fix(ai-chat): ревью-раунд #500 — красный серверный сьют + регрессия Redis-health

B1: переименование строки ошибки (#394) не обновило 4 предсуществующих
share-ассерта → полный серверный сьют был красный. Обновлены ассерты в
public-share-chat.spec.ts и public-share-chat-tools.service.spec.ts под
новую классифицированную строку.

B2: /health первая проба после старта врала DOWN при живом Redis
(lazyConnect + enableOfflineQueue:false + maxRetriesPerRequest:1 → первый
ping до открытия сокета). Добавлен ensureConnected() с bounded-таймаутом
перед первым ping; покрывает и путь пересоздания после onModuleDestroy.
Тест UP-с-первой-пробы против реального ioredis (mutation-verified).

Устранён open-handle leak в redis.health.spec (drain ioredis
force-destroy-таймера в teardown; без forceExit) и окно ложного DOWN при
конкурентных пробах (мемоизированный connectingPromise).

B3: комментарий про orphan-чат при провале beginRun (insert до begin).
B4: описание listPages упоминает поле truncated в tree-режиме.

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

.env.example

@@ -335,6 +335,20 @@ MCP_DOCMOST_PASSWORD=
 #    VictoriaMetrics/Prometheus reaching it as <host>:<port>/metrics.
 # METRICS_PORT=9464
 #
+#    METRICS_BIND — interface the /metrics listener binds to. DEFAULT 127.0.0.1
+#    (loopback only), so the unauthenticated endpoint is NOT exposed on all
+#    interfaces. If the scraper runs in a SEPARATE container and reaches this as
+#    docmost:9464, set METRICS_BIND=0.0.0.0 — but then also set METRICS_TOKEN
+#    and/or keep the port on a private network, since /metrics is otherwise open.
+# METRICS_BIND=127.0.0.1
+#
+#    METRICS_TOKEN — optional Bearer token guarding /metrics. When set, every
+#    scrape MUST send `Authorization: Bearer <token>` (others get 401). Configure
+#    the scraper with the same bearer token (e.g. VictoriaMetrics/vmagent
+#    `bearer_token`, Prometheus `authorization.credentials`). Leave unset only
+#    when the endpoint is bound to loopback or an otherwise-trusted network.
+# METRICS_TOKEN=
+#
 # 2) CLIENT_TELEMETRY_ENABLED — the public client perf-telemetry sink.
 #    OFF by default. When true, the unauthenticated POST /api/telemetry/vitals
 #    endpoint is registered and browsers collect + send web-vitals / editor

AGENTS.md

@@ -470,7 +470,7 @@ Vite SPA. Code is organized by feature under `apps/client/src/features/*` (mirro
 - **Errors must never be swallowed or shown as generic messages.** Every caught error MUST (1) be logged in full to the console/logger — error name, message, stack, `cause`, and (for HTTP/provider failures) the status code and response body — and (2) be surfaced to the user with a *specific, human-readable explanation of what actually went wrong*, never a bare generic string like "Something went wrong" / "Could not start recording" / "Transcription failed". Include the real reason (the underlying error/provider message) in the user-facing text. On the server, wrap third-party/provider failures with `describeProviderError` (or equivalent) and rethrow as a meaningful HTTP status + message — never let them collapse into an opaque 500. On the client, `console.error(<context>, err)` the raw error AND show the extracted reason (e.g. `err.response?.data?.message`, or the error `name: message`) in the notification.
 - The version string shown in the UI comes from `APP_VERSION` (CI/Docker) or `git describe --tags --always` (local), resolved in `vite.config.ts` — not from `package.json`.
 - Server TS config is permissive (`noImplicitAny: false`, `strictNullChecks: false`, `no-explicit-any` lint disabled). Follow the existing relaxed style rather than tightening types broadly.
-- Dependency versions are heavily pinned via `pnpm.overrides` and `pnpm.patchedDependencies` (`scimmy`, `yjs`, `ai`) in the root `package.json`. Don't bump pinned/patched deps casually; the patches and overrides exist for compatibility/security reasons. The `ai@6.0.134` patch disables the SDK's O(n²) cumulative `partialOutput` accumulation when no output strategy is requested (server heap OOM on long agent runs, #184; tripwire test: `apps/server/src/integrations/ai/ai-sdk-partial-output.patch.spec.ts`) — it MUST be re-created via `pnpm patch` when bumping `ai`.
+- Dependency versions are heavily pinned via `pnpm.overrides` and `pnpm.patchedDependencies` (`scimmy`, `yjs`, `ai`) in the root `package.json`. Don't bump pinned/patched deps casually; the patches and overrides exist for compatibility/security reasons. The `ai@6.0.134` patch carries TWO independent server fixes, each with its own tripwire test: (1) it disables the SDK's O(n²) cumulative `partialOutput` accumulation when no output strategy is requested (server heap OOM on long agent runs, #184; tripwire: `apps/server/src/integrations/ai/ai-sdk-partial-output.patch.spec.ts`); (2) it fixes `writeToServerResponse`'s drain-hang — the loop awaited only `"drain"` under backpressure, so a mid-write client disconnect parked the pipe forever and leaked the reader/buffers until restart; it now races `"drain"` against `"close"`/`"error"`, cancels the reader on disconnect, and swallows the fire-and-forget read rejection (#486; tripwire: `apps/server/src/integrations/ai/ai-sdk-drain-hang.patch.spec.ts`). Both tripwires assert BOTH installed dist builds carry their patch marker. The patch MUST be re-created via `pnpm patch` when bumping `ai`.
 - **The MCP tool inventory in `SERVER_INSTRUCTIONS` is GENERATED from the registry** (`packages/mcp/src/server-instructions.ts`: `buildToolInventory()` over `SHARED_TOOL_SPECS`) and spliced into the hand-written routing prose (`ROUTING_PROSE`). So adding/renaming/removing a **shared** spec in `packages/mcp/src/tool-specs.ts` auto-updates the `<tool_inventory>` — no manual `SERVER_INSTRUCTIONS` edit needed. Only an **inline** MCP-only tool (those registered via `server.registerTool(...)` in `index.ts`, not through the registry) needs a one-line entry in `INLINE_MCP_INVENTORY`. Enforced by `packages/mcp/test/unit/tool-inventory.test.mjs`, which fails when a registered tool is missing from the generated inventory (there is no `EXCEPTIONS` opt-out anymore — every tool must appear). Update `ROUTING_PROSE` when a tool's *intent guidance* (when-to-use) changes. `packages/mcp/build/` is gitignored and rebuilt in CI/Docker via `pnpm build` (same convention as `git-sync`/`prosemirror-markdown`) — never commit it; rebuild locally after editing to run the tests.
 
 ## CI / release

CHANGELOG.md

@@ -115,6 +115,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   the old ProseMirror-JSON output. Released together with the `#411`/`#412`
   breaking window so external configs break exactly once. (#413)
 
+- **The Prometheus `/metrics` listener now binds to `127.0.0.1` (loopback) by
+  default instead of `0.0.0.0` (all interfaces).** This closes an unauthenticated
+  endpoint that was previously reachable on every interface. **DEPLOY MIGRATION —
+  cross-container scraping breaks silently otherwise:** if your scraper runs in a
+  SEPARATE container and reaches the app as `docmost:9464` (the exact topology the
+  old `0.0.0.0` hardcode served), you MUST now set `METRICS_BIND=0.0.0.0` — and,
+  because that re-exposes the endpoint, also set `METRICS_TOKEN=<secret>` and
+  configure the scraper with a matching Bearer token. Without `METRICS_BIND`, the
+  scraper can no longer connect and metrics go dark with no error. See the
+  `METRICS_BIND` / `METRICS_TOKEN` block in `.env.example` for the migration.
+  Same-host (loopback) scrapers need no change. (#486)
+
 ### Added
 
 - **Place several images side by side in a row.** A new "Inline (side by
@@ -310,6 +322,39 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   `tee()` branch of the stream result — a ~20-step, ~28k-chunk agent run
   retained ~1.7 GB and OOM'd the 2 GB JS heap. Streaming granularity is
   unchanged; the patch must be re-created if `ai` is ever bumped. (#184)
+
+- **The server no longer leaks a hung stream pipe on every mid-run client
+  disconnect.** The same `ai@6.0.134` pnpm patch now also fixes the SDK's
+  `writeToServerResponse`, which awaited only a `"drain"` event under
+  backpressure: when a client disconnected mid-write the socket never drained, so
+  the write loop parked forever, `response.end()` was unreachable, and the stream
+  reader plus buffered chunks were pinned until process restart (every mid-run
+  disconnect in autonomous mode leaked one). The patch races `"drain"` against
+  `"close"`/`"error"`, cancels the reader and ends the response on disconnect, and
+  swallows the fire-and-forget read rejection instead of crashing on an
+  unhandledRejection. (#486)
+
+- **A failed autonomous agent-run start no longer becomes an unstoppable ghost
+  run.** When `beginRun` failed for a transient reason (e.g. a DB-pool blip),
+  the turn previously continued with NO run row — invisible to `/stop`, not
+  aborted on disconnect, and able to slip a second run past the one-run-per-chat
+  gate, leaving an unstoppable run until restart. The turn now fails fast with an
+  honest `503 A_RUN_BEGIN_FAILED` before the first byte (no orphan state), and the
+  client shows a "temporary — please try again" message instead of a misleading
+  "provider not configured". (#486)
+
+- **A pathological draw.io graph can no longer wedge the whole server.** The ELK
+  auto-layout (`layout:"elk"`) ran elkjs synchronously on the main event loop, so
+  a graph at the node/edge cap blocked ALL HTTP/SSE/loopback traffic while it
+  churned — and the old `setTimeout` "timeout" could never fire because the same
+  thread was blocked. Layout now runs in a worker thread with the timeout enforced
+  by `worker.terminate()`; the main loop stays responsive. (#486)
+
+- **The `/health` Redis probe no longer leaks a client on every tick while Redis
+  is down.** It built a new `ioredis` client per probe and disconnected it only on
+  success, so during an outage each health tick added another forever-reconnecting
+  client (an unbounded handle leak). A single long-lived probe client is now
+  reused and closed on shutdown. (#486)
 - **Internal links in exported Markdown no longer lose their visible text.** A
   link whose target page name had no file extension (e.g. a bare title) was
   collapsed to empty text during export, producing an unclickable, label-less
@@ -386,6 +431,24 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   share); any other value now returns the generic "not found" instead of
   serving the page. (#218)
 
+- **Tool and provider error text no longer leaks to anonymous readers in the
+  public-share AI chat.** A failing tool's raw error (which could carry an
+  internal page title or a stack fragment) and a provider error (which bundles the
+  provider `statusCode` and response body — potentially the internal baseUrl or
+  model name) were streamed verbatim to the anonymous reader over SSE. Errors are
+  now sanitized at the source: the share toolset collapses any unclassified tool
+  error to a safe generic string (safe, classified tool messages still pass
+  through for the model's self-correction), and the anonymous stream `onError`
+  maps provider failures to a fixed set of neutral strings — the full detail goes
+  only to the server log. A UI render gate is layered on top. (closes #394)
+
+- **The Prometheus `/metrics` endpoint can now require Bearer authentication and
+  is loopback-bound by default.** Previously it listened on all interfaces with no
+  auth. Setting `METRICS_TOKEN` requires every scrape to present
+  `Authorization: Bearer <token>` (compared in constant time), and the listener
+  defaults to `127.0.0.1` (see the Breaking Changes entry for the cross-container
+  migration). (#486)
+
 ## [0.94.0] - 2026-06-26
 
 This release makes AI chat durable and fast: assistant turns are persisted to

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

@@ -203,6 +203,52 @@ describe("ChatThread — send now (#198)", () => {
   });
 });
 
+// #486: the final onFinish -> flushNext() must be gated on the live-mount flag.
+// A clean onFinish can land AFTER the thread unmounts (New-chat / chat-switch
+// mid-stream — the async attach/resume settles late); flushing then dequeues and
+// re-POSTs a queued message from an abandoned thread (a "ghost" send).
+describe("ChatThread — onFinish flush gated on mount (#486)", () => {
+  beforeEach(resetState);
+  afterEach(cleanup);
+
+  it("a clean onFinish WHILE MOUNTED flushes the queued message (control)", () => {
+    renderThread();
+    fireEvent.click(screen.getByTestId("queue-btn")); // enqueue "queued text"
+    expect(h.state.sendMessage).not.toHaveBeenCalled();
+
+    act(() => {
+      h.state.onFinish?.({
+        message: { id: "a", role: "assistant", parts: [] },
+        isAbort: false,
+        isDisconnect: false,
+        isError: false,
+      });
+    });
+    // Mounted: the queue flushes normally.
+    expect(h.state.sendMessage).toHaveBeenCalledWith({ text: "queued text" });
+  });
+
+  it("a clean onFinish AFTER unmount does NOT flush (no ghost send)", () => {
+    const { unmount } = renderThread();
+    fireEvent.click(screen.getByTestId("queue-btn")); // enqueue "queued text"
+    h.state.sendMessage.mockClear();
+
+    // Chat switched away mid-stream: the streamer unmounts...
+    unmount();
+    // ...and a late, clean onFinish lands on the abandoned thread.
+    act(() => {
+      h.state.onFinish?.({
+        message: { id: "a", role: "assistant", parts: [] },
+        isAbort: false,
+        isDisconnect: false,
+        isError: false,
+      });
+    });
+    // Gated on mountedRef: NOTHING is sent from the dead thread.
+    expect(h.state.sendMessage).not.toHaveBeenCalled();
+  });
+});
+
 // #396: in autonomous mode a live sendNow must additionally request the
 // AUTHORITATIVE server stop of the detached run (a local abort is only a client
 // disconnect the server ignores) and arm a bounded 409 retry so the re-POST

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

@@ -659,7 +659,13 @@ export default function ChatThread({
         return;
       }
       if (isAbort || isDisconnect || isError) return;
-      flushNext();
+      // Gate the final flush on the live-mount flag (#486): a clean onFinish can
+      // land AFTER this thread unmounted (a New-chat / chat-switch mid-stream —
+      // the async attach/resume settles late). Flushing then dequeues and POSTs a
+      // queued message from an abandoned thread — a "ghost" send / ghost chat.
+      // Every other queue side effect already guards on mountedRef; this last one
+      // was the gap.
+      if (mountedRef.current) flushNext();
     },
     // `onError` runs in addition to `onFinish` (which ai@6 also calls on error).
     // Log the raw failure here for devtools; the UI shows a friendly classified

apps/client/src/features/ai-chat/components/message-item.tsx

@@ -47,6 +47,13 @@ interface MessageItemProps {
    * agent's raw query/argument text.
    */
   showInput?: boolean;
+  /**
+   * Forwarded to ToolCallCard: whether a failed tool card renders its raw
+   * errorText. Defaults to true (internal chat). The public share passes false so
+   * internal detail in a tool error is never painted (belt to the server-side
+   * byte sanitization).
+   */
+  showErrors?: boolean;
   /**
    * Neutralize internal/relative markdown links in the rendered answer (drop
    * their href so they become inert text). Defaults to false (internal chat,
@@ -125,6 +132,7 @@ function MessageItem({
   message,
   showCitations = true,
   showInput = true,
+  showErrors = true,
   neutralizeInternalLinks = false,
   assistantName,
   turnStreaming = false,
@@ -219,6 +227,7 @@ function MessageItem({
               part={part as unknown as ToolUiPart}
               showCitations={showCitations}
               showInput={showInput}
+              showErrors={showErrors}
             />
           );
         }
@@ -284,6 +293,7 @@ export function arePropsEqual(
     prev.signature === next.signature &&
     prev.showCitations === next.showCitations &&
     prev.showInput === next.showInput &&
+    prev.showErrors === next.showErrors &&
     prev.neutralizeInternalLinks === next.neutralizeInternalLinks &&
     prev.assistantName === next.assistantName &&
     // The turn-end flip re-renders every row once (cheap, terminal event) —

apps/client/src/features/ai-chat/components/message-list.tsx

@@ -32,6 +32,12 @@ interface MessageListProps {
    * doesn't see the agent's raw query/argument text.
    */
   showInput?: boolean;
+  /**
+   * Forwarded to MessageItem -> ToolCallCard: whether a failed tool card renders
+   * its raw errorText. Defaults to true (internal chat). The public share passes
+   * false so internal detail in a tool error is never painted.
+   */
+  showErrors?: boolean;
   /**
    * Forwarded to MessageItem: neutralize internal/relative markdown links in
    * the rendered answers (drop their href so they render as inert text).
@@ -127,6 +133,7 @@ export default function MessageList({
   emptyState,
   showCitations = true,
   showInput = true,
+  showErrors = true,
   neutralizeInternalLinks = false,
   assistantName,
 }: MessageListProps) {
@@ -217,6 +224,7 @@ export default function MessageList({
             signature={messageSignature(message)}
             showCitations={showCitations}
             showInput={showInput}
+            showErrors={showErrors}
             neutralizeInternalLinks={neutralizeInternalLinks}
             assistantName={assistantName}
             // Turn-level liveness, gated to the TAIL row: only the tail message

apps/client/src/features/ai-chat/components/tool-call-card.tsx

@@ -30,6 +30,16 @@ interface ToolCallCardProps {
    * the extra summary line, leaving the card (the action log) intact.
    */
   showInput?: boolean;
+  /**
+   * Whether to render the tool's raw errorText on a failed call. Defaults to true
+   * (the internal chat, where the operator may debug). The public share passes
+   * false: a tool error string can carry internal detail (an internal page title,
+   * a stack fragment, a provider message). This is the RENDER gate only — the
+   * authoritative fix also sanitizes the bytes server-side (see
+   * PublicShareChatToolsService.forShare), so a share reader never receives raw
+   * error text over the wire, not just never sees it painted (#394).
+   */
+  showErrors?: boolean;
 }
 
 /**
@@ -41,6 +51,7 @@ export default function ToolCallCard({
   part,
   showCitations = true,
   showInput = true,
+  showErrors = true,
 }: ToolCallCardProps) {
   const { t } = useTranslation();
   const toolName = getToolName(part);
@@ -74,7 +85,7 @@ export default function ToolCallCard({
         </Text>
       )}
 
-      {state === "error" && part.errorText && (
+      {state === "error" && showErrors && part.errorText && (
         <Text size="xs" c="red" mt={2}>
           {part.errorText}
         </Text>

apps/client/src/features/ai-chat/utils/error-message.test.ts

@@ -23,6 +23,25 @@ describe("describeChatError", () => {
     });
   });
 
+  it("classifies an A_RUN_BEGIN_FAILED 503 as a temporary run-start failure, NOT provider-not-configured (#486)", () => {
+    // The FULL real body the server writes for a beginRun failure: a
+    // ServiceUnavailableException(object) whose response is serialized verbatim
+    // onto the raw socket, self-describing statusCode 503 + the run-start code.
+    const body =
+      '{"message":"Could not start the agent run. This is usually temporary — please try again.","code":"A_RUN_BEGIN_FAILED","statusCode":503}';
+    expect(describeChatError(body, t)).toEqual({
+      title: "Could not start the run",
+      detail:
+        "The agent run could not be started. This is usually temporary — please try again.",
+    });
+    // ORDER GUARD: even though the body ALSO carries statusCode 503 (which the
+    // generic branch matches), the A_RUN_BEGIN_FAILED branch runs first, so it is
+    // never mislabeled "AI provider not configured".
+    expect(describeChatError(body, t).title).not.toBe(
+      "AI provider not configured",
+    );
+  });
+
   it("classifies a dropped connection (ECONNRESET) as a lost-connection error", () => {
     expect(
       describeChatError("Cannot connect to API: read ECONNRESET", t).title,

apps/client/src/features/ai-chat/utils/error-message.ts

@@ -24,6 +24,21 @@ export function describeChatError(
 ): ChatErrorView {
   const msg = message ?? "";
 
+  // Our own "could not start the run" gate (A_RUN_BEGIN_FAILED, #486): a 503
+  // whose body carries this code is a TEMPORARY server-side failure while
+  // starting the run (e.g. a DB-pool blip), NOT an unconfigured provider. It MUST
+  // be matched STRICTLY BEFORE the generic 503 branch below, which would
+  // otherwise mislabel it "The AI provider is not configured" and tell the user
+  // to call an admin instead of just retrying.
+  if (/"code"\s*:\s*"A_RUN_BEGIN_FAILED"/.test(msg)) {
+    return {
+      title: t("Could not start the run"),
+      detail: t(
+        "The agent run could not be started. This is usually temporary — please try again.",
+      ),
+    };
+  }
+
   if (/"statusCode"\s*:\s*403\b/.test(msg)) {
     return {
       title: t("AI chat is disabled"),

apps/client/src/features/share/components/share-ai-widget.tsx

@@ -168,6 +168,10 @@ export default function ShareAiWidget({
             // Anonymous reader: suppress the tool-argument summary line so the
             // agent's raw query/argument text isn't shown on the public share.
             showInput={false}
+            // Anonymous reader: never paint a tool's raw errorText (it can carry
+            // internal detail). This is the render gate; the bytes are also
+            // sanitized server-side in PublicShareChatToolsService.forShare (#394).
+            showErrors={false}
             // Anonymous reader: neutralize internal/relative links in the
             // assistant's markdown so internal UUIDs/auth-gated routes don't
             // leak as clickable links (external http(s) links are kept).

