] = "300000";
+  let n = 0;
+  const provider = async () => {
+    n++;
+    if (n === 1) {
+      // The FIRST mint fails with an auth error; the internal reauth retry must
+      // re-invoke the provider (bypassing the empty cache) for a fresh token.
+      const err = new Error("collab token expired");
+      err.status = 401;
+      throw err;
+    }
+    return `provider-token-${n}`;
+  };
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+    getCollabToken: provider,
+  });
+
+  // Cache is empty: mint #1 401s -> the reauth retry mints #2 and caches it.
+  const tok = await client.getCollabTokenWithReauth();
+  assert.equal(tok, "provider-token-2", "the post-401 retry token wins");
+  assert.equal(n, 2, "exactly one failed mint + one retry, no loop");
+
+  // The retried token is what got cached (no extra mint on a cache hit).
+  const cached = await client.getCollabTokenWithReauth();
+  assert.equal(cached, "provider-token-2");
+  assert.equal(n, 2, "served from cache, provider not re-invoked");
+});
+
+test("forceRefresh=true bypasses a warm cache and mints a fresh token (provider path)", async () => {
+  process.env[ENV_KEY] = "300000";
+  const p = countingProvider();
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+    getCollabToken: p.fn,
+  });
+
+  const first = await client.getCollabTokenWithReauth(); // caches token-1
+  assert.equal(first, "provider-token-1");
+
+  // A forced refresh (what the reauth path passes) must NOT return the cached
+  // token-1; it mints a fresh token-2 and replaces the cache.
+  const forced = await client.getCollabTokenWithReauth(true);
+  assert.equal(forced, "provider-token-2", "cache bypassed on forceRefresh");
+  assert.equal(p.calls, 2);
+
+  const cached = await client.getCollabTokenWithReauth();
+  assert.equal(cached, "provider-token-2", "the fresh token replaced the cache");
+  assert.equal(p.calls, 2);
+});
+
+test("two consecutive mutations keep the SAME token, so the session key is stable (provider path)", async () => {
+  // The whole point of #435: acquireCollabSession keys on the token, so two
+  // acquire calls in a burst must be handed the identical token string.
+  process.env[ENV_KEY] = "300000";
+  const p = countingProvider();
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+    getCollabToken: p.fn,
+  });
+
+  const t1 = await client.getCollabTokenWithReauth();
+  const t2 = await client.getCollabTokenWithReauth();
+  assert.equal(t1, t2, "identical token across two mutations -> one session key");
+  assert.equal(p.calls, 1);
+});
+
+// ===========================================================================
+// REST /auth/collab-token path (external MCP)
+// ===========================================================================
+
+test("within TTL, the REST /auth/collab-token endpoint is hit ONCE", async () => {
+  process.env[ENV_KEY] = "300000";
+  const state = { collabCalls: 0, loginCalls: 0 };
+  const baseURL = await spawnCollabServer(state);
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+
+  const a = await client.getCollabTokenWithReauth();
+  const b = await client.getCollabTokenWithReauth();
+
+  assert.equal(a, "collab-1");
+  assert.equal(b, a, "cached token reused");
+  assert.equal(state.collabCalls, 1, "POST /auth/collab-token called once");
+});
+
+test("TTL=0 hits the REST endpoint on every call", async () => {
+  process.env[ENV_KEY] = "0";
+  const state = { collabCalls: 0, loginCalls: 0 };
+  const baseURL = await spawnCollabServer(state);
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+
+  await client.getCollabTokenWithReauth();
+  await client.getCollabTokenWithReauth();
+
+  assert.equal(state.collabCalls, 2, "cache disabled -> fetch each call");
+});
+
+test("401 on REST collab-token re-logs-in and refetches (cache bypassed)", async () => {
+  process.env[ENV_KEY] = "300000";
+  const state = { collabCalls: 0, loginCalls: 0 };
+  // The first collab-token mint 401s; the reauth path logs in and retries.
+  const baseURL = await spawnCollabServer(state, { collabAuthFailsFor: 1 });
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  // Pre-seed a token so the initial call does not perform an initial login.
+  client.token = "seed";
+  client.client.defaults.headers.common["Authorization"] = "Bearer seed";
+
+  const tok = await client.getCollabTokenWithReauth();
+  assert.equal(tok, "collab-2", "the post-reauth mint wins, not the failed one");
+  assert.equal(state.loginCalls, 1, "re-login happened exactly once");
+  assert.equal(state.collabCalls, 2, "one failed mint + one successful retry");
+});
+
+test("a fresh login clears the cache so a collab token cannot outlive the identity", async () => {
+  process.env[ENV_KEY] = "300000";
+  const state = { collabCalls: 0, loginCalls: 0 };
+  const baseURL = await spawnCollabServer(state);
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+
+  const before = await client.getCollabTokenWithReauth();
+  assert.equal(before, "collab-1");
+
+  // Simulate an identity change (the 401 interceptor / re-login path calls
+  // login(), which must drop the cached collab token).
+  await client.login();
+
+  const after = await client.getCollabTokenWithReauth();
+  assert.equal(after, "collab-2", "cache was invalidated by login(); refetched");
+  assert.equal(state.collabCalls, 2);
+});

packages/mcp/test/mock/create-comment.test.mjs

@@ -548,3 +548,94 @@ test("suggestedText: the stored selection is the doc's RAW typographic substring
   );
   assert.equal(createPayload.suggestedText, "goodbye");
 });
+
+// -----------------------------------------------------------------------------
+// 8) #408: a not-found selection error QUOTES the closest block text so the
+//    model can self-correct instead of blind-retrying.
+// -----------------------------------------------------------------------------
+test("a not-found selection error includes a 'Closest block text' hint", async () => {
+  let createCalls = 0;
+  const { baseURL } = await spawn(async (req, res) => {
+    await readBody(req);
+    if (req.url === "/api/auth/login") {
+      sendJson(res, 200, { success: true }, {
+        "Set-Cookie": "authToken=t; Path=/; HttpOnly",
+      });
+      return;
+    }
+    if (req.url === "/api/pages/info") {
+      sendJson(res, 200, {
+        data: {
+          id: "page-1",
+          content: {
+            type: "doc",
+            content: [
+              { type: "paragraph", content: [{ type: "text", text: "The quick brown fox jumps" }] },
+            ],
+          },
+        },
+      });
+      return;
+    }
+    if (req.url === "/api/comments/create") {
+      createCalls++;
+      sendJson(res, 200, { data: { id: "should-not-happen" } });
+      return;
+    }
+    sendJson(res, 404, { message: "not found" });
+  });
+
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  await assert.rejects(
+    () => client.createComment("page-1", "body", "inline", "quick brown cat"),
+    /Closest block text: "The quick brown fox jumps"/,
+    "a not-found selection must quote the closest block text",
+  );
+  assert.equal(createCalls, 0, "/comments/create must NOT be called on a miss");
+});
+
+// -----------------------------------------------------------------------------
+// 9) #408: a selection that straddles two blocks gets the explicit
+//    "spans multiple blocks" message instead of a bare not-found.
+// -----------------------------------------------------------------------------
+test("a selection spanning multiple blocks gets the explicit spans-multiple-blocks message", async () => {
+  let createCalls = 0;
+  const { baseURL } = await spawn(async (req, res) => {
+    await readBody(req);
+    if (req.url === "/api/auth/login") {
+      sendJson(res, 200, { success: true }, {
+        "Set-Cookie": "authToken=t; Path=/; HttpOnly",
+      });
+      return;
+    }
+    if (req.url === "/api/pages/info") {
+      sendJson(res, 200, {
+        data: {
+          id: "page-1",
+          content: {
+            type: "doc",
+            content: [
+              { type: "paragraph", content: [{ type: "text", text: "the quick brown" }] },
+              { type: "paragraph", content: [{ type: "text", text: "fox jumps over" }] },
+            ],
+          },
+        },
+      });
+      return;
+    }
+    if (req.url === "/api/comments/create") {
+      createCalls++;
+      sendJson(res, 200, { data: { id: "should-not-happen" } });
+      return;
+    }
+    sendJson(res, 404, { message: "not found" });
+  });
+
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  await assert.rejects(
+    () => client.createComment("page-1", "body", "inline", "brown fox"),
+    /spans multiple blocks/,
+    "a cross-block selection must report the spans-multiple-blocks hint",
+  );
+  assert.equal(createCalls, 0, "/comments/create must NOT be called on a miss");
+});