apps/server/src/core/ai-chat/ai-chat.begin-run-failure.spec.ts

@@ -0,0 +1,109 @@
+import {
+  ConflictException,
+  Logger,
+  ServiceUnavailableException,
+} from '@nestjs/common';
+import { AiChatService } from './ai-chat.service';
+import { RunAlreadyActiveError } from './ai-chat-run.service';
+
+/**
+ * Fail-fast guard for beginRun failures (#486, commit 4).
+ *
+ * When runHooks.begin() rejects for a reason OTHER than RunAlreadyActiveError
+ * (e.g. a DB-pool blip), the turn must NOT continue untracked. The old code
+ * logged and streamed anyway, leaving a run with NO run-row: in autonomous mode
+ * nobody could abort it (/stop can't see it, disconnect doesn't abort it, and the
+ * one-run gate would admit a SECOND run) — an unstoppable invisible run until
+ * restart. The fix throws A_RUN_BEGIN_FAILED (503) BEFORE the first byte and
+ * before the user row is persisted.
+ *
+ * We drive `stream()` directly on a prototype instance wired with only the
+ * collaborators it touches before the throw, so the assertion is on the REAL
+ * control flow, not a mock of it.
+ */
+describe('AiChatService beginRun failure (#486)', () => {
+  function makeService(insertSpy: jest.Mock): AiChatService {
+    // Bypass the (heavy) DI constructor: exercise the real stream() method on a
+    // bare prototype instance with just the fields reached before the throw.
+    // `any` because the private `logger` field makes a typed intersection collapse.
+    const svc = Object.create(AiChatService.prototype);
+    svc.aiChatRepo = {
+      // Existing chat -> no insert path; chatId is kept as-is.
+      findById: jest.fn().mockResolvedValue({ id: 'chat1' }),
+    };
+    svc.aiChatMessageRepo = { insert: insertSpy };
+    svc.logger = new Logger('test');
+    return svc as AiChatService;
+  }
+
+  const baseArgs = () => {
+    const write = jest.fn();
+    const res = {
+      raw: { write, writableEnded: false, headersSent: false },
+    };
+    return {
+      user: { id: 'u1' } as never,
+      workspace: { id: 'w1' } as never,
+      sessionId: 's1',
+      // openPage undefined -> resolveOpenPageContext returns null without any DB
+      // call; chatId present -> the existing-chat path.
+      body: { chatId: 'chat1', messages: [] } as never,
+      res: res as never,
+      signal: new AbortController().signal,
+      model: {} as never,
+      role: null,
+      write,
+    };
+  };
+
+  it('throws A_RUN_BEGIN_FAILED (503) before the first byte and before persisting the user turn', async () => {
+    const insertSpy = jest.fn();
+    const svc = makeService(insertSpy);
+    const { write, ...args } = baseArgs();
+
+    const runHooks = {
+      begin: jest.fn().mockRejectedValue(new Error('DB pool exhausted')),
+    } as never;
+
+    let caught: unknown;
+    try {
+      await svc.stream({ ...args, runHooks });
+    } catch (e) {
+      caught = e;
+    }
+
+    expect(caught).toBeInstanceOf(ServiceUnavailableException);
+    const http = caught as ServiceUnavailableException;
+    expect(http.getStatus()).toBe(503);
+    expect(http.getResponse()).toMatchObject({ code: 'A_RUN_BEGIN_FAILED' });
+
+    // Fail-fast: nothing was written to the socket and NO user message row was
+    // persisted, so the turn left no orphan state to clean up.
+    expect(write).not.toHaveBeenCalled();
+    expect(insertSpy).not.toHaveBeenCalled();
+  });
+
+  it('still maps a lost-the-race RunAlreadyActiveError to a 409, not A_RUN_BEGIN_FAILED', async () => {
+    const insertSpy = jest.fn();
+    const svc = makeService(insertSpy);
+    const { write, ...args } = baseArgs();
+
+    const runHooks = {
+      begin: jest.fn().mockRejectedValue(new RunAlreadyActiveError('chat1')),
+    } as never;
+
+    let caught: unknown;
+    try {
+      await svc.stream({ ...args, runHooks });
+    } catch (e) {
+      caught = e;
+    }
+
+    expect(caught).toBeInstanceOf(ConflictException);
+    expect((caught as ConflictException).getResponse()).toMatchObject({
+      code: 'A_RUN_ALREADY_ACTIVE',
+    });
+    expect(write).not.toHaveBeenCalled();
+    expect(insertSpy).not.toHaveBeenCalled();
+  });
+});

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

@@ -1,4 +1,8 @@
-import { ConflictException, Logger } from '@nestjs/common';
+import {
+  ConflictException,
+  Logger,
+  ServiceUnavailableException,
+} from '@nestjs/common';
 
 // Mock the AI SDK so we can PROVE no provider call is made for the turn we are
 // about to reject. The race rejection happens at runHooks.begin(), long before
@@ -360,22 +364,22 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
 });
 
 /**
- * F14 — the begin-failure RESILIENCE branch (the `else` of the run-race guard).
+ * F14 — the begin-failure branch (the `else` of the run-race guard).
  *
  * stream() wraps runHooks.begin in try/catch with TWO branches:
  *   - RunAlreadyActiveError  -> 409 ConflictException (pinned above).
- *   - ANY OTHER begin failure -> SWALLOW + continue UNTRACKED on the socket signal
+ *   - ANY OTHER begin failure -> throw ServiceUnavailableException(A_RUN_BEGIN_FAILED)
- *     (legacy fallback): it logs "...streaming without run tracking", leaves
+ *     BEFORE the first byte (#486, commit 4).
- *     `effectiveSignal = signal` (runId undefined) and serves the turn anyway.
  *
- * The contract: a transient beginRun failure (e.g. a non-unique DB error inserting
+ * POLICY CHANGE (#486): the OLD contract here was "SWALLOW + stream the turn
- * the run row) must STILL serve the user's turn — it must NOT re-throw and must NOT
+ * UNTRACKED on the socket signal". That was reversed: an untracked run is
- * be misclassified as a 409. A regression that re-threw here would break EVERY turn
+ * invisible to /stop, is not aborted on disconnect, and slips past the one-run
- * on a begin failure with nothing to catch it. This branch is otherwise undriven by
+ * gate — an unstoppable ghost run in autonomous mode. Now a plain begin failure
- * any spec, so it is pinned here SEPARATELY from the 409 path: a plain begin error
+ * FAILS the turn fast with a 503 A_RUN_BEGIN_FAILED, before any user row is
- * proceeds to streamText with the SOCKET signal and still persists the user turn.
+ * persisted and before streamText runs. This case is INVERTED (not deleted) so
+ * the "plain begin failure" path stays explicitly pinned under the new policy.
  */
-describe('AiChatService.stream — begin-failure resilience / legacy fallback (#184 F14)', () => {
+describe('AiChatService.stream — begin-failure fails the turn (#184 F14 / #486)', () => {
   const streamTextMock = streamText as unknown as jest.Mock;
 
   function makeStreamResult() {
@@ -455,7 +459,7 @@ describe('AiChatService.stream — begin-failure resilience / legacy fallback (#
 
   afterEach(() => jest.restoreAllMocks());
 
-  it('a PLAIN begin() failure (NOT RunAlreadyActiveError) does NOT 409 — it swallows, logs, and streams the turn UNTRACKED on the socket signal', async () => {
+  it('a PLAIN begin() failure (NOT RunAlreadyActiveError) FAILS the turn with a 503 A_RUN_BEGIN_FAILED before the first byte — NO untracked stream (#486)', async () => {
     const errorSpy = jest
       .spyOn(Logger.prototype, 'error')
       .mockImplementation(() => undefined as never);
@@ -487,28 +491,26 @@ describe('AiChatService.stream — begin-failure resilience / legacy fallback (#
       } as never,
     });
 
-    // The turn proceeds: NO throw at all (in particular NOT a 409).
+    // NEW POLICY: the turn is REJECTED with a 503 A_RUN_BEGIN_FAILED (not a 409,
-    await expect(promise).resolves.toBeUndefined();
+    // and NOT swallowed into an untracked stream).
+    await expect(promise).rejects.toBeInstanceOf(ServiceUnavailableException);
+    const err = (await promise.catch(
+      (e) => e,
+    )) as ServiceUnavailableException;
+    expect(err.getStatus()).toBe(503);
+    expect(err.getResponse()).toMatchObject({ code: 'A_RUN_BEGIN_FAILED' });
 
     expect(begin).toHaveBeenCalledTimes(1);
 
-    // The resilience branch logged the legacy-fallback warning.
+    // It logged the fail-the-turn line.
     expect(errorSpy).toHaveBeenCalledWith(
-      expect.stringContaining('streaming without run tracking'),
+      expect.stringContaining('failing the turn'),
       expect.anything(),
     );
 
-    // The turn really streamed: the user message was persisted and streamText ran.
+    // Fail-fast: the turn NEVER streamed — no user row persisted, no streamText
-    expect(aiChatMessageRepo.insert).toHaveBeenCalled();
+    // call, so no orphan/untracked run was left behind.
-    expect(streamTextMock).toHaveBeenCalledTimes(1);
+    expect(aiChatMessageRepo.insert).not.toHaveBeenCalled();
-
+    expect(streamTextMock).not.toHaveBeenCalled();
-    // The decisive wiring: with no run handle, the fallback uses the SOCKET signal
-    // (effectiveSignal = signal, runId undefined) — not a run-bound signal. #444:
-    // the signal is unioned with the degeneration controller via AbortSignal.any,
-    // so assert the socket abort still reaches the turn rather than identity.
-    const passed = streamTextMock.mock.calls[0][0].abortSignal as AbortSignal;
-    expect(passed.aborted).toBe(false);
-    socketController.abort();
-    expect(passed.aborted).toBe(true);
   });
 });

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

@@ -4,6 +4,7 @@ import {
   Injectable,
   Logger,
   OnModuleInit,
+  ServiceUnavailableException,
 } from '@nestjs/common';
 import { FastifyReply } from 'fastify';
 import {
@@ -55,6 +56,7 @@ import {
 import {
   isDegenerateOutput,
   truncateDegeneratedTail,
+  shouldCheckDegeneration,
 } from './output-degeneration';
 
 // Max agent steps per turn. One step = one model generation; a step that calls
@@ -756,6 +758,13 @@ export class AiChatService implements OnModuleInit {
       // or violate the page_id FK on insert (this runs after res.hijack(), so a
       // DB error would break the stream).
       const originPageId: string | null = openPageContext?.id ?? null;
+      // ORPHAN-ON-BEGIN-FAILURE tradeoff (#486, B3): the chat row is inserted
+      // HERE, before runHooks.begin below. If begin fails (e.g. a 503 / run-slot
+      // rejection) the turn aborts before the client is told this new chatId, so
+      // an empty chat is left behind and a retry mints ANOTHER one. We accept this
+      // over reordering: begin needs a chatId to bind the run to, and inserting
+      // the chat first keeps the id stable + the FK/history-join invariants above
+      // intact. Orphan empty chats are cheap and swept by normal chat cleanup.
       const chat = await this.aiChatRepo.insert({
         creatorId: user.id,
         workspaceId: workspace.id,
@@ -797,12 +806,32 @@ export class AiChatService implements OnModuleInit {
             code: 'A_RUN_ALREADY_ACTIVE',
           });
         }
-        // Any OTHER run-start failure must not break the turn — fall back to the
+        // Any OTHER run-start failure (e.g. a DB-pool blip) must FAIL THE TURN,
-        // socket signal (legacy behavior) and stream anyway.
+        // not silently stream without a run-row. The old fallback let the turn
+        // continue untracked: in autonomous mode nobody could then abort it —
+        // /stop can't see a run that doesn't exist, a client disconnect doesn't
+        // abort it, and the one-run-per-chat gate would let a SECOND run in. That
+        // is an unstoppable, invisible run until process restart. Reject NOW,
+        // BEFORE the first byte (nothing is written yet, no user row inserted, no
+        // MCP lease taken), so the controller's post-hijack catch turns this
+        // HttpException into an honest 503 on the raw socket. Same policy for BOTH
+        // modes — #487 inherits it (no mode-branching here).
         this.logger.error(
-          `Failed to begin agent run (chat ${chatId}); streaming without run tracking`,
+          `Failed to begin agent run (chat ${chatId}); failing the turn`,
           err as Error,
         );
+        throw new ServiceUnavailableException({
+          message:
+            'Could not start the agent run. This is usually temporary — please try again.',
+          code: 'A_RUN_BEGIN_FAILED',
+          // Self-describe the status in the body: the controller's post-hijack
+          // catch writes getResponse() verbatim onto the raw socket, and an
+          // object-arg HttpException does NOT inject statusCode. Without it the
+          // client's 503 classifier (which reads the body JSON) could not see the
+          // status. With it present, the client's A_RUN_BEGIN_FAILED branch (which
+          // runs strictly before the generic-503 branch) shows "temporary, retry".
+          statusCode: 503,
+        });
       }
     }
 
@@ -1080,7 +1109,6 @@ export class AiChatService implements OnModuleInit {
       const degenerationController = new AbortController();
       let degenerationDetected = false;
       let lastDegenerationCheckLen = 0;
-      const DEGENERATION_CHECK_STEP = 2000;
 
       // Step-granular durability (#183): create the assistant row UPFRONT in the
       // 'streaming' state (before any token), then UPDATE it as each step finishes
@@ -1254,8 +1282,10 @@ export class AiChatService implements OnModuleInit {
               // trigger, abort the run ONCE with a distinguishable reason.
               if (
                 !degenerationDetected &&
-                inProgressText.length - lastDegenerationCheckLen >=
+                shouldCheckDegeneration(
-                  DEGENERATION_CHECK_STEP
+                  inProgressText.length,
+                  lastDegenerationCheckLen,
+                )
               ) {
                 lastDegenerationCheckLen = inProgressText.length;
                 if (isDegenerateOutput(inProgressText)) {
@@ -1275,6 +1305,13 @@ export class AiChatService implements OnModuleInit {
             // the in-progress accumulator for the next step.
             capturedSteps.push(step as StepLike);
             inProgressText = '';
+            // Reset the degeneration-check watermark too (#486): it tracks a byte
+            // offset INTO inProgressText, so once that resets to '' a stale (large)
+            // mark makes `inProgressText.length - lastDegenerationCheckLen` go
+            // negative and the throttled detector stays silent until a later step's
+            // text re-grows past the old offset — a whole degenerate step could slip
+            // through undetected. Zeroing it re-arms the check from the next byte.
+            lastDegenerationCheckLen = 0;
             // Step-granular durability (#183): persist this finished step (its text +
             // tool calls + tool RESULTS) the moment it ends, so a process death after
             // this point still recovers the step. Not awaited here (never block the

apps/server/src/core/ai-chat/output-degeneration.spec.ts

@@ -1,11 +1,24 @@
+import { Logger } from '@nestjs/common';
+import { streamText } from 'ai';
 import {
   hasRepeatedLineRun,
   hasPeriodicTail,
   isDegenerateOutput,
   truncateDegeneratedTail,
+  shouldCheckDegeneration,
+  DEGENERATION_CHECK_STEP,
   REPEATED_LINES_THRESHOLD,
   MIN_PERIOD_REPEATS,
 } from './output-degeneration';
+import { AiChatService } from './ai-chat.service';
+
+// Mock ONLY streamText so we can capture the onChunk/onStepFinish callbacks the
+// service registers and drive them by hand; every other `ai` export the service
+// uses (convertToModelMessages, stepCountIs, …) stays real.
+jest.mock('ai', () => {
+  const actual = jest.requireActual('ai');
+  return { ...actual, streamText: jest.fn() };
+});
 
 /**
  * Unit tests for the token-degeneration detector (#444) — the sole anti-babble
@@ -180,3 +193,188 @@ describe('truncateDegeneratedTail', () => {
     expect(truncateDegeneratedTail(text)).toBe(text);
   });
 });
+
+/**
+ * Throttle + step-boundary reset (#486). The stream keeps a watermark
+ * (`lastDegenerationCheckLen`) that is an OFFSET into the accumulated step text.
+ * On a step boundary the accumulator resets to '', so the watermark MUST reset to
+ * 0 too — otherwise the throttle goes silent for the whole next step. These tests
+ * pin the pure decision AND the reset property that ai-chat.service.onStepFinish
+ * now enforces.
+ */
+describe('shouldCheckDegeneration (throttle) + step-boundary reset (#486)', () => {
+  it('fires once the text grows a full DEGENERATION_CHECK_STEP past the mark', () => {
+    expect(shouldCheckDegeneration(DEGENERATION_CHECK_STEP, 0)).toBe(true);
+    expect(shouldCheckDegeneration(DEGENERATION_CHECK_STEP - 1, 0)).toBe(false);
+    expect(shouldCheckDegeneration(5000, 3000)).toBe(true); // grew 2000 since mark
+    expect(shouldCheckDegeneration(4000, 3000)).toBe(false); // grew only 1000
+  });
+
+  it('BUG (no reset): a stale large watermark silences the next step', () => {
+    // End of a long step: the watermark sits at 5000. The step ends and the
+    // accumulator resets to '' — but if the watermark is NOT reset, a fresh short
+    // degenerate burst (length 2000) never triggers a check: 2000 - 5000 < STEP.
+    const staleWatermark = 5000;
+    const nextStepLen = DEGENERATION_CHECK_STEP; // a fresh 2KB burst
+    expect(shouldCheckDegeneration(nextStepLen, staleWatermark)).toBe(false);
+  });
+
+  it('FIX (reset to 0): the same short degenerate burst IS checked and detected', () => {
+    // onStepFinish now zeroes the watermark, so the fresh burst re-arms the check.
+    const resetWatermark = 0;
+    const degenerateBurst = 'loadTools.\n'.repeat(300); // real degeneration
+    expect(degenerateBurst.length).toBeGreaterThanOrEqual(DEGENERATION_CHECK_STEP);
+    // The throttle now fires...
+    expect(
+      shouldCheckDegeneration(degenerateBurst.length, resetWatermark),
+    ).toBe(true);
+    // ...and the detector catches the loop that would otherwise stream unchecked.
+    expect(isDegenerateOutput(degenerateBurst)).toBe(true);
+  });
+});
+
+/**
+ * BEHAVIOR guard for the ACTUAL fix (#486, ai-chat.service.onStepFinish resets
+ * lastDegenerationCheckLen to 0). The pure tests above use a hard-coded
+ * resetWatermark, so a REVERT of the real `lastDegenerationCheckLen = 0` line
+ * would not redden any of them. This drives the REAL onChunk/onStepFinish
+ * closures from stream() end to end and asserts the run is aborted when a fresh
+ * degenerate burst arrives in the step AFTER a long clean step — which only
+ * happens if the watermark was actually zeroed on the step boundary.
+ */
+describe('AiChatService: onStepFinish re-arms the degeneration watermark (#486)', () => {
+  const streamTextMock = streamText as unknown as jest.Mock;
+
+  function makeRes() {
+    return {
+      raw: {
+        writeHead: jest.fn(),
+        write: jest.fn(),
+        once: jest.fn(),
+        on: jest.fn(),
+        flushHeaders: jest.fn(),
+        writableEnded: false,
+        destroyed: false,
+      },
+    };
+  }
+
+  function makeService() {
+    const aiChatRepo = {
+      findById: jest.fn(async () => ({ id: 'chat-1', workspaceId: 'ws-1' })),
+      insert: jest.fn(),
+    };
+    const aiChatMessageRepo = {
+      insert: jest.fn(async () => ({ id: 'msg-1' })),
+      findAllByChat: jest.fn(async () => []),
+      update: jest.fn(async () => ({ id: 'msg-1' })),
+    };
+    const aiSettings = { resolve: jest.fn(async () => ({})) };
+    const tools = { forUser: jest.fn(async () => ({})) };
+    const mcpClients = {
+      toolsFor: jest.fn(async () => ({
+        tools: {},
+        clients: [],
+        outcomes: [],
+        instructions: [],
+      })),
+    };
+    return new AiChatService(
+      {} as never, // ai
+      aiChatRepo as never,
+      aiChatMessageRepo as never,
+      {} as never, // aiChatPageSnapshotRepo
+      aiSettings as never,
+      tools as never,
+      mcpClients as never,
+      {} as never, // aiAgentRoleRepo
+      {} as never, // pageRepo
+      {} as never, // pageAccess
+      {
+        isAiChatDeferredToolsEnabled: () => false,
+        // Lockdown OFF -> the degeneration guard is the active anti-babble path.
+        isAiChatFinalStepLockdownEnabled: () => false,
+      } as never, // environment
+    );
+  }
+
+  beforeEach(() => {
+    streamTextMock.mockReset();
+    jest.spyOn(Logger.prototype, 'log').mockImplementation(() => undefined as never);
+    jest.spyOn(Logger.prototype, 'warn').mockImplementation(() => undefined as never);
+  });
+
+  afterEach(() => jest.restoreAllMocks());
+
+  it('aborts on a fresh degenerate burst in the NEXT step (reverting the reset line reddens this)', async () => {
+    let captured:
+      | {
+          onChunk?: (e: { chunk: { type: string; text: string } }) => void;
+          onStepFinish?: (step: unknown) => void;
+          abortSignal?: AbortSignal;
+        }
+      | undefined;
+    streamTextMock.mockImplementation((opts: never) => {
+      captured = opts;
+      return {
+        consumeStream: jest.fn(),
+        pipeUIMessageStreamToResponse: jest.fn(),
+      };
+    });
+
+    const svc = makeService();
+    await svc.stream({
+      user: { id: 'user-1' } as never,
+      workspace: { id: 'ws-1' } as never,
+      sessionId: 'sess-1',
+      body: {
+        chatId: 'chat-1',
+        messages: [
+          { id: 'm1', role: 'user', parts: [{ type: 'text', text: 'hi' }] },
+        ],
+      } as never,
+      res: makeRes() as never,
+      signal: new AbortController().signal,
+      model: {} as never,
+      role: null,
+      // No runHooks -> legacy path (socket signal), degeneration guard active.
+    });
+
+    expect(streamTextMock).toHaveBeenCalledTimes(1);
+    const onChunk = captured!.onChunk!;
+    const onStepFinish = captured!.onStepFinish!;
+    const abortSignal = captured!.abortSignal!;
+    expect(abortSignal.aborted).toBe(false);
+
+    // STEP 1: a LONG, non-degenerate first step. Distinct lines never trip the
+    // detector, but they advance the throttle watermark far past the burst size
+    // that follows (to ~5x the step). This is the stale watermark that, WITHOUT
+    // the reset, would silence step 2.
+    let counter = 0;
+    let accumulated = 0;
+    while (accumulated < DEGENERATION_CHECK_STEP * 5) {
+      const line = `unique clean line number ${counter++} with distinct words\n`;
+      accumulated += line.length;
+      onChunk({ chunk: { type: 'text-delta', text: line } });
+    }
+    expect(abortSignal.aborted).toBe(false); // clean step must not abort
+
+    // STEP BOUNDARY: the real onStepFinish resets inProgressText AND (the fix)
+    // zeroes lastDegenerationCheckLen.
+    onStepFinish({ text: 'a clean first step', toolCalls: [], toolResults: [] });
+
+    // STEP 2: a FRESH, short degenerate burst (~3.3KB). Its length is far below
+    // the step-1 stale watermark (~10KB), so WITHOUT the reset the throttle stays
+    // silent and this streams unchecked. WITH the reset (watermark 0) it re-arms,
+    // the detector fires, and the run aborts.
+    const burst = 'loadTools.\n'.repeat(300);
+    expect(burst.length).toBeGreaterThanOrEqual(DEGENERATION_CHECK_STEP);
+    expect(burst.length).toBeLessThan(DEGENERATION_CHECK_STEP * 5);
+    onChunk({ chunk: { type: 'text-delta', text: burst } });
+
+    // The decisive assertion: the composed abortSignal (unioned with the
+    // degeneration controller) is now aborted. Reverting `lastDegenerationCheckLen
+    // = 0` in onStepFinish makes this stay false.
+    expect(abortSignal.aborted).toBe(true);
+  });
+});

apps/server/src/core/ai-chat/output-degeneration.ts

@@ -131,6 +131,32 @@ export function isDegenerateOutput(text: string): boolean {
   return hasRepeatedLineRun(text) || hasPeriodicTail(text);
 }
 