packages/mcp/test/mock/drawio-tools.test.mjs

@@ -0,0 +1,467 @@
+// Contract tests for the drawio_get / drawio_create / drawio_update client
+// methods (issue #423). Follows the repo's seam-override pattern (see
+// full-doc-write-canonicalize.test.mjs): a DocmostClient subclass stubs the I/O
+// seams (auth, collab token, page read, attachment upload/fetch, the mutatePage
+// write) so the tool logic is exercised without a live Docmost or collab socket.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import pako from "pako";
+import { DocmostClient } from "../../build/client.js";
+import {
+  buildDrawioSvg,
+  encodeDrawioFile,
+  normalizeXml,
+  mxHash,
+  decodeDrawioSvg,
+} from "../../build/lib/drawio-xml.js";
+
+const MODEL =
+  '<mxGraphModel><root>' +
+  '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+  '<mxCell id="2" value="Hi" style="rounded=1;" vertex="1" parent="1">' +
+  '<mxGeometry x="20" y="20" width="120" height="60" as="geometry"/></mxCell>' +
+  '</root></mxGraphModel>';
+
+// Build a Docmost-style `.drawio.svg` (base64 content) for a model.
+function svgFor(model) {
+  return buildDrawioSvg(normalizeXml(model), "<g/>", { width: 200, height: 120 });
+}
+
+// Build a human/compressed-export `.drawio.svg` (base64 content wrapping a
+// compressed <diagram> payload), mimicking a diagram a person saved.
+function compressedSvgFor(model) {
+  const compressed = Buffer.from(
+    pako.deflateRaw(encodeURIComponent(normalizeXml(model))),
+  ).toString("base64");
+  const file = `<mxfile host="Electron"><diagram id="a" name="Page-1">${compressed}</diagram></mxfile>`;
+  const content = Buffer.from(file, "utf-8").toString("base64");
+  return `<svg xmlns="http://www.w3.org/2000/svg" content="${content}"><image href="x"/></svg>`;
+}
+
+// The vendored `drawio` node schema (diagramAttributes) declares ONLY these
+// attributes; PMNode.fromJSON drops anything else on save. Mirror that here so
+// the mock write path behaves like the real one — in particular, a block `id`
+// set on a drawio node does NOT survive the save, so a handle keyed on it is
+// un-resolvable. This is exactly what the production bug (issue #423 Fix 1) was.
+const DRAWIO_SCHEMA_ATTRS = new Set([
+  "src",
+  "title",
+  "alt",
+  "width",
+  "height",
+  "size",
+  "aspectRatio",
+  "align",
+  "attachmentId",
+]);
+
+function applyDrawioSchemaDrop(node) {
+  if (!node || typeof node !== "object") return;
+  if (node.type === "drawio" && node.attrs && typeof node.attrs === "object") {
+    for (const key of Object.keys(node.attrs)) {
+      if (!DRAWIO_SCHEMA_ATTRS.has(key)) delete node.attrs[key];
+    }
+  }
+  if (Array.isArray(node.content)) for (const c of node.content) applyDrawioSchemaDrop(c);
+}
+
+function makeClient({ pageDoc, attachmentSvg } = {}) {
+  const calls = { uploads: [], mutations: [] };
+  class TestClient extends DocmostClient {
+    async ensureAuthenticated() {}
+    async getCollabTokenWithReauth() {
+      return "collab-token";
+    }
+    async resolvePageId(pageId) {
+      return `uuid-${pageId}`;
+    }
+    async getPageRaw(pageId) {
+      return {
+        id: pageId,
+        slugId: "s",
+        title: "P",
+        spaceId: "sp",
+        content: pageDoc ?? { type: "doc", content: [] },
+      };
+    }
+    async uploadAttachmentBuffer(pageId, buffer, fileName, mime) {
+      const id = `att-${calls.uploads.length + 1}`;
+      calls.uploads.push({ pageId, fileName, mime, svg: buffer.toString("utf-8") });
+      return { id, fileName, fileSize: buffer.length };
+    }
+    async fetchAttachmentText(src) {
+      return attachmentSvg;
+    }
+    mutatePage(pageId, token, apiUrl, transform) {
+      // Run the transform against a clone of the source doc, capture the result.
+      const clone = structuredClone(pageDoc ?? { type: "doc", content: [] });
+      const doc = transform(clone);
+      // Mirror the real schema: unknown drawio attrs (e.g. a block `id`) are
+      // dropped on save, so callers can never rely on them to address the node.
+      if (doc) applyDrawioSchemaDrop(doc);
+      calls.mutations.push({ pageId, doc });
+      return Promise.resolve({ doc, verify: { changed: doc != null } });
+    }
+  }
+  const client = new TestClient("http://127.0.0.1:1/api", "e@x.com", "pw");
+  return { client, calls };
+}
+
+function findDrawio(node, acc = []) {
+  if (!node || typeof node !== "object") return acc;
+  if (node.type === "drawio") acc.push(node);
+  if (Array.isArray(node.content)) for (const c of node.content) findDrawio(c, acc);
+  return acc;
+}
+
+// --- drawio_create ---------------------------------------------------------
+
+test("drawio_create: lints, builds the .drawio.svg, uploads and inserts a node", async () => {
+  const pageDoc = {
+    type: "doc",
+    content: [{ type: "paragraph", attrs: { id: "p1" }, content: [] }],
+  };
+  const { client, calls } = makeClient({ pageDoc });
+  const res = await client.drawioCreate("page1", { position: "append" }, MODEL, "My diagram");
+
+  assert.equal(res.success, true);
+  // The returned handle is an index-based "#<index>" ref (drawio nodes carry no
+  // persisted attrs.id), addressing the appended top-level block (index 1, after
+  // the existing paragraph).
+  assert.equal(res.nodeId, "#1");
+  assert.equal(res.attachmentId, "att-1");
+  assert.equal(calls.uploads.length, 1);
+  assert.equal(calls.uploads[0].fileName, "diagram.drawio.svg");
+  assert.equal(calls.uploads[0].mime, "image/svg+xml");
+  // The uploaded SVG carries the model back (round-trips through the decode chain).
+  assert.equal(decodeDrawioSvg(calls.uploads[0].svg), normalizeXml(MODEL));
+
+  // A drawio node was appended with src/attachmentId/dimensions and the title.
+  const drawios = findDrawio(calls.mutations[0].doc);
+  assert.equal(drawios.length, 1);
+  const n = drawios[0];
+  // No `id` attribute is set/persisted on the node (schema has none).
+  assert.equal(n.attrs.id, undefined);
+  assert.equal(n.attrs.attachmentId, "att-1");
+  assert.match(n.attrs.src, /^\/api\/files\/att-1\//);
+  assert.ok(n.attrs.width > 0 && n.attrs.height > 0);
+  assert.equal(n.attrs.title, "My diagram");
+});
+
+test("drawio_create: a lint violation throws before any upload", async () => {
+  const { client, calls } = makeClient({ pageDoc: { type: "doc", content: [] } });
+  // Edge with no child geometry -> edge-geometry rule.
+  const bad =
+    '<mxGraphModel><root><mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '<mxCell id="3" edge="1" parent="1" source="2" target="2"/></root></mxGraphModel>';
+  await assert.rejects(
+    () => client.drawioCreate("page1", { position: "append" }, bad, undefined),
+    /edge-geometry/,
+  );
+  assert.equal(calls.uploads.length, 0, "no attachment uploaded on lint failure");
+});
+
+test("drawio_create: before/after requires exactly one anchor", async () => {
+  const { client } = makeClient({ pageDoc: { type: "doc", content: [] } });
+  await assert.rejects(
+    () => client.drawioCreate("page1", { position: "before" }, MODEL),
+    /exactly one of anchorNodeId or anchorText/,
+  );
+});
+
+// --- drawio_get ------------------------------------------------------------
+
+test("drawio_get: decodes the model and returns meta with a hash", async () => {
+  const pageDoc = {
+    type: "doc",
+    content: [
+      {
+        type: "drawio",
+        attrs: {
+          id: "d1",
+          src: "/api/files/att-1/diagram.drawio.svg",
+          attachmentId: "att-1",
+          title: "T",
+          width: 200,
+          height: 120,
+        },
+      },
+    ],
+  };
+  const { client } = makeClient({ pageDoc, attachmentSvg: svgFor(MODEL) });
+  const res = await client.drawioGet("page1", "d1", "xml");
+  assert.equal(res.content, normalizeXml(MODEL));
+  assert.equal(res.meta.attachmentId, "att-1");
+  assert.equal(res.meta.title, "T");
+  assert.equal(res.meta.cellCount, 1);
+  assert.equal(res.meta.hash, mxHash(normalizeXml(MODEL)));
+});
+
+test("drawio_get: format=svg returns the raw .drawio.svg", async () => {
+  const svg = svgFor(MODEL);
+  const pageDoc = {
+    type: "doc",
+    content: [
+      { type: "drawio", attrs: { id: "d1", src: "/api/files/att-1/x.svg", attachmentId: "att-1" } },
+    ],
+  };
+  const { client } = makeClient({ pageDoc, attachmentSvg: svg });
+  const res = await client.drawioGet("page1", "d1", "svg");
+  assert.equal(res.content, svg);
+});
+
+test("drawio_get: reads a HUMAN-saved compressed diagram losslessly (pako)", async () => {
+  const pageDoc = {
+    type: "doc",
+    content: [
+      { type: "drawio", attrs: { id: "d1", src: "/api/files/att-1/x.svg", attachmentId: "att-1" } },
+    ],
+  };
+  const { client } = makeClient({ pageDoc, attachmentSvg: compressedSvgFor(MODEL) });
+  const res = await client.drawioGet("page1", "d1", "xml");
+  assert.equal(res.content, normalizeXml(MODEL));
+});
+
+// --- drawio_update ---------------------------------------------------------
+
+const UPDATED_MODEL =
+  '<mxGraphModel><root>' +
+  '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+  '<mxCell id="2" value="Changed" style="rounded=1;" vertex="1" parent="1">' +
+  '<mxGeometry x="20" y="20" width="300" height="200" as="geometry"/></mxCell>' +
+  '</root></mxGraphModel>';
+
+function updatePageDoc() {
+  return {
+    type: "doc",
+    content: [
+      {
+        type: "drawio",
+        attrs: {
+          id: "d1",
+          src: "/api/files/att-1/diagram.drawio.svg",
+          attachmentId: "att-1",
+          width: 200,
+          height: 120,
+        },
+      },
+    ],
+  };
+}
+
+test("drawio_update: stale baseHash -> conflict, no upload", async () => {
+  const { client, calls } = makeClient({
+    pageDoc: updatePageDoc(),
+    attachmentSvg: svgFor(MODEL),
+  });
+  await assert.rejects(
+    () => client.drawioUpdate("page1", "d1", UPDATED_MODEL, "deadbeef-stale"),
+    /conflict/,
+  );
+  assert.equal(calls.uploads.length, 0, "no upload on conflict");
+});
+
+test("drawio_update: current baseHash -> uploads new attachment and repoints node dims", async () => {
+  const currentHash = mxHash(normalizeXml(MODEL));
+  const { client, calls } = makeClient({
+    pageDoc: updatePageDoc(),
+    attachmentSvg: svgFor(MODEL),
+  });
+  const res = await client.drawioUpdate("page1", "d1", UPDATED_MODEL, currentHash);
+  assert.equal(res.success, true);
+  assert.equal(res.attachmentId, "att-1"); // fresh id from the stub sequence
+  assert.equal(calls.uploads.length, 1);
+  // The uploaded SVG carries the NEW model.
+  assert.equal(decodeDrawioSvg(calls.uploads[0].svg), normalizeXml(UPDATED_MODEL));
+  // The node was repointed with the new bounding-box dimensions:
+  // vertex maxX=320,maxY=220 + the 20px preview margin -> 340 x 240.
+  const n = findDrawio(calls.mutations[0].doc)[0];
+  assert.equal(n.attrs.attachmentId, "att-1");
+  assert.equal(n.attrs.width, 340);
+  assert.equal(n.attrs.height, 240);
+  // The block `id` used as the legacy resolution handle is dropped on save
+  // (schema declares no `id`); the update still targeted the correct node.
+  assert.equal(n.attrs.id, undefined);
+});
+
+test("drawio_update: baseHash is mandatory", async () => {
+  const { client } = makeClient({ pageDoc: updatePageDoc(), attachmentSvg: svgFor(MODEL) });
+  await assert.rejects(
+    () => client.drawioUpdate("page1", "d1", UPDATED_MODEL, ""),
+    /baseHash is mandatory/,
+  );
+});
+
+// --- Fix 1: the create handle must resolve on the SAVED doc (no id) ---------
+
+test("drawio_create -> get/update: returned #<index> handle resolves on the saved doc (id dropped)", async () => {
+  // Create appends a drawio node after the existing paragraph.
+  const createDoc = {
+    type: "doc",
+    content: [{ type: "paragraph", attrs: { id: "p1" }, content: [] }],
+  };
+  const create = makeClient({ pageDoc: createDoc });
+  const res = await create.client.drawioCreate(
+    "page1",
+    { position: "append" },
+    MODEL,
+    "T",
+  );
+  // The handle is index-based, not a block id.
+  assert.equal(res.nodeId, "#1");
+
+  // Take the document EXACTLY as it was saved: the schema drop stripped the
+  // node's id, so no id-based handle could ever resolve against it.
+  const savedDoc = create.calls.mutations[0].doc;
+  assert.equal(findDrawio(savedDoc)[0].attrs.id, undefined);
+
+  // drawio_get with the returned handle resolves the just-created node.
+  const getClient = makeClient({ pageDoc: savedDoc, attachmentSvg: svgFor(MODEL) });
+  const got = await getClient.client.drawioGet("page1", res.nodeId, "xml");
+  assert.equal(got.nodeId, res.nodeId);
+  assert.equal(got.content, normalizeXml(MODEL));
+
+  // drawio_update with the same handle + the hash from get repoints that node.
+  const upClient = makeClient({ pageDoc: savedDoc, attachmentSvg: svgFor(MODEL) });
+  const upd = await upClient.client.drawioUpdate(
+    "page1",
+    res.nodeId,
+    UPDATED_MODEL,
+    got.meta.hash,
+  );
+  assert.equal(upd.success, true);
+  assert.equal(upd.nodeId, res.nodeId);
+  const updated = findDrawio(upClient.calls.mutations[0].doc)[0];
+  assert.equal(
+    decodeDrawioSvg(upClient.calls.uploads[0].svg),
+    normalizeXml(UPDATED_MODEL),
+  );
+  assert.equal(updated.attrs.width, 340);
+});
+
+// --- error paths: the LLM must get a clean error, not a crash --------------
+
+test("drawio_get: a bad node ref -> clean 'no node found' error", async () => {
+  // Page has one paragraph; the requested ref resolves to nothing.
+  const pageDoc = {
+    type: "doc",
+    content: [{ type: "paragraph", attrs: { id: "p1" }, content: [] }],
+  };
+  const { client } = makeClient({ pageDoc, attachmentSvg: svgFor(MODEL) });
+  await assert.rejects(
+    () => client.drawioGet("page1", "does-not-exist", "xml"),
+    /no node found for "does-not-exist"/,
+  );
+});
+
+test("drawio_get: a drawio node with no src -> clean 'has no src to read' error", async () => {
+  const pageDoc = {
+    type: "doc",
+    content: [
+      // A drawio node that carries no `src` (e.g. a half-written node).
+      { type: "drawio", attrs: { id: "d1", attachmentId: "att-1" } },
+    ],
+  };
+  const { client } = makeClient({ pageDoc, attachmentSvg: svgFor(MODEL) });
+  await assert.rejects(
+    () => client.drawioGet("page1", "d1", "xml"),
+    /node "d1" on page page1 has no src to read/,
+  );
+});
+
+test("drawio_update: the resolved node is NOT a drawio node -> clean error, no upload", async () => {
+  // "#0" resolves to a paragraph. The update must refuse cleanly rather than
+  // crash or repoint the wrong node.
+  const pageDoc = {
+    type: "doc",
+    content: [{ type: "paragraph", attrs: { id: "p1" }, content: [] }],
+  };
+  const { client, calls } = makeClient({ pageDoc, attachmentSvg: svgFor(MODEL) });
+  await assert.rejects(
+    () => client.drawioUpdate("page1", "#0", UPDATED_MODEL, "any-nonempty-hash"),
+    /node "#0" on page page1 is a paragraph, not a drawio diagram/,
+  );
+  assert.equal(calls.uploads.length, 0, "no upload when the node is not a diagram");
+  assert.equal(calls.mutations.length, 0, "no write when the node is not a diagram");
+});
+
+test("drawio_create: anchor not found -> clean error that reports the orphan attachment", async () => {
+  // The upload happens before the mutate transform; when the anchor cannot be
+  // found the write is skipped and the (now unreferenced) attachment is named
+  // in the error, exactly as the code documents.
+  const pageDoc = {
+    type: "doc",
+    content: [{ type: "paragraph", attrs: { id: "p1" }, content: [] }],
+  };
+  const { client, calls } = makeClient({ pageDoc });
+  await assert.rejects(
+    () =>
+      client.drawioCreate(
+        "page1",
+        { position: "after", anchorNodeId: "nope" },
+        MODEL,
+        "T",
+      ),
+    (err) =>
+      /anchor not found/.test(err.message) &&
+      /unreferenced orphan/.test(err.message) &&
+      /att-1/.test(err.message),
+  );
+  // The orphan was uploaded (and reported), but no node was written.
+  assert.equal(calls.uploads.length, 1, "attachment uploaded before the failed insert");
+  const drawios = calls.mutations.length ? findDrawio(calls.mutations[0].doc) : [];
+  assert.equal(drawios.length, 0, "no drawio node written when the anchor is missing");
+});
+
+// --- Fix 2: update targets ONLY the resolved node --------------------------
+
+test("drawio_update: repoints ONLY the addressed node, not siblings sharing an attachmentId", async () => {
+  // A copied diagram: two drawio nodes share one attachmentId. Updating via the
+  // "#0" handle must touch node #0 only, never the sibling copy.
+  const shared = {
+    type: "doc",
+    content: [
+      {
+        type: "drawio",
+        attrs: {
+          src: "/api/files/shared/x.svg",
+          attachmentId: "shared",
+          width: 200,
+          height: 120,
+        },
+      },
+      {
+        type: "drawio",
+        attrs: {
+          src: "/api/files/shared/x.svg",
+          attachmentId: "shared",
+          width: 200,
+          height: 120,
+        },
+      },
+    ],
+  };
+  const { client, calls } = makeClient({
+    pageDoc: shared,
+    attachmentSvg: svgFor(MODEL),
+  });
+  const res = await client.drawioUpdate(
+    "page1",
+    "#0",
+    UPDATED_MODEL,
+    mxHash(normalizeXml(MODEL)),
+  );
+  assert.equal(res.success, true);
+
+  const drawios = findDrawio(calls.mutations[0].doc);
+  assert.equal(drawios.length, 2);
+  // Node #0 repointed to the NEW attachment ("att-1" from the stub) and dims.
+  assert.equal(drawios[0].attrs.attachmentId, "att-1");
+  assert.equal(drawios[0].attrs.width, 340);
+  assert.match(drawios[0].attrs.src, /^\/api\/files\/att-1\//);
+  // Node #1 (the sibling copy) is untouched despite sharing the old attachmentId.
+  assert.equal(drawios[1].attrs.attachmentId, "shared");
+  assert.equal(drawios[1].attrs.width, 200);
+  assert.equal(drawios[1].attrs.src, "/api/files/shared/x.svg");
+});