+/**
+ * How many bytes the in-progress text must grow before the (amortized) tail
+ * heuristics are re-run. Shared with ai-chat.service so the throttle the stream
+ * applies is the SAME one the unit test drives.
+ */
+export const DEGENERATION_CHECK_STEP = 2000;
+
+/**
+ * Throttle decision for the degeneration guard (#444/#486). Returns true when
+ * the accumulated text has grown at least DEGENERATION_CHECK_STEP bytes past the
+ * last-checked offset, so the pure rules only fire every ~2KB. Pure; the caller
+ * updates its watermark to `textLen` when this returns true.
+ *
+ * The watermark is an offset INTO the accumulator, so when the accumulator is
+ * reset to '' on a step boundary the caller MUST reset the watermark to 0 too
+ * (#486). Otherwise `textLen - lastCheckLen` goes negative after the reset and
+ * this returns false until a later step re-grows past the stale offset — a whole
+ * degenerate step could stream unchecked.
+ */
+export function shouldCheckDegeneration(
+  textLen: number,
+  lastCheckLen: number,
+): boolean {
+  return textLen - lastCheckLen >= DEGENERATION_CHECK_STEP;
+}
+
 /**
  * Truncate a degenerated tail before persist so hundreds of KB of garbage never
  * reach the DB / replay (#444). Keeps everything up to and including the FIRST

apps/server/src/core/ai-chat/public-share-chat.error-leak.spec.ts

@@ -0,0 +1,241 @@
+// Break the editor-ext import chain (share.service -> collaboration.util ->
+// @docmost/editor-ext -> @tiptap/core) that is unresolvable in this jest env and
+// pre-existingly breaks these specs. jsonToMarkdown is never reached in these
+// tests (the tools fail before rendering markdown).
+jest.mock('../../collaboration/collaboration.util', () => ({
+  jsonToMarkdown: () => '',
+}));
+
+import { Logger } from '@nestjs/common';
+import { MockLanguageModelV3, simulateReadableStream } from 'ai/test';
+import { PublicShareChatService } from './public-share-chat.service';
+import { PublicShareChatToolsService } from './tools/public-share-chat-tools.service';
+
+/**
+ * SECURITY integration guard for #394 (commit 5): a tool's or the provider's raw
+ * error text must NOT leak to an anonymous public-share reader.
+ *
+ * The render gate (ToolCallCard showErrors=false) hides the text in the DOM but
+ * NOT on the wire, so this test asserts on the RAW SSE BYTES the server writes —
+ * exactly the channel the render gate masks. We drive the real
+ * PublicShareChatService.stream() with a real share toolset (its underlying
+ * services mocked to fail) and a mock model, then inspect every byte piped to the
+ * fake socket.
+ */
+
+// A minimal ServerResponse stand-in that records every written chunk.
+class FakeSocket {
+  chunks: string[] = [];
+  statusCode = 200;
+  writableEnded = false;
+  destroyed = false;
+  headersSent = false;
+  writeHead(): this {
+    this.headersSent = true;
+    return this;
+  }
+  setHeader(): void {}
+  removeHeader(): void {}
+  getHeader(): undefined {
+    return undefined;
+  }
+  flushHeaders(): void {}
+  write(chunk: unknown): boolean {
+    this.chunks.push(
+      typeof chunk === 'string' ? chunk : Buffer.from(chunk as never).toString('utf8'),
+    );
+    return true;
+  }
+  end(chunk?: unknown): void {
+    if (chunk) this.write(chunk);
+    this.writableEnded = true;
+  }
+  on(): this {
+    return this;
+  }
+  once(): this {
+    return this;
+  }
+  get body(): string {
+    return this.chunks.join('');
+  }
+}
+
+/** Mock model that issues one getSharePage tool call, then finishes with text. */
+function toolCallingModel(): MockLanguageModelV3 {
+  let call = 0;
+  return new MockLanguageModelV3({
+    doStream: async () => {
+      call++;
+      if (call === 1) {
+        return {
+          stream: simulateReadableStream({
+            chunks: [
+              { type: 'stream-start' as const, warnings: [] },
+              { type: 'tool-input-start' as const, id: 't1', toolName: 'getSharePage' },
+              { type: 'tool-input-end' as const, id: 't1' },
+              {
+                type: 'tool-call' as const,
+                toolCallId: 't1',
+                toolName: 'getSharePage',
+                input: '{"pageId":"secret-page"}',
+              },
+              {
+                type: 'finish' as const,
+                finishReason: { unified: 'tool-calls' as const, raw: 'tool_calls' },
+                usage: {
+                  inputTokens: { total: 1, noCache: undefined, cacheRead: undefined, cacheWrite: undefined },
+                  outputTokens: { total: 1, text: 1, reasoning: undefined },
+                },
+              },
+            ],
+          }),
+        };
+      }
+      return {
+        stream: simulateReadableStream({
+          chunks: [
+            { type: 'stream-start' as const, warnings: [] },
+            { type: 'text-start' as const, id: '1' },
+            { type: 'text-delta' as const, id: '1', delta: 'Sorry.' },
+            { type: 'text-end' as const, id: '1' },
+            {
+              type: 'finish' as const,
+              finishReason: { unified: 'stop' as const, raw: 'stop' },
+              usage: {
+                inputTokens: { total: 1, noCache: undefined, cacheRead: undefined, cacheWrite: undefined },
+                outputTokens: { total: 1, text: 1, reasoning: undefined },
+              },
+            },
+          ],
+        }),
+      };
+    },
+  });
+}
+
+/** Mock model whose stream emits a provider error carrying an internal secret. */
+function providerErrorModel(secret: string): MockLanguageModelV3 {
+  return new MockLanguageModelV3({
+    doStream: async () => ({
+      stream: simulateReadableStream({
+        chunks: [
+          { type: 'stream-start' as const, warnings: [] },
+          {
+            type: 'error' as const,
+            error: {
+              statusCode: 503,
+              message: 'Service Unavailable',
+              responseBody: `upstream ${secret} model=internal-gpt`,
+            },
+          },
+        ],
+      }),
+    }),
+  });
+}
+
+function makeService(toolsService: PublicShareChatToolsService): {
+  svc: PublicShareChatService;
+  logSpy: jest.SpyInstance;
+} {
+  const svc = Object.create(PublicShareChatService.prototype);
+  const logger = new Logger('test');
+  const logSpy = jest.spyOn(logger, 'error').mockImplementation(() => undefined);
+  jest.spyOn(logger, 'warn').mockImplementation(() => undefined);
+  svc.tools = toolsService;
+  svc.logger = logger;
+  svc.tokenBudget = { record: jest.fn().mockResolvedValue(undefined) };
+  return { svc, logSpy };
+}
+
+async function runStream(
+  svc: PublicShareChatService,
+  model: MockLanguageModelV3,
+): Promise<FakeSocket> {
+  const socket = new FakeSocket();
+  await svc.stream({
+    workspaceId: 'ws1',
+    shareId: 'share1',
+    share: { id: 'share1', pageId: 'p1', sharedPage: { id: 'p1', title: 'Docs' } },
+    openedPage: null,
+    messages: [
+      { id: 'm1', role: 'user', parts: [{ type: 'text', text: 'read the page' }] } as never,
+    ],
+    res: { raw: socket } as never,
+    signal: new AbortController().signal,
+    model: model as never,
+    role: null,
+  });
+  // Let the piped stream drain fully.
+  await new Promise((r) => setTimeout(r, 300));
+  return socket;
+}
+
+describe('public share chat error leak (#394)', () => {
+  afterEach(() => jest.restoreAllMocks());
+
+  it('does NOT leak a tool\'s raw internal error to the SSE bytes (generic classified string instead)', async () => {
+    const SECRET = 'INTERNAL_baseUrl_http://provider.internal:8080/v1';
+    const shareService = {
+      // The canonical boundary throws a RAW internal error (with a secret).
+      resolveReadableSharePage: jest
+        .fn()
+        .mockRejectedValue(new Error(`db failed at ${SECRET} stack@line42`)),
+    };
+    const tools = new PublicShareChatToolsService(
+      shareService as never,
+      {} as never,
+      {} as never,
+    );
+    const { svc } = makeService(tools);
+
+    const socket = await runStream(svc, toolCallingModel());
+
+    // The tool-output-error frame is present on the wire...
+    expect(socket.body).toContain('tool-output-error');
+    // ...but it carries ONLY the generic classified string — never the secret,
+    // the raw driver message, or a stack fragment.
+    expect(socket.body).toContain('The tool could not complete the request.');
+    expect(socket.body).not.toContain(SECRET);
+    expect(socket.body).not.toContain('stack@line42');
+    expect(socket.body).not.toContain('db failed');
+  });
+
+  it('passes a SAFE ShareToolError message (page not available) through to the bytes', async () => {
+    const shareService = {
+      // Not found in this share -> the tool throws the classified SAFE message.
+      resolveReadableSharePage: jest.fn().mockResolvedValue(null),
+    };
+    const tools = new PublicShareChatToolsService(
+      shareService as never,
+      {} as never,
+      {} as never,
+    );
+    const { svc } = makeService(tools);
+
+    const socket = await runStream(svc, toolCallingModel());
+    expect(socket.body).toContain('tool-output-error');
+    expect(socket.body).toContain('not available in this share');
+  });
+
+  it('does NOT leak a provider error (statusCode + response body) to the SSE bytes', async () => {
+    const SECRET = 'http://provider.internal:8080';
+    const tools = new PublicShareChatToolsService(
+      {} as never,
+      {} as never,
+      {} as never,
+    );
+    const { svc, logSpy } = makeService(tools);
+
+    const socket = await runStream(svc, providerErrorModel(SECRET));
+
+    // The anon sees a fixed classified string, not the provider body/baseUrl/model.
+    expect(socket.body).toContain('temporarily unavailable');
+    expect(socket.body).not.toContain(SECRET);
+    expect(socket.body).not.toContain('internal-gpt');
+    // The FULL provider detail is logged server-side only.
+    const logged = logSpy.mock.calls.map((c) => String(c[0])).join('\n');
+    expect(logged).toContain(SECRET);
+  });
+});

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

@@ -12,7 +12,10 @@ import { AiAgentRoleRepo } from '@docmost/db/repos/ai-agent-roles/ai-agent-roles
 import { AiAgentRole } from '@docmost/db/types/entity.types';
 import { AiService } from '../../integrations/ai/ai.service';
 import { AiSettingsService } from '../../integrations/ai/ai-settings.service';
-import { PublicShareChatToolsService } from './tools/public-share-chat-tools.service';
+import {
+  PublicShareChatToolsService,
+  ShareToolError,
+} from './tools/public-share-chat-tools.service';
 import { buildShareSystemPrompt } from './public-share-chat.prompt';
 import { roleModelOverride } from './roles/role-model-config';
 import {
@@ -102,6 +105,30 @@ export function filterShareTranscript(messages: UIMessage[]): UIMessage[] {
   );
 }
 
+/**
+ * Fixed, classified strings an ANONYMOUS share reader may see when the assistant
+ * stream fails (#394). These reveal NOTHING about the internal provider, its
+ * baseUrl, the model name, or the raw response body — unlike describeProviderError
+ * (which is for the server log / the authenticated operator only). We classify by
+ * HTTP status where available so the reader still gets a useful hint (retry vs.
+ * give up) without any internal detail.
+ */
+export function classifyAnonStreamError(error: unknown): string {
+  const status =
+    typeof error === 'object' && error !== null
+      ? (error as { statusCode?: number }).statusCode
+      : undefined;
+  if (status === 429) {
+    return 'The assistant is receiving too many requests right now. Please try again shortly.';
+  }
+  if (typeof status === 'number' && status >= 500) {
+    return 'The assistant is temporarily unavailable. Please try again.';
+  }
+  // Any other failure (including a bare connection error with no status): a
+  // single neutral line. No provider identity, no config, no response body.
+  return 'The assistant could not complete your request. Please try again.';
+}
+
 /**
  * Anonymous, read-only AI assistant for a single PUBLIC share tree.
  *
@@ -318,11 +345,28 @@ export class PublicShareChatService {
       result.pipeUIMessageStreamToResponse(res.raw, {
         headers: { 'X-Accel-Buffering': 'no' },
         onError: (error: unknown) => {
-          // Reuse the shared formatter so provider error formatting stays
+          // SECURITY (#394): the string this returns is written verbatim into the
-          // unified between the log line and the streamed error message — a
+          // SSE error frame delivered to an ANONYMOUS reader (for a tool failure
-          // share reader sees 402/429/503 causes consistently with the
+          // it becomes the atomic `tool-output-error` frame's errorText; for a
-          // authenticated path.
+          // stream/provider failure, the terminal error frame).
-          return describeProviderError(error, 'AI stream error');
+          //
+          // A ShareToolError is already a classified, safe tool message (see
+          // PublicShareChatToolsService.wrapToolErrors) — pass it through so the
+          // reader still gets the useful "page not available in this share" hint.
+          if (error instanceof ShareToolError) {
+            return error.message;
+          }
+          // Anything else is a provider/stream error. describeProviderError
+          // bundles the provider statusCode AND response body, which can carry the
+          // internal baseUrl or model name — NEVER expose that to the public. Log
+          // the full detail server-side only and return a fixed classified string.
+          this.logger.error(
+            `Public share chat pipe error: ${describeProviderError(
+              error,
+              'AI stream error',
+            )}`,
+          );
+          return classifyAnonStreamError(error);
         },
       });
 

apps/server/src/core/ai-chat/public-share-chat.spec.ts

@@ -808,7 +808,7 @@ describe('PublicShareChatToolsService share scoping', () => {
     };
 
     await expect(getSharePage.execute({ pageId: 'p-outside' })).rejects.toThrow(
-      /not part of this published share/i,
+      /not available in this share/i,
     );
     // The tool delegated the resolve to the canonical boundary with the
     // forShare-scoped shareId, and returned NO content for a non-resolving page.
@@ -841,7 +841,7 @@ describe('PublicShareChatToolsService share scoping', () => {
 
     await expect(
       getSharePage.execute({ pageId: 'p-restricted' }),
-    ).rejects.toThrow(/not part of this published share/i);
+    ).rejects.toThrow(/not available in this share/i);
     // No content was ever sanitized/returned for the blocked page.
     expect(shareService.updatePublicAttachments).not.toHaveBeenCalled();
   });
@@ -1003,7 +1003,7 @@ describe('public-share assistant boundary locks (red-team regression guards)', (
     };
     await expect(
       getSharePage.execute({ pageId: 'p-elsewhere' }),
-    ).rejects.toThrow(/not part of this published share/i);
+    ).rejects.toThrow(/not available in this share/i);
     // The forged share id is the scope the boundary re-derivation rejects against.
     expect(shareService.resolveReadableSharePage).toHaveBeenCalledWith(
       'FORGED-SHARE',

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

@@ -1,7 +1,15 @@
 import { createHash } from 'node:crypto';
-import { mkdtempSync, mkdirSync, writeFileSync, rmSync } from 'node:fs';
+import {
+  mkdtempSync,
+  mkdirSync,
+  writeFileSync,
+  rmSync,
+  readdirSync,
+  statSync,
+  readFileSync,
+} from 'node:fs';
 import { tmpdir } from 'node:os';
-import { join } from 'node:path';
+import { dirname, join, relative, sep } from 'node:path';
 
 import { computeSrcRegistryStamp } from './docmost-client.loader';
 
@@ -30,10 +38,14 @@ function assertStaleGuard(
   }
 }
 
-// Build a throwaway `<pkg>/build/index.js` + optional `<pkg>/src/tool-specs.ts`
+// Build a throwaway `<pkg>/build/index.js` + optional `<pkg>/src/` tree so
-// layout so `computeSrcRegistryStamp(<pkg>/build/index.js)` resolves src the same
+// `computeSrcRegistryStamp(<pkg>/build/index.js)` resolves src the same way the
-// way the loader does (dirname(dirname(entry))/src/tool-specs.ts).
+// loader does (dirname(dirname(entry))/src). Since #486 the stamp hashes the WHOLE
-function makeFakePackage(toolSpecsSource: string | null): {
+// src tree, so a fixture is a { relPath: content } map. A bare string is sugar for
+// a single `tool-specs.ts`; `null` means "no src tree" (the prod no-op path).
+function makeFakePackage(
+  src: string | Record<string, string> | null,
+): {
   entry: string;
   cleanup: () => void;
 } {
@@ -42,10 +54,15 @@ function makeFakePackage(toolSpecsSource: string | null): {
   mkdirSync(buildDir, { recursive: true });
   const entry = join(buildDir, 'index.js');
   writeFileSync(entry, '// fake @docmost/mcp build entry\n', 'utf8');
-  if (toolSpecsSource !== null) {
+  if (src !== null) {
+    const files =
+      typeof src === 'string' ? { 'tool-specs.ts': src } : src;
     const srcDir = join(root, 'src');
-    mkdirSync(srcDir, { recursive: true });
+    for (const [rel, content] of Object.entries(files)) {
-    writeFileSync(join(srcDir, 'tool-specs.ts'), toolSpecsSource, 'utf8');
+      const full = join(srcDir, rel);
+      mkdirSync(dirname(full), { recursive: true });
+      writeFileSync(full, content, 'utf8');
+    }
   }
   return { entry, cleanup: () => rmSync(root, { recursive: true, force: true }) };
 }
@@ -93,34 +110,109 @@ describe('computeSrcRegistryStamp (#447 stale-build guard)', () => {
     }
   });
 
-  // CROSS-IMPL EQUALITY (covers reviewer suggestion 2). The SAME fixed input and
+  // #486 CORE (negative): an edit to a NON-tool-specs src file (client.ts) with a
+  // rebuild NOT run must move the src stamp away from the built REGISTRY_STAMP, so
+  // the loader's stale-check refuses. Under the old tool-specs.ts-only hash this
+  // edit was invisible and a stale build/ served the old client.ts silently.
+  it('a client.ts edit (no rebuild) moves the src stamp -> loader refuses (#486)', () => {
+    // "Built" state: the package as it was compiled.
+    const built = makeFakePackage({
+      'tool-specs.ts': 'export const SPECS = 1;\n',
+      'client.ts': "export const impl = 'v1';\n",
+    });
+    // "Dev edited src, forgot to rebuild": client.ts changed, tool-specs.ts not.
+    const edited = makeFakePackage({
+      'tool-specs.ts': 'export const SPECS = 1;\n',
+      'client.ts': "export const impl = 'v2';\n",
+    });
+    try {
+      const builtStamp = computeSrcRegistryStamp(built.entry);
+      const editedStamp = computeSrcRegistryStamp(edited.entry);
+      expect(builtStamp).not.toBeNull();
+      expect(editedStamp).not.toBe(builtStamp);
+      // build/ still carries builtStamp; src now hashes to editedStamp -> refuse.
+      expect(() => assertStaleGuard(editedStamp, builtStamp as string)).toThrow(
+        STALE_BUILD_MESSAGE,
+      );
+    } finally {
+      built.cleanup();
+      edited.cleanup();
+    }
+  });
+
+  // *.generated.ts is excluded (the codegen's own output — a fixed-point cycle
+  // otherwise): its presence/content must not move the stamp.
+  it('excludes *.generated.ts from the stamp', () => {
+    const without = makeFakePackage({ 'tool-specs.ts': 'x\n' });
+    const withGen = makeFakePackage({
+      'tool-specs.ts': 'x\n',
+      'registry-stamp.generated.ts': 'export const REGISTRY_STAMP = "abc";\n',
+    });
+    try {
+      expect(computeSrcRegistryStamp(withGen.entry)).toBe(
+        computeSrcRegistryStamp(without.entry),
+      );
+    } finally {
+      without.cleanup();
+      withGen.cleanup();
+    }
+  });
+
+  // CROSS-IMPL EQUALITY (covers reviewer suggestion 2). The SAME fixed tree and
   // EXPECTED hash are asserted in the mcp-side node test
   // (packages/mcp/test/unit/registry-stamp.test.mjs) against the codegen's
   // `computeRegistryStamp`. Asserting the SAME pair here against the loader's
-  // `computeSrcRegistryStamp` proves both implementations normalize+hash
+  // `computeSrcRegistryStamp` proves both implementations enumerate+normalize+hash
   // identically; a divergence in EITHER side reddens one of the two tests.
-  it('matches the documented cross-impl hash for a fixed input', () => {
+  const CROSS_IMPL_TREE = {
-    const FIXED_INPUT = 'line1\r\nline2\n';
+    'tool-specs.ts': 'line1\r\nline2\n',
-    const EXPECTED =
+    'client/read.ts': 'export const R = 1;\n',
-      '683376e290829b482c2655745caffa7a1dccfa10afaa62dac2b42dd6c68d0f83';
+    'registry-stamp.generated.ts': 'export const REGISTRY_STAMP="ignored";\n',
-    const { entry, cleanup } = makeFakePackage(FIXED_INPUT);
+  };
+  const CROSS_IMPL_EXPECTED =
+    '131c1b9e4e2f5a7d6cef91ca8df619822b442f52bc45ebd09474a4c1d6728616';
+
+  it('matches the documented cross-impl hash for a fixed tree', () => {
+    const { entry, cleanup } = makeFakePackage(CROSS_IMPL_TREE);
     try {
-      expect(computeSrcRegistryStamp(entry)).toBe(EXPECTED);
+      expect(computeSrcRegistryStamp(entry)).toBe(CROSS_IMPL_EXPECTED);
     } finally {
       cleanup();
     }
   });
 
-  it('the documented EXPECTED is the normalize+sha256 of the fixed input', () => {
+  it('the documented EXPECTED is the enumerate+normalize+sha256 of the tree', () => {
-    // Proves EXPECTED is not a magic constant but the documented computation.
+    // Proves EXPECTED is not a magic constant but the documented computation — a
-    const FIXED_INPUT = 'line1\r\nline2\n';
+    // local re-implementation of the loader's tree walk.
-    const normalized = FIXED_INPUT.replace(/\r\n/g, '\n').replace(/\n$/, '');
+    const { entry, cleanup } = makeFakePackage(CROSS_IMPL_TREE);
-    const expected = createHash('sha256')
-      .update(normalized, 'utf8')
-      .digest('hex');
-    const { entry, cleanup } = makeFakePackage(FIXED_INPUT);
     try {
-      expect(computeSrcRegistryStamp(entry)).toBe(expected);
+      const srcDir = join(dirname(dirname(entry)), 'src');
+      const collect = (dir: string): string[] => {
+        const out: string[] = [];
+        for (const e of readdirSync(dir)) {
+          const f = join(dir, e);
+          if (statSync(f).isDirectory()) out.push(...collect(f));
+          else if (e.endsWith('.ts') && !e.endsWith('.generated.ts'))
+            out.push(f);
+        }
+        return out;
+      };
+      const files = collect(srcDir)
+        .map((abs) => ({ rel: relative(srcDir, abs).split(sep).join('/'), abs }))
+        .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0));
+      const h = createHash('sha256');
+      for (const { rel, abs } of files) {
+        const n = readFileSync(abs, 'utf8')
+          .replace(/\r\n/g, '\n')
+          .replace(/\n$/, '');
+        h.update(rel, 'utf8');
+        h.update('\0', 'utf8');
+        h.update(n, 'utf8');
+        h.update('\0', 'utf8');
+      }
+      const localHash = h.digest('hex');
+      expect(computeSrcRegistryStamp(entry)).toBe(localHash);
+      expect(localHash).toBe(CROSS_IMPL_EXPECTED);
     } finally {
       cleanup();
     }

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

@@ -1,6 +1,6 @@
 import { createHash } from 'node:crypto';
-import { existsSync, readFileSync } from 'node:fs';
+import { existsSync, readdirSync, readFileSync, statSync } from 'node:fs';
-import { dirname, join } from 'node:path';
+import { dirname, join, relative, sep } from 'node:path';
 import { pathToFileURL } from 'node:url';
 import type { DocmostClient, SharedToolSpec } from '@docmost/mcp';
 
@@ -191,33 +191,52 @@ interface DocmostMcpModule {
  * present. Returns the stamp string, or `null` when the source is absent (a prod
  * image ships only build/, no src/). MUST stay byte-for-byte identical to
  * packages/mcp/scripts/gen-registry-stamp.mjs's `computeRegistryStamp` so the
- * build-time and src-time hashes agree: same input file (src/tool-specs.ts), same
+ * build-time and src-time hashes agree: same file set (every src/**\/*.ts except
- * normalization (CRLF -> LF, strip a single trailing newline), same sha256.
+ * *.generated.ts), same POSIX-relative sort, same per-file normalization (CRLF ->
+ * LF, strip a single trailing newline) with the same path+content framing, same
+ * sha256. Hashing the WHOLE src tree (not just tool-specs.ts) is #486: an edit to
+ * client.ts / a client/* module / comment-signal / drawio-* without a rebuild
+ * must also be caught, otherwise build/ silently serves the old code.
  *
  * DEV vs PROD detection is by FILE EXISTENCE, not NODE_ENV: we resolve the
  * package's own directory from `require.resolve('@docmost/mcp')` (which points at
- * build/index.js) and look for ../src/tool-specs.ts next to it. In a dev/test
+ * build/index.js) and look for ../src next to it. In a dev/test worktree that
- * worktree that file exists; in a prod image (build/ only, src/ stripped) it does
+ * directory exists; in a prod image (build/ only, src/ stripped) it does not, so
- * not, so this returns null and the caller skips the check. Any error (ENOENT, a
+ * this returns null and the caller skips the check. Any error (ENOENT, a bad
- * bad resolve) is swallowed to null — the stale-check must NEVER break startup.
+ * resolve) is swallowed to null — the stale-check must NEVER break startup.
  *
  * Exported for unit testing (docmost-client.loader.spec.ts): the export keyword
  * is behaviourally a no-op — the module-internal caller `loadDocmostMcp` is
  * unaffected. The test drives the null (no-src) path and asserts this
- * normalize+sha256 stays identical to the codegen's `computeRegistryStamp`.
+ * enumerate+normalize+sha256 stays identical to the codegen's
+ * `computeRegistryStamp`.
  */
 export function computeSrcRegistryStamp(packageEntry: string): string | null {
   try {
     // packageEntry is <pkg>/build/index.js; the source lives at <pkg>/src/.
-    const toolSpecsPath = join(
+    const srcDir = join(dirname(dirname(packageEntry)), 'src');
-      dirname(dirname(packageEntry)),
+    if (!existsSync(srcDir)) return null; // prod: no src tree -> skip.
-      'src',
+    // Enumerate every src/**\/*.ts except the codegen's own *.generated.ts
-      'tool-specs.ts',
+    // output (including it would be a fixed-point cycle). Sort by POSIX-relative
-    );
+    // path so ordering is platform-independent, then fold each file's relative
-    if (!existsSync(toolSpecsPath)) return null; // prod: no src tree -> skip.
+    // path + normalized content into one hash — identical to the codegen.
-    const source = readFileSync(toolSpecsPath, 'utf8');
+    const files = collectStampFiles(srcDir)
-    const normalized = source.replace(/\r\n/g, '\n').replace(/\n$/, '');
+      .map((abs) => ({
-    return createHash('sha256').update(normalized, 'utf8').digest('hex');
+        rel: relative(srcDir, abs).split(sep).join('/'),
+        abs,
+      }))
+      .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0));
+    const hash = createHash('sha256');
+    for (const { rel, abs } of files) {
+      const normalized = readFileSync(abs, 'utf8')
+        .replace(/\r\n/g, '\n')
+        .replace(/\n$/, '');
+      hash.update(rel, 'utf8');
+      hash.update('\0', 'utf8');
+      hash.update(normalized, 'utf8');
+      hash.update('\0', 'utf8');
+    }
+    return hash.digest('hex');
   } catch {
     // Never let a resolution/read hiccup break server startup — treat as "no
     // src available" and skip the check (identical to the prod no-op path).
@@ -225,6 +244,24 @@ export function computeSrcRegistryStamp(packageEntry: string): string | null {
   }
 }
 
+/**
+ * Recursively enumerate every `*.ts` under `dir`, EXCLUDING `*.generated.ts`.
+ * Mirror of the codegen's `collectStampFiles` (packages/mcp/scripts/
+ * gen-registry-stamp.mjs) — keep the two walk/filter rules identical.
+ */
+function collectStampFiles(dir: string): string[] {
+  const out: string[] = [];
+  for (const entry of readdirSync(dir)) {
+    const full = join(dir, entry);
+    if (statSync(full).isDirectory()) {
+      out.push(...collectStampFiles(full));
+    } else if (entry.endsWith('.ts') && !entry.endsWith('.generated.ts')) {
+      out.push(full);
+    }
+  }
+  return out;
+}
+
 // TS with module:commonjs downlevels a literal `import()` to `require()`, which
 // cannot load the ESM-only `@docmost/mcp` package. Indirect through Function so
 // the real dynamic `import()` survives compilation and can load ESM from

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

@@ -224,7 +224,7 @@ describe('PublicShareChatToolsService.forShare', () => {
         (tools.getSharePage as unknown as ToolExec).execute({
           pageId: 'page-1',
         }),
-      ).rejects.toThrow('That page is not part of this published share.');
+      ).rejects.toThrow('The requested page is not available in this share.');
 
       // No content is ever fetched/returned for a non-resolving page.
       expect(shareService.updatePublicAttachments).not.toHaveBeenCalled();

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

@@ -7,6 +7,22 @@ import { PageRepo } from '@docmost/db/repos/page/page.repo';
 import { jsonToMarkdown } from '../../../collaboration/collaboration.util';
 import { modelFriendlyInput } from './model-friendly-input';
 
+/**
+ * A tool error whose message is DELIBERATELY safe to expose to an anonymous
+ * share reader (and to the model, for self-correction). Every OTHER thrown error
+ * is treated as internal and replaced with a generic string by `wrapToolErrors`,
+ * so a raw exception message — an internal page title, a DB/stack fragment, a
+ * driver detail — never rides the public UI stream (#394).
+ */
+export class ShareToolError extends Error {}
+
+// The only two classified strings an anonymous reader may ever see from a tool
+// failure. The specific one keeps the model's self-correction useful ("try a
+// different page"); the generic one reveals nothing about the internal fault.
+const SHARE_TOOL_ERROR_NOT_AVAILABLE =
+  'The requested page is not available in this share.';
+const SHARE_TOOL_ERROR_GENERIC = 'The tool could not complete the request.';
+
 /**
  * Isolated, READ-ONLY toolset for the ANONYMOUS public-share assistant.
  *
@@ -44,7 +60,7 @@ export class PublicShareChatToolsService {
    * are NO write tools, NO comments/history, NO cross-space or external tools.
    */
   forShare(shareId: string, workspaceId: string): Record<string, Tool> {
-    return {
+    return this.wrapToolErrors({
       searchSharePages: tool({
         description:
           'Search the pages of THIS published documentation share for a ' +
@@ -96,7 +112,7 @@ export class PublicShareChatToolsService {
         execute: async ({ pageId }) => {
           const id = (pageId ?? '').trim();
           if (!id) {
-            throw new Error('A pageId is required.');
+            throw new ShareToolError('A pageId is required.');
           }
           // Resolve via the SINGLE canonical share-access boundary: confirms the
           // page resolves to THIS share (recursive CTE up the tree, honouring
@@ -112,7 +128,7 @@ export class PublicShareChatToolsService {
             workspaceId,
           );
           if (!resolved) {
-            throw new Error('That page is not part of this published share.');
+            throw new ShareToolError(SHARE_TOOL_ERROR_NOT_AVAILABLE);
           }
           const { page } = resolved;
 
@@ -193,6 +209,57 @@ export class PublicShareChatToolsService {
           }
         },
       }),
-    };
+    });
+  }
+
+  /**
+   * Wrap every tool's `execute` so a THROWN error is sanitized in ONE place —
+   * closing the byte leak, the render, and the model context at once (#394).
+   *
+   * The AI SDK surfaces a tool-execution throw as an atomic `tool-output-error`
+   * frame on the v6 UI stream whose `errorText` is the thrown message; on the
+   * public share that frame goes straight to an anonymous reader. Unwrapped, a
+   * raw exception (an internal page title, a DB/stack fragment, a driver detail)
+   * would ride that frame verbatim. Here we catch it, LOG the full detail
+   * server-side only, and re-throw a CLASSIFIED, safe error: the tool's own
+   * intentional ShareToolError messages pass through (they keep the model's
+   * self-correction useful), everything else collapses to a generic string.
+   */
+  private wrapToolErrors(
+    tools: Record<string, Tool>,
+  ): Record<string, Tool> {
+    const wrapped: Record<string, Tool> = {};
+    for (const [name, t] of Object.entries(tools)) {
+      const original = t.execute;
+      if (typeof original !== 'function') {
+        wrapped[name] = t;
+        continue;
+      }
+      wrapped[name] = {
+        ...t,
+        execute: async (args: unknown, options: unknown) => {
+          try {
+            return await (
+              original as (a: unknown, o: unknown) => Promise<unknown>
+            )(args, options);
+          } catch (err) {
+            const safe =
+              err instanceof ShareToolError
+                ? err.message
+                : SHARE_TOOL_ERROR_GENERIC;
+            // Full detail to the server log ONLY — never to the anon.
+            this.logger.warn(
+              `Public share tool "${name}" failed: ${
+                err instanceof Error ? err.message : String(err)
+              }`,
+            );
+            // This safe string is ALL that rides the tool-output-error frame,
+            // becomes model context, and could be rendered — one choke point.
+            throw new ShareToolError(safe);
+          }
+        },
+      } as Tool;
+    }
+    return wrapped;
   }
 }

apps/server/src/core/auth/strategies/jwt.strategy.spec.ts

@@ -120,3 +120,102 @@ describe('JwtStrategy — provenance derivation', () => {
     expect(req.raw.actor).toBeUndefined();
   });
 });
+
+/**
+ * Provenance derivation on the API-KEY path (jwt.strategy.validateApiKey, #486).
+ *
+ * The access-token path stamped provenance; the API-key path returned early
+ * WITHOUT it, so an is_agent API key's REST writes recorded no 'agent' marker.
+ * The API-key payload carries no signed claim, so provenance is resolved from the
+ * SERVER-SIDE user returned by ApiKeyService.validateApiKey: isAgent -> 'agent',
+ * otherwise 'user'; aiChatId is always null (an API key has no ai_chats row).
+ *
+ * The enterprise ApiKeyService is not bundled in the OSS build, so the strategy
+ * loads it through an overridable `resolveApiKeyService` seam that we stub here.
+ */
+describe('JwtStrategy — API-key provenance derivation (#486)', () => {
+  function makeApiKeyStrategy(validateApiKeyImpl: (p: any) => Promise<any>) {
+    const userRepo: any = { findById: jest.fn() };
+    const workspaceRepo: any = { findById: jest.fn() };
+    const userSessionRepo: any = { findActiveById: jest.fn() };
+    const sessionActivityService: any = { trackActivity: jest.fn() };
+    const environmentService: any = { getAppSecret: () => 'test-secret' };
+    const moduleRef: any = {};
+
+    const strategy = new JwtStrategy(
+      userRepo,
+      workspaceRepo,
+      userSessionRepo,
+      sessionActivityService,
+      environmentService,
+      moduleRef,
+    );
+    // Stub the EE ApiKeyService seam (the real module is not in the OSS build).
+    const validateApiKey = jest.fn(validateApiKeyImpl);
+    jest
+      .spyOn(strategy as any, 'resolveApiKeyService')
+      .mockReturnValue({ validateApiKey });
+    return { strategy, validateApiKey };
+  }
+
+  const makeReq = () => ({ raw: {} as Record<string, any> });
+  const apiKeyPayload = () => ({
+    sub: 'svc-1',
+    workspaceId: 'ws-1',
+    apiKeyId: 'key-1',
+    type: JwtType.API_KEY,
+  });
+
+  it("stamps actor='agent' for an is_agent API key (from the validated user)", async () => {
+    const validated = {
+      user: { id: 'svc-1', isAgent: true },
+      workspace: { id: 'ws-1' },
+    };
+    const { strategy, validateApiKey } = makeApiKeyStrategy(
+      async () => validated,
+    );
+    const req = makeReq();
+
+    const result = await strategy.validate(req, apiKeyPayload() as any);
+
+    expect(validateApiKey).toHaveBeenCalledTimes(1);
+    expect(req.raw.actor).toBe('agent');
+    // API keys carry no internal ai_chats row -> null.
+    expect(req.raw.aiChatId).toBeNull();
+    // The validated auth object is returned unchanged (req.user shape preserved).
+    expect(result).toBe(validated);
+  });
+
+  it("stamps actor='user' for an ordinary (non-agent) API key", async () => {
+    const { strategy } = makeApiKeyStrategy(async () => ({
+      user: { id: 'u-1', isAgent: false },
+      workspace: { id: 'ws-1' },
+    }));
+    const req = makeReq();
+
+    await strategy.validate(req, apiKeyPayload() as any);
+
+    expect(req.raw.actor).toBe('user');
+    expect(req.raw.aiChatId).toBeNull();
+  });
+
+  it('throws Unauthorized (and stamps nothing) when the EE module is missing', async () => {
+    const userRepo: any = { findById: jest.fn() };
+    const strategy = new JwtStrategy(
+      userRepo,
+      { findById: jest.fn() } as any,
+      { findActiveById: jest.fn() } as any,
+      { trackActivity: jest.fn() } as any,
+      { getAppSecret: () => 'test-secret' } as any,
+      {} as any,
+    );
+    // EE not bundled: the seam returns null.
+    jest.spyOn(strategy as any, 'resolveApiKeyService').mockReturnValue(null);
+    const req = makeReq();
+
+    await expect(
+      strategy.validate(req, apiKeyPayload() as any),
+    ).rejects.toThrow(UnauthorizedException);
+    expect(req.raw.actor).toBeUndefined();
+  });
+});

apps/server/src/core/auth/strategies/jwt.strategy.ts

@@ -102,28 +102,49 @@ export class JwtStrategy extends PassportStrategy(Strategy, 'jwt') {
   }
 
   private async validateApiKey(req: any, payload: JwtApiKeyPayload) {
-    let ApiKeyModule: any;
+    const apiKeyService = this.resolveApiKeyService();
-    let isApiKeyModuleReady = false;
+    if (!apiKeyService) {
+      throw new UnauthorizedException('Enterprise API Key module missing');
+    }
 
+    const result = await apiKeyService.validateApiKey(payload);
+
+    // Stamp the agent-edit provenance for the API-KEY path too (#486). Unlike the
+    // access-token path above, it CANNOT be resolved before this point: the
+    // API-key payload carries no signed actor/aiChatId claim, and the user (with
+    // its isAgent flag) is unknown until the key is validated. Claim semantics for
+    // API keys: an is_agent API key (an agent service account) stamps 'agent' on
+    // every REST write; an ordinary API key resolves to 'user'. An API key has no
+    // internal ai_chats row, so aiChatId is always null. Derived from the
+    // SERVER-SIDE user (never a client field), so an 'agent' badge is unspoofable
+    // — mirroring the access-token path. Passing `null` for the claim means the
+    // actor is decided solely by user.isAgent.
+    const provenance = resolveProvenance((result as any)?.user, null);
+    req.raw.actor = provenance.actor;
+    req.raw.aiChatId = provenance.aiChatId;
+
+    return result;
+  }
+
+  /**
+   * Resolve the enterprise ApiKeyService, or `null` when the EE module is not
+   * bundled in this build (community build). Extracted as an overridable seam so
+   * the API-key provenance stamping can be unit-tested without the EE package
+   * present (docmost is OSS + a separate EE bundle; `require` of the EE path
+   * throws here). Any load/resolve error is treated as "module missing".
+   */
+  protected resolveApiKeyService(): {
+    validateApiKey: (payload: JwtApiKeyPayload) => Promise<unknown>;
+  } | null {
     try {
       // eslint-disable-next-line @typescript-eslint/no-require-imports
-      ApiKeyModule = require('./../../../ee/api-key/api-key.service');
+      const ApiKeyModule = require('./../../../ee/api-key/api-key.service');
-      isApiKeyModuleReady = true;
+      return this.moduleRef.get(ApiKeyModule.ApiKeyService, { strict: false });
     } catch (err) {
       this.logger.debug(
         'API Key module requested but enterprise module not bundled in this build',
       );
-      isApiKeyModuleReady = false;
+      return null;
-    }
+    }
-
-    if (isApiKeyModuleReady) {
-      const ApiKeyService = this.moduleRef.get(ApiKeyModule.ApiKeyService, {
-        strict: false,
-      });
-
-      return ApiKeyService.validateApiKey(payload);
-    }
-
-    throw new UnauthorizedException('Enterprise API Key module missing');
   }
 }

apps/server/src/integrations/ai/ai-sdk-drain-hang.patch.spec.ts

@@ -0,0 +1,133 @@
+import { readFileSync } from 'fs';
+import { EventEmitter } from 'node:events';
+import { streamText } from 'ai';
+import { MockLanguageModelV3, simulateReadableStream } from 'ai/test';
+
+/**
+ * Regression tests for the writeToServerResponse drain-hang fix in
+ * patches/ai@6.0.134.patch (#486, commit 6).
+ *
+ * Unpatched ai@6.0.134's writeToServerResponse awaits ONLY `once("drain")` when
+ * response.write() returns false (backpressure). If the client disconnects
+ * mid-write the socket never drains, so that await never resolves: the read loop
+ * parks FOREVER, its `finally { response.end() }` is unreachable, and the stream
+ * reader + buffered chunks are pinned until process restart. In autonomous mode
+ * the run keeps producing output after the disconnect, so EVERY mid-run
+ * disconnect leaks a hung pipe. The patch races drain against close/error, and on
+ * a terminal socket event cancels the reader and breaks so `finally` always runs.
+ *
+ * This drives the REAL patched writeToServerResponse through the public
+ * pipeUIMessageStreamToResponse API with a response that never drains and closes
+ * mid-write — exactly the leak scenario.
+ */
+
+/** A ServerResponse-like emitter whose first write() stalls (returns false) and
+ *  then "closes" like a disconnecting client — never firing 'drain'. */
+class DisconnectingResponse extends EventEmitter {
+  ended = false;
+  writeCount = 0;
+  statusCode = 200;
+  writableEnded = false;
+  destroyed = false;
+  writeHead(): this {
+    return this;
+  }
+  setHeader(): void {}
+  flushHeaders(): void {}
+  write(): boolean {
+    this.writeCount++;
+    if (this.writeCount === 1) {
+      // Simulate the client vanishing mid-write: backpressure (false) and then a
+      // 'close' on the next tick, and CRUCIALLY never a 'drain'. Unpatched, the
+      // loop would await drain forever here.
+      setImmediate(() => this.emit('close'));
+      return false;
+    }
+    return true;
+  }
+  end(): void {
+    this.ended = true;
+    this.writableEnded = true;
+    this.emit('finish');
+  }
+}
+
+function makeModel() {
+  return new MockLanguageModelV3({
+    doStream: async () => ({
+      stream: simulateReadableStream({
+        chunks: [
+          { type: 'stream-start' as const, warnings: [] },
+          { type: 'text-start' as const, id: '1' },
+          { type: 'text-delta' as const, id: '1', delta: 'hello ' },
+          { type: 'text-delta' as const, id: '1', delta: 'world' },
+          { type: 'text-end' as const, id: '1' },
+          {
+            type: 'finish' as const,
+            finishReason: { unified: 'stop' as const, raw: 'stop' },
+            usage: {
+              inputTokens: { total: 1, noCache: undefined, cacheRead: undefined, cacheWrite: undefined },
+              outputTokens: { total: 1, text: 1, reasoning: undefined },
+            },
+          },
+        ],
+      }),
+    }),
+  });
+}
+
+describe('ai@6.0.134 pnpm patch: writeToServerResponse drain-hang (#486)', () => {
+  it('ends the response (does NOT hang) when the socket closes mid-write without draining', async () => {
+    const result = streamText({ model: makeModel(), prompt: 'hi' });
+    const res = new DisconnectingResponse();
+    // Drain the SDK stream independently, like the production detached path.
+    void result.consumeStream({ onError: () => undefined });
+    result.pipeUIMessageStreamToResponse(res as never);
+
+    // TRIPWIRE: the patched loop exits on 'close' and runs finally -> end().
+    // Unpatched, it awaits 'drain' forever and this never becomes true.
+    await new Promise<void>((resolve, reject) => {
+      const started = Date.now();
+      const poll = setInterval(() => {
+        if (res.ended) {
+          clearInterval(poll);
+          resolve();
+        } else if (Date.now() - started > 3000) {
+          clearInterval(poll);
+          reject(new Error('writeToServerResponse hung: response never ended'));
+        }
+      }, 20);
+    });
+
+    expect(res.ended).toBe(true);
+  });
+
+  it('does not emit an unhandledRejection when the fire-and-forget read() throws', async () => {
+    // The patch swallows read()'s rejection (fire-and-forget) with a log instead
+    // of letting it surface as a process-killing unhandledRejection.
+    const rejections: unknown[] = [];
+    const onUnhandled = (e: unknown) => rejections.push(e);
+    process.on('unhandledRejection', onUnhandled);
+    // Silence the patch's diagnostic console.error for the throwing read().
+    const errSpy = jest.spyOn(console, 'error').mockImplementation(() => undefined);
+    try {
+      const result = streamText({ model: makeModel(), prompt: 'hi' });
+      const res = new DisconnectingResponse();
+      void result.consumeStream({ onError: () => undefined });
+      result.pipeUIMessageStreamToResponse(res as never);
+      await new Promise((r) => setTimeout(r, 300));
+    } finally {
+      process.off('unhandledRejection', onUnhandled);
+      errSpy.mockRestore();
+    }
+    expect(rejections).toEqual([]);
+  });
+
+  it('both installed dist builds (CJS and ESM) carry the #486 patch marker', () => {
+    const cjsPath = require.resolve('ai');
+    const mjsPath = cjsPath.replace(/index\.js$/, 'index.mjs');
+    expect(cjsPath).toMatch(/index\.js$/);
+    expect(readFileSync(cjsPath, 'utf8')).toContain('PATCH(docmost #486)');
+    expect(readFileSync(mjsPath, 'utf8')).toContain('PATCH(docmost #486)');
+  });
+});