packages/mcp/test/mock/resolve-page-id-collab-doc-name.test.mjs

@@ -19,6 +19,7 @@ import { WebSocketServer } from "ws";
 import { Hocuspocus } from "@hocuspocus/server";
 import { DocmostClient } from "../../build/client.js";
 import { buildYDoc } from "../../build/lib/collaboration.js";
+import { destroyAllSessions } from "../../build/lib/collab-session.js";
 // Import the SAME page-lock module instance that build/client.js imports. ESM
 // caches modules by resolved URL, so this `withPageLock` shares the very
 // per-page mutex map (`chains`) the client uses — letting the replaceImage test
@@ -188,6 +189,10 @@ async function spawnCollabStack(opts = {}) {
 
 const openStacks = [];
 after(async () => {
+  // #400: tests now leave a cached live CollabSession per page. Destroy them
+  // first (closes the client ws) so the server.close() below is not racing an
+  // open collab connection.
+  destroyAllSessions();
   await Promise.all(
     openStacks.map(
       ({ server, hocuspocus }) =>
@@ -270,17 +275,23 @@ test("a UUID input is passed through unchanged and triggers NO /pages/info fetch
   );
 });
 
-test("a repeated slugId edit resolves the UUID only once (cache)", async () => {
+test("repeated slugId edits reuse ONE live collab session and resolve the UUID only once (#400 cache)", async () => {
   const { state, baseURL } = await spawnCollabStack();
   const client = new DocmostClient(baseURL, "user@example.com", "pw");
 
-  // Each mock connection re-seeds a fresh "hello world" doc (the mock does not
+  // #400: a series of edits on the same page reuses ONE live CollabSession, so
-  // persist across connects), so both edits target "hello". The cache assertion
+  // the connect/handshake happens once and the collab doc is OPENED a single
-  // only concerns the slugId->uuid resolution, not the document content.
+  // time (not per edit). The live ydoc persists between edits (the whole point),
+  // so the second edit sees the first edit's result: after "hello" -> "hi world"
+  // it targets the still-present "world".
   await client.editPageText(SLUG, [{ find: "hello", replace: "hi" }]);
-  await client.editPageText(SLUG, [{ find: "hello", replace: "hey" }]);
+  await client.editPageText(SLUG, [{ find: "world", replace: "planet" }]);
 
-  assert.deepEqual(state.docNames, [`page.${UUID}`, `page.${UUID}`]);
+  assert.deepEqual(
+    state.docNames,
+    [`page.${UUID}`],
+    "the two edits must reuse one live collab session -> a single collab-doc open (#400)",
+  );
   assert.equal(
     state.pagesInfoCalls.length,
     1,
@@ -325,8 +336,9 @@ test("replaceImage opens by the resolved UUID AND keys its page lock by that UUI
   await uploadStarted; // deterministic: replaceImage now holds its page lock.
 
   // (a) OPEN BY UUID: the only collab doc opened so far (the scan pass) used the
-  // canonical UUID, never the slugId. (The write pass opens a second time after
+  // canonical UUID, never the slugId. (#400: the write pass will REUSE this same
-  // we release the gate; asserted at the end.)
+  // live session rather than reopen, so docNames stays a single entry — asserted
+  // at the end.)
   assert.deepEqual(
     state.docNames,
     [`page.${UUID}`],
@@ -378,8 +390,9 @@ test("replaceImage opens by the resolved UUID AND keys its page lock by that UUI
 
   assert.equal(res.success, true);
   assert.equal(res.replaced, 1, "the one seeded image must be repointed");
-  // Both opens (scan pass + write pass) used the UUID; the slugId never appears.
+  // #400: the write pass REUSES the scan pass's live session, so the collab doc
-  assert.deepEqual(state.docNames, [`page.${UUID}`, `page.${UUID}`]);
+  // is opened ONCE across both passes (never reopened, never by the slugId).
+  assert.deepEqual(state.docNames, [`page.${UUID}`]);
   assert.ok(
     !state.docNames.includes(`page.${SLUG}`),
     "replaceImage must NEVER open the collab doc by the slugId (the #260 bug)",

packages/mcp/test/unit/client-host-contract.test.mjs

@@ -81,6 +81,10 @@ const HOST_CONTRACT_METHODS = [
   "insertImage",
   "replaceImage",
   "insertFootnote",
+  // draw.io diagrams (#423, stage 1) — read + create + optimistic-locked update
+  "drawioGet",
+  "drawioCreate",
+  "drawioUpdate",
   // write (comment)
   "createComment",
   "resolveComment",