apps/server/src/integrations/health/redis.health.spec.ts

@@ -0,0 +1,211 @@
+import type { HealthIndicatorService } from '@nestjs/terminus';
+import type { EnvironmentService } from '../environment/environment.service';
+
+/**
+ * Integration guard for the /health Redis-probe handle leak (#486, commit 2).
+ *
+ * The bug: `pingCheck` built `new Redis(...)` per call and only disconnected on
+ * the SUCCESS path, so when Redis is DOWN every probe tick added ANOTHER
+ * forever-reconnecting client — an unbounded handle/client leak for the duration
+ * of the outage. The fix reuses ONE long-lived probe client.
+ *
+ * This is an OBSERVABLE-property test, not an assertion on a mocked return value:
+ * we point the indicator at a REAL, refused TCP endpoint (a dead port) so ioredis
+ * genuinely fails to connect, run many probes, and assert the number of live
+ * Redis CLIENTS created stays at exactly ONE. `ioredis` is delegated to its real
+ * implementation (requireActual) — only the constructor is wrapped to COUNT the
+ * real clients it creates, which is precisely the leaking resource.
+ */
+import type { Redis } from 'ioredis';
+
+const mockLiveClients: Redis[] = [];
+
+/**
+ * Fully tear a REAL ioredis client down so NO timer survives jest's 1s exit
+ * window (this suite must exit cleanly WITHOUT forceExit; see #382).
+ *
+ * `connector.disconnect()` arms a ~12s "force-destroy the stream" `setTimeout`
+ * that is cleared ONLY by the stream's 'close' event — but only when the
+ * connector still holds a stream. Two problem cases:
+ *  - a LIVE/connecting socket: disconnect arms the timer and 'close' may lag
+ *    past jest's window, so we destroy the socket to make 'close' fire NOW;
+ *  - a client BETWEEN reconnect attempts to a dead port: the held socket is
+ *    ALREADY destroyed (its 'close' fired long ago), so disconnect would arm a
+ *    timer whose clearing 'close' can never come again. We drop that dead stream
+ *    reference BEFORE disconnect so the doomed timer is never armed.
+ * `disconnect()` itself also clears ioredis' own reconnect backoff timer.
+ */
+type DrainableStream = { destroyed?: boolean; destroy?: () => void } | null;
+type DrainableClient = {
+  removeAllListeners: (event: string) => void;
+  disconnect: () => void;
+  stream?: DrainableStream;
+  connector?: { stream?: DrainableStream };
+};
+
+async function drainClient(client: Redis): Promise<void> {
+  if (!client || client.status === 'end') return;
+  const c = client as unknown as DrainableClient;
+  c.removeAllListeners('error');
+
+  // Drop an already-dead held socket so disconnect() can't arm a timer whose
+  // clearing 'close' will never fire again.
+  if (c.connector?.stream && c.connector.stream.destroyed) {
+    c.connector.stream = null;
+  }
+  if (c.stream && c.stream.destroyed) {
+    c.stream = null;
+  }
+
+  await new Promise<void>((resolve) => {
+    let done = false;
+    const finish = () => {
+      if (done) return;
+      done = true;
+      resolve();
+    };
+    client.once('end', finish);
+    // reconnect=false (the default): stop the retry loop and close the socket.
+    client.disconnect();
+    // Force any still-live socket closed NOW so the connector's stream-destroy
+    // timer clears inside jest's window instead of lagging behind a real 'close'.
+    if (c.stream && !c.stream.destroyed) {
+      c.stream.destroy?.();
+    }
+    // Fallback for a client with no live stream to emit 'end' (unref'd so it
+    // can never itself hold the loop open).
+    const fallback = setTimeout(finish, 500);
+    (fallback as { unref?: () => void }).unref?.();
+  });
+}
+
+async function drainAll(): Promise<void> {
+  await Promise.all(mockLiveClients.map((c) => drainClient(c)));
+}
+
+jest.mock('ioredis', () => {
+  const actual = jest.requireActual('ioredis');
+  const RealRedis = actual.Redis ?? actual.default ?? actual;
+  class CountingRedis extends RealRedis {
+    constructor(...args: unknown[]) {
+      super(...(args as []));
+      mockLiveClients.push(this as never);
+    }
+  }
+  return { ...actual, Redis: CountingRedis, default: CountingRedis };
+});
+
+// Import AFTER the mock is registered so the class picks up the counting client.
+import { RedisHealthIndicator } from './redis.health';
+
+describe('RedisHealthIndicator handle leak (#486)', () => {
+  const indicatorService = {
+    check: (key: string) => ({
+      up: () => ({ [key]: { status: 'up' } }),
+      down: (message: string) => ({ [key]: { status: 'down', message } }),
+    }),
+  } as unknown as HealthIndicatorService;
+
+  // A port with (almost certainly) nothing listening -> connection refused fast.
+  const environmentService = {
+    getRedisUrl: () => 'redis://127.0.0.1:6399/0',
+  } as unknown as EnvironmentService;
+
+  let indicator: RedisHealthIndicator;
+
+  beforeEach(() => {
+    mockLiveClients.length = 0;
+    indicator = new RedisHealthIndicator(indicatorService, environmentService);
+  });
+
+  afterEach(async () => {
+    // Drain (destroy socket + AWAIT 'end') every client the test created FIRST,
+    // so each is fully 'end' before onModuleDestroy's disconnect runs — that way
+    // no ioredis reconnect / stream-destroy timer outlives jest's exit window.
+    await drainAll();
+    indicator.onModuleDestroy();
+  });
+
+  it('creates exactly ONE Redis client across many probes while Redis is DOWN', async () => {
+    const N = 8;
+    for (let i = 0; i < N; i++) {
+      const result = await indicator.pingCheck('redis');
+      // Down endpoint -> every probe reports "down" (not an unhandled crash).
+      expect(result.redis.status).toBe('down');
+    }
+
+    // THE OBSERVABLE LEAK: on the buggy code this is N (a fresh, never-cleaned
+    // reconnecting client per probe). The fix reuses one shared client.
+    expect(mockLiveClients).toHaveLength(1);
+  });
+
+  it('onModuleDestroy releases the probe client (a later probe builds a fresh one)', async () => {
+    await indicator.pingCheck('redis');
+    expect(mockLiveClients).toHaveLength(1);
+
+    indicator.onModuleDestroy();
+    // A second destroy is a safe no-op (probeClient was nulled).
+    indicator.onModuleDestroy();
+
+    // After shutdown the indicator lazily builds a NEW client on the next probe,
+    // proving the old one was truly released rather than reused.
+    await indicator.pingCheck('redis');
+    expect(mockLiveClients).toHaveLength(2);
+  });
+});
+
+/**
+ * Happy-path regression guard (#486, B2): the FIRST probe against a LIVE Redis
+ * must report UP.
+ *
+ * With `lazyConnect: true` + `enableOfflineQueue: false`, a freshly-built client
+ * is in the `wait` state and the socket opens lazily. If the very first `ping()`
+ * is issued before an explicit `connect()`, ioredis rejects it instantly with
+ * "Stream isn't writeable and enableOfflineQueue options is false" — a FALSE
+ * DOWN even though Redis is alive. The fix opens the socket before the first
+ * ping. This exercises a REAL ioredis client against a REAL TCP redis server
+ * (not a mock), so a regression genuinely reddens it.
+ */
+describe('RedisHealthIndicator live Redis first-probe (#486, B2)', () => {
+  const indicatorService = {
+    check: (key: string) => ({
+      up: () => ({ [key]: { status: 'up' } }),
+      down: (message: string) => ({ [key]: { status: 'down', message } }),
+    }),
+  } as unknown as HealthIndicatorService;
+
+  // A REAL running redis (see the neighboring harness / CI env).
+  const environmentService = {
+    getRedisUrl: () => 'redis://127.0.0.1:6379/0',
+  } as unknown as EnvironmentService;
+
+  let indicator: RedisHealthIndicator;
+
+  beforeEach(() => {
+    mockLiveClients.length = 0;
+    indicator = new RedisHealthIndicator(indicatorService, environmentService);
+  });
+
+  afterEach(async () => {
+    // Await full socket close of every live client (see drainClient) BEFORE
+    // onModuleDestroy: a real, connected ioredis client MUST be drained to 'end'
+    // or its stream-destroy timer keeps the jest worker alive past the 1s window.
+    await drainAll();
+    indicator.onModuleDestroy();
+  });
+
+  it('reports UP on the FIRST probe against a live Redis', async () => {
+    // The VERY FIRST probe — no warm-up ping — must be UP.
+    const result = await indicator.pingCheck('redis');
+    expect(result.redis.status).toBe('up');
+  });
+
+  it('stays UP on a probe AFTER onModuleDestroy re-creates the client', async () => {
+    await indicator.pingCheck('redis');
+    indicator.onModuleDestroy();
+    // The re-created client is again in `wait`; the first ping on it must still
+    // open the socket (the false-DOWN also recurs on the post-destroy path).
+    const result = await indicator.pingCheck('redis');
+    expect(result.redis.status).toBe('up');
+  });
+});

apps/server/src/integrations/health/redis.health.ts

@@ -2,33 +2,173 @@ import {
   HealthIndicatorResult,
   HealthIndicatorService,
 } from '@nestjs/terminus';
-import { Injectable, Logger } from '@nestjs/common';
+import { Injectable, Logger, OnModuleDestroy } from '@nestjs/common';
 import { EnvironmentService } from '../environment/environment.service';
 import { Redis } from 'ioredis';
 
 @Injectable()
-export class RedisHealthIndicator {
+export class RedisHealthIndicator implements OnModuleDestroy {
   private readonly logger = new Logger(RedisHealthIndicator.name);
 
+  /**
+   * ONE long-lived probe connection, reused across every /health tick. The old
+   * code built `new Redis(...)` per call and only `disconnect()`d on the SUCCESS
+   * path, so while Redis was DOWN every probe added a fresh, forever-reconnecting
+   * client — a handle leak that grew without bound for as long as the outage (and
+   * the health checker keeps polling) lasted. A single shared client keeps at most
+   * ONE background reconnect loop regardless of how many probes run.
+   */
+  private probeClient: Redis | null = null;
+
+  /**
+   * How long the first-ping `connect()` may take before a probe gives up and
+   * reports DOWN. A `connect()` against a truly-down Redis never settles on its
+   * own (ioredis retries the socket indefinitely per its retryStrategy), so the
+   * probe MUST bound it or the /health handler would hang. Kept short so a real
+   * outage is reported fast; localhost/live Redis connects well within it.
+   */
+  private static readonly CONNECT_TIMEOUT_MS = 2000;
+
+  /**
+   * The single in-flight first-`connect()`, memoized so CONCURRENT probes share
+   * it. k8s liveness+readiness hit /health in parallel on startup: without this,
+   * probe A drives `connect()` (the client leaves the `wait` state) and probe B,
+   * seeing a not-`wait`/not-`ready` client, would skip connect and fire `ping()`
+   * at a still-opening socket → an instant FALSE DOWN. With the memo, B awaits
+   * the SAME connect. Cleared once it settles so a later disconnect / re-create
+   * starts a fresh connect.
+   */
+  private connectingPromise: Promise<void> | null = null;
+
   constructor(
     private readonly healthIndicatorService: HealthIndicatorService,
     private environmentService: EnvironmentService,
   ) {}
 
+  private getProbeClient(): Redis {
+    if (!this.probeClient) {
+      this.probeClient = new Redis(this.environmentService.getRedisUrl(), {
+        // Constructing must never throw or eagerly connect; the first ping opens
+        // the socket. This lets us build the client once and reuse it.
+        lazyConnect: true,
+        // A health probe must fail FAST, not queue behind a stuck reconnect: one
+        // retry per request, and no offline queue so a ping while disconnected
+        // rejects immediately instead of buffering commands that pile up in RAM.
+        maxRetriesPerRequest: 1,
+        enableOfflineQueue: false,
+      });
+      // ioredis emits 'error' on every failed (re)connect; with no listener that
+      // surfaces as an unhandled 'error' event and can crash the process. Swallow
+      // it here — pingCheck already reports health — and log at debug so a Redis
+      // outage does not flood the logs.
+      this.probeClient.on('error', (err) => {
+        this.logger.debug(
+          `Redis probe connection error: ${
+            err instanceof Error ? err.message : String(err)
+          }`,
+        );
+      });
+    }
+    return this.probeClient;
+  }
+
+  /**
+   * Open the probe socket BEFORE the first ping. `lazyConnect: true` leaves a
+   * freshly-built (or post-destroy re-built) client in the `wait` state: the
+   * socket is NOT open yet, so with `enableOfflineQueue: false` the very first
+   * `ping()` rejects instantly with "Stream isn't writeable and
+   * enableOfflineQueue options is false" even when Redis is perfectly alive — a
+   * false DOWN on the happy path. We drive `connect()` ONLY from `wait`; once
+   * the client is connected, ioredis owns its own (re)connect loop and a ping
+   * issued while it reconnects still fast-fails to a correct DOWN (offline queue
+   * stays off). A failed/timed-out connect rejects → reported DOWN, which is the
+   * right signal for a truly-down Redis.
+   */
+  private ensureConnected(client: Redis): Promise<void> {
+    // Already open — steady state, nothing to do.
+    if (client.status === 'ready') return Promise.resolve();
+    // A first-connect is already in flight (possibly started by a CONCURRENT
+    // probe): await the SAME one instead of racing a second connect() (ioredis
+    // throws "already connecting") or firing ping() at a not-yet-open socket.
+    if (this.connectingPromise) return this.connectingPromise;
+    // Only DRIVE connect() from the initial `wait` state (fresh / post-destroy
+    // re-created client). In any other non-ready state ioredis already owns its
+    // (re)connect loop; a ping there fast-fails to a correct DOWN, so we must not
+    // start a competing connect.
+    if (client.status !== 'wait') return Promise.resolve();
+
+    const promise = this.connectWithTimeout(client).finally(() => {
+      // Clear only if still ours, so a later disconnect / re-create can connect
+      // again. Whether it resolved or rejected, the memo has served its window.
+      if (this.connectingPromise === promise) {
+        this.connectingPromise = null;
+      }
+    });
+    this.connectingPromise = promise;
+    return promise;
+  }
+
+  private connectWithTimeout(client: Redis): Promise<void> {
+    return new Promise<void>((resolve, reject) => {
+      let settled = false;
+      const timer = setTimeout(() => {
+        if (settled) return;
+        settled = true;
+        reject(new Error('Redis probe connect timed out'));
+      }, RedisHealthIndicator.CONNECT_TIMEOUT_MS);
+      // Never let THIS timer alone keep the event loop (or a jest worker) alive;
+      // it is cleared on settle anyway, this is belt-and-braces.
+      timer.unref?.();
+      // `.catch` is always attached, so a connect() that rejects AFTER we have
+      // already timed out is handled here (guarded by `settled`) and never
+      // surfaces as an unhandled rejection.
+      client
+        .connect()
+        .then(() => {
+          if (settled) return;
+          settled = true;
+          clearTimeout(timer);
+          resolve();
+        })
+        .catch((err) => {
+          if (settled) return;
+          settled = true;
+          clearTimeout(timer);
+          reject(err);
+        });
+    });
+  }
+
   async pingCheck(key: string): Promise<HealthIndicatorResult> {
     const indicator = this.healthIndicatorService.check(key);
 
     try {
-      const redis = new Redis(this.environmentService.getRedisUrl(), {
+      const redis = this.getProbeClient();
-        maxRetriesPerRequest: 15,
+      // Open the socket before the first ping (see ensureConnected); without
-      });
+      // this the first probe after (re)creation falsely reports DOWN on a live
-
+      // Redis because lazyConnect defers the connect past the first ping.
+      await this.ensureConnected(redis);
       await redis.ping();
-      redis.disconnect();
       return indicator.up();
     } catch (e) {
       this.logger.error(e);
       return indicator.down(`${key} is not available`);
     }
   }
+
+  onModuleDestroy(): void {
+    if (this.probeClient) {
+      // disconnect() (not quit()) tears the socket + reconnect loop down
+      // immediately without waiting on a round-trip to a possibly-down server.
+      // Do NOT removeAllListeners() with no event name — that would also strip
+      // ioredis' OWN internal listeners and break its teardown; our 'error'
+      // listener is harmless and dies with the dropped client reference.
+      this.probeClient.disconnect();
+      this.probeClient = null;
+    }
+    // Drop any in-flight first-connect memo so the NEXT client (lazily rebuilt on
+    // the next probe) starts a fresh connect rather than awaiting a promise tied
+    // to the client we just tore down.
+    this.connectingPromise = null;
+  }
 }

apps/server/src/integrations/mcp/mcp.service.spec.ts

@@ -16,6 +16,7 @@ import {
 } from './mcp-auth.helpers';
 import { JwtType } from '../../core/auth/dto/jwt-payload';
 import { CREDENTIALS_MISMATCH_MESSAGE } from '../../core/auth/auth.constants';
+import { McpService } from './mcp.service';
 
 // The /mcp per-user auth decision logic is tested through the framework-free
 // `resolveMcpSessionConfig` helper that McpService delegates to. McpService
@@ -1179,3 +1180,46 @@ describe('mapAuthResultToResponse (handle status/body mapping, refactor R2)', ()
     });
   });
 });
+
+// #486: onModuleDestroy must ALSO tear down the live loopback CollabSessions, not
+// just clear the sweep timer — otherwise the embedded MCP's collab sockets keep
+// docs pinned open on the collab server past process exit. The teardown goes
+// through an overridable seam (destroyAllMcpSessions) so it can be spied without
+// loading the ESM-only @docmost/mcp package.
+describe('McpService.onModuleDestroy — CollabSession teardown (#486)', () => {
+  function makeService(): McpService {
+    // The constructor only stores its deps and starts the (unref'd) sweep timer,
+    // so bare stubs suffice. onModuleDestroy clears that timer, so no leak.
+    return new McpService(
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+      {} as any,
+    );
+  }
+
+  it('destroys all sessions AND clears the sweep timer on shutdown', async () => {
+    const svc = makeService();
+    const destroy = jest.fn().mockResolvedValue(undefined);
+    (svc as any).destroyAllMcpSessions = destroy;
+    const clearSpy = jest.spyOn(global, 'clearInterval');
+
+    await svc.onModuleDestroy();
+
+    expect(destroy).toHaveBeenCalledTimes(1);
+    expect(clearSpy).toHaveBeenCalledWith((svc as any).sweepTimer);
+    clearSpy.mockRestore();
+  });
+
+  it('swallows a teardown failure so shutdown never throws', async () => {
+    const svc = makeService();
+    (svc as any).destroyAllMcpSessions = jest
+      .fn()
+      .mockRejectedValue(new Error('collab teardown boom'));
+
+    await expect(svc.onModuleDestroy()).resolves.toBeUndefined();
+  });
+});

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

@@ -119,9 +119,41 @@ export class McpService implements OnModuleDestroy {
     this.sweepTimer.unref?.();
   }
 
-  onModuleDestroy(): void {
+  async onModuleDestroy(): Promise<void> {
     clearInterval(this.sweepTimer);
+    // Tear down any live loopback CollabSession providers at shutdown (#486). The
+    // embedded MCP (and the in-app AI agent) open Hocuspocus collab sockets against
+    // THIS process; without an explicit teardown those sessions keep their docs
+    // "open" on the collab server and hold providers/buffers until they idle out,
+    // so a restart can race a doc still pinned by the dying worker. Best-effort:
+    // any failure is logged, never allowed to break shutdown.
+    try {
+      await this.destroyAllMcpSessions();
+    } catch (err) {
+      this.logger.error(
+        'MCP CollabSession teardown on shutdown failed',
+        err as Error,
+      );
     }
+  }
+
+  /**
+   * Resolve @docmost/mcp's `destroyAllSessions` and invoke it (#486). The live
+   * CollabSession registry is a module-level singleton in the ESM package, shared
+   * by every entry (`.`/`./http`), so this tears down ALL sessions regardless of
+   * which surface opened them. The module is already loaded whenever MCP was used;
+   * if it was never loaded (or is absent) the import + no-op is harmless.
+   *
+   * Held as an overridable field so a unit test can spy the teardown without
+   * loading the ESM-only package or standing up the DI graph.
+   */
+  private destroyAllMcpSessions: () => Promise<void> = async () => {
+    const entry = require.resolve('@docmost/mcp');
+    const mod = (await esmImport(pathToFileURL(entry).href)) as {
+      destroyAllSessions?: () => void;
+    };
+    mod.destroyAllSessions?.();
+  };
 