packages/mcp/test/unit/collab-session.test.mjs

@@ -0,0 +1,348 @@
+import { test, beforeEach, afterEach, mock } from "node:test";
+import assert from "node:assert/strict";
+import { EventEmitter } from "node:events";
+
+import {
+  acquireCollabSession,
+  destroyAllSessions,
+  __setCollabProviderFactory,
+  __sessionCountForTests,
+} from "../../build/lib/collab-session.js";
+import { withPageLock } from "../../build/lib/page-lock.js";
+
+// A stand-in for HocuspocusProvider: it shares the ydoc (so the real yjs
+// read/transform/write in CollabSession.mutate runs unchanged), auto-completes
+// the connect+sync handshake on a microtask (unaffected by mock timers), and
+// exposes hooks to drive disconnect/close/auth-failure and the unsyncedChanges
+// ack. There is no collaboration server in the test env, so every test drives
+// the provider through this fake.
+class FakeProvider extends EventEmitter {
+  static instances = [];
+  static connectCount = 0;
+
+  static reset() {
+    FakeProvider.instances = [];
+    FakeProvider.connectCount = 0;
+  }
+
+  static last() {
+    return FakeProvider.instances[FakeProvider.instances.length - 1];
+  }
+
+  constructor(config, opts = {}) {
+    super();
+    this.config = config;
+    this.ydoc = config.document;
+    this.synced = false;
+    this.unsyncedChanges = opts.unsynced ?? 0;
+    this.destroyed = false;
+    FakeProvider.instances.push(this);
+    FakeProvider.connectCount += 1;
+    if (opts.autoSync !== false) {
+      // Real HocuspocusProvider fires onSynced asynchronously after the
+      // handshake; a microtask reproduces that without depending on timers.
+      queueMicrotask(() => {
+        if (this.destroyed) return;
+        this.config.onConnect?.();
+        this.synced = true;
+        this.config.onSynced?.();
+      });
+    }
+  }
+
+  destroy() {
+    this.destroyed = true;
+  }
+
+  // --- test drivers ---
+  _disconnect() {
+    this.config.onDisconnect?.();
+  }
+  _close() {
+    this.config.onClose?.();
+  }
+  _authFail() {
+    this.config.onAuthenticationFailed?.();
+  }
+  _ack() {
+    this.unsyncedChanges = 0;
+    this.emit("unsyncedChanges", { number: 0 });
+  }
+}
+
+/** Build a provider factory that stamps every provider with the given opts. */
+function factory(opts = {}) {
+  return (config) => new FakeProvider(config, opts);
+}
+
+/** A minimal, schema-valid ProseMirror doc for a write. */
+function docWith(text) {
+  return {
+    type: "doc",
+    content: [
+      { type: "paragraph", content: [{ type: "text", text: String(text) }] },
+    ],
+  };
+}
+
+const ENV_KEYS = [
+  "MCP_COLLAB_SESSION_IDLE_MS",
+  "MCP_COLLAB_SESSION_MAX_AGE_MS",
+  "MCP_COLLAB_SESSION_MAX_ENTRIES",
+];
+let savedEnv;
+
+beforeEach(() => {
+  savedEnv = {};
+  for (const k of ENV_KEYS) savedEnv[k] = process.env[k];
+  FakeProvider.reset();
+  __setCollabProviderFactory(factory());
+});
+
+afterEach(() => {
+  destroyAllSessions();
+  __setCollabProviderFactory(null);
+  mock.timers.reset();
+  for (const k of ENV_KEYS) {
+    if (savedEnv[k] === undefined) delete process.env[k];
+    else process.env[k] = savedEnv[k];
+  }
+});
+
+test("acquire opens one provider; reuse returns the SAME live session", async () => {
+  const a = await acquireCollabSession("page-1", "tok", "http://h/api");
+  const b = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.equal(a, b, "same (wsUrl,page,token) must reuse the session");
+  assert.equal(FakeProvider.connectCount, 1, "exactly one connect/sync");
+  assert.equal(__sessionCountForTests(), 1);
+});
+
+test("N mutates on one page open the provider ONCE (coalesced series)", async () => {
+  const session = await acquireCollabSession("page-1", "tok", "http://h/api");
+  for (let i = 0; i < 20; i++) {
+    const r = await session.mutate(() => docWith(`edit ${i}`));
+    assert.ok(r && r.doc, "each mutate resolves a MutationResult");
+  }
+  assert.equal(
+    FakeProvider.connectCount,
+    1,
+    "20 mutates must not reconnect — one live provider for the whole series",
+  );
+});
+
+test("mutate preserves the transform-abort no-op report (null transform)", async () => {
+  const session = await acquireCollabSession("page-1", "tok", "http://h/api");
+  const r = await session.mutate(() => null);
+  assert.equal(r.verify.changed, false);
+  assert.equal(r.verify.summary, "no changes (transform aborted)");
+});
+
+test("registry key includes the token: different tokens => different sessions", async () => {
+  const a = await acquireCollabSession("page-1", "tok-A", "http://h/api");
+  const b = await acquireCollabSession("page-1", "tok-B", "http://h/api");
+  assert.notEqual(a, b, "different tokens must never share a session");
+  assert.equal(FakeProvider.connectCount, 2);
+  assert.equal(__sessionCountForTests(), 2);
+  // Same token reuses.
+  const a2 = await acquireCollabSession("page-1", "tok-A", "http://h/api");
+  assert.equal(a, a2);
+  assert.equal(FakeProvider.connectCount, 2);
+});
+
+test("disconnect at any time kills the session and removes it from the registry", async () => {
+  const a = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.equal(__sessionCountForTests(), 1);
+  FakeProvider.last()._disconnect();
+  assert.equal(__sessionCountForTests(), 0, "dead session is deregistered");
+  // Re-acquire opens a brand new provider.
+  const b = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.notEqual(a, b);
+  assert.equal(FakeProvider.connectCount, 2);
+});
+
+test("an in-flight mutate rejects with the connection-closed text on disconnect", async () => {
+  __setCollabProviderFactory(factory({ unsynced: 1 })); // stay pending after write
+  const session = await acquireCollabSession("page-1", "tok", "http://h/api");
+  const p = session.mutate(() => docWith("x"));
+  // The write happened synchronously; persistence is pending. Now the socket drops.
+  FakeProvider.last()._disconnect();
+  await assert.rejects(
+    p,
+    /Collaboration connection closed before the update was persisted\/synced/,
+  );
+});
+
+test("auth failure rejects the pending open with the auth error text", async () => {
+  __setCollabProviderFactory(factory({ autoSync: false }));
+  const p = acquireCollabSession("page-1", "tok", "http://h/api");
+  // Let the provider be constructed, then fail auth.
+  await Promise.resolve();
+  FakeProvider.last()._authFail();
+  await assert.rejects(
+    p,
+    /Authentication failed for collaboration connection/,
+  );
+  assert.equal(__sessionCountForTests(), 0);
+});
+
+test("mutate-error invalidates the session (caller destroys; re-acquire is fresh)", async () => {
+  const session = await acquireCollabSession("page-1", "tok", "http://h/api");
+  await assert.rejects(
+    session.mutate(() => {
+      throw new Error("afterText not found");
+    }),
+    /afterText not found/,
+  );
+  // The production caller destroys on failure; mirror that here.
+  session.destroy("mutate failed");
+  assert.equal(__sessionCountForTests(), 0);
+  const fresh = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.notEqual(fresh, session);
+  assert.equal(FakeProvider.connectCount, 2);
+});
+
+test("a pending write resolves when the server acks (unsyncedChanges -> 0)", async () => {
+  __setCollabProviderFactory(factory({ unsynced: 1 }));
+  const session = await acquireCollabSession("page-1", "tok", "http://h/api");
+  const p = session.mutate(() => docWith("y"));
+  FakeProvider.last()._ack();
+  const r = await p;
+  assert.ok(r.doc);
+});
+
+test("concurrent mutate on one session: the second rejects, the first is unaffected", async () => {
+  __setCollabProviderFactory(factory({ unsynced: 1 })); // first stays pending after write
+  const session = await acquireCollabSession("page-1", "tok", "http://h/api");
+  // First mutate: write happens synchronously, persistence ack still pending.
+  const first = session.mutate(() => docWith("first"));
+  // Second overlapping mutate on the SAME session must fail fast.
+  await assert.rejects(
+    session.mutate(() => docWith("second")),
+    /mutate already in-flight; caller must serialize \(hold the page lock\)/,
+  );
+  // The first op is untouched: its rejector was not clobbered. Ack it now.
+  FakeProvider.last()._ack();
+  const r = await first;
+  assert.ok(r.doc, "the first in-flight mutate still resolves on its own ack");
+  assert.equal(FakeProvider.connectCount, 1, "no reconnect from the rejected overlap");
+});
+
+test("sequential mutates on one session both succeed (the guard doesn't break serialized use)", async () => {
+  const session = await acquireCollabSession("page-1", "tok", "http://h/api");
+  // Await the first fully, then run the second: inflightReject was cleared by
+  // localFinish before the first settled, so the guard is clear for the second.
+  const r1 = await session.mutate(() => docWith("one"));
+  assert.ok(r1.doc, "first sequential mutate resolves");
+  const r2 = await session.mutate(() => docWith("two"));
+  assert.ok(r2.doc, "second sequential mutate resolves — guard not tripped");
+  assert.equal(FakeProvider.connectCount, 1, "sequential mutates reuse one provider");
+});
+
+test("connect timeout rejects with the connect-timeout text and fires the metric hook", async () => {
+  mock.timers.enable({ apis: ["setTimeout"] });
+  __setCollabProviderFactory(factory({ autoSync: false }));
+  let metricFired = 0;
+  const p = acquireCollabSession("page-1", "tok", "http://h/api", {
+    onConnectTimeout: () => {
+      metricFired += 1;
+    },
+  });
+  mock.timers.tick(25000);
+  await assert.rejects(p, /Connection timeout to collaboration server/);
+  assert.equal(metricFired, 1);
+  assert.equal(__sessionCountForTests(), 0);
+});
+
+test("idle TTL destroys the session; re-acquire reconnects", async () => {
+  mock.timers.enable({ apis: ["setTimeout"] });
+  process.env.MCP_COLLAB_SESSION_IDLE_MS = "1000";
+  const a = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.equal(__sessionCountForTests(), 1);
+  mock.timers.tick(1000); // idle fires
+  assert.equal(__sessionCountForTests(), 0, "idle timeout destroyed it");
+  const b = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.notEqual(a, b);
+  assert.equal(FakeProvider.connectCount, 2);
+});
+
+test("max age is enforced at acquire (destroy + open fresh), idle held off", async () => {
+  mock.timers.enable({ apis: ["setTimeout", "Date"] });
+  process.env.MCP_COLLAB_SESSION_MAX_AGE_MS = "1000";
+  process.env.MCP_COLLAB_SESSION_IDLE_MS = "10000000"; // never fires in this test
+  const a = await acquireCollabSession("page-1", "tok", "http://h/api");
+  mock.timers.tick(2000); // past max age, below idle
+  const b = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.notEqual(a, b, "a session past its max age must be replaced");
+  assert.equal(FakeProvider.connectCount, 2);
+});
+
+test("registry cap: least-recently-used session is destroy-evicted", async () => {
+  process.env.MCP_COLLAB_SESSION_MAX_ENTRIES = "2";
+  const s1 = await acquireCollabSession("page-1", "tok", "http://h/api");
+  const p1prov = FakeProvider.last();
+  const s2 = await acquireCollabSession("page-2", "tok", "http://h/api");
+  const p2prov = FakeProvider.last();
+  assert.equal(__sessionCountForTests(), 2);
+
+  // Touch page-1 so page-2 becomes the least-recently-used entry.
+  assert.equal(await acquireCollabSession("page-1", "tok", "http://h/api"), s1);
+  assert.equal(FakeProvider.connectCount, 2, "reuse must not reconnect");
+
+  // A third distinct page (cap 2) must evict the LRU (page-2), not page-1.
+  const s3 = await acquireCollabSession("page-3", "tok", "http://h/api");
+  assert.equal(__sessionCountForTests(), 2);
+  assert.equal(FakeProvider.connectCount, 3);
+  assert.equal(p2prov.destroyed, true, "the LRU provider was destroy-evicted");
+  assert.equal(p1prov.destroyed, false, "the MRU page-1 survived");
+
+  // page-1 still reused (survived), page-3 still reused.
+  assert.equal(await acquireCollabSession("page-1", "tok", "http://h/api"), s1);
+  assert.equal(await acquireCollabSession("page-3", "tok", "http://h/api"), s3);
+  assert.equal(FakeProvider.connectCount, 3, "no extra reconnects");
+});
+
+test("MCP_COLLAB_SESSION_IDLE_MS=0 disables the cache (legacy provider-per-op)", async () => {
+  process.env.MCP_COLLAB_SESSION_IDLE_MS = "0";
+  const s1 = await acquireCollabSession("page-1", "tok", "http://h/api");
+  // Never registered (ephemeral).
+  assert.equal(__sessionCountForTests(), 0);
+  const r1 = await s1.mutate(() => docWith("a"));
+  assert.ok(r1.doc);
+  // After its single op the ephemeral session self-destroyed.
+  assert.equal(FakeProvider.instances[0].destroyed, true);
+  // A second op opens a BRAND NEW provider (no reuse).
+  const s2 = await acquireCollabSession("page-1", "tok", "http://h/api");
+  assert.notEqual(s1, s2);
+  assert.equal(FakeProvider.connectCount, 2);
+});
+
+test("replaceImage-shaped flow: acquire under an EXTERNAL page lock does not deadlock and reuses one session", async () => {
+  const pageId = "page-lock";
+  // Mirror replaceImage: hold ONE withPageLock across scan (read-only) + write,
+  // each going through the non-locking acquireCollabSession.
+  const result = await withPageLock(pageId, async () => {
+    // pass 1: read-only scan (transform returns null -> no write)
+    const scan = await acquireCollabSession(pageId, "tok", "http://h/api");
+    const scanRes = await scan.mutate(() => null);
+    assert.equal(scanRes.verify.changed, false);
+    // pass 2: the actual write, same held lock
+    const write = await acquireCollabSession(pageId, "tok", "http://h/api");
+    return write.mutate(() => docWith("repointed"));
+  });
+  assert.ok(result.doc, "the locked scan+write completed without deadlock");
+  assert.equal(
+    FakeProvider.connectCount,
+    1,
+    "both passes under the held lock reuse ONE live session",
+  );
+});
+
+test("destroyAllSessions tears down every cached session", async () => {
+  await acquireCollabSession("page-1", "tok", "http://h/api");
+  await acquireCollabSession("page-2", "tok", "http://h/api");
+  assert.equal(__sessionCountForTests(), 2);
+  const provs = [...FakeProvider.instances];
+  destroyAllSessions();
+  assert.equal(__sessionCountForTests(), 0);
+  assert.ok(provs.every((p) => p.destroyed), "all providers destroyed");
+});

packages/mcp/test/unit/comment-anchor.test.mjs

@@ -8,6 +8,7 @@ import {
   applyAnchorInDoc,
   countAnchorMatches,
   getAnchoredText,
+  resolveAnchorSelection,
 } from "../../build/lib/comment-anchor.js";
 
 const COMMENT_ID = "cmt-123";
@@ -308,3 +309,70 @@ test("getAnchoredText returns null when the selection does not anchor", () => {
   const doc = paragraphDoc([{ type: "text", text: "hello world" }]);
   assert.equal(getAnchoredText(doc, "not present"), null);
 });
+
+// ---------------------------------------------------------------------------
+// #408 MARKDOWN-STRIP FALLBACK. A selection copied with inline markdown still
+// carries `**`/`` ` ``/`[t](u)` markers the plain document text lacks. When the
+// verbatim selection anchors nowhere, all four entry points retry with the
+// markdown stripped — consistently, so the suggestion-uniqueness gate stays
+// coherent — while what gets STORED remains the raw document substring.
+// ---------------------------------------------------------------------------
+test("a markdown-styled selection anchors against plain doc text via the strip fallback", () => {
+  const doc = paragraphDoc([{ type: "text", text: "a bold word here" }]);
+  // The agent quoted "**bold** word" from a styled view; the doc is plain text.
+  const sel = "**bold** word";
+  const resolved = resolveAnchorSelection(doc, sel);
+  assert.equal(resolved.found, true, "strip fallback finds the anchor");
+  assert.equal(resolved.normalized, true, "reports the soft-warning flag");
+  assert.equal(canAnchorInDoc(doc, sel), true);
+  assert.equal(countAnchorMatches(doc, sel), 1);
+
+  const ok = applyAnchorInDoc(doc, sel, COMMENT_ID);
+  assert.equal(ok, true);
+  const marked = doc.content[0].content.filter((p) => commentMark(p));
+  assert.equal(marked.map((m) => m.text).join(""), "bold word",
+    "the mark lands on the plain-text span");
+});
+
+test("getAnchoredText stores the RAW doc substring even when matched via the strip fallback", () => {
+  // Doc uses a smart apostrophe; the agent typed ASCII + markdown emphasis.
+  const doc = paragraphDoc([{ type: "text", text: "it’s bold now" }]);
+  const stored = getAnchoredText(doc, "it's **bold**");
+  assert.equal(stored, "it’s bold",
+    "stored selection is the raw document text, not the stripped/ASCII locator");
+});
+
+test("the strip fallback does not flip a raw-unique selection to ambiguous", () => {
+  // "config" appears twice, but the raw phrase "config value" appears once.
+  const doc = {
+    type: "doc",
+    content: [
+      { type: "paragraph", content: [{ type: "text", text: "the config value here" }] },
+      { type: "paragraph", content: [{ type: "text", text: "another config here" }] },
+    ],
+  };
+  // Raw phrase is unique -> exactly 1, and no strip happens (nothing to strip).
+  assert.equal(countAnchorMatches(doc, "config value"), 1);
+  assert.equal(resolveAnchorSelection(doc, "config value").normalized, false);
+});
+
+test("EXACT WINS: a raw match short-circuits the strip fallback (count reflects raw)", () => {
+  // A literal "**" run exists raw once; its stripped form would also appear.
+  const doc = paragraphDoc([{ type: "text", text: "use **stars** and stars" }]);
+  // Raw "**stars**" occurs once -> count 1 from the verbatim locator; the
+  // fallback (which would find two "stars") never runs.
+  assert.equal(countAnchorMatches(doc, "**stars**"), 1);
+  assert.equal(resolveAnchorSelection(doc, "**stars**").normalized, false);
+});
+
+test("a markdown selection whose stripped form is ambiguous is counted as ambiguous", () => {
+  const doc = {
+    type: "doc",
+    content: [
+      { type: "paragraph", content: [{ type: "text", text: "first config here" }] },
+      { type: "paragraph", content: [{ type: "text", text: "second config here" }] },
+    ],
+  };
+  // Verbatim "**config**" matches nothing; stripped "config" matches twice.
+  assert.equal(countAnchorMatches(doc, "**config**"), 2);
+});

packages/mcp/test/unit/drawio-preview.test.mjs

@@ -0,0 +1,84 @@
+// Unit tests for the pure-TS schematic SVG preview (issue #423). Asserts that
+// the preview emits well-formed SVG covering each primitive (rect/ellipse/
+// rhombus/edge), resolves container-relative coordinates to absolute, and that
+// the full `.drawio.svg` wrapper parses.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { JSDOM } from "jsdom";
+import { renderDiagramShapes } from "../../build/lib/drawio-preview.js";
+import {
+  parseCells,
+  computeBBox,
+  buildDrawioSvg,
+  normalizeXml,
+} from "../../build/lib/drawio-xml.js";
+
+const { window } = new JSDOM("");
+function parseSvg(svg) {
+  const doc = new window.DOMParser().parseFromString(svg, "application/xml");
+  const err = doc.getElementsByTagName("parsererror");
+  assert.equal(err.length, 0, `SVG did not parse: ${err[0]?.textContent}`);
+  return doc;
+}
+
+const MODEL =
+  '<mxGraphModel><root>' +
+  '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+  '<mxCell id="2" value="Box &amp; Co" style="rounded=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">' +
+  '<mxGeometry x="40" y="40" width="120" height="60" as="geometry"/></mxCell>' +
+  '<mxCell id="3" value="Circle" style="ellipse;fillColor=#d5e8d4;" vertex="1" parent="1">' +
+  '<mxGeometry x="240" y="40" width="80" height="80" as="geometry"/></mxCell>' +
+  '<mxCell id="5" value="Dec" style="rhombus;" vertex="1" parent="1">' +
+  '<mxGeometry x="40" y="160" width="100" height="60" as="geometry"/></mxCell>' +
+  '<mxCell id="4" value="link" edge="1" parent="1" source="2" target="3"><mxGeometry relative="1" as="geometry"/></mxCell>' +
+  '</root></mxGraphModel>';
+
+test("renderDiagramShapes emits rect, ellipse, polygon and a line", () => {
+  const cells = parseCells(MODEL);
+  const inner = renderDiagramShapes(cells, computeBBox(cells));
+  assert.ok(inner.includes("<rect"), "has a rect");
+  assert.ok(inner.includes("<ellipse"), "has an ellipse");
+  assert.ok(inner.includes("<polygon"), "has a polygon (rhombus)");
+  assert.ok(inner.includes("<line"), "has an edge line");
+  // Labels are HTML-escaped.
+  assert.ok(inner.includes("Box &amp; Co"));
+});
+
+test("the full .drawio.svg wrapper parses as valid XML", () => {
+  const cells = parseCells(MODEL);
+  const bbox = computeBBox(cells);
+  const inner = renderDiagramShapes(cells, bbox);
+  const svg = buildDrawioSvg(normalizeXml(MODEL), inner, bbox);
+  const doc = parseSvg(svg);
+  assert.equal(doc.documentElement.tagName, "svg");
+  assert.ok(doc.documentElement.getAttribute("content"), "carries content=");
+  // The visible children exist.
+  assert.ok(doc.getElementsByTagName("rect").length >= 1);
+});
+
+test("container children resolve to absolute coordinates", () => {
+  // A group at (100,100) with a child rect at relative (10,10,20,20) -> abs 110,110.
+  const model =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="g" style="group;" vertex="1" parent="1"><mxGeometry x="100" y="100" width="200" height="200" as="geometry"/></mxCell>' +
+    '<mxCell id="c" value="in" vertex="1" parent="g"><mxGeometry x="10" y="10" width="20" height="20" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const cells = parseCells(model);
+  const inner = renderDiagramShapes(cells, computeBBox(cells));
+  // The child rect must be placed at absolute x=110,y=110.
+  assert.ok(/<rect x="110" y="110"/.test(inner), `expected abs child rect, got: ${inner}`);
+});
+
+test("unknown stencil (shape=mxgraph.*) degrades to a labeled rectangle", () => {
+  const model =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" value="AWS" style="shape=mxgraph.aws4.lambda;" vertex="1" parent="1">' +
+    '<mxGeometry x="0" y="0" width="60" height="60" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const cells = parseCells(model);
+  const inner = renderDiagramShapes(cells, computeBBox(cells));
+  assert.ok(inner.includes("<rect"), "unknown stencil -> rect");
+  assert.ok(inner.includes(">AWS<"), "keeps the label");
+});