   // Service account the embedded MCP uses to talk back to this Docmost
   // instance over loopback REST + the collaboration WebSocket. Now OPTIONAL:

apps/server/src/integrations/metrics/metrics.server.spec.ts

@@ -0,0 +1,148 @@
+import { get as httpGet } from 'node:http';
+import { AddressInfo } from 'node:net';
+import { createServer } from 'node:http';
+
+// Drive the metrics HTTP server without the load-time METRICS_PORT gate: mock the
+// registry so isMetricsEnabled()/getMetricsRegistry() are always satisfied. What
+// we assert is observed over a REAL socket (bind address, status codes), not on
+// the mock.
+jest.mock('./metrics.registry', () => ({
+  isMetricsEnabled: () => true,
+  getMetricsRegistry: () => ({
+    metrics: async () => '# HELP up test\nup 1\n',
+    contentType: 'text/plain; version=0.0.4',
+  }),
+}));
+
+import {
+  startMetricsServer,
+  closeMetricsServer,
+  resolveMetricsBind,
+  resolveMetricsToken,
+} from './metrics.server';
+
+/** Find a free TCP port (the metrics server requires METRICS_PORT > 0). */
+function freePort(): Promise<number> {
+  return new Promise((resolve, reject) => {
+    const s = createServer();
+    s.once('error', reject);
+    s.listen(0, '127.0.0.1', () => {
+      const p = (s.address() as AddressInfo).port;
+      s.close(() => resolve(p));
+    });
+  });
+}
+
+/** Minimal GET against 127.0.0.1:port with optional Authorization header. */
+function req(
+  port: number,
+  headers: Record<string, string> = {},
+): Promise<{ status: number; body: string }> {
+  return new Promise((resolve, reject) => {
+    const r = httpGet(
+      { host: '127.0.0.1', port, path: '/metrics', headers },
+      (res) => {
+        let body = '';
+        res.on('data', (c) => (body += c));
+        res.on('end', () =>
+          resolve({ status: res.statusCode ?? 0, body }),
+        );
+      },
+    );
+    r.on('error', reject);
+  });
+}
+
+describe('metrics server bind + auth (#486)', () => {
+  const saved = {
+    bind: process.env.METRICS_BIND,
+    token: process.env.METRICS_TOKEN,
+    port: process.env.METRICS_PORT,
+  };
+
+  afterEach(async () => {
+    await closeMetricsServer();
+    process.env.METRICS_BIND = saved.bind;
+    process.env.METRICS_TOKEN = saved.token;
+    process.env.METRICS_PORT = saved.port;
+    delete process.env.METRICS_BIND;
+    delete process.env.METRICS_TOKEN;
+  });
+
+  describe('resolveMetricsBind', () => {
+    it('defaults to loopback 127.0.0.1', () => {
+      delete process.env.METRICS_BIND;
+      expect(resolveMetricsBind()).toBe('127.0.0.1');
+    });
+    it('honours the METRICS_BIND override', () => {
+      process.env.METRICS_BIND = '0.0.0.0';
+      expect(resolveMetricsBind()).toBe('0.0.0.0');
+    });
+    it('treats a blank override as unset (loopback)', () => {
+      process.env.METRICS_BIND = '   ';
+      expect(resolveMetricsBind()).toBe('127.0.0.1');
+    });
+  });
+
+  describe('resolveMetricsToken', () => {
+    it('is null when unset', () => {
+      delete process.env.METRICS_TOKEN;
+      expect(resolveMetricsToken()).toBeNull();
+    });
+    it('returns the trimmed token when set', () => {
+      process.env.METRICS_TOKEN = ' s3cret ';
+      expect(resolveMetricsToken()).toBe('s3cret');
+    });
+  });
+
+  it('binds to loopback by default and serves /metrics without auth when no token', async () => {
+    delete process.env.METRICS_BIND;
+    delete process.env.METRICS_TOKEN;
+    const port = await freePort();
+    process.env.METRICS_PORT = String(port);
+
+    const server = startMetricsServer();
+    expect(server).not.toBeNull();
+    await new Promise<void>((resolve) => {
+      if (server!.listening) resolve();
+      else server!.once('listening', () => resolve());
+    });
+    // OBSERVABLE: the listener bound to loopback, not 0.0.0.0.
+    expect((server!.address() as AddressInfo).address).toBe('127.0.0.1');
+
+    const res = await req(port);
+    expect(res.status).toBe(200);
+    expect(res.body).toContain('up 1');
+  });
+
+  it('rejects unauthenticated scrapes with 401 and accepts the exact Bearer token', async () => {
+    delete process.env.METRICS_BIND;
+    process.env.METRICS_TOKEN = 'topsecret';
+    const port = await freePort();
+    process.env.METRICS_PORT = String(port);
+
+    const server = startMetricsServer();
+    expect(server).not.toBeNull();
+
+    // No auth -> 401.
+    const noAuth = await req(port);
+    expect(noAuth.status).toBe(401);
+
+    // Wrong token, DIFFERENT length -> 401 (short-circuits on the length guard).
+    const wrong = await req(port, { authorization: 'Bearer nope' });
+    expect(wrong.status).toBe(401);
+
+    // Wrong token, SAME length -> 401. This drives the timingSafeEqual compare
+    // itself (the length guard passes: 'Bearer topsecreX' has the same length as
+    // 'Bearer topsecret'). Pins the constant-time compare: a regression that made
+    // it return true would let this equal-length wrong token through — the
+    // different-length case above would NOT catch that.
+    const sameLen = await req(port, { authorization: 'Bearer topsecreX' });
+    expect(sameLen.status).toBe(401);
+
+    // Correct token -> 200 with the metrics body.
+    const ok = await req(port, { authorization: 'Bearer topsecret' });
+    expect(ok.status).toBe(200);
+    expect(ok.body).toContain('up 1');
+  });
+});

apps/server/src/integrations/metrics/metrics.server.ts

@@ -1,7 +1,27 @@
 import { createServer, Server } from 'node:http';
+import { timingSafeEqual } from 'node:crypto';
 import { Logger } from '@nestjs/common';
 import { getMetricsRegistry, isMetricsEnabled } from './metrics.registry';
 
+/**
+ * Constant-time compare of the presented Authorization header against the
+ * expected `Bearer <token>`. This is the ONLY auth layer for the metrics
+ * endpoint, so a naive `!==` would leak the token byte-by-byte via timing.
+ * timingSafeEqual requires equal-length buffers, so a length mismatch short-
+ * circuits to "not equal" (its own length is not itself a useful oracle: the
+ * expected string length is fixed by config, not secret-derived).
+ */
+function bearerMatches(
+  presented: string | undefined,
+  expected: string,
+): boolean {
+  if (typeof presented !== 'string') return false;
+  const a = Buffer.from(presented);
+  const b = Buffer.from(expected);
+  if (a.length !== b.length) return false;
+  return timingSafeEqual(a, b);
+}
+
 /**
  * Start the Prometheus scrape endpoint on a SEPARATE port, taken from
  * `METRICS_PORT`. There is NO default port: when `METRICS_PORT` is unset the
@@ -16,6 +36,30 @@ import { getMetricsRegistry, isMetricsEnabled } from './metrics.registry';
  */
 let metricsServer: Server | null = null;
 
+/**
+ * Interface the metrics endpoint binds to. Defaults to LOOPBACK (127.0.0.1) so
+ * the unauthenticated `/metrics` surface is NOT exposed on all interfaces by
+ * default — the old `0.0.0.0` bind put an auth-less endpoint on every interface.
+ * Deployments where the scraper runs in a SEPARATE container (and reaches this as
+ * `docmost:9464`) set `METRICS_BIND=0.0.0.0`, ideally together with METRICS_TOKEN
+ * and/or a private network so the port is not world-readable.
+ */
+export function resolveMetricsBind(): string {
+  const raw = (process.env.METRICS_BIND ?? '').trim();
+  return raw.length > 0 ? raw : '127.0.0.1';
+}
+
+/**
+ * Optional Bearer token guarding `/metrics`. When `METRICS_TOKEN` is set, every
+ * scrape must present `Authorization: Bearer <token>`; unset (default) leaves the
+ * endpoint open (safe when bound to loopback / a trusted network). Returns the
+ * trimmed token or null when unset/blank.
+ */
+export function resolveMetricsToken(): string | null {
+  const raw = (process.env.METRICS_TOKEN ?? '').trim();
+  return raw.length > 0 ? raw : null;
+}
+
 export function startMetricsServer(): Server | null {
   if (!isMetricsEnabled()) return null;
 
@@ -31,8 +75,22 @@ export function startMetricsServer(): Server | null {
     return null;
   }
 
+  const bind = resolveMetricsBind();
+  const token = resolveMetricsToken();
+
   const server = createServer(async (req, res) => {
     if (req.method === 'GET' && req.url === '/metrics') {
+      // Optional Bearer auth: reject scrapes without the exact token when one is
+      // configured. This is the auth layer the old all-interfaces bind lacked.
+      if (token) {
+        const auth = req.headers['authorization'];
+        if (!bearerMatches(auth, `Bearer ${token}`)) {
+          res.statusCode = 401;
+          res.setHeader('WWW-Authenticate', 'Bearer');
+          res.end();
+          return;
+        }
+      }
       try {
         const body = await register.metrics();
         res.setHeader('Content-Type', register.contentType);
@@ -48,10 +106,14 @@ export function startMetricsServer(): Server | null {
     res.end();
   });
 
-  // Bind on all interfaces: the scraper (VictoriaMetrics) reaches this from
+  // Bind to loopback by default so the auth-less endpoint is not exposed on all
-  // another container as docmost:9464. The port is not published to the host.
+  // interfaces. Set METRICS_BIND=0.0.0.0 (ideally with METRICS_TOKEN) when the
-  server.listen(port, '0.0.0.0', () => {
+  // scraper runs in a separate container and reaches this as docmost:9464.
-    logger.log(`Metrics endpoint listening on :${port}/metrics`);
+  server.listen(port, bind, () => {
+    logger.log(
+      `Metrics endpoint listening on ${bind}:${port}/metrics` +
+        (token ? ' (Bearer auth required)' : ''),
+    );
   });
 
   server.on('error', (err) => {

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

@@ -31,9 +31,6 @@ export enum QueueJob {
   IMPORT_TASK = 'import-task',
   EXPORT_TASK = 'export-task',
 
-  SEARCH_REMOVE_PAGE = 'search-remove-page',
-  SEARCH_REMOVE_ASSET = 'search-remove-attachment',
-  SEARCH_REMOVE_FACE = 'search-remove-comment',
   TYPESENSE_FLUSH = 'typesense-flush',
 
   PAGE_CREATED = 'page-created',

packages/mcp/scripts/gen-registry-stamp.mjs

@@ -1,60 +1,100 @@
 // Codegen: emit src/registry-stamp.generated.ts with a REGISTRY_STAMP hash of
-// the tool-specs REGISTRY CONTENT, so a build/ vs src/ skew (issue #447) is
+// the ENTIRE src/ tree, so a build/ vs src/ skew (issue #447) is detectable at
-// detectable at runtime.
+// runtime for ANY source file — not just tool-specs.ts.
 //
-// WHY hash the raw source text (not extracted structured data):
+// WHY hash the whole src tree (not just tool-specs.ts): the runtime tools are
-// SHARED_TOOL_SPECS carries `buildShape` functions (the input SCHEMAS) which are
+// assembled from far more than the spec registry — client.ts, the client/*
-// NOT serializable. The input schema is exactly one of the things that MUST stay
+// domain modules, comment-signal.ts and the drawio-* helpers all ship in build/
-// in sync between build/ and src/, so we cannot drop it from the hash. Rather
+// and are loaded by the in-app server. Hashing ONLY tool-specs.ts meant an edit
-// than probe zod with a fragile shim to reconstruct the schema shape, we hash the
+// to any of those (e.g. a behavioural fix in client.ts) left the stamp unchanged,
-// STABLE, deterministic source TEXT of tool-specs.ts. That text fully captures
+// so a stale build/ served the OLD code silently (issue #486). Hashing every
-// every field that must stay in sync — mcpName, inAppKey, description, tier,
+// src/**/*.ts closes that gap: any source edit changes the stamp.
-// catalogLine AND the buildShape bodies (input schemas) — with zero probing
-// fragility. Any edit to a spec (a renamed tool, a reworded description, a
-// changed schema field) changes the text and therefore the stamp.
 //
-// DETERMINISM: the hash is computed over the file bytes with line endings
+// WHY hash the raw source text (not extracted structured data): the tool input
-// normalized to LF and a single trailing newline stripped, so a CRLF checkout or
+// SCHEMAS live as `buildShape` functions which are NOT serializable, so we cannot
-// an editor's trailing-newline habit cannot make build/ and src/ disagree. No
+// reduce them to structured data without a fragile zod shim. Hashing the STABLE,
-// Date.now / randomness. The loader's dev-only stale-check (docmost-client.loader.ts)
+// deterministic source TEXT captures every field that must stay in sync with zero
-// re-runs THIS SAME normalization + sha256 over src/tool-specs.ts and compares to
+// probing fragility. Any edit to any source file changes the text → the stamp.
-// the built REGISTRY_STAMP; the two must compute identically.
+//
+// DETERMINISM: files are enumerated recursively, filtered to *.ts EXCLUDING
+// *.generated.ts (the codegen's OWN output — including it would create a
+// fixed-point cycle), and sorted by their POSIX-normalized path relative to src/
+// so the order is platform-independent. Each file contributes its relative path
+// AND its content with line endings normalized to LF and a single trailing
+// newline stripped, so a CRLF checkout or an editor's trailing-newline habit
+// cannot make build/ and src/ disagree. No Date.now / randomness. The loader's
+// dev-only stale-check (docmost-client.loader.ts) re-runs THIS SAME enumeration +
+// normalization + sha256 and compares to the built REGISTRY_STAMP; the two must
+// compute identically.
 //
 // This script runs from the `build` and `pretest` npm scripts BEFORE tsc, so
-// build/ always carries a stamp derived from the tool-specs.ts that was compiled.
+// build/ always carries a stamp derived from the src/ tree that was compiled.
 
 import { createHash } from 'node:crypto';
-import { readFileSync, writeFileSync } from 'node:fs';
+import { readdirSync, readFileSync, statSync, writeFileSync } from 'node:fs';
 import { fileURLToPath } from 'node:url';
-import { dirname, join } from 'node:path';
+import { dirname, join, relative, sep } from 'node:path';
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 const SRC_DIR = join(__dirname, '..', 'src');
-const TOOL_SPECS_PATH = join(SRC_DIR, 'tool-specs.ts');
 const OUT_PATH = join(SRC_DIR, 'registry-stamp.generated.ts');
 
 /**
- * Deterministic stamp of the tool-specs registry content. Kept as a plain
+ * Recursively enumerate every `*.ts` file under `dir`, EXCLUDING the codegen's
- * function (exported) so the algorithm has a single home; the loader duplicates
+ * own `*.generated.ts` output (a self-referential cycle otherwise). Returns
- * only the tiny normalize+sha256 steps because it lives in the CJS server build
+ * absolute paths, unsorted (the caller sorts by relative path for determinism).
- * and cannot import this ESM script. If you change the normalization here, mirror
+ * Kept as a plain exported function so the algorithm has a single home; the
- * it in apps/server/src/core/ai-chat/tools/docmost-client.loader.ts.
+ * loader duplicates it because it lives in the CJS server build and cannot import
+ * this ESM script. If you change the walk/filter here, mirror it in
+ * apps/server/src/core/ai-chat/tools/docmost-client.loader.ts.
  */
-export function computeRegistryStamp(toolSpecsSource) {
+export function collectStampFiles(dir) {
-  const normalized = toolSpecsSource.replace(/\r\n/g, '\n').replace(/\n$/, '');
+  const out = [];
-  return createHash('sha256').update(normalized, 'utf8').digest('hex');
+  for (const entry of readdirSync(dir)) {
+    const full = join(dir, entry);
+    if (statSync(full).isDirectory()) {
+      out.push(...collectStampFiles(full));
+    } else if (entry.endsWith('.ts') && !entry.endsWith('.generated.ts')) {
+      out.push(full);
+    }
+  }
+  return out;
+}
+
+/**
+ * Deterministic stamp of the whole src/ tree. Enumerate + sort by POSIX-relative
+ * path, then fold each file's relative path AND normalized content into one
+ * sha256. MUST stay byte-for-byte identical to the loader's recompute.
+ */
+export function computeRegistryStamp(srcDir) {
+  const files = collectStampFiles(srcDir)
+    .map((abs) => ({
+      rel: relative(srcDir, abs).split(sep).join('/'),
+      abs,
+    }))
+    .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0));
+  const hash = createHash('sha256');
+  for (const { rel, abs } of files) {
+    const normalized = readFileSync(abs, 'utf8')
+      .replace(/\r\n/g, '\n')
+      .replace(/\n$/, '');
+    hash.update(rel, 'utf8');
+    hash.update('\0', 'utf8');
+    hash.update(normalized, 'utf8');
+    hash.update('\0', 'utf8');
+  }
+  return hash.digest('hex');
 }
 
 function main() {
-  const source = readFileSync(TOOL_SPECS_PATH, 'utf8');
+  const stamp = computeRegistryStamp(SRC_DIR);
-  const stamp = computeRegistryStamp(source);
   const out =
     '// AUTO-GENERATED by scripts/gen-registry-stamp.mjs — DO NOT EDIT BY HAND.\n' +
-    '// A deterministic hash of src/tool-specs.ts content (tool names, descriptions,\n' +
+    '// A deterministic hash of the whole src/ tree (every src/**/*.ts except\n' +
-    '// tiers, catalog lines and input schemas). Regenerated on every build/pretest\n' +
+    '// *.generated.ts). Regenerated on every build/pretest so build/ always\n' +
-    '// so build/ always matches the compiled src. The in-app loader recomputes this\n' +
+    '// matches the compiled src. The in-app loader recomputes this from src and\n' +
-    '// from src and refuses to run on a mismatch (issue #447). This file is\n' +
+    '// refuses to run on a mismatch (issue #447/#486). This file is gitignored\n' +
-    '// gitignored and produced by the build — see .gitignore.\n' +
+    '// and produced by the build — see .gitignore.\n' +
     `export const REGISTRY_STAMP = ${JSON.stringify(stamp)};\n`;
   writeFileSync(OUT_PATH, out, 'utf8');
   // eslint-disable-next-line no-console

packages/mcp/src/client/context.ts

@@ -19,7 +19,10 @@ import {
   assertYjsEncodable,
   MutationResult,
 } from "../lib/collaboration.js";
-import { acquireCollabSession } from "../lib/collab-session.js";
+import {
+  acquireCollabSession,
+  isCollabAuthFailedError,
+} from "../lib/collab-session.js";
 import { withPageLock, isUuid } from "../lib/page-lock.js";
 import { getCollabToken, performLogin } from "../lib/auth-utils.js";
 import { formatDocmostAxiosError } from "./errors.js";
@@ -492,6 +495,37 @@ export abstract class DocmostClientContext {
     );
   }
 
+  /**
+   * Run a collab write and, on a Hocuspocus HANDSHAKE auth failure, self-heal
+   * once (#486). Symmetric to the HTTP-401 path in getCollabTokenWithReauth: the
+   * REST interceptor and login() already drop the cached collab token on a 401/
+   * 403, but a rejected WEBSOCKET handshake left the stale token in the cache, so
+   * every subsequent mutation kept re-presenting the same bad token for up to the
+   * collab-token TTL (minutes) with no self-heal. Here, when the write rejects
+   * with the tagged collab-auth error, we invalidate the cached token and retry
+   * the write EXACTLY once with a force-refreshed token. Not a loop: a second
+   * failure (or any non-auth error) propagates unchanged.
+   *
+   * `write` receives the token to use, so the retry can hand it a genuinely fresh
+   * one rather than re-running with the same stale string.
+   */
+  protected async writeWithCollabAuthRetry<T>(
+    collabToken: string,
+    write: (token: string) => Promise<T>,
+  ): Promise<T> {
+    try {
+      return await write(collabToken);
+    } catch (e) {
+      if (!isCollabAuthFailedError(e)) throw e;
+      // The WS handshake rejected our token: drop it from the cache so it can't
+      // be reused for the rest of the TTL, mint a fresh one (forceRefresh bypasses
+      // the cache and re-invokes the provider/login), and retry the write once.
+      this.collabTokenCache = null;
+      const fresh = await this.getCollabTokenWithReauth(true);
+      return await write(fresh);
+    }
+  }
+
   /**
    * Connect to the collaboration websocket, read the live doc, apply
    * `transform`, write the result, and wait for the server to persist it —
@@ -526,7 +560,11 @@ export abstract class DocmostClientContext {
     // unsyncedChanges/connectionLost ack logic live in CollabSession.mutate,
     // preserved verbatim from the old inline machine (incl. the #152 structural
     // diff that keeps a live editor's cursor anchored).
-    const session = await acquireCollabSession(pageId, collabToken, this.apiUrl, {
+    // Wrap in the collab-auth self-heal (#486): a rejected WS handshake drops the
+    // cached collab token and retries once with a fresh one (the retry passes the
+    // refreshed token down to acquireCollabSession via `token`).
+    return this.writeWithCollabAuthRetry(collabToken, async (token) => {
+      const session = await acquireCollabSession(pageId, token, this.apiUrl, {
         // Only the actual 25s collab connect timeout emits this — the connect-vs-
         // unload signal; the other failure paths must NOT emit it.
         onConnectTimeout: () =>
@@ -539,6 +577,7 @@ export abstract class DocmostClientContext {
         session.destroy("mutate failed");
         throw e;
       }
+    });
   }
 
   /**
@@ -667,7 +706,11 @@ export abstract class DocmostClientContext {
     transform: (doc: any) => any,
   ): Promise<{ doc?: any; verify?: any }> {
     const pageUuid = await this.resolvePageId(pageId);
-    return mutatePageContent(pageUuid, collabToken, apiUrl, transform);
+    // #486: on a rejected collab-WS handshake, invalidate + refresh the token and
+    // retry the write once (symmetric to the HTTP-401 reauth path).
+    return this.writeWithCollabAuthRetry(collabToken, (token) =>
+      mutatePageContent(pageUuid, token, apiUrl, transform),
+    );
   }
 
   /**
@@ -687,7 +730,11 @@ export abstract class DocmostClientContext {
     apiUrl: string,
   ): Promise<{ doc?: any; verify?: any }> {
     const pageUuid = await this.resolvePageId(pageId);
-    return replacePageContent(pageUuid, doc, collabToken, apiUrl);
+    // #486: on a rejected collab-WS handshake, invalidate + refresh the token and
+    // retry the write once (symmetric to the HTTP-401 reauth path).
+    return this.writeWithCollabAuthRetry(collabToken, (token) =>
+      replacePageContent(pageUuid, doc, token, apiUrl),
+    );
   }
 
   /**

packages/mcp/src/client/read.ts

@@ -127,8 +127,13 @@ export function ReadMixin<TBase extends GConstructor<DocmostClientContext>>(Base
           "listPages: tree mode requires a spaceId (a page tree is scoped to one space). Pass spaceId, or omit tree to get the recent-pages list.",
         );
       }
-      const { pages } = await this.enumerateSpacePages(spaceId);
+      // #486: propagate `truncated` (same pattern as check_new_comments). The old
-      return buildPageTree(pages);
+      // code dropped it, so a caller handed an INCOMPLETE tree (the stdio-fallback
+      // BFS hit its node cap) had no way to know pages were missing. Return the
+      // tree alongside the flag; the primary /pages/tree path is uncapped so this
+      // is false there.
+      const { pages, truncated } = await this.enumerateSpacePages(spaceId);
+      return { tree: buildPageTree(pages), truncated };
     }
 
     const clampedLimit = Math.max(1, Math.min(100, limit));

packages/mcp/src/index.ts

@@ -4,7 +4,6 @@ import { readFileSync } from "fs";
 import { fileURLToPath } from "url";
 import { dirname, join } from "path";
 import { DocmostClient, DocmostMcpConfig } from "./client.js";
-import { parseNodeArg } from "@docmost/prosemirror-markdown";
 import { searchShapes } from "./lib/drawio-shapes.js";
 import { getGuideSection } from "./lib/drawio-guide.js";
 import { SHARED_TOOL_SPECS, SharedToolSpec } from "./tool-specs.js";

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

@@ -37,6 +37,25 @@ const CONNECT_TIMEOUT_MS = 25000;
 /** Time we wait for the server to acknowledge our write before giving up. */
 const PERSIST_TIMEOUT_MS = 20000;
 
+/**
+ * Marker property set on the Error thrown when the Hocuspocus handshake REJECTS
+ * our collab token (onAuthenticationFailed). The client wraps content writes so
+ * that on this specific failure it invalidates its cached collab token and
+ * retries once with a fresh one — symmetric to the HTTP-401 reauth path (#486).
+ * A plain message-match would be brittle; a tagged property is unambiguous and
+ * survives teardown (which rejects pending ops with this SAME error object).
+ */
+const COLLAB_AUTH_FAILED_MARKER = "collabAuthFailed";
+
+/** True when `e` is the tagged collab-WS auth-failure error (see marker above). */
+export function isCollabAuthFailedError(e: unknown): boolean {
+  return !!(
+    e &&
+    typeof e === "object" &&
+    (e as Record<string, unknown>)[COLLAB_AUTH_FAILED_MARKER] === true
+  );
+}
+
 /**
  * Tunables, read fresh from the environment on every acquire so tests (and a
  * live rollback) can change them without reloading the module. Mirrors how
@@ -302,10 +321,13 @@ export class CollabSession {
           this.openResolve?.();
         },
         onAuthenticationFailed: () => {
-          this.teardown(
+          // Tag the error so the client can tell a REJECTED collab token apart
-            new Error("Authentication failed for collaboration connection"),
+          // from a generic disconnect and invalidate + refresh it (#486).
-            true,
+          const err = new Error(
-          );
+            "Authentication failed for collaboration connection",
+          ) as Error & { [COLLAB_AUTH_FAILED_MARKER]?: boolean };
+          err[COLLAB_AUTH_FAILED_MARKER] = true;
+          this.teardown(err, true);
         },
       });
     });

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

@@ -10,7 +10,7 @@
 // exactly mxGraph's convention for a child of a container, so they map across
 // directly. Container sizes are computed by ELK; leaf sizes are preserved.
 