packages/mcp/test/unit/drawio-xml.test.mjs

@@ -0,0 +1,360 @@
+// Unit tests for the drawio-xml module (issue #423): the linter (a positive
+// baseline + a negative case per rule, each asserting rule + cellId), the
+// decode chain (plain nested XML AND draw.io's compressed <diagram> via pako),
+// encode/round-trip byte-stability, hash stability, style parsing and the
+// bounding box.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import pako from "pako";
+import {
+  parseStyle,
+  lintModel,
+  prepareModel,
+  normalizeInput,
+  normalizeXml,
+  mxHash,
+  computeBBox,
+  parseCells,
+  decodeDrawioSvg,
+  decodeDrawioFileToModel,
+  buildDrawioSvg,
+  encodeDrawioFile,
+  countUserCells,
+  DrawioLintError,
+  inflateDiagramPayload,
+  MAX_INFLATED_DIAGRAM_BYTES,
+} from "../../build/lib/drawio-xml.js";
+
+// A well-formed model with one vertex and a valid edge to it.
+const VALID_MODEL =
+  '<mxGraphModel><root>' +
+  '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+  '<mxCell id="2" value="Hello" style="rounded=1;whiteSpace=wrap;html=1;" vertex="1" parent="1">' +
+  '<mxGeometry x="100" y="100" width="120" height="60" as="geometry"/></mxCell>' +
+  '<mxCell id="3" value="Two" style="ellipse;" vertex="1" parent="1">' +
+  '<mxGeometry x="300" y="100" width="80" height="80" as="geometry"/></mxCell>' +
+  '<mxCell id="4" edge="1" parent="1" source="2" target="3">' +
+  '<mxGeometry relative="1" as="geometry"/></mxCell>' +
+  '</root></mxGraphModel>';
+
+function issuesOf(fn) {
+  try {
+    fn();
+    return null;
+  } catch (e) {
+    assert.ok(e instanceof DrawioLintError, `expected DrawioLintError, got ${e}`);
+    return e.issues;
+  }
+}
+const hasRule = (issues, rule, cellId) =>
+  issues.some(
+    (i) => i.rule === rule && (cellId === undefined || i.cellId === cellId),
+  );
+
+// --- style parsing ---------------------------------------------------------
+
+test("parseStyle: base stylename + key=value pairs", () => {
+  const r = parseStyle("ellipse;fillColor=#ff0000;whiteSpace=wrap;");
+  assert.equal(r.baseStyle, "ellipse");
+  assert.equal(r.map.fillColor, "#ff0000");
+  assert.equal(r.map.whiteSpace, "wrap");
+  assert.equal(r.badSegment, undefined);
+});
+
+test("parseStyle: flags a segment with two '='", () => {
+  const r = parseStyle("a=b=c;");
+  assert.equal(r.badSegment, "a=b=c");
+});
+
+test("parseStyle: a second bare token is malformed", () => {
+  const r = parseStyle("rounded=1;bareword");
+  assert.equal(r.badSegment, "bareword");
+});
+
+// --- linter: positive baseline ---------------------------------------------
+
+test("lintModel: the canonical valid model passes", () => {
+  const { cells } = lintModel(VALID_MODEL);
+  assert.equal(cells.length, 5);
+});
+
+// --- linter: one negative case per rule ------------------------------------
+
+test("rule well-formed-xml: malformed XML", () => {
+  const issues = issuesOf(() => lintModel("<mxGraphModel><root><mxCell id=\"0\"></root>"));
+  assert.ok(hasRule(issues, "well-formed-xml"));
+  assert.ok(issues[0].position, "carries a line:col position");
+});
+
+test("rule structure: root is not mxGraphModel", () => {
+  const issues = issuesOf(() => lintModel("<foo><root/></foo>"));
+  assert.ok(hasRule(issues, "structure"));
+});
+
+test("rule sentinel-cells: missing id=0 / id=1", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "sentinel-cells", "0"));
+  assert.ok(hasRule(issues, "sentinel-cells", "1"));
+});
+
+test("rule duplicate-id: two cells share an id", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "duplicate-id", "2"));
+});
+
+test("rule vertex-edge-exclusive: cell is both vertex and edge", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" vertex="1" edge="1" parent="1"><mxGeometry as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "vertex-edge-exclusive", "2"));
+});
+
+test("rule edge-geometry: self-closed edge without child mxGeometry", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '<mxCell id="3" vertex="1" parent="1"><mxGeometry x="30" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '<mxCell id="4" edge="1" parent="1" source="2" target="3"/>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "edge-geometry", "4"));
+});
+
+test("rule edge-endpoint: source/target does not resolve", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '<mxCell id="4" edge="1" parent="1" source="2" target="99"><mxGeometry relative="1" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "edge-endpoint", "4"));
+});
+
+test("rule parent-exists: parent points at a missing id", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" vertex="1" parent="42"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "parent-exists", "2"));
+});
+
+test("rule no-comments: XML comment present", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<!-- a comment --><mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "no-comments"));
+});
+
+test("rule style-format: malformed style segment (cellId reported)", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" style="rounded=1;a=b=c;" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "style-format", "2"));
+});
+
+test("rule value-newline: literal newline in a value (cellId reported)", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" value="line1\nline2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "value-newline", "2"));
+});
+
+test("rule value-escaping: unescaped ampersand in a value (cellId reported)", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" value="A & B" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  const issues = issuesOf(() => lintModel(m));
+  assert.ok(hasRule(issues, "value-escaping", "2"));
+});
+
+test("rule reserved id: escaped entity value passes (no false positive)", () => {
+  const m =
+    '<mxGraphModel><root>' +
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" value="A &amp; B &lt;ok&gt;" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>' +
+    '</root></mxGraphModel>';
+  assert.doesNotThrow(() => lintModel(m));
+});
+
+// --- input normalization ---------------------------------------------------
+
+test("normalizeInput: a list of <mxCell> is wrapped and sentinels added", () => {
+  const frag =
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>';
+  const model = normalizeInput(frag);
+  assert.ok(model.startsWith("<mxGraphModel"));
+  assert.ok(model.includes('<mxCell id="0"/>'));
+  assert.ok(model.includes('<mxCell id="1" parent="0"/>'));
+  // And it lints clean.
+  assert.doesNotThrow(() => lintModel(model));
+});
+
+test("normalizeInput: an existing sentinel is not duplicated", () => {
+  const frag =
+    '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    '<mxCell id="2" vertex="1" parent="1"><mxGeometry x="0" y="0" width="10" height="10" as="geometry"/></mxCell>';
+  const model = normalizeInput(frag);
+  const count0 = (model.match(/id="0"/g) || []).length;
+  assert.equal(count0, 1);
+});
+
+test("prepareModel: returns bbox, cellCount, hash and lints", () => {
+  const p = prepareModel(VALID_MODEL);
+  assert.equal(p.cellCount, 3); // 2, 3, 4 (sentinels excluded)
+  assert.ok(p.bbox.width > 0 && p.bbox.height > 0);
+  assert.equal(p.hash, mxHash(normalizeXml(VALID_MODEL)));
+});
+
+// --- decode chain: plain -----------------------------------------------------
+
+test("decode chain (plain): buildDrawioSvg -> decodeDrawioSvg round-trips byte-stable", () => {
+  const model = normalizeXml(VALID_MODEL);
+  const svg = buildDrawioSvg(model, "<g/>", { width: 400, height: 200 });
+  const decoded = decodeDrawioSvg(svg);
+  assert.equal(decoded, model);
+});
+
+test("decode chain: entity-encoded content= (draw.io export style) is read directly", () => {
+  const model = normalizeXml(VALID_MODEL);
+  const file = encodeDrawioFile(model, "Page-1");
+  const escaped = file
+    .replace(/&/g, "&amp;")
+    .replace(/</g, "&lt;")
+    .replace(/>/g, "&gt;")
+    .replace(/"/g, "&quot;");
+  const svg = `<svg xmlns="http://www.w3.org/2000/svg" content="${escaped}"></svg>`;
+  const decoded = decodeDrawioSvg(svg);
+  assert.equal(decoded, model);
+});
+
+// --- decode chain: compressed (pako) ---------------------------------------
+
+test("decode chain (compressed pako): human-saved <diagram> payload decodes losslessly", () => {
+  const model = normalizeXml(VALID_MODEL);
+  // Reproduce draw.io's compression: encodeURIComponent -> raw deflate -> base64.
+  const compressed = Buffer.from(
+    pako.deflateRaw(encodeURIComponent(model)),
+  ).toString("base64");
+  const file = `<mxfile host="Electron"><diagram id="abc" name="Page-1">${compressed}</diagram></mxfile>`;
+  // Docmost stores the file base64 in content=.
+  const contentB64 = Buffer.from(file, "utf-8").toString("base64");
+  const svg = `<svg xmlns="http://www.w3.org/2000/svg" content="${contentB64}"><image href="x"/></svg>`;
+  const decoded = decodeDrawioSvg(svg);
+  assert.equal(decoded, model);
+});
+
+test("decodeDrawioFileToModel: bare mxGraphModel file returns the model substring", () => {
+  const model = normalizeXml(VALID_MODEL);
+  assert.equal(decodeDrawioFileToModel(model), model);
+});
+
+// --- hash stability --------------------------------------------------------
+
+test("mxHash: stable across inter-tag whitespace, sensitive to content", () => {
+  const a = VALID_MODEL;
+  const b = VALID_MODEL.replace(/></g, ">\n  <"); // reformat only
+  assert.equal(mxHash(a), mxHash(b));
+  const c = VALID_MODEL.replace('value="Hello"', 'value="Changed"');
+  assert.notEqual(mxHash(a), mxHash(c));
+});
+
+// --- bounding box + cell count ---------------------------------------------
+
+test("computeBBox + countUserCells", () => {
+  const cells = parseCells(VALID_MODEL);
+  const bbox = computeBBox(cells);
+  // Vertex 3 spans to x=380,y=180; plus the 20px margin.
+  assert.equal(bbox.width, 400);
+  assert.equal(bbox.height, 200);
+  assert.equal(countUserCells(VALID_MODEL), 3);
+});
+
+// --- decompression-bomb guard (Fix 3) --------------------------------------
+
+test("inflateDiagramPayload: a small legitimate payload inflates fine", () => {
+  const xml = normalizeXml(VALID_MODEL);
+  const base64 = Buffer.from(
+    pako.deflateRaw(encodeURIComponent(xml)),
+  ).toString("base64");
+  assert.equal(inflateDiagramPayload(base64), xml);
+});
+
+test("inflateDiagramPayload: rejects an over-cap decompression bomb", () => {
+  // A tiny compressed payload that inflates to just over the cap. Highly
+  // compressible (all one byte) -> the base64 is small, but the inflated output
+  // exceeds MAX_INFLATED_DIAGRAM_BYTES and must be refused before it is fully
+  // materialised.
+  const bombSize = MAX_INFLATED_DIAGRAM_BYTES + 1024;
+  const base64 = Buffer.from(
+    pako.deflateRaw(Buffer.alloc(bombSize, 0x41 /* 'A' */)),
+  ).toString("base64");
+  assert.ok(
+    base64.length < 1024 * 1024,
+    "the compressed bomb is tiny relative to its inflated size",
+  );
+  assert.throws(
+    () => inflateDiagramPayload(base64),
+    /decompression bomb/,
+  );
+});
+
+test("encode/build: a title with < > \" & round-trips without corrupting the SVG", async () => {
+  // A user-supplied title full of XML metacharacters must be escaped so the
+  // inner <mxfile> stays well-formed and the outer content="..." attribute is
+  // never broken out of. Prove it survives the encode -> build -> decode chain.
+  const title = 'A < B > C " D & E';
+  const model = normalizeXml(VALID_MODEL);
+  const svg = buildDrawioSvg(model, "<g/>", { width: 200, height: 120 }, title);
+
+  // The outer content="..." attribute must not be broken by the title: the raw
+  // title metacharacters never appear literally in the SVG markup (they are
+  // base64-encoded inside content=, and escaped inside the file XML).
+  const contentMatch = /content="([^"]*)"/.exec(svg);
+  assert.ok(contentMatch, "SVG has a single well-formed content= attribute");
+
+  // The diagram model still decodes losslessly despite the exotic title.
+  assert.equal(decodeDrawioSvg(svg), model);
+
+  // The file XML is well-formed: the title lives in name="..." as escaped
+  // entities, so unescaping recovers the original title byte-for-byte.
+  const fileXml = Buffer.from(contentMatch[1], "base64").toString("utf-8");
+  const nameMatch = /<diagram id="[^"]*" name="([^"]*)">/.exec(fileXml);
+  assert.ok(nameMatch, "the diagram name attribute is intact and quote-safe");
+  const decodedTitle = nameMatch[1]
+    .replace(/&lt;/g, "<")
+    .replace(/&gt;/g, ">")
+    .replace(/&quot;/g, '"')
+    .replace(/&amp;/g, "&");
+  assert.equal(decodedTitle, title);
+
+  // encodeDrawioFile alone produces the same escaped, well-formed envelope.
+  const file = encodeDrawioFile(model, title);
+  assert.match(file, /name="A &lt; B &gt; C &quot; D &amp; E">/);
+});

patches/@hocuspocus__server@3.4.4.patch

@@ -0,0 +1,62 @@
+diff --git a/dist/hocuspocus-server.cjs b/dist/hocuspocus-server.cjs
+index b24ff6d091c32f733089eeaa47b03f7b37cf5964..f003af304fc751b7edc1aee17f3651282d70666a 100644
+--- a/dist/hocuspocus-server.cjs
++++ b/dist/hocuspocus-server.cjs
+@@ -2426,6 +2426,26 @@ class Hocuspocus {
+      * Create a new document by the given request
+      */
+     async createDocument(documentName, request, socketId, connection, context) {
++        // PATCH(gitmost #401): close the connect-vs-unload race. When the last
++        // client disconnects, storeDocumentHooks' finally schedules an unload;
++        // unloadDocument runs its beforeUnloadDocument hooks asynchronously and
++        // records an in-flight promise in `unloadingDocuments`. A NEW connection
++        // arriving in that window would otherwise fall straight through to the
++        // `documents`/`loadingDocuments` checks and could either reuse a doc that
++        // is about to be destroyed, or start loading a fresh doc concurrently
++        // with the destroy. Awaiting the in-flight unload first makes the decision
++        // deterministic: once it settles, either the doc was fully unloaded
++        // (removed from `documents`, so we do a clean fresh load below) or the
++        // unload aborted because work/connections reappeared (the healthy doc is
++        // still in `documents`, so we reuse it). Either way the new connection can
++        // never hand-shake onto an orphaned, about-to-be-destroyed Document.
++        const existingUnloadingDoc = this.unloadingDocuments.get(documentName);
++        if (existingUnloadingDoc) {
++            // Wait for the in-flight unload to settle so we never hand-shake onto a dying
++            // Document. Swallow a rejected unload — fall through to a fresh load, matching
++            // pre-patch behavior (the doc is already removed from `documents` by then).
++            try { await existingUnloadingDoc; } catch { /* unload rejected — fresh load */ }
++        }
+         const existingLoadingDoc = this.loadingDocuments.get(documentName);
+         if (existingLoadingDoc) {
+             return existingLoadingDoc;
+diff --git a/dist/hocuspocus-server.esm.js b/dist/hocuspocus-server.esm.js
+index 1f4dd80244e899128e2c4e5dad8eab7cfc1cbad6..8c2411747bba27fb9486e1df81678a14e41e884e 100644
+--- a/dist/hocuspocus-server.esm.js
++++ b/dist/hocuspocus-server.esm.js
+@@ -2406,6 +2406,26 @@ class Hocuspocus {
+      * Create a new document by the given request
+      */
+     async createDocument(documentName, request, socketId, connection, context) {
++        // PATCH(gitmost #401): close the connect-vs-unload race. When the last
++        // client disconnects, storeDocumentHooks' finally schedules an unload;
++        // unloadDocument runs its beforeUnloadDocument hooks asynchronously and
++        // records an in-flight promise in `unloadingDocuments`. A NEW connection
++        // arriving in that window would otherwise fall straight through to the
++        // `documents`/`loadingDocuments` checks and could either reuse a doc that
++        // is about to be destroyed, or start loading a fresh doc concurrently
++        // with the destroy. Awaiting the in-flight unload first makes the decision
++        // deterministic: once it settles, either the doc was fully unloaded
++        // (removed from `documents`, so we do a clean fresh load below) or the
++        // unload aborted because work/connections reappeared (the healthy doc is
++        // still in `documents`, so we reuse it). Either way the new connection can
++        // never hand-shake onto an orphaned, about-to-be-destroyed Document.
++        const existingUnloadingDoc = this.unloadingDocuments.get(documentName);
++        if (existingUnloadingDoc) {
++            // Wait for the in-flight unload to settle so we never hand-shake onto a dying
++            // Document. Swallow a rejected unload — fall through to a fresh load, matching
++            // pre-patch behavior (the doc is already removed from `documents` by then).
++            try { await existingUnloadingDoc; } catch { /* unload rejected — fresh load */ }
++        }
+         const existingLoadingDoc = this.loadingDocuments.get(documentName);
+         if (existingLoadingDoc) {
+             return existingLoadingDoc;

pnpm-lock.yaml

@@ -44,6 +44,9 @@ overrides:
   ip-address: 10.1.1
 
 patchedDependencies:
+  '@hocuspocus/server@3.4.4':
+    hash: d8dc66e5ec3b9d23a876b979f493b6aa901fd2d965be54729495da5136296a42
+    path: patches/@hocuspocus__server@3.4.4.patch
   ai@6.0.134:
     hash: f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9
     path: patches/ai@6.0.134.patch
@@ -75,7 +78,7 @@ importers:
         version: 3.4.4(y-protocols@1.0.6(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810)))(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810))
       '@hocuspocus/server':
         specifier: 3.4.4
-        version: 3.4.4(y-protocols@1.0.6(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810)))(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810))
+        version: 3.4.4(patch_hash=d8dc66e5ec3b9d23a876b979f493b6aa901fd2d965be54729495da5136296a42)(y-protocols@1.0.6(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810)))(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810))
       '@hocuspocus/transformer':
         specifier: 3.4.4
         version: 3.4.4(@tiptap/core@3.20.4(@tiptap/pm@3.20.4))(@tiptap/pm@3.20.4)(y-prosemirror@1.3.7(prosemirror-model@1.25.1)(prosemirror-state@1.4.3)(prosemirror-view@1.40.0)(y-protocols@1.0.6(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810)))(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810)))(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810))
@@ -1041,6 +1044,9 @@ importers:
       marked:
         specifier: ^17.0.1
         version: 17.0.5
+      pako:
+        specifier: ^2.0.3
+        version: 2.0.3
       re2:
         specifier: ^1.21.0
         version: 1.25.0
@@ -13055,7 +13061,7 @@ snapshots:
       - bufferutil
       - utf-8-validate
 
-  '@hocuspocus/server@3.4.4(y-protocols@1.0.6(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810)))(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810))':
+  '@hocuspocus/server@3.4.4(patch_hash=d8dc66e5ec3b9d23a876b979f493b6aa901fd2d965be54729495da5136296a42)(y-protocols@1.0.6(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810)))(yjs@13.6.30(patch_hash=1ceeb66dba1f86545c98a3ff7f5152aff9b35caf409091cef9caedb5e65c8810))':
     dependencies:
       '@hocuspocus/common': 3.4.4
       async-lock: 1.4.1