-import ELK from "elkjs/lib/elk.bundled.js";
+import { Worker } from "node:worker_threads";
 import { JSDOM } from "jsdom";
 import { normalizeInput, parseCells, type DrawioCell } from "./drawio-xml.js";
 
@@ -18,22 +18,33 @@ import { normalizeInput, parseCells, type DrawioCell } from "./drawio-xml.js";
 const DEFAULT_W = 140;
 const DEFAULT_H = 60;
 
-// DoS bounds for the in-process ELK layout. The mxGraph XML is LLM-supplied
+// DoS bounds for the ELK layout. The mxGraph XML is LLM-supplied (layout:"elk"
-// (layout:"elk" in drawioCreate/drawioUpdate) and elkjs runs synchronously on
+// in drawioCreate/drawioUpdate). elkjs' layout() returns a Promise but runs the
-// the MCP server's event loop, so an unbounded graph would block it for
+// crossing-minimisation SYNCHRONOUSLY — it blocks whatever thread it runs on for
-// seconds-to-minutes. A ~1MB XML (well under the stage-1 16MB cap) can carry
+// the whole pass. A ~1MB XML (well under the stage-1 16MB cap) can carry
-// thousands of nodes. We cap the graph size and race the layout against a
+// thousands of nodes. We (a) cap the graph size before ever calling ELK and
-// wall-clock timeout; on either bound we fall back to the ORIGINAL model, the
+// (b) run the layout in a WORKER THREAD so the main event loop stays free, with
-// same best-effort contract the catch already honours.
+// the wall-clock timeout enforced by terminating that worker. On either bound we
+// fall back to the ORIGINAL model, the same best-effort contract the catch honours.
 //   - 500 nodes lays out in well under a second; beyond that ELK cost climbs
 //     steeply, so refuse and leave the (already-valid) model untouched.
 //   - Edges dominate the layered-crossing cost, so allow a bit more headroom
 //     (1000) than nodes but still bound them.
-//   - 5s is generous for any graph within the caps yet short enough that a
+//   - The timeout is a HARD kill of the worker thread — the only way to interrupt
-//     pathological input can never wedge the server.
+//     synchronous JS. The in-process setTimeout race we used before was an
+//     illusion: the timer could never fire while the SAME thread was blocked
+//     inside elkjs, so it "protected" nothing. Now the timer runs on the main
+//     thread while ELK runs on the worker, so it can actually fire and terminate.
 const ELK_MAX_NODES = 500;
 const ELK_MAX_EDGES = 1000;
-const ELK_TIMEOUT_MS = 5000;
+// Wall-clock ceiling for a single layout pass. Overridable for tests (a tiny
+// value forces the terminate-on-timeout path deterministically); a non-positive
+// or unparseable override falls back to the default.
+const ELK_TIMEOUT_DEFAULT_MS = 5000;
+function resolveElkTimeoutMs(): number {
+  const raw = Number(process.env.DRAWIO_ELK_TIMEOUT_MS);
+  return Number.isFinite(raw) && raw > 0 ? Math.floor(raw) : ELK_TIMEOUT_DEFAULT_MS;
+}
 
 // Spacing is set >=150px on purpose so an ELK layout never trips the linter's
 // "gap between adjacent shapes < 150px" quality warning (acceptance #3).
@@ -78,13 +89,57 @@ interface ElkGraph extends ElkNode {
   edges?: ElkEdge[];
 }
 
+/**
+ * Run one ELK layered layout on a worker thread and resolve with the laid-out
+ * graph. The timeout is enforced by `worker.terminate()` — a HARD kill, which is
+ * the only way to interrupt elkjs' synchronous crossing-minimisation once it has
+ * started. Rejects on timeout, worker error, or an early exit; the caller treats
+ * any rejection as "keep the original model" (best-effort layout).
+ */
+function layoutInWorker(graph: ElkGraph, timeoutMs: number): Promise<ElkGraph> {
+  return new Promise((resolve, reject) => {
+    const worker = new Worker(
+      new URL("./drawio-layout.worker.js", import.meta.url),
+      { workerData: { graph } },
+    );
+    let settled = false;
+    const finish = (fn: () => void) => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+      // Always tear the worker down: on the happy path so it does not linger,
+      // on timeout so the blocked synchronous ELK run is actually interrupted.
+      void worker.terminate();
+      fn();
+    };
+    const timer = setTimeout(
+      () => finish(() => reject(new Error("ELK layout timed out"))),
+      timeoutMs,
+    );
+    worker.once("message", (msg: { ok?: boolean; laid?: ElkGraph; error?: string }) => {
+      finish(() =>
+        msg?.ok
+          ? resolve(msg.laid as ElkGraph)
+          : reject(new Error(msg?.error ?? "ELK layout failed")),
+      );
+    });
+    worker.once("error", (err) => finish(() => reject(err)));
+    worker.once("exit", (code) => {
+      // A clean exit after we already settled is normal (terminate()); only an
+      // unexpected early exit while still pending is a failure.
+      if (settled) return;
+      finish(() => reject(new Error(`ELK worker exited early (code ${code})`)));
+    });
+  });
+}
+
 /**
  * Apply an ELK layered layout to a drawio input and return a full mxGraphModel
  * string with rewritten geometry. Accepts the same three input forms as
  * drawioCreate (a bare model, an <mxfile>, or a <mxCell> list). Async because
- * elkjs' layout() is promise-based. On any layout failure the ORIGINAL
+ * the layout runs on a worker thread. On any layout failure (including a
- * (normalized) model is returned unchanged — layout is best-effort polish, never
+ * terminate-on-timeout) the ORIGINAL (normalized) model is returned unchanged —
- * a reason to fail the write.
+ * layout is best-effort polish, never a reason to fail the write.
  */
 export async function applyElkLayout(inputXml: string): Promise<string> {
   const modelXml = normalizeInput(inputXml);
@@ -150,26 +205,14 @@ export async function applyElkLayout(inputXml: string): Promise<string> {
   };
 
   let laid: ElkGraph;
-  let timer: ReturnType<typeof setTimeout> | undefined;
   try {
-    // elkjs ships a CJS default export whose interop shape varies across
+    // Run the (synchronous-under-the-hood) ELK pass on a worker thread so the
-    // module systems; resolve the real constructor at runtime, then cast (the
+    // main event loop is never blocked, and enforce the wall-clock ceiling by
-    // runtime call is verified — see the layout unit test).
+    // terminating that worker on timeout. A graph under the node/edge caps but
-    const Ctor: any = (ELK as any).default ?? ELK;
+    // still pathologically slow is hard-killed instead of wedging anything.
-    const elk = new Ctor();
+    laid = await layoutInWorker(graph, resolveElkTimeoutMs());
-    // Race the layout against a wall-clock timeout so a graph that is under the
-    // node/edge caps but still pathologically slow can never wedge the server.
-    const timeout = new Promise<never>((_, reject) => {
-      timer = setTimeout(
-        () => reject(new Error("ELK layout timed out")),
-        ELK_TIMEOUT_MS,
-      );
-    });
-    laid = (await Promise.race([elk.layout(graph as any), timeout])) as ElkGraph;
   } catch {
     return modelXml; // best-effort: keep the model as-is on timeout or ELK failure
-  } finally {
-    if (timer) clearTimeout(timer);
   }
 
   // Collect computed geometry per node id (coords are parent-relative already).

packages/mcp/src/lib/drawio-layout.worker.ts

@@ -0,0 +1,36 @@
+// Worker-thread entry for the ELK layered layout (issue #486, commit 1).
+//
+// elkjs' layout() returns a Promise but runs the actual crossing-minimisation
+// SYNCHRONOUSLY — it blocks whatever thread it runs on for the whole pass. On
+// the in-app MCP host that thread used to be the main NestJS event loop, so a
+// pathological graph at the node/edge cap could wedge ALL HTTP/SSE/loopback
+// traffic while it churned. Running it HERE, on a dedicated worker thread, keeps
+// the main loop free; the parent enforces the wall-clock timeout by calling
+// `worker.terminate()` — the only way to interrupt synchronous JS — since the
+// in-process `setTimeout` race the parent used before could never fire while the
+// same thread was blocked inside elkjs.
+import { parentPort, workerData } from "node:worker_threads";
+import ELK from "elkjs/lib/elk.bundled.js";
+
+interface WorkerInput {
+  graph: unknown;
+}
+
+const { graph } = (workerData ?? {}) as WorkerInput;
+
+// elkjs ships a CJS default export whose interop shape varies across module
+// systems; resolve the real constructor at runtime (same as the parent did).
+const Ctor: any = (ELK as any).default ?? ELK;
+const elk = new Ctor();
+
+elk
+  .layout(graph as any)
+  .then((laid: unknown) => {
+    parentPort?.postMessage({ ok: true, laid });
+  })
+  .catch((err: unknown) => {
+    parentPort?.postMessage({
+      ok: false,
+      error: err instanceof Error ? err.message : String(err),
+    });
+  });

packages/mcp/src/tool-specs.ts

@@ -923,9 +923,10 @@ export const SHARED_TOOL_SPECS = {
       'List the most recent pages (ordered by updatedAt, descending), ' +
       'optionally scoped to a single space. Returns a bounded list (default ' +
       '50, max 100) — use search for lookups in large spaces. tree:true (with ' +
-      "spaceId) returns the space's full page hierarchy as a nested tree, but " +
+      "spaceId) returns { tree, truncated } — the space's full page hierarchy " +
-      'is DEPRECATED — use getTree instead (leaner nodes, plus rootPageId / ' +
+      'as a nested tree, plus a `truncated` flag that is true when the tree was ' +
-      'maxDepth).',
+      'capped and is INCOMPLETE — but is DEPRECATED, use getTree instead ' +
+      '(leaner nodes, plus rootPageId / maxDepth).',
     tier: 'core',
     catalogLine:
       "listPages — list recent pages (tree:true is deprecated; use getTree for the hierarchy).",

packages/mcp/test/mock/collab-auth-failure-reset.test.mjs

@@ -0,0 +1,137 @@
+// Unit tests for the collab-token reset on a Hocuspocus WS auth failure (#486).
+//
+// Before this fix the cached collab token (#435) was dropped ONLY on an HTTP
+// 401/403 (the REST interceptor + login()); a rejected collab-WEBSOCKET handshake
+// left the stale token in the cache, so every subsequent mutation re-presented
+// the SAME bad token for up to the collab-token TTL (minutes) with no self-heal.
+//
+// The fix wraps collab writes in `writeWithCollabAuthRetry`: when the write
+// rejects with the tagged collab-auth error (collab-session.ts's
+// onAuthenticationFailed), it invalidates the cached token and retries the write
+// ONCE with a force-refreshed token — symmetric to the HTTP-401 path.
+//
+// writeWithCollabAuthRetry / getCollabTokenWithReauth are protected in TS but
+// plain methods on the compiled build, so the tests call them directly (same
+// convention as collab-token-cache.test.mjs).
+import { test, afterEach } from "node:test";
+import assert from "node:assert/strict";
+import { DocmostClient } from "../../build/client.js";
+
+const ENV_KEY = "MCP_COLLAB_TOKEN_TTL_MS";
+afterEach(() => {
+  delete process.env[ENV_KEY];
+});
+
+// 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-${n}`;
+  };
+  return {
+    fn,
+    get calls() {
+      return n;
+    },
+  };
+}
+
+// The tagged error collab-session.ts throws on a rejected WS handshake.
+function collabAuthError() {
+  const err = new Error("Authentication failed for collaboration connection");
+  err.collabAuthFailed = true;
+  return err;
+}
+
+test("a WS auth failure clears the cached token and retries the write with a FRESH one (#486)", async () => {
+  process.env[ENV_KEY] = "300000"; // 5 min: the cache is warm across the burst.
+  const p = countingProvider();
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+    getCollabToken: p.fn,
+  });
+
+  // Warm the cache the way a real write would (mints provider-token-1).
+  const initial = await client.getCollabTokenWithReauth();
+  assert.equal(initial, "provider-token-1");
+  assert.equal(p.calls, 1);
+
+  const tokensSeen = [];
+  const write = async (token) => {
+    tokensSeen.push(token);
+    // The FIRST attempt (with the stale cached token) fails the WS handshake;
+    // the retry (with a fresh token) succeeds.
+    if (tokensSeen.length === 1) throw collabAuthError();
+    return `written-with:${token}`;
+  };
+
+  const result = await client.writeWithCollabAuthRetry(initial, write);
+
+  assert.equal(tokensSeen.length, 2, "write attempted exactly twice (one retry)");
+  assert.equal(tokensSeen[0], "provider-token-1", "first attempt used the stale token");
+  assert.equal(
+    tokensSeen[1],
+    "provider-token-2",
+    "retry used a FRESH force-refreshed token, not the stale cached one",
+  );
+  assert.equal(result, "written-with:provider-token-2", "the retry's result wins");
+  assert.equal(p.calls, 2, "exactly one extra mint for the retry — no loop");
+
+  // The cache now holds the fresh token, so a subsequent op reuses it (proving
+  // the stale token was evicted and the fresh one cached, not re-minted).
+  const next = await client.getCollabTokenWithReauth();
+  assert.equal(next, "provider-token-2", "the fresh token replaced the stale cache");
+  assert.equal(p.calls, 2, "served from cache — provider not re-invoked");
+});
+
+test("a successful write is NOT retried and mints nothing extra", 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 initial = await client.getCollabTokenWithReauth(); // provider-token-1
+  let attempts = 0;
+  const result = await client.writeWithCollabAuthRetry(initial, async (token) => {
+    attempts++;
+    return `ok:${token}`;
+  });
+
+  assert.equal(attempts, 1, "no retry on success");
+  assert.equal(result, "ok:provider-token-1");
+  assert.equal(p.calls, 1, "no extra mint");
+});
+
+test("a NON-auth write error propagates unchanged (no reset, no retry)", 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 initial = await client.getCollabTokenWithReauth(); // provider-token-1
+  let attempts = 0;
+  await assert.rejects(
+    client.writeWithCollabAuthRetry(initial, async () => {
+      attempts++;
+      throw new Error("collab connection closed before persist"); // NOT tagged.
+    }),
+    /closed before persist/,
+  );
+
+  assert.equal(attempts, 1, "a non-auth error is not retried");
+  assert.equal(p.calls, 1, "the cache is untouched -> no fresh mint");
+
+  // Cache still holds the original token (was never invalidated).
+  const still = await client.getCollabTokenWithReauth();
+  assert.equal(still, "provider-token-1");
+  assert.equal(p.calls, 1);
+});

packages/mcp/test/mock/list-pages-tree-truncated.test.mjs

@@ -0,0 +1,55 @@
+// Unit test: listPages tree mode must propagate the `truncated` flag (#486).
+//
+// enumerateSpacePages returns { pages, truncated } — truncated is true ONLY when
+// the stdio-fallback BFS hit its node cap (the primary /pages/tree path is
+// uncapped). The old tree-mode listPages destructured only `pages` and returned a
+// bare tree, dropping `truncated`, so a caller handed an INCOMPLETE tree had no
+// way to know pages were missing. The fix returns { tree, truncated } (same
+// pattern check_new_comments uses).
+//
+// Reaching the real cap (MAX_NODES = 10000) in a mock is impractical, so we stub
+// enumerateSpacePages directly to assert the flag is threaded through verbatim.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { DocmostClient } from "../../build/client.js";
+
+function stubClient() {
+  const client = new DocmostClient({
+    apiUrl: "http://127.0.0.1:1/api",
+    getToken: async () => "access",
+  });
+  // No network: the tree path only calls ensureAuthenticated + enumerateSpacePages.
+  client.ensureAuthenticated = async () => {};
+  return client;
+}
+
+const onePage = [{ id: "r1", title: "Root", parentPageId: null }];
+
+test("tree mode carries truncated:true when the enumeration truncated (#486)", async () => {
+  const client = stubClient();
+  client.enumerateSpacePages = async () => ({ pages: onePage, truncated: true });
+
+  const res = await client.listPages("space-1", 50, true);
+
+  assert.equal(res.truncated, true, "the truncated flag is threaded through");
+  assert.ok(Array.isArray(res.tree), "the built tree rides alongside the flag");
+  assert.equal(res.tree[0].id, "r1");
+});
+
+test("tree mode carries truncated:false for a complete enumeration", async () => {
+  const client = stubClient();
+  client.enumerateSpacePages = async () => ({ pages: onePage, truncated: false });
+
+  const res = await client.listPages("space-1", 50, true);
+
+  assert.equal(res.truncated, false);
+  assert.equal(res.tree[0].id, "r1");
+});
+
+test("tree mode still requires a spaceId", async () => {
+  const client = stubClient();
+  await assert.rejects(
+    client.listPages(undefined, 50, true),
+    /tree mode requires a spaceId/,
+  );
+});

packages/mcp/test/mock/pagination-cursor.test.mjs

@@ -184,12 +184,14 @@ test("enumerateSpacePages (via listPages tree) uses one /pages/tree request", as
   });
 
   const client = new DocmostClient(baseURL, "user@example.com", "pw");
-  // listPages tree:true -> enumerateSpacePages(spaceId) -> buildPageTree.
+  // listPages tree:true -> enumerateSpacePages(spaceId) -> { tree, truncated }.
-  const tree = await client.listPages("space-1", 50, true);
+  const { tree, truncated } = await client.listPages("space-1", 50, true);
 
   assert.equal(treeRequests, 1, "exactly one /pages/tree request for the space");
   assert.equal(sidebarRequests, 0, "no per-node sidebar BFS requests");
   assert.deepEqual(treeBody, { spaceId: "space-1" }, "space scope posts spaceId only");
+  // The uncapped /pages/tree path is never truncated (#486).
+  assert.equal(truncated, false, "primary /pages/tree path is not truncated");
   // buildPageTree nests c1 under r1; two roots at the top level.
   assert.equal(tree.length, 2, "two root nodes");
   const r1 = tree.find((n) => n.id === "r1");
@@ -249,7 +251,7 @@ test("enumerateSpacePages falls back to the cursor BFS on /pages/tree 404", asyn
   });
 
   const client = new DocmostClient(baseURL, "user@example.com", "pw");
-  const tree = await client.listPages("space-1", 50, true);
+  const { tree, truncated } = await client.listPages("space-1", 50, true);
 
   assert.ok(treeRequests >= 1, "the tree endpoint was attempted first");
   assert.deepEqual(
@@ -257,6 +259,8 @@ test("enumerateSpacePages falls back to the cursor BFS on /pages/tree 404", asyn
     ["<root>", "r1"],
     "fell back to the sidebar BFS: roots then the root's children",
   );
+  // Small fallback walk well under the node cap -> not truncated (#486).
+  assert.equal(truncated, false, "fallback BFS below the cap is not truncated");
   assert.equal(tree.length, 1, "one root in the built tree");
   assert.equal(tree[0].children[0].id, "c1", "leaf nested via the BFS");
 });

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

@@ -101,6 +101,88 @@ test("DoS guard: a graph over the node cap is returned unchanged, quickly", asyn
   assert.ok(dt < 2000, `cap path should be fast, took ${dt}ms`);
 });
 
+/** Build a layered DAG near the caps: `n` vertices, up to ~2 edges each into the
+ * next layer of `layerSize`. Used as a real worst-case graph for the benchmark. */
+function layeredGraph(n, layerSize) {
+  let cells = "";
+  for (let i = 2; i < 2 + n; i++) {
+    cells +=
+      `<mxCell id="${i}" value="N${i}" style="rounded=1;html=1;" vertex="1" parent="1">` +
+      `<mxGeometry x="10" y="10" width="120" height="60" as="geometry"/></mxCell>`;
+  }
+  let ei = 0;
+  for (let i = 2; i < 2 + n; i++) {
+    for (const off of [layerSize, layerSize + 1]) {
+      const t = i + off;
+      if (t < 2 + n) cells += `<mxCell id="e${ei++}" edge="1" parent="1" source="${i}" target="${t}"><mxGeometry relative="1" as="geometry"/></mxCell>`;
+    }
+  }
+  return (
+    '<mxGraphModel><root><mxCell id="0"/><mxCell id="1" parent="0"/>' +
+    cells +
+    "</root></mxGraphModel>"
+  );
+}
+
+test("terminate-on-timeout: a layout that exceeds the wall-clock ceiling is hard-killed and the original model is returned (#486)", async () => {
+  // A 1ms ceiling fires before the worker can even finish loading elkjs, so the
+  // parent must terminate() the worker and fall back to the ORIGINAL model. On
+  // the OLD in-process race this timer could never fire while the SAME thread was
+  // blocked inside elkjs — the fallback path was unreachable; here it works.
+  const prev = process.env.DRAWIO_ELK_TIMEOUT_MS;
+  process.env.DRAWIO_ELK_TIMEOUT_MS = "1";
+  try {
+    const model = layeredGraph(400, 20);
+    const t0 = Date.now();
+    const laid = await applyElkLayout(model);
+    const dt = Date.now() - t0;
+    // Original geometry is preserved verbatim: every vertex is still stacked at
+    // (10,10), proving NO ELK coordinates were applied (the pass was killed).
+    const verts = parseCells(laid).filter((c) => c.vertex);
+    assert.equal(verts.length, 400, "all vertices survived the fallback");
+    for (const v of verts) {
+      assert.equal(v.geometry.x, 10, "x untouched -> layout was terminated");
+      assert.equal(v.geometry.y, 10, "y untouched -> layout was terminated");
+    }
+    // The kill is prompt: terminate() returns the call well under the natural
+    // layout time for a 400-node graph.
+    assert.ok(dt < 2000, `terminate path should be prompt, took ${dt}ms`);
+  } finally {
+    if (prev === undefined) delete process.env.DRAWIO_ELK_TIMEOUT_MS;
+    else process.env.DRAWIO_ELK_TIMEOUT_MS = prev;
+  }
+});
+
+test("benchmark guard: a worst-case graph AT the cap lays out without wedging the main event loop (#486)", async () => {
+  // ~500 nodes / ~1000 edges — a real worst case at the node/edge caps. The
+  // layout runs on a WORKER thread, so the MAIN event loop must stay responsive
+  // throughout: a timer scheduled on the main thread keeps firing while ELK
+  // churns. On the OLD synchronous-on-main-thread code this counter would be
+  // pinned at 0 for the whole layout (event loop wedged) — exactly the prod fire.
+  const model = layeredGraph(500, 20);
+  let mainLoopTicks = 0;
+  const iv = setInterval(() => {
+    mainLoopTicks++;
+  }, 2);
+  const t0 = Date.now();
+  const laid = await applyElkLayout(model);
+  const dt = Date.now() - t0;
+  clearInterval(iv);
+
+  assert.ok(
+    mainLoopTicks > 0,
+    "main event loop must stay responsive while ELK runs on the worker",
+  );
+  // Benchmark guard: the worst-case graph actually LAYS OUT within the default
+  // ceiling (it did not fall back). At least one vertex moved off the stack.
+  const verts = parseCells(laid).filter((c) => c.vertex);
+  assert.equal(verts.length, 500, "all vertices survived");
+  const moved = verts.some((v) => v.geometry.x !== 10 || v.geometry.y !== 10);
+  assert.ok(moved, "layout was applied (did not time out / fall back)");
+  // Sanity ceiling well under the 5s wall-clock timeout.
+  assert.ok(dt < 5000, `worst-case layout should be under the ceiling, took ${dt}ms`);
+});
+
 test("layout is best-effort: an empty/degenerate model is returned intact", async () => {
   const model =
     '<mxGraphModel><root><mxCell id="0"/><mxCell id="1" parent="0"/></root></mxGraphModel>';

packages/mcp/test/unit/registry-stamp.test.mjs

@@ -1,101 +1,220 @@
 import { test } from "node:test";
 import assert from "node:assert/strict";
-import { createHash } from "node:crypto";
+import {
-import { readFileSync } from "node:fs";
+  mkdtempSync,
+  mkdirSync,
+  writeFileSync,
+  rmSync,
+  readdirSync,
+  statSync,
+  readFileSync,
+} from "node:fs";
+import { tmpdir } from "node:os";
 import { fileURLToPath } from "node:url";
-import { dirname, join } from "node:path";
+import { dirname, join, relative, sep } from "node:path";
+import { createHash } from "node:crypto";
 
 import { computeRegistryStamp } from "../../scripts/gen-registry-stamp.mjs";
 import { REGISTRY_STAMP } from "../../build/index.js";
 
-// Guard tests for the build/src-skew stamp (issue #447). The codegen script
+// Guard tests for the build/src-skew stamp (issues #447/#486). The codegen script
-// exports `computeRegistryStamp(sourceText)` — a sha256 over normalized source
+// exports `computeRegistryStamp(srcDir)` — a sha256 over the WHOLE src/ tree
-// text (CRLF->LF, single trailing newline stripped). The in-app loader
+// (every src/**/*.ts EXCEPT *.generated.ts), each file folded in as its
-// (apps/server/.../docmost-client.loader.ts) DUPLICATES that normalize+sha256 to
+// POSIX-relative path + its normalized content (CRLF->LF, single trailing newline
-// recompute the stamp from src and refuse a stale build. These tests pin the
+// stripped). Hashing the whole tree (not just tool-specs.ts) is #486: an edit to
-// algorithm's behaviour AND assert the built stamp matches the current src, so a
+// client.ts / a client/* module without a rebuild must ALSO redden. The in-app
-// stale generated file OR a normalize divergence reddens.
+// loader (apps/server/.../docmost-client.loader.ts) DUPLICATES this enumerate+
+// normalize+sha256 to refuse a stale build. These tests pin the algorithm and
+// assert the built stamp matches the current src.
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
-const TOOL_SPECS_PATH = join(__dirname, "..", "..", "src", "tool-specs.ts");
+const SRC_DIR = join(__dirname, "..", "..", "src");
 
-test("computeRegistryStamp is deterministic: same input -> same hash", () => {
+// Build a throwaway src/ tree from a { relPath: content } map and return its dir.
-  const input = "export const X = 1;\nexport const Y = 2;\n";
+function makeSrcTree(files) {
-  assert.equal(computeRegistryStamp(input), computeRegistryStamp(input));
+  const root = mkdtempSync(join(tmpdir(), "mcp-stamp-tree-"));
+  const src = join(root, "src");
+  for (const [rel, content] of Object.entries(files)) {
+    const full = join(src, rel);
+    mkdirSync(dirname(full), { recursive: true });
+    writeFileSync(full, content, "utf8");
+  }
+  return { src, cleanup: () => rmSync(root, { recursive: true, force: true }) };
+}
+
+test("computeRegistryStamp is deterministic: same tree -> same hash", () => {
+  const a = makeSrcTree({ "tool-specs.ts": "export const X = 1;\n" });
+  const b = makeSrcTree({ "tool-specs.ts": "export const X = 1;\n" });
+  try {
+    assert.equal(computeRegistryStamp(a.src), computeRegistryStamp(b.src));
+  } finally {
+    a.cleanup();
+    b.cleanup();
+  }
 });
 
 test("computeRegistryStamp returns a 64-char lowercase hex sha256", () => {
-  const stamp = computeRegistryStamp("anything");
+  const t = makeSrcTree({ "tool-specs.ts": "anything\n" });
-  assert.match(stamp, /^[0-9a-f]{64}$/);
+  try {
+    assert.match(computeRegistryStamp(t.src), /^[0-9a-f]{64}$/);
+  } finally {
+    t.cleanup();
+  }
 });
 
-test("normalizes CRLF vs LF: the same content hashes equal", () => {
+// #486 CORE: an edit to a NON-tool-specs source file (client.ts) must change the
-  const lf = "line1\nline2\nline3";
+// stamp. Under the old single-file (tool-specs.ts only) hash this edit was
-  const crlf = "line1\r\nline2\r\nline3";
+// invisible and a stale build/ served the old client.ts silently.
-  assert.equal(computeRegistryStamp(crlf), computeRegistryStamp(lf));
+test("editing client.ts (not tool-specs.ts) changes the stamp (#486)", () => {
+  const before = makeSrcTree({
+    "tool-specs.ts": "export const SPECS = 1;\n",
+    "client.ts": "export const impl = 'v1';\n",
   });
-
+  const after = makeSrcTree({
-test("normalizes a trailing newline: with/without a final \\n hashes equal", () => {
+    "tool-specs.ts": "export const SPECS = 1;\n",
-  const noTrailing = "alpha\nbeta";
+    "client.ts": "export const impl = 'v2';\n",
-  const trailing = "alpha\nbeta\n";
-  assert.equal(computeRegistryStamp(trailing), computeRegistryStamp(noTrailing));
   });
-
+  try {
-test("a CRLF checkout WITH a trailing CRLF still hashes equal to bare LF", () => {
-  // A worst-case Windows checkout: CRLF line endings + a trailing CRLF. Both the
-  // \r\n->\n replace and the trailing-newline strip must apply for parity.
-  const bare = "alpha\nbeta";
-  const crlfTrailing = "alpha\r\nbeta\r\n";
-  assert.equal(
-    computeRegistryStamp(crlfTrailing),
-    computeRegistryStamp(bare),
-  );
-});
-
-test("a real content change hashes differently", () => {
-  const before = "export const description = 'search a page';\n";
-  const after = "export const description = 'search a PAGE';\n";
-  assert.notEqual(computeRegistryStamp(before), computeRegistryStamp(after));
-});
-
-// Only a SINGLE trailing newline is stripped — a second blank line is content and
-// must change the hash. This pins the exact `/\n$/` semantics the loader mirrors.
-test("only ONE trailing newline is stripped (two differ from one)", () => {
     assert.notEqual(
-    computeRegistryStamp("x\n"),
+      computeRegistryStamp(before.src),
-    computeRegistryStamp("x\n\n"),
+      computeRegistryStamp(after.src),
+      "a client.ts edit with an unchanged tool-specs.ts must move the stamp",
     );
+  } finally {
+    before.cleanup();
+    after.cleanup();
+  }
 });
 
-// Cross-impl equality against a fixed, documented input. The SAME literal input
+test("editing a nested client/* module changes the stamp", () => {
-// and expected hash are asserted in the server-side jest test
+  const before = makeSrcTree({
-// (docmost-client.loader.spec.ts). If either side's normalize+sha256 ever
+    "tool-specs.ts": "x\n",
-// diverges, one of the two tests reddens. Input exercises BOTH normalize steps.
+    "client/read.ts": "export const READ = 1;\n",
-test("fixed-input hash matches the documented cross-impl value", () => {
+  });
-  const FIXED_INPUT = "line1\r\nline2\n";
+  const after = makeSrcTree({
-  const EXPECTED =
+    "tool-specs.ts": "x\n",
-    "683376e290829b482c2655745caffa7a1dccfa10afaa62dac2b42dd6c68d0f83";
+    "client/read.ts": "export const READ = 2;\n",
-  assert.equal(computeRegistryStamp(FIXED_INPUT), EXPECTED);
+  });
+  try {
+    assert.notEqual(
+      computeRegistryStamp(before.src),
+      computeRegistryStamp(after.src),
+    );
+  } finally {
+    before.cleanup();
+    after.cleanup();
+  }
 });
 
-// DESYNC GUARD (covers reviewer suggestion 2). Recompute the stamp from the
+// *.generated.ts is EXCLUDED (else the codegen's own output is a fixed-point
-// actual src/tool-specs.ts and assert it equals the REGISTRY_STAMP baked into the
+// cycle): adding/removing/changing it must not move the stamp.
-// freshly-built build/index.js. This reddens if the generated file is stale OR if
+test("*.generated.ts is excluded from the stamp", () => {
-// the codegen normalize ever diverges from what produced the built stamp.
+  const without = makeSrcTree({ "tool-specs.ts": "x\n" });
-test("built REGISTRY_STAMP equals the stamp recomputed from src/tool-specs.ts", () => {
+  const withGen = makeSrcTree({
-  const source = readFileSync(TOOL_SPECS_PATH, "utf8");
+    "tool-specs.ts": "x\n",
-  assert.equal(computeRegistryStamp(source), REGISTRY_STAMP);
+    "registry-stamp.generated.ts": 'export const REGISTRY_STAMP = "abc";\n',
+  });
+  try {
+    assert.equal(
+      computeRegistryStamp(without.src),
+      computeRegistryStamp(withGen.src),
+      "a *.generated.ts file must not affect the stamp",
+    );
+  } finally {
+    without.cleanup();
+    withGen.cleanup();
+  }
 });
 
-// Sanity: the fixed-input helper computes the SAME way the codegen does, proving
+test("a CRLF checkout WITH trailing CRLF hashes equal to bare LF", () => {
-// the EXPECTED constant above is not an arbitrary magic value but the documented
+  const bare = makeSrcTree({ "tool-specs.ts": "alpha\nbeta" });
-// normalize+sha256 of FIXED_INPUT. Belt-and-braces so a bad EXPECTED can't hide a
+  const crlfTrailing = makeSrcTree({ "tool-specs.ts": "alpha\r\nbeta\r\n" });
-// real regression.
+  try {
-test("the documented EXPECTED constant is the normalize+sha256 of FIXED_INPUT", () => {
+    assert.equal(
-  const FIXED_INPUT = "line1\r\nline2\n";
+      computeRegistryStamp(crlfTrailing.src),
-  const normalized = FIXED_INPUT.replace(/\r\n/g, "\n").replace(/\n$/, "");
+      computeRegistryStamp(bare.src),
-  const expected = createHash("sha256")
+    );
-    .update(normalized, "utf8")
+  } finally {
-    .digest("hex");
+    bare.cleanup();
-  assert.equal(computeRegistryStamp(FIXED_INPUT), expected);
+    crlfTrailing.cleanup();
+  }
+});
+
+// Only a SINGLE trailing newline is stripped — a second blank line is content.
+test("only ONE trailing newline is stripped (two differ from one)", () => {
+  const one = makeSrcTree({ "tool-specs.ts": "x\n" });
+  const two = makeSrcTree({ "tool-specs.ts": "x\n\n" });
+  try {
+    assert.notEqual(
+      computeRegistryStamp(one.src),
+      computeRegistryStamp(two.src),
+    );
+  } finally {
+    one.cleanup();
+    two.cleanup();
+  }
+});
+
+// Cross-impl equality against a fixed, documented tree. The SAME literal tree and
+// expected hash are asserted in the server-side jest test
+// (docmost-client.loader.spec.ts). If either side's enumerate+normalize+sha256
+// ever diverges, one of the two tests reddens. The tree exercises: a nested file,
+// BOTH normalize steps (tool-specs.ts uses CRLF + trailing \n) and the
+// *.generated.ts exclusion.
+const CROSS_IMPL_TREE = {
+  "tool-specs.ts": "line1\r\nline2\n",
+  "client/read.ts": "export const R = 1;\n",
+  "registry-stamp.generated.ts": 'export const REGISTRY_STAMP="ignored";\n',
+};
+const CROSS_IMPL_EXPECTED =
+  "131c1b9e4e2f5a7d6cef91ca8df619822b442f52bc45ebd09474a4c1d6728616";
+
+test("fixed-tree hash matches the documented cross-impl value", () => {
+  const t = makeSrcTree(CROSS_IMPL_TREE);
+  try {
+    assert.equal(computeRegistryStamp(t.src), CROSS_IMPL_EXPECTED);
+  } finally {
+    t.cleanup();
+  }
+});
+
+// Sanity: the EXPECTED constant is not a magic value but the documented
+// enumerate+normalize+sha256 of CROSS_IMPL_TREE (a local re-implementation).
+test("the documented EXPECTED is the enumerate+normalize+sha256 of the tree", () => {
+  const t = makeSrcTree(CROSS_IMPL_TREE);
+  try {
+    const collect = (dir) => {
+      const out = [];
+      for (const e of readdirSync(dir)) {
+        const f = join(dir, e);
+        if (statSync(f).isDirectory()) out.push(...collect(f));
+        else if (e.endsWith(".ts") && !e.endsWith(".generated.ts")) out.push(f);
+      }
+      return out;
+    };
+    const files = collect(t.src)
+      .map((abs) => ({ rel: relative(t.src, abs).split(sep).join("/"), abs }))
+      .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0));
+    const h = createHash("sha256");
+    for (const { rel, abs } of files) {
+      const n = readFileSync(abs, "utf8")
+        .replace(/\r\n/g, "\n")
+        .replace(/\n$/, "");
+      h.update(rel, "utf8");
+      h.update("\0", "utf8");
+      h.update(n, "utf8");
+      h.update("\0", "utf8");
+    }
+    assert.equal(h.digest("hex"), CROSS_IMPL_EXPECTED);
+  } finally {
+    t.cleanup();
+  }
+});
+
+// DESYNC GUARD. Recompute the stamp from the REAL src/ tree and assert it equals
+// the REGISTRY_STAMP baked into the freshly-built build/index.js. This reddens if
+// the generated file is stale OR if the codegen ever diverges from what produced
+// the built stamp.
+test("built REGISTRY_STAMP equals the stamp recomputed from src/", () => {
+  assert.equal(computeRegistryStamp(SRC_DIR), REGISTRY_STAMP);
 });

patches/ai@6.0.134.patch

@@ -1,8 +1,62 @@
 diff --git a/dist/index.js b/dist/index.js
-index ae447a12f7823ec0a00837ee9f0eb809a610d5f8..a3402b2c2d021ef432cfa76e35d370073d525135 100644
+index ae447a12f7823ec0a00837ee9f0eb809a610d5f8..210b5a9009e5cf1537cb2007c524007c1a01253c 100644
 --- a/dist/index.js
 +++ b/dist/index.js
-@@ -6578,9 +6578,19 @@ function createOutputTransformStream(output) {
+@@ -5036,9 +5036,40 @@ function writeToServerResponse({
+           break;
+         const canContinue = response.write(value);
+         if (!canContinue) {
+-          await new Promise((resolve3) => {
+-            response.once("drain", resolve3);
++          // PATCH(docmost #486): race "drain" against "close"/"error". The
++          // original awaited ONLY "drain", so a client that disconnected mid-write
++          // (the socket never drains) parked this loop FOREVER: the finally never
++          // ran, response.end() was unreachable, and the reader + buffered chunks
++          // were held until process restart. On close/error we cancel the reader
++          // and break so the finally always runs (safe for detached runs:
++          // consumeStream drains the SDK stream independently). Listener hygiene:
++          // all three once-listeners are removed on the first settle, so they
++          // cannot pile up one-per-stall.
++          const closed = await new Promise((resolve3) => {
++            function finish(isClosed) {
++              response.removeListener("drain", onDrain);
++              response.removeListener("close", onClose);
++              response.removeListener("error", onError);
++              resolve3(isClosed);
++            }
++            function onDrain() {
++              finish(false);
++            }
++            function onClose() {
++              finish(true);
++            }
++            function onError() {
++              finish(true);
++            }
++            response.once("drain", onDrain);
++            response.once("close", onClose);
++            response.once("error", onError);
+           });
++          if (closed) {
++            await reader.cancel().catch(() => {
++            });
++            break;
++          }
+         }
+       }
+     } catch (error) {
+@@ -5047,7 +5078,9 @@ function writeToServerResponse({
+       response.end();
+     }
+   };
+-  read();
++  read().catch((error) => {
++    console.error("ai writeToServerResponse read() failed:", error);
++  });
+ }
+ 
+ // src/text-stream/pipe-text-stream-to-response.ts
+@@ -6578,9 +6611,19 @@ function createOutputTransformStream(output) {
          controller.enqueue({ part: chunk, partialOutput: void 0 });
          return;
        }
@@ -23,7 +77,7 @@ index ae447a12f7823ec0a00837ee9f0eb809a610d5f8..a3402b2c2d021ef432cfa76e35d37007
        const result = await output.parsePartialOutput({ text: text2 });
        if (result !== void 0) {
          const currentJson = JSON.stringify(result.partial);
-@@ -6959,7 +6969,7 @@ var DefaultStreamTextResult = class {
+@@ -6959,7 +7002,7 @@ var DefaultStreamTextResult = class {
          })
        );
      }
@@ -33,10 +87,64 @@ index ae447a12f7823ec0a00837ee9f0eb809a610d5f8..a3402b2c2d021ef432cfa76e35d37007
        maxRetries: maxRetriesArg,
        abortSignal
 diff --git a/dist/index.mjs b/dist/index.mjs
-index 663875332e3f9a9bd167c25583c515876f42951b..b840b0502c9894df983e0154805abb80e70e6331 100644
+index 663875332e3f9a9bd167c25583c515876f42951b..a51514a390d22811a407edd8b703e21793586cc8 100644
 --- a/dist/index.mjs
 +++ b/dist/index.mjs
-@@ -6501,9 +6501,19 @@ function createOutputTransformStream(output) {
+@@ -4957,9 +4957,40 @@ function writeToServerResponse({
+           break;
+         const canContinue = response.write(value);
+         if (!canContinue) {
+-          await new Promise((resolve3) => {
+-            response.once("drain", resolve3);
++          // PATCH(docmost #486): race "drain" against "close"/"error". The
++          // original awaited ONLY "drain", so a client that disconnected mid-write
++          // (the socket never drains) parked this loop FOREVER: the finally never
++          // ran, response.end() was unreachable, and the reader + buffered chunks
++          // were held until process restart. On close/error we cancel the reader
++          // and break so the finally always runs (safe for detached runs:
++          // consumeStream drains the SDK stream independently). Listener hygiene:
++          // all three once-listeners are removed on the first settle, so they
++          // cannot pile up one-per-stall.
++          const closed = await new Promise((resolve3) => {
++            function finish(isClosed) {
++              response.removeListener("drain", onDrain);
++              response.removeListener("close", onClose);
++              response.removeListener("error", onError);
++              resolve3(isClosed);
++            }
++            function onDrain() {
++              finish(false);
++            }
++            function onClose() {
++              finish(true);
++            }
++            function onError() {
++              finish(true);
++            }
++            response.once("drain", onDrain);
++            response.once("close", onClose);
++            response.once("error", onError);
+           });
++          if (closed) {
++            await reader.cancel().catch(() => {
++            });
++            break;
++          }
+         }
+       }
+     } catch (error) {
+@@ -4968,7 +4999,9 @@ function writeToServerResponse({
+       response.end();
+     }
+   };
+-  read();
++  read().catch((error) => {
++    console.error("ai writeToServerResponse read() failed:", error);
++  });
+ }
+ 
+ // src/text-stream/pipe-text-stream-to-response.ts
+@@ -6501,9 +6534,19 @@ function createOutputTransformStream(output) {
          controller.enqueue({ part: chunk, partialOutput: void 0 });
          return;
        }
@@ -57,7 +165,7 @@ index 663875332e3f9a9bd167c25583c515876f42951b..b840b0502c9894df983e0154805abb80
        const result = await output.parsePartialOutput({ text: text2 });
        if (result !== void 0) {
          const currentJson = JSON.stringify(result.partial);
-@@ -6882,7 +6892,7 @@ var DefaultStreamTextResult = class {
+@@ -6882,7 +6925,7 @@ var DefaultStreamTextResult = class {
          })
        );
      }

pnpm-lock.yaml

@@ -48,7 +48,7 @@ patchedDependencies:
     hash: d8dc66e5ec3b9d23a876b979f493b6aa901fd2d965be54729495da5136296a42
     path: patches/@hocuspocus__server@3.4.4.patch
   ai@6.0.134:
-    hash: f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9
+    hash: e8c599b3963eb01b9ed1481683b7b795ce94137aa1a0c951917c20c8b870299b
     path: patches/ai@6.0.134.patch
   scimmy@1.3.5:
     hash: 775d80f86830b2c5dd1a250c9802c10f8fc3da3c7898373de5aa0c23993d1673
@@ -644,10 +644,10 @@ importers:
         version: 8.3.0(socket.io-adapter@2.5.4)
       ai:
         specifier: ^6.0.134
-        version: 6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6)
+        version: 6.0.134(patch_hash=e8c599b3963eb01b9ed1481683b7b795ce94137aa1a0c951917c20c8b870299b)(zod@4.3.6)
       ai-sdk-ollama:
         specifier: ^3.8.1
-        version: 3.8.1(ai@6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6))(zod@4.3.6)
+        version: 3.8.1(ai@6.0.134(patch_hash=e8c599b3963eb01b9ed1481683b7b795ce94137aa1a0c951917c20c8b870299b)(zod@4.3.6))(zod@4.3.6)
       bcrypt:
         specifier: ^6.0.0
         version: 6.0.0
@@ -16455,17 +16455,17 @@ snapshots:
 
   agent-base@7.1.4: {}
 
-  ai-sdk-ollama@3.8.1(ai@6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6))(zod@4.3.6):
+  ai-sdk-ollama@3.8.1(ai@6.0.134(patch_hash=e8c599b3963eb01b9ed1481683b7b795ce94137aa1a0c951917c20c8b870299b)(zod@4.3.6))(zod@4.3.6):
     dependencies:
       '@ai-sdk/provider': 3.0.8
       '@ai-sdk/provider-utils': 4.0.21(zod@4.3.6)
-      ai: 6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6)
+      ai: 6.0.134(patch_hash=e8c599b3963eb01b9ed1481683b7b795ce94137aa1a0c951917c20c8b870299b)(zod@4.3.6)
       jsonrepair: 3.13.3
       ollama: 0.6.3
     transitivePeerDependencies:
       - zod
 
-  ai@6.0.134(patch_hash=f60bfc3357e01e1f3978c6c40fdd65aeb33fefaad7179cde8676465b6c5ff4d9)(zod@4.3.6):
+  ai@6.0.134(patch_hash=e8c599b3963eb01b9ed1481683b7b795ce94137aa1a0c951917c20c8b870299b)(zod@4.3.6):
     dependencies:
       '@ai-sdk/gateway': 3.0.77(zod@4.3.6)
       '@ai-sdk/provider': 3.0.